add example_dataset in training scripts (#1358)

* add example_dataset in training scripts

* fix example datasets

.gitignore

@@ -2,6 +2,7 @@
 /models
 /scripts
 /diffusers
+/.vscode
 *.pkl
 *.safetensors
 *.pth

README.md

@@ -32,6 +32,14 @@ We believe that a well-developed open-source code framework can lower the thresh
 > DiffSynth-Studio has undergone major version updates, and some old features are no longer maintained. If you need to use old features, please switch to the [last historical version](https://github.com/modelscope/DiffSynth-Studio/tree/afd101f3452c9ecae0c87b79adfa2e22d65ffdc3) before the major version update.
 
 > Currently, the development personnel of this project are limited, with most of the work handled by [Artiprocher](https://github.com/Artiprocher). Therefore, the progress of new feature development will be relatively slow, and the speed of responding to and resolving issues is limited. We apologize for this and ask developers to understand.
+- **January 19, 2026**: Added support for [openmoss/MOVA-720p](https://modelscope.cn/models/openmoss/MOVA-720p) and [openmoss/MOVA-360p](https://modelscope.cn/models/openmoss/MOVA-360p) models, including training and inference capabilities. [Documentation](/docs/en/Model_Details/Wan.md) and [example code](/examples/mova/) are now available.
+
+- **March 12, 2026**: We have added support for the [LTX-2.3](https://modelscope.cn/models/Lightricks/LTX-2.3) audio-video generation model. The features includes text-to-audio/video, image-to-audio/video, IC-LoRA control, audio-to-video, and audio-video inpainting. We have supported the complete inference and training functionalities. For details, please refer to the [documentation](/docs/en/Model_Details/LTX-2.md) and [code](/examples/ltx2/).
+
+- **March 3, 2026**: We released the [DiffSynth-Studio/Qwen-Image-Layered-Control-V2](https://www.modelscope.cn/models/DiffSynth-Studio/Qwen-Image-Layered-Control-V2) model, which is an updated version of Qwen-Image-Layered-Control. In addition to the originally supported text-guided functionality, it adds brush-controlled layer separation capabilities.
+
+- **March 2, 2026** Added support for [Anima](https://modelscope.cn/models/circlestone-labs/Anima). For details, please refer to the [documentation](docs/en/Model_Details/Anima.md). This is an interesting anime-style image generation model. We look forward to its future updates.
+
 - **February 26, 2026** Added full and lora training support for the LTX-2 audio-video generation model. See the [documentation](/docs/en/Model_Details/LTX-2.md) for details.
 
 - **February 10, 2026** Added inference support for the LTX-2 audio-video generation model. See the [documentation](/docs/en/Model_Details/LTX-2.md) for details. Support for model training will be implemented in the future.
@@ -343,6 +351,60 @@ Example code for FLUX.2 is available at: [/examples/flux2/](/examples/flux2/)
 
 </details>
 
+#### Anima: [/docs/en/Model_Details/Anima.md](/docs/en/Model_Details/Anima.md)
+
+<details>
+
+<summary>Quick Start</summary>
+
+Run the following code to quickly load the [circlestone-labs/Anima](https://www.modelscope.cn/models/circlestone-labs/Anima) model and perform inference. VRAM management is enabled, and the framework will automatically control the loading of model parameters based on available VRAM. The model can run with a minimum of 8GB VRAM.
+
+```python
+from diffsynth.pipelines.anima_image import AnimaImagePipeline, ModelConfig
+import torch
+
+vram_config = {
+    "offload_dtype": "disk",
+    "offload_device": "disk",
+    "onload_dtype": "disk",
+    "onload_device": "disk",
+    "preparing_dtype": torch.bfloat16,
+    "preparing_device": "cuda",
+    "computation_dtype": torch.bfloat16,
+    "computation_device": "cuda",
+}
+pipe = AnimaImagePipeline.from_pretrained(
+    torch_dtype=torch.bfloat16,
+    device="cuda",
+    model_configs=[
+        ModelConfig(model_id="circlestone-labs/Anima", origin_file_pattern="split_files/diffusion_models/anima-preview.safetensors", **vram_config),
+        ModelConfig(model_id="circlestone-labs/Anima", origin_file_pattern="split_files/text_encoders/qwen_3_06b_base.safetensors", **vram_config),
+        ModelConfig(model_id="circlestone-labs/Anima", origin_file_pattern="split_files/vae/qwen_image_vae.safetensors", **vram_config),
+    ],
+    tokenizer_config=ModelConfig(model_id="Qwen/Qwen3-0.6B", origin_file_pattern="./"),
+    tokenizer_t5xxl_config=ModelConfig(model_id="stabilityai/stable-diffusion-3.5-large", origin_file_pattern="tokenizer_3/"),
+    vram_limit=torch.cuda.mem_get_info("cuda")[1] / (1024 ** 3) - 0.5,
+)
+prompt = "Masterpiece, best quality, solo, long hair, wavy hair, silver hair, blue eyes, blue dress, medium breasts, dress, underwater, air bubble, floating hair, refraction, portrait."
+negative_prompt = "worst quality, low quality, monochrome, zombie, interlocked fingers, Aissist, cleavage, nsfw,"
+image = pipe(prompt, seed=0, num_inference_steps=50)
+image.save("image.jpg")
+```
+
+</details>
+
+<details>
+
+<summary>Examples</summary>
+
+Example code for Anima is located at: [/examples/anima/](/examples/anima/)
+
+| Model ID | Inference | Low VRAM Inference | Full Training | Validation after Full Training | LoRA Training | Validation after LoRA Training |
+|-|-|-|-|-|-|-|
+|[circlestone-labs/Anima](https://www.modelscope.cn/models/circlestone-labs/Anima)|[code](/examples/anima/model_inference/anima-preview.py)|[code](/examples/anima/model_inference_low_vram/anima-preview.py)|[code](/examples/anima/model_training/full/anima-preview.sh)|[code](/examples/anima/model_training/validate_full/anima-preview.py)|[code](/examples/anima/model_training/lora/anima-preview.sh)|[code](/examples/anima/model_training/validate_lora/anima-preview.py)|
+
+</details>
+
 #### Qwen-Image: [/docs/en/Model_Details/Qwen-Image.md](/docs/en/Model_Details/Qwen-Image.md)
 
 <details>
@@ -423,9 +485,11 @@ Example code for Qwen-Image is available at: [/examples/qwen_image/](/examples/q
 |[Qwen/Qwen-Image-Edit-2509](https://www.modelscope.cn/models/Qwen/Qwen-Image-Edit-2509)|[code](/examples/qwen_image/model_inference/Qwen-Image-Edit-2509.py)|[code](/examples/qwen_image/model_inference_low_vram/Qwen-Image-Edit-2509.py)|[code](/examples/qwen_image/model_training/full/Qwen-Image-Edit-2509.sh)|[code](/examples/qwen_image/model_training/validate_full/Qwen-Image-Edit-2509.py)|[code](/examples/qwen_image/model_training/lora/Qwen-Image-Edit-2509.sh)|[code](/examples/qwen_image/model_training/validate_lora/Qwen-Image-Edit-2509.py)|
 |[Qwen/Qwen-Image-Edit-2511](https://www.modelscope.cn/models/Qwen/Qwen-Image-Edit-2511)|[code](/examples/qwen_image/model_inference/Qwen-Image-Edit-2511.py)|[code](/examples/qwen_image/model_inference_low_vram/Qwen-Image-Edit-2511.py)|[code](/examples/qwen_image/model_training/full/Qwen-Image-Edit-2511.sh)|[code](/examples/qwen_image/model_training/validate_full/Qwen-Image-Edit-2511.py)|[code](/examples/qwen_image/model_training/lora/Qwen-Image-Edit-2511.sh)|[code](/examples/qwen_image/model_training/validate_lora/Qwen-Image-Edit-2511.py)|
 |[FireRedTeam/FireRed-Image-Edit-1.0](https://www.modelscope.cn/models/FireRedTeam/FireRed-Image-Edit-1.0)|[code](/examples/qwen_image/model_inference/FireRed-Image-Edit-1.0.py)|[code](/examples/qwen_image/model_inference_low_vram/FireRed-Image-Edit-1.0.py)|[code](/examples/qwen_image/model_training/full/FireRed-Image-Edit-1.0.sh)|[code](/examples/qwen_image/model_training/validate_full/FireRed-Image-Edit-1.0.py)|[code](/examples/qwen_image/model_training/lora/FireRed-Image-Edit-1.0.sh)|[code](/examples/qwen_image/model_training/validate_lora/FireRed-Image-Edit-1.0.py)|
+|[FireRedTeam/FireRed-Image-Edit-1.1](https://www.modelscope.cn/models/FireRedTeam/FireRed-Image-Edit-1.1)|[code](/examples/qwen_image/model_inference/FireRed-Image-Edit-1.1.py)|[code](/examples/qwen_image/model_inference_low_vram/FireRed-Image-Edit-1.1.py)|[code](/examples/qwen_image/model_training/full/FireRed-Image-Edit-1.1.sh)|[code](/examples/qwen_image/model_training/validate_full/FireRed-Image-Edit-1.1.py)|[code](/examples/qwen_image/model_training/lora/FireRed-Image-Edit-1.1.sh)|[code](/examples/qwen_image/model_training/validate_lora/FireRed-Image-Edit-1.1.py)|
 |[lightx2v/Qwen-Image-Edit-2511-Lightning](https://modelscope.cn/models/lightx2v/Qwen-Image-Edit-2511-Lightning)|[code](/examples/qwen_image/model_inference/Qwen-Image-Edit-2511-Lightning.py)|[code](/examples/qwen_image/model_inference_low_vram/Qwen-Image-Edit-2511-Lightning.py)|-|-|-|-|
 |[Qwen/Qwen-Image-Layered](https://www.modelscope.cn/models/Qwen/Qwen-Image-Layered)|[code](/examples/qwen_image/model_inference/Qwen-Image-Layered.py)|[code](/examples/qwen_image/model_inference_low_vram/Qwen-Image-Layered.py)|[code](/examples/qwen_image/model_training/full/Qwen-Image-Layered.sh)|[code](/examples/qwen_image/model_training/validate_full/Qwen-Image-Layered.py)|[code](/examples/qwen_image/model_training/lora/Qwen-Image-Layered.sh)|[code](/examples/qwen_image/model_training/validate_lora/Qwen-Image-Layered.py)|
 |[DiffSynth-Studio/Qwen-Image-Layered-Control](https://www.modelscope.cn/models/DiffSynth-Studio/Qwen-Image-Layered-Control)|[code](/examples/qwen_image/model_inference/Qwen-Image-Layered-Control.py)|[code](/examples/qwen_image/model_inference_low_vram/Qwen-Image-Layered-Control.py)|[code](/examples/qwen_image/model_training/full/Qwen-Image-Layered-Control.sh)|[code](/examples/qwen_image/model_training/validate_full/Qwen-Image-Layered-Control.py)|[code](/examples/qwen_image/model_training/lora/Qwen-Image-Layered-Control.sh)|[code](/examples/qwen_image/model_training/validate_lora/Qwen-Image-Layered-Control.py)|
+|[DiffSynth-Studio/Qwen-Image-Layered-Control-V2](https://www.modelscope.cn/models/DiffSynth-Studio/Qwen-Image-Layered-Control-V2)|[code](/examples/qwen_image/model_inference/Qwen-Image-Layered-Control-V2.py)|[code](/examples/qwen_image/model_inference_low_vram/Qwen-Image-Layered-Control-V2.py)|-|-|[code](/examples/qwen_image/model_training/lora/Qwen-Image-Layered-Control-V2.sh)|[code](/examples/qwen_image/model_training/validate_lora/Qwen-Image-Layered-Control-V2.py)|
 |[DiffSynth-Studio/Qwen-Image-EliGen](https://www.modelscope.cn/models/DiffSynth-Studio/Qwen-Image-EliGen)|[code](/examples/qwen_image/model_inference/Qwen-Image-EliGen.py)|[code](/examples/qwen_image/model_inference_low_vram/Qwen-Image-EliGen.py)|-|-|[code](/examples/qwen_image/model_training/lora/Qwen-Image-EliGen.sh)|[code](/examples/qwen_image/model_training/validate_lora/Qwen-Image-EliGen.py)|
 |[DiffSynth-Studio/Qwen-Image-EliGen-V2](https://www.modelscope.cn/models/DiffSynth-Studio/Qwen-Image-EliGen-V2)|[code](/examples/qwen_image/model_inference/Qwen-Image-EliGen-V2.py)|[code](/examples/qwen_image/model_inference_low_vram/Qwen-Image-EliGen-V2.py)|-|-|[code](/examples/qwen_image/model_training/lora/Qwen-Image-EliGen.sh)|[code](/examples/qwen_image/model_training/validate_lora/Qwen-Image-EliGen.py)|
 |[DiffSynth-Studio/Qwen-Image-EliGen-Poster](https://www.modelscope.cn/models/DiffSynth-Studio/Qwen-Image-EliGen-Poster)|[code](/examples/qwen_image/model_inference/Qwen-Image-EliGen-Poster.py)|[code](/examples/qwen_image/model_inference_low_vram/Qwen-Image-EliGen-Poster.py)|-|-|[code](/examples/qwen_image/model_training/lora/Qwen-Image-EliGen-Poster.sh)|[code](/examples/qwen_image/model_training/validate_lora/Qwen-Image-EliGen-Poster.py)|
@@ -644,7 +708,19 @@ Example code for LTX-2 is available at: [/examples/ltx2/](/examples/ltx2/)
 
 | Model ID | Extra Args | Inference | Low-VRAM Inference | Full Training | Full Training Validation | LoRA Training | LoRA Training Validation |
 |-|-|-|-|-|-|-|-|
+|[Lightricks/LTX-2.3: OneStagePipeline-I2AV](https://www.modelscope.cn/models/Lightricks/LTX-2.3)|`input_images`|[code](/examples/ltx2/model_inference/LTX-2.3-I2AV-OneStage.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2.3-I2AV-OneStage.py)|[code](/examples/ltx2/model_training/full/LTX-2.3-I2AV-splited.sh)|[code](/examples/ltx2/model_training/validate_full/LTX-2.3-I2AV.py)|[code](/examples/ltx2/model_training/lora/LTX-2.3-I2AV-splited.sh)|[code](/examples/ltx2/model_training/validate_lora/LTX-2.3-I2AV.py)|
+|[Lightricks/LTX-2.3: TwoStagePipeline-I2AV](https://www.modelscope.cn/models/Lightricks/LTX-2.3)|`input_images`|[code](/examples/ltx2/model_inference/LTX-2.3-I2AV-TwoStage.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2.3-I2AV-TwoStage.py)|-|-|-|-|
+|[Lightricks/LTX-2.3: DistilledPipeline-I2AV](https://www.modelscope.cn/models/Lightricks/LTX-2.3)|`input_images`|[code](/examples/ltx2/model_inference/LTX-2.3-I2AV-DistilledPipeline.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2.3-I2AV-DistilledPipeline.py)|-|-|-|-|
+|[Lightricks/LTX-2.3: OneStagePipeline-T2AV](https://www.modelscope.cn/models/Lightricks/LTX-2.3)||[code](/examples/ltx2/model_inference/LTX-2.3-T2AV-OneStage.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2.3-T2AV-OneStage.py)|[code](/examples/ltx2/model_training/full/LTX-2.3-T2AV-splited.sh)|[code](/examples/ltx2/model_training/validate_full/LTX-2.3-T2AV.py)|[code](/examples/ltx2/model_training/lora/LTX-2.3-T2AV-splited.sh)|[code](/examples/ltx2/model_training/validate_lora/LTX-2.3-T2AV.py)|
+|[Lightricks/LTX-2.3: TwoStagePipeline-T2AV](https://www.modelscope.cn/models/Lightricks/LTX-2.3)||[code](/examples/ltx2/model_inference/LTX-2.3-T2AV-TwoStage.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2.3-T2AV-TwoStage.py)|-|-|-|-|
+|[Lightricks/LTX-2.3: DistilledPipeline-T2AV](https://www.modelscope.cn/models/Lightricks/LTX-2.3)||[code](/examples/ltx2/model_inference/LTX-2.3-T2AV-DistilledPipeline.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2.3-T2AV-DistilledPipeline.py)|-|-|-|-|
+|[Lightricks/LTX-2.3: A2V](https://www.modelscope.cn/models/Lightricks/LTX-2.3)|`retake_audio`,`audio_sample_rate`,`retake_audio_regions`|[code](/examples/ltx2/model_inference/LTX-2.3-A2V-TwoStage.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2.3-A2V-TwoStage.py)|-|-|-|-|
+|[Lightricks/LTX-2.3: Retake](https://www.modelscope.cn/models/Lightricks/LTX-2.3)|`retake_video`,`retake_video_regions`,`retake_audio`,`audio_sample_rate`,`retake_audio_regions`|[code](/examples/ltx2/model_inference/LTX-2.3-T2AV-TwoStage-Retake.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2.3-T2AV-TwoStage-Retake.py)|-|-|-|-|
+|[Lightricks/LTX-2.3-22b-IC-LoRA-Union-Control](https://www.modelscope.cn/models/Lightricks/LTX-2.3-22b-IC-LoRA-Union-Control)|`in_context_videos`,`in_context_downsample_factor`|[code](/examples/ltx2/model_inference/LTX-2.3-T2AV-IC-LoRA-Union-Control.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2.3-T2AV-IC-LoRA-Union-Control.py)|-|-|[code](/examples/ltx2/model_training/lora/LTX-2.3-T2AV-IC-LoRA-splited.sh)|[code](/examples/ltx2/model_training/validate_lora/LTX-2.3-T2AV-IC-LoRA.py)|
+|[Lightricks/LTX-2.3-22b-IC-LoRA-Motion-Track-Control](https://www.modelscope.cn/models/Lightricks/LTX-2.3-22b-IC-LoRA-Motion-Track-Control)|`in_context_videos`,`in_context_downsample_factor`|[code](/examples/ltx2/model_inference/LTX-2.3-T2AV-IC-LoRA-Motion-Track-Control.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2.3-T2AV-IC-LoRA-Motion-Track-Control.py)|-|-|[code](/examples/ltx2/model_training/lora/LTX-2.3-T2AV-IC-LoRA-splited.sh)|[code](/examples/ltx2/model_training/validate_lora/LTX-2.3-T2AV-IC-LoRA.py)|
 |[Lightricks/LTX-2: OneStagePipeline-T2AV](https://www.modelscope.cn/models/Lightricks/LTX-2)||[code](/examples/ltx2/model_inference/LTX-2-T2AV-OneStage.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2-T2AV-OneStage.py)|[code](/examples/ltx2/model_training/full/LTX-2-T2AV-splited.sh)|[code](/examples/ltx2/model_training/validate_full/LTX-2-T2AV.py)|[code](/examples/ltx2/model_training/lora/LTX-2-T2AV-splited.sh)|[code](/examples/ltx2/model_training/validate_lora/LTX-2-T2AV.py)|
+|[Lightricks/LTX-2-19b-IC-LoRA-Union-Control](https://www.modelscope.cn/models/Lightricks/LTX-2-19b-IC-LoRA-Union-Control)|`in_context_videos`,`in_context_downsample_factor`|[code](/examples/ltx2/model_inference/LTX-2-T2AV-IC-LoRA-Union-Control.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2-T2AV-IC-LoRA-Union-Control.py)|-|-|[code](/examples/ltx2/model_training/lora/LTX-2-T2AV-IC-LoRA-splited.sh)|[code](/examples/ltx2/model_training/validate_lora/LTX-2-T2AV-IC-LoRA.py)|
+|[Lightricks/LTX-2-19b-IC-LoRA-Detailer](https://www.modelscope.cn/models/Lightricks/LTX-2-19b-IC-LoRA-Detailer)|`in_context_videos`,`in_context_downsample_factor`|[code](/examples/ltx2/model_inference/LTX-2-T2AV-IC-LoRA-Detailer.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2-T2AV-IC-LoRA-Detailer.py)|-|-|[code](/examples/ltx2/model_training/lora/LTX-2-T2AV-IC-LoRA-splited.sh)|[code](/examples/ltx2/model_training/validate_lora/LTX-2-T2AV-IC-LoRA.py)|
 |[Lightricks/LTX-2: TwoStagePipeline-T2AV](https://www.modelscope.cn/models/Lightricks/LTX-2)||[code](/examples/ltx2/model_inference/LTX-2-T2AV-TwoStage.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2-T2AV-TwoStage.py)|-|-|-|-|
 |[Lightricks/LTX-2: DistilledPipeline-T2AV](https://www.modelscope.cn/models/Lightricks/LTX-2)||[code](/examples/ltx2/model_inference/LTX-2-T2AV-DistilledPipeline.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2-T2AV-DistilledPipeline.py)|-|-|-|-|
 |[Lightricks/LTX-2: OneStagePipeline-I2AV](https://www.modelscope.cn/models/Lightricks/LTX-2)|`input_images`|[code](/examples/ltx2/model_inference/LTX-2-I2AV-OneStage.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2-I2AV-OneStage.py)|-|-|-|-|
@@ -792,6 +868,8 @@ Example code for Wan is available at: [/examples/wanvideo/](/examples/wanvideo/)
 |[PAI/Wan2.2-Fun-A14B-InP](https://modelscope.cn/models/PAI/Wan2.2-Fun-A14B-InP)|`input_image`, `end_image`|[code](/examples/wanvideo/model_inference/Wan2.2-Fun-A14B-InP.py)|[code](/examples/wanvideo/model_training/full/Wan2.2-Fun-A14B-InP.sh)|[code](/examples/wanvideo/model_training/validate_full/Wan2.2-Fun-A14B-InP.py)|[code](/examples/wanvideo/model_training/lora/Wan2.2-Fun-A14B-InP.sh)|[code](/examples/wanvideo/model_training/validate_lora/Wan2.2-Fun-A14B-InP.py)|
 |[PAI/Wan2.2-Fun-A14B-Control](https://modelscope.cn/models/PAI/Wan2.2-Fun-A14B-Control)|`control_video`, `reference_image`|[code](/examples/wanvideo/model_inference/Wan2.2-Fun-A14B-Control.py)|[code](/examples/wanvideo/model_training/full/Wan2.2-Fun-A14B-Control.sh)|[code](/examples/wanvideo/model_training/validate_full/Wan2.2-Fun-A14B-Control.py)|[code](/examples/wanvideo/model_training/lora/Wan2.2-Fun-A14B-Control.sh)|[code](/examples/wanvideo/model_training/validate_lora/Wan2.2-Fun-A14B-Control.py)|
 |[PAI/Wan2.2-Fun-A14B-Control-Camera](https://modelscope.cn/models/PAI/Wan2.2-Fun-A14B-Control-Camera)|`control_camera_video`, `input_image`|[code](/examples/wanvideo/model_inference/Wan2.2-Fun-A14B-Control-Camera.py)|[code](/examples/wanvideo/model_training/full/Wan2.2-Fun-A14B-Control-Camera.sh)|[code](/examples/wanvideo/model_training/validate_full/Wan2.2-Fun-A14B-Control-Camera.py)|[code](/examples/wanvideo/model_training/lora/Wan2.2-Fun-A14B-Control-Camera.sh)|[code](/examples/wanvideo/model_training/validate_lora/Wan2.2-Fun-A14B-Control-Camera.py)|
+| [openmoss/MOVA-360p](https://modelscope.cn/models/openmoss/MOVA-360p) | `input_image` | [code](/examples/mova/model_inference/MOVA-360p-I2AV.py) | [code](/examples/mova/model_training/full/MOVA-360P-I2AV.sh) | [code](/examples/mova/model_training/validate_full/MOVA-360p-I2AV.py) | [code](/examples/mova/model_training/lora/MOVA-360P-I2AV.sh) | [code](/examples/mova/model_training/validate_lora/MOVA-360p-I2AV.py) |
+| [openmoss/MOVA-720p](https://modelscope.cn/models/openmoss/MOVA-720p) | `input_image` | [code](/examples/mova/model_inference/MOVA-720p-I2AV.py) | [code](/examples/mova/model_training/full/MOVA-720P-I2AV.sh) | [code](/examples/mova/model_training/validate_full/MOVA-720p-I2AV.py) | [code](/examples/mova/model_training/lora/MOVA-720P-I2AV.sh) | [code](/examples/mova/model_training/validate_lora/MOVA-720p-I2AV.py) |
 
 </details>
 
@@ -805,7 +883,7 @@ DiffSynth-Studio is not just an engineered model framework, but also an incubato
 
 - Paper: [Spectral Evolution Search: Efficient Inference-Time Scaling for Reward-Aligned Image Generation
 ](https://arxiv.org/abs/2602.03208)
-- Sample Code: coming soon
+- Sample Code: [/docs/en/Research_Tutorial/inference_time_scaling.md](/docs/en/Research_Tutorial/inference_time_scaling.md)
 
 |FLUX.1-dev|FLUX.1-dev + SES|Qwen-Image|Qwen-Image + SES|
 |-|-|-|-|

README_zh.md

@@ -32,6 +32,14 @@ DiffSynth 目前包括两个开源项目：
 > DiffSynth-Studio 经历了大版本更新，部分旧功能已停止维护，如需使用旧版功能，请切换到大版本更新前的[最后一个历史版本](https://github.com/modelscope/DiffSynth-Studio/tree/afd101f3452c9ecae0c87b79adfa2e22d65ffdc3)。
 
 > 目前本项目的开发人员有限，大部分工作由 [Artiprocher](https://github.com/Artiprocher) 负责，因此新功能的开发进展会比较缓慢，issue 的回复和解决速度有限，我们对此感到非常抱歉，请各位开发者理解。
+- **2026年1月19日** 新增对 [openmoss/MOVA-720p](https://modelscope.cn/models/openmoss/MOVA-720p) 和 [openmoss/MOVA-360p](https://modelscope.cn/models/openmoss/MOVA-360p) 模型的支持，包括完整的训练和推理功能。[文档](/docs/zh/Model_Details/Wan.md)和[示例代码](/examples/mova/)现已可用。
+
+- **2026年3月12日** 我们新增了 [LTX-2.3](https://modelscope.cn/models/Lightricks/LTX-2.3) 音视频生成模型的支持，模型支持的功能包括文生音视频、图生音视频、IC-LoRA控制、音频生视频、音视频局部Inpainting，框架支持完整的推理和训练功能。详细信息请参考 [文档](/docs/zh/Model_Details/LTX-2.md) 和 [示例代码](/examples/ltx2/)。
+
+- **2026年3月3日** 我们发布了 [DiffSynth-Studio/Qwen-Image-Layered-Control-V2](https://www.modelscope.cn/models/DiffSynth-Studio/Qwen-Image-Layered-Control-V2) 模型，这是 Qwen-Image-Layered-Control 的更新版本。除了原本就支持的文本引导功能，新增了画笔控制的图层拆分能力。
+
+- **2026年3月2日** 新增对[Anima](https://modelscope.cn/models/circlestone-labs/Anima)的支持，详见[文档](docs/zh/Model_Details/Anima.md)。这是一个有趣的动漫风格图像生成模型，我们期待其后续的模型更新。
+
 - **2026年2月26日** 新增对[LTX-2](https://www.modelscope.cn/models/Lightricks/LTX-2)音视频生成模型全量微调与LoRA训练支持，详见[文档](docs/zh/Model_Details/LTX-2.md)。
 
 - **2026年2月10日** 新增对[LTX-2](https://www.modelscope.cn/models/Lightricks/LTX-2)音视频生成模型的推理支持，详见[文档](docs/zh/Model_Details/LTX-2.md)，后续将推进模型训练的支持。
@@ -343,6 +351,60 @@ FLUX.2 的示例代码位于：[/examples/flux2/](/examples/flux2/)
 
 </details>
 
+#### Anima: [/docs/zh/Model_Details/Anima.md](/docs/zh/Model_Details/Anima.md)
+
+<details>
+
+<summary>快速开始</summary>
+
+运行以下代码可以快速加载 [circlestone-labs/Anima](https://www.modelscope.cn/models/circlestone-labs/Anima) 模型并进行推理。显存管理已启动，框架会自动根据剩余显存控制模型参数的加载，最低 8G 显存即可运行。
+
+```python
+from diffsynth.pipelines.anima_image import AnimaImagePipeline, ModelConfig
+import torch
+
+vram_config = {
+    "offload_dtype": "disk",
+    "offload_device": "disk",
+    "onload_dtype": "disk",
+    "onload_device": "disk",
+    "preparing_dtype": torch.bfloat16,
+    "preparing_device": "cuda",
+    "computation_dtype": torch.bfloat16,
+    "computation_device": "cuda",
+}
+pipe = AnimaImagePipeline.from_pretrained(
+    torch_dtype=torch.bfloat16,
+    device="cuda",
+    model_configs=[
+        ModelConfig(model_id="circlestone-labs/Anima", origin_file_pattern="split_files/diffusion_models/anima-preview.safetensors", **vram_config),
+        ModelConfig(model_id="circlestone-labs/Anima", origin_file_pattern="split_files/text_encoders/qwen_3_06b_base.safetensors", **vram_config),
+        ModelConfig(model_id="circlestone-labs/Anima", origin_file_pattern="split_files/vae/qwen_image_vae.safetensors", **vram_config),
+    ],
+    tokenizer_config=ModelConfig(model_id="Qwen/Qwen3-0.6B", origin_file_pattern="./"),
+    tokenizer_t5xxl_config=ModelConfig(model_id="stabilityai/stable-diffusion-3.5-large", origin_file_pattern="tokenizer_3/"),
+    vram_limit=torch.cuda.mem_get_info("cuda")[1] / (1024 ** 3) - 0.5,
+)
+prompt = "Masterpiece, best quality, solo, long hair, wavy hair, silver hair, blue eyes, blue dress, medium breasts, dress, underwater, air bubble, floating hair, refraction, portrait."
+negative_prompt = "worst quality, low quality, monochrome, zombie, interlocked fingers, Aissist, cleavage, nsfw,"
+image = pipe(prompt, seed=0, num_inference_steps=50)
+image.save("image.jpg")
+```
+
+</details>
+
+<details>
+
+<summary>示例代码</summary>
+
+Anima 的示例代码位于：[/examples/anima/](/examples/anima/)
+
+|模型 ID|推理|低显存推理|全量训练|全量训练后验证|LoRA 训练|LoRA 训练后验证|
+|-|-|-|-|-|-|-|
+|[circlestone-labs/Anima](https://www.modelscope.cn/models/circlestone-labs/Anima)|[code](/examples/anima/model_inference/anima-preview.py)|[code](/examples/anima/model_inference_low_vram/anima-preview.py)|[code](/examples/anima/model_training/full/anima-preview.sh)|[code](/examples/anima/model_training/validate_full/anima-preview.py)|[code](/examples/anima/model_training/lora/anima-preview.sh)|[code](/examples/anima/model_training/validate_lora/anima-preview.py)|
+
+</details>
+
 #### Qwen-Image: [/docs/zh/Model_Details/Qwen-Image.md](/docs/zh/Model_Details/Qwen-Image.md)
 
 <details>
@@ -423,9 +485,11 @@ Qwen-Image 的示例代码位于：[/examples/qwen_image/](/examples/qwen_image/
 |[Qwen/Qwen-Image-Edit-2509](https://www.modelscope.cn/models/Qwen/Qwen-Image-Edit-2509)|[code](/examples/qwen_image/model_inference/Qwen-Image-Edit-2509.py)|[code](/examples/qwen_image/model_inference_low_vram/Qwen-Image-Edit-2509.py)|[code](/examples/qwen_image/model_training/full/Qwen-Image-Edit-2509.sh)|[code](/examples/qwen_image/model_training/validate_full/Qwen-Image-Edit-2509.py)|[code](/examples/qwen_image/model_training/lora/Qwen-Image-Edit-2509.sh)|[code](/examples/qwen_image/model_training/validate_lora/Qwen-Image-Edit-2509.py)|
 |[Qwen/Qwen-Image-Edit-2511](https://www.modelscope.cn/models/Qwen/Qwen-Image-Edit-2511)|[code](/examples/qwen_image/model_inference/Qwen-Image-Edit-2511.py)|[code](/examples/qwen_image/model_inference_low_vram/Qwen-Image-Edit-2511.py)|[code](/examples/qwen_image/model_training/full/Qwen-Image-Edit-2511.sh)|[code](/examples/qwen_image/model_training/validate_full/Qwen-Image-Edit-2511.py)|[code](/examples/qwen_image/model_training/lora/Qwen-Image-Edit-2511.sh)|[code](/examples/qwen_image/model_training/validate_lora/Qwen-Image-Edit-2511.py)|
 |[FireRedTeam/FireRed-Image-Edit-1.0](https://www.modelscope.cn/models/FireRedTeam/FireRed-Image-Edit-1.0)|[code](/examples/qwen_image/model_inference/FireRed-Image-Edit-1.0.py)|[code](/examples/qwen_image/model_inference_low_vram/FireRed-Image-Edit-1.0.py)|[code](/examples/qwen_image/model_training/full/FireRed-Image-Edit-1.0.sh)|[code](/examples/qwen_image/model_training/validate_full/FireRed-Image-Edit-1.0.py)|[code](/examples/qwen_image/model_training/lora/FireRed-Image-Edit-1.0.sh)|[code](/examples/qwen_image/model_training/validate_lora/FireRed-Image-Edit-1.0.py)|
+|[FireRedTeam/FireRed-Image-Edit-1.1](https://www.modelscope.cn/models/FireRedTeam/FireRed-Image-Edit-1.1)|[code](/examples/qwen_image/model_inference/FireRed-Image-Edit-1.1.py)|[code](/examples/qwen_image/model_inference_low_vram/FireRed-Image-Edit-1.1.py)|[code](/examples/qwen_image/model_training/full/FireRed-Image-Edit-1.1.sh)|[code](/examples/qwen_image/model_training/validate_full/FireRed-Image-Edit-1.1.py)|[code](/examples/qwen_image/model_training/lora/FireRed-Image-Edit-1.1.sh)|[code](/examples/qwen_image/model_training/validate_lora/FireRed-Image-Edit-1.1.py)|
 |[lightx2v/Qwen-Image-Edit-2511-Lightning](https://modelscope.cn/models/lightx2v/Qwen-Image-Edit-2511-Lightning)|[code](/examples/qwen_image/model_inference/Qwen-Image-Edit-2511-Lightning.py)|[code](/examples/qwen_image/model_inference_low_vram/Qwen-Image-Edit-2511-Lightning.py)|-|-|-|-|
 |[Qwen/Qwen-Image-Layered](https://www.modelscope.cn/models/Qwen/Qwen-Image-Layered)|[code](/examples/qwen_image/model_inference/Qwen-Image-Layered.py)|[code](/examples/qwen_image/model_inference_low_vram/Qwen-Image-Layered.py)|[code](/examples/qwen_image/model_training/full/Qwen-Image-Layered.sh)|[code](/examples/qwen_image/model_training/validate_full/Qwen-Image-Layered.py)|[code](/examples/qwen_image/model_training/lora/Qwen-Image-Layered.sh)|[code](/examples/qwen_image/model_training/validate_lora/Qwen-Image-Layered.py)|
 |[DiffSynth-Studio/Qwen-Image-Layered-Control](https://www.modelscope.cn/models/DiffSynth-Studio/Qwen-Image-Layered-Control)|[code](/examples/qwen_image/model_inference/Qwen-Image-Layered-Control.py)|[code](/examples/qwen_image/model_inference_low_vram/Qwen-Image-Layered-Control.py)|[code](/examples/qwen_image/model_training/full/Qwen-Image-Layered-Control.sh)|[code](/examples/qwen_image/model_training/validate_full/Qwen-Image-Layered-Control.py)|[code](/examples/qwen_image/model_training/lora/Qwen-Image-Layered-Control.sh)|[code](/examples/qwen_image/model_training/validate_lora/Qwen-Image-Layered-Control.py)|
+|[DiffSynth-Studio/Qwen-Image-Layered-Control-V2](https://www.modelscope.cn/models/DiffSynth-Studio/Qwen-Image-Layered-Control-V2)|[code](/examples/qwen_image/model_inference/Qwen-Image-Layered-Control-V2.py)|[code](/examples/qwen_image/model_inference_low_vram/Qwen-Image-Layered-Control-V2.py)|-|-|[code](/examples/qwen_image/model_training/lora/Qwen-Image-Layered-Control-V2.sh)|[code](/examples/qwen_image/model_training/validate_lora/Qwen-Image-Layered-Control-V2.py)|
 |[DiffSynth-Studio/Qwen-Image-EliGen](https://www.modelscope.cn/models/DiffSynth-Studio/Qwen-Image-EliGen)|[code](/examples/qwen_image/model_inference/Qwen-Image-EliGen.py)|[code](/examples/qwen_image/model_inference_low_vram/Qwen-Image-EliGen.py)|-|-|[code](/examples/qwen_image/model_training/lora/Qwen-Image-EliGen.sh)|[code](/examples/qwen_image/model_training/validate_lora/Qwen-Image-EliGen.py)|
 |[DiffSynth-Studio/Qwen-Image-EliGen-V2](https://www.modelscope.cn/models/DiffSynth-Studio/Qwen-Image-EliGen-V2)|[code](/examples/qwen_image/model_inference/Qwen-Image-EliGen-V2.py)|[code](/examples/qwen_image/model_inference_low_vram/Qwen-Image-EliGen-V2.py)|-|-|[code](/examples/qwen_image/model_training/lora/Qwen-Image-EliGen.sh)|[code](/examples/qwen_image/model_training/validate_lora/Qwen-Image-EliGen.py)|
 |[DiffSynth-Studio/Qwen-Image-EliGen-Poster](https://www.modelscope.cn/models/DiffSynth-Studio/Qwen-Image-EliGen-Poster)|[code](/examples/qwen_image/model_inference/Qwen-Image-EliGen-Poster.py)|[code](/examples/qwen_image/model_inference_low_vram/Qwen-Image-EliGen-Poster.py)|-|-|[code](/examples/qwen_image/model_training/lora/Qwen-Image-EliGen-Poster.sh)|[code](/examples/qwen_image/model_training/validate_lora/Qwen-Image-EliGen-Poster.py)|
@@ -644,7 +708,19 @@ LTX-2 的示例代码位于：[/examples/ltx2/](/examples/ltx2/)
 
 |模型 ID|额外参数|推理|低显存推理|全量训练|全量训练后验证|LoRA 训练|LoRA 训练后验证|
 |-|-|-|-|-|-|-|-|
+|[Lightricks/LTX-2.3: OneStagePipeline-I2AV](https://www.modelscope.cn/models/Lightricks/LTX-2.3)|`input_images`|[code](/examples/ltx2/model_inference/LTX-2.3-I2AV-OneStage.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2.3-I2AV-OneStage.py)|[code](/examples/ltx2/model_training/full/LTX-2.3-I2AV-splited.sh)|[code](/examples/ltx2/model_training/validate_full/LTX-2.3-I2AV.py)|[code](/examples/ltx2/model_training/lora/LTX-2.3-I2AV-splited.sh)|[code](/examples/ltx2/model_training/validate_lora/LTX-2.3-I2AV.py)|
+|[Lightricks/LTX-2.3: TwoStagePipeline-I2AV](https://www.modelscope.cn/models/Lightricks/LTX-2.3)|`input_images`|[code](/examples/ltx2/model_inference/LTX-2.3-I2AV-TwoStage.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2.3-I2AV-TwoStage.py)|-|-|-|-|
+|[Lightricks/LTX-2.3: DistilledPipeline-I2AV](https://www.modelscope.cn/models/Lightricks/LTX-2.3)|`input_images`|[code](/examples/ltx2/model_inference/LTX-2.3-I2AV-DistilledPipeline.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2.3-I2AV-DistilledPipeline.py)|-|-|-|-|
+|[Lightricks/LTX-2.3: OneStagePipeline-T2AV](https://www.modelscope.cn/models/Lightricks/LTX-2.3)||[code](/examples/ltx2/model_inference/LTX-2.3-T2AV-OneStage.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2.3-T2AV-OneStage.py)|[code](/examples/ltx2/model_training/full/LTX-2.3-T2AV-splited.sh)|[code](/examples/ltx2/model_training/validate_full/LTX-2.3-T2AV.py)|[code](/examples/ltx2/model_training/lora/LTX-2.3-T2AV-splited.sh)|[code](/examples/ltx2/model_training/validate_lora/LTX-2.3-T2AV.py)|
+|[Lightricks/LTX-2.3: TwoStagePipeline-T2AV](https://www.modelscope.cn/models/Lightricks/LTX-2.3)||[code](/examples/ltx2/model_inference/LTX-2.3-T2AV-TwoStage.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2.3-T2AV-TwoStage.py)|-|-|-|-|
+|[Lightricks/LTX-2.3: DistilledPipeline-T2AV](https://www.modelscope.cn/models/Lightricks/LTX-2.3)||[code](/examples/ltx2/model_inference/LTX-2.3-T2AV-DistilledPipeline.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2.3-T2AV-DistilledPipeline.py)|-|-|-|-|
+|[Lightricks/LTX-2.3: A2V](https://www.modelscope.cn/models/Lightricks/LTX-2.3)|`retake_audio`,`audio_sample_rate`,`retake_audio_regions`|[code](/examples/ltx2/model_inference/LTX-2.3-A2V-TwoStage.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2.3-A2V-TwoStage.py)|-|-|-|-|
+|[Lightricks/LTX-2.3: Retake](https://www.modelscope.cn/models/Lightricks/LTX-2.3)|`retake_video`,`retake_video_regions`,`retake_audio`,`audio_sample_rate`,`retake_audio_regions`|[code](/examples/ltx2/model_inference/LTX-2.3-T2AV-TwoStage-Retake.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2.3-T2AV-TwoStage-Retake.py)|-|-|-|-|
+|[Lightricks/LTX-2.3-22b-IC-LoRA-Union-Control](https://www.modelscope.cn/models/Lightricks/LTX-2.3-22b-IC-LoRA-Union-Control)|`in_context_videos`,`in_context_downsample_factor`|[code](/examples/ltx2/model_inference/LTX-2.3-T2AV-IC-LoRA-Union-Control.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2.3-T2AV-IC-LoRA-Union-Control.py)|-|-|[code](/examples/ltx2/model_training/lora/LTX-2.3-T2AV-IC-LoRA-splited.sh)|[code](/examples/ltx2/model_training/validate_lora/LTX-2.3-T2AV-IC-LoRA.py)|
+|[Lightricks/LTX-2.3-22b-IC-LoRA-Motion-Track-Control](https://www.modelscope.cn/models/Lightricks/LTX-2.3-22b-IC-LoRA-Motion-Track-Control)|`in_context_videos`,`in_context_downsample_factor`|[code](/examples/ltx2/model_inference/LTX-2.3-T2AV-IC-LoRA-Motion-Track-Control.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2.3-T2AV-IC-LoRA-Motion-Track-Control.py)|-|-|[code](/examples/ltx2/model_training/lora/LTX-2.3-T2AV-IC-LoRA-splited.sh)|[code](/examples/ltx2/model_training/validate_lora/LTX-2.3-T2AV-IC-LoRA.py)|
 |[Lightricks/LTX-2: OneStagePipeline-T2AV](https://www.modelscope.cn/models/Lightricks/LTX-2)||[code](/examples/ltx2/model_inference/LTX-2-T2AV-OneStage.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2-T2AV-OneStage.py)|[code](/examples/ltx2/model_training/full/LTX-2-T2AV-splited.sh)|[code](/examples/ltx2/model_training/validate_full/LTX-2-T2AV.py)|[code](/examples/ltx2/model_training/lora/LTX-2-T2AV-splited.sh)|[code](/examples/ltx2/model_training/validate_lora/LTX-2-T2AV.py)|
+|[Lightricks/LTX-2-19b-IC-LoRA-Union-Control](https://www.modelscope.cn/models/Lightricks/LTX-2-19b-IC-LoRA-Union-Control)|`in_context_videos`,`in_context_downsample_factor`|[code](/examples/ltx2/model_inference/LTX-2-T2AV-IC-LoRA-Union-Control.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2-T2AV-IC-LoRA-Union-Control.py)|-|-|[code](/examples/ltx2/model_training/lora/LTX-2-T2AV-IC-LoRA-splited.sh)|[code](/examples/ltx2/model_training/validate_lora/LTX-2-T2AV-IC-LoRA.py)|
+|[Lightricks/LTX-2-19b-IC-LoRA-Detailer](https://www.modelscope.cn/models/Lightricks/LTX-2-19b-IC-LoRA-Detailer)|`in_context_videos`,`in_context_downsample_factor`|[code](/examples/ltx2/model_inference/LTX-2-T2AV-IC-LoRA-Detailer.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2-T2AV-IC-LoRA-Detailer.py)|-|-|[code](/examples/ltx2/model_training/lora/LTX-2-T2AV-IC-LoRA-splited.sh)|[code](/examples/ltx2/model_training/validate_lora/LTX-2-T2AV-IC-LoRA.py)|
 |[Lightricks/LTX-2: TwoStagePipeline-T2AV](https://www.modelscope.cn/models/Lightricks/LTX-2)||[code](/examples/ltx2/model_inference/LTX-2-T2AV-TwoStage.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2-T2AV-TwoStage.py)|-|-|-|-|
 |[Lightricks/LTX-2: DistilledPipeline-T2AV](https://www.modelscope.cn/models/Lightricks/LTX-2)||[code](/examples/ltx2/model_inference/LTX-2-T2AV-DistilledPipeline.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2-T2AV-DistilledPipeline.py)|-|-|-|-|
 |[Lightricks/LTX-2: OneStagePipeline-I2AV](https://www.modelscope.cn/models/Lightricks/LTX-2)|`input_images`|[code](/examples/ltx2/model_inference/LTX-2-I2AV-OneStage.py)|[code](/examples/ltx2/model_inference_low_vram/LTX-2-I2AV-OneStage.py)|-|-|-|-|
@@ -792,6 +868,8 @@ Wan 的示例代码位于：[/examples/wanvideo/](/examples/wanvideo/)
 |[PAI/Wan2.2-Fun-A14B-InP](https://modelscope.cn/models/PAI/Wan2.2-Fun-A14B-InP)|`input_image`, `end_image`|[code](/examples/wanvideo/model_inference/Wan2.2-Fun-A14B-InP.py)|[code](/examples/wanvideo/model_training/full/Wan2.2-Fun-A14B-InP.sh)|[code](/examples/wanvideo/model_training/validate_full/Wan2.2-Fun-A14B-InP.py)|[code](/examples/wanvideo/model_training/lora/Wan2.2-Fun-A14B-InP.sh)|[code](/examples/wanvideo/model_training/validate_lora/Wan2.2-Fun-A14B-InP.py)|
 |[PAI/Wan2.2-Fun-A14B-Control](https://modelscope.cn/models/PAI/Wan2.2-Fun-A14B-Control)|`control_video`, `reference_image`|[code](/examples/wanvideo/model_inference/Wan2.2-Fun-A14B-Control.py)|[code](/examples/wanvideo/model_training/full/Wan2.2-Fun-A14B-Control.sh)|[code](/examples/wanvideo/model_training/validate_full/Wan2.2-Fun-A14B-Control.py)|[code](/examples/wanvideo/model_training/lora/Wan2.2-Fun-A14B-Control.sh)|[code](/examples/wanvideo/model_training/validate_lora/Wan2.2-Fun-A14B-Control.py)|
 |[PAI/Wan2.2-Fun-A14B-Control-Camera](https://modelscope.cn/models/PAI/Wan2.2-Fun-A14B-Control-Camera)|`control_camera_video`, `input_image`|[code](/examples/wanvideo/model_inference/Wan2.2-Fun-A14B-Control-Camera.py)|[code](/examples/wanvideo/model_training/full/Wan2.2-Fun-A14B-Control-Camera.sh)|[code](/examples/wanvideo/model_training/validate_full/Wan2.2-Fun-A14B-Control-Camera.py)|[code](/examples/wanvideo/model_training/lora/Wan2.2-Fun-A14B-Control-Camera.sh)|[code](/examples/wanvideo/model_training/validate_lora/Wan2.2-Fun-A14B-Control-Camera.py)|
+| [openmoss/MOVA-360p](https://modelscope.cn/models/openmoss/MOVA-360p) | `input_image` | [code](/examples/mova/model_inference/MOVA-360p-I2AV.py) | [code](/examples/mova/model_training/full/MOVA-360P-I2AV.sh) | [code](/examples/mova/model_training/validate_full/MOVA-360p-I2AV.py) | [code](/examples/mova/model_training/lora/MOVA-360P-I2AV.sh) | [code](/examples/mova/model_training/validate_lora/MOVA-360p-I2AV.py) |
+| [openmoss/MOVA-720p](https://modelscope.cn/models/openmoss/MOVA-720p) | `input_image` | [code](/examples/mova/model_inference/MOVA-720p-I2AV.py) | [code](/examples/mova/model_training/full/MOVA-720P-I2AV.sh) | [code](/examples/mova/model_training/validate_full/MOVA-720p-I2AV.py) | [code](/examples/mova/model_training/lora/MOVA-720P-I2AV.sh) | [code](/examples/mova/model_training/validate_lora/MOVA-720p-I2AV.py) |
 
 </details>
 
@@ -805,7 +883,7 @@ DiffSynth-Studio 不仅仅是一个工程化的模型框架，更是创新成果
 
 - 论文：[Spectral Evolution Search: Efficient Inference-Time Scaling for Reward-Aligned Image Generation
 ](https://arxiv.org/abs/2602.03208)
-- 代码样例：coming soon
+- 代码样例：[/docs/en/Research_Tutorial/inference_time_scaling.md](/docs/en/Research_Tutorial/inference_time_scaling.md)
 
 |FLUX.1-dev|FLUX.1-dev + SES|Qwen-Image|Qwen-Image + SES|
 |-|-|-|-|

diffsynth/configs/__init__.py

@@ -1,2 +1,2 @@
 from .model_configs import MODEL_CONFIGS
-from .vram_management_module_maps import VRAM_MANAGEMENT_MODULE_MAPS
+from .vram_management_module_maps import VRAM_MANAGEMENT_MODULE_MAPS, VERSION_CHECKER_MAPS

diffsynth/configs/model_configs.py

@@ -718,5 +718,156 @@ ltx2_series = [
         "model_name": "ltx2_latent_upsampler",
         "model_class": "diffsynth.models.ltx2_upsampler.LTX2LatentUpsampler",
     },
+    {
+        # Example: ModelConfig(model_id="Lightricks/LTX-2.3", origin_file_pattern="ltx-2.3-22b-dev.safetensors")
+        "model_hash": "f3a83ecf3995dcc4fae2d27e08ad5767",
+        "model_name": "ltx2_dit",
+        "model_class": "diffsynth.models.ltx2_dit.LTXModel",
+        "extra_kwargs": {"apply_gated_attention": True, "cross_attention_adaln": True, "caption_channels": None},
+        "state_dict_converter": "diffsynth.utils.state_dict_converters.ltx2_dit.LTXModelStateDictConverter",
+    },
+    {
+        # Example: ModelConfig(model_id="Lightricks/LTX-2.3", origin_file_pattern="ltx-2.3-22b-dev.safetensors")
+        "model_hash": "f3a83ecf3995dcc4fae2d27e08ad5767",
+        "model_name": "ltx2_video_vae_encoder",
+        "model_class": "diffsynth.models.ltx2_video_vae.LTX2VideoEncoder",
+        "extra_kwargs": {"encoder_version": "ltx-2.3"},
+        "state_dict_converter": "diffsynth.utils.state_dict_converters.ltx2_video_vae.LTX2VideoEncoderStateDictConverter",
+    },
+    {
+        # Example: ModelConfig(model_id="Lightricks/LTX-2.3", origin_file_pattern="ltx-2.3-22b-dev.safetensors")
+        "model_hash": "f3a83ecf3995dcc4fae2d27e08ad5767",
+        "model_name": "ltx2_video_vae_decoder",
+        "model_class": "diffsynth.models.ltx2_video_vae.LTX2VideoDecoder",
+        "extra_kwargs": {"decoder_version": "ltx-2.3"},
+        "state_dict_converter": "diffsynth.utils.state_dict_converters.ltx2_video_vae.LTX2VideoDecoderStateDictConverter",
+    },
+    {
+        # Example: ModelConfig(model_id="Lightricks/LTX-2.3", origin_file_pattern="ltx-2.3-22b-dev.safetensors")
+        "model_hash": "f3a83ecf3995dcc4fae2d27e08ad5767",
+        "model_name": "ltx2_audio_vae_decoder",
+        "model_class": "diffsynth.models.ltx2_audio_vae.LTX2AudioDecoder",
+        "state_dict_converter": "diffsynth.utils.state_dict_converters.ltx2_audio_vae.LTX2AudioDecoderStateDictConverter",
+    },
+    {
+        # Example: ModelConfig(model_id="Lightricks/LTX-2.3", origin_file_pattern="ltx-2.3-22b-dev.safetensors")
+        "model_hash": "f3a83ecf3995dcc4fae2d27e08ad5767",
+        "model_name": "ltx2_audio_vocoder",
+        "model_class": "diffsynth.models.ltx2_audio_vae.LTX2VocoderWithBWE",
+        "state_dict_converter": "diffsynth.utils.state_dict_converters.ltx2_audio_vae.LTX2VocoderStateDictConverter",
+    },
+    {
+        # Example: ModelConfig(model_id="Lightricks/LTX-2.3", origin_file_pattern="ltx-2.3-22b-dev.safetensors")
+        "model_hash": "f3a83ecf3995dcc4fae2d27e08ad5767",
+        "model_name": "ltx2_audio_vae_encoder",
+        "model_class": "diffsynth.models.ltx2_audio_vae.LTX2AudioEncoder",
+        "state_dict_converter": "diffsynth.utils.state_dict_converters.ltx2_audio_vae.LTX2AudioEncoderStateDictConverter",
+    },
+    {
+        # Example: ModelConfig(model_id="Lightricks/LTX-2.3", origin_file_pattern="ltx-2.3-22b-dev.safetensors")
+        "model_hash": "f3a83ecf3995dcc4fae2d27e08ad5767",
+        "model_name": "ltx2_text_encoder_post_modules",
+        "model_class": "diffsynth.models.ltx2_text_encoder.LTX2TextEncoderPostModules",
+        "extra_kwargs": {"separated_audio_video": True, "embedding_dim_gemma": 3840, "num_layers_gemma": 49, "video_attention_heads": 32, "video_attention_head_dim": 128, "audio_attention_heads": 32, "audio_attention_head_dim": 64, "num_connector_layers": 8, "apply_gated_attention": True},
+        "state_dict_converter": "diffsynth.utils.state_dict_converters.ltx2_text_encoder.LTX2TextEncoderPostModulesStateDictConverter",
+    },
+    {
+        # Example: ModelConfig(model_id="Lightricks/LTX-2.3", origin_file_pattern="ltx-2.3-spatial-upscaler-x2-1.0.safetensors")
+        "model_hash": "aed408774d694a2452f69936c32febb5",
+        "model_name": "ltx2_latent_upsampler",
+        "model_class": "diffsynth.models.ltx2_upsampler.LTX2LatentUpsampler",
+        "extra_kwargs": {"rational_resampler": False},
+    },
+    {
+        # Example: ModelConfig(model_id="DiffSynth-Studio/LTX-2.3-Repackage", origin_file_pattern="transformer.safetensors")
+        "model_hash": "1c55afad76ed33c112a2978550b524d1",
+        "model_name": "ltx2_dit",
+        "model_class": "diffsynth.models.ltx2_dit.LTXModel",
+        "extra_kwargs": {"apply_gated_attention": True, "cross_attention_adaln": True, "caption_channels": None},
+        "state_dict_converter": "diffsynth.utils.state_dict_converters.ltx2_dit.LTXModelStateDictConverter",
+    },
+    {
+        # Example: ModelConfig(model_id="DiffSynth-Studio/LTX-2.3-Repackage", origin_file_pattern="video_vae_encoder.safetensors")
+        "model_hash": "eecdc07c2ec30863b8a2b8b2134036cf",
+        "model_name": "ltx2_video_vae_encoder",
+        "model_class": "diffsynth.models.ltx2_video_vae.LTX2VideoEncoder",
+        "extra_kwargs": {"encoder_version": "ltx-2.3"},
+        "state_dict_converter": "diffsynth.utils.state_dict_converters.ltx2_video_vae.LTX2VideoEncoderStateDictConverter",
+    },
+    {
+        # Example: ModelConfig(model_id="DiffSynth-Studio/LTX-2.3-Repackage", origin_file_pattern="video_vae_decoder.safetensors")
+        "model_hash": "deda2f542e17ee25bc8c38fd605316ea",
+        "model_name": "ltx2_video_vae_decoder",
+        "model_class": "diffsynth.models.ltx2_video_vae.LTX2VideoDecoder",
+        "extra_kwargs": {"decoder_version": "ltx-2.3"},
+        "state_dict_converter": "diffsynth.utils.state_dict_converters.ltx2_video_vae.LTX2VideoDecoderStateDictConverter",
+    },
+    {
+        # Example: ModelConfig(model_id="DiffSynth-Studio/LTX-2.3-Repackage", origin_file_pattern="audio_vocoder.safetensors")
+        "model_hash": "7d7823dde8f1ea0b50fb07ac329dd4cb",
+        "model_name": "ltx2_audio_vae_decoder",
+        "model_class": "diffsynth.models.ltx2_audio_vae.LTX2AudioDecoder",
+        "state_dict_converter": "diffsynth.utils.state_dict_converters.ltx2_audio_vae.LTX2AudioDecoderStateDictConverter",
+    },
+    {
+        # Example: ModelConfig(model_id="DiffSynth-Studio/LTX-2.3-Repackage", origin_file_pattern="audio_vae_encoder.safetensors")
+        "model_hash": "29338f3b95e7e312a3460a482e4f4554",
+        "model_name": "ltx2_audio_vae_encoder",
+        "model_class": "diffsynth.models.ltx2_audio_vae.LTX2AudioEncoder",
+        "state_dict_converter": "diffsynth.utils.state_dict_converters.ltx2_audio_vae.LTX2AudioEncoderStateDictConverter",
+    },
+    {
+        # Example: ModelConfig(model_id="DiffSynth-Studio/LTX-2.3-Repackage", origin_file_pattern="audio_vocoder.safetensors")
+        "model_hash": "cd436c99e69ec5c80f050f0944f02a15",
+        "model_name": "ltx2_audio_vocoder",
+        "model_class": "diffsynth.models.ltx2_audio_vae.LTX2VocoderWithBWE",
+        "state_dict_converter": "diffsynth.utils.state_dict_converters.ltx2_audio_vae.LTX2VocoderStateDictConverter",
+    },
+    {
+        # Example: ModelConfig(model_id="DiffSynth-Studio/LTX-2.3-Repackage", origin_file_pattern="text_encoder_post_modules.safetensors")
+        "model_hash": "05da2aab1c4b061f72c426311c165a43",
+        "model_name": "ltx2_text_encoder_post_modules",
+        "model_class": "diffsynth.models.ltx2_text_encoder.LTX2TextEncoderPostModules",
+        "extra_kwargs": {"separated_audio_video": True, "embedding_dim_gemma": 3840, "num_layers_gemma": 49, "video_attention_heads": 32, "video_attention_head_dim": 128, "audio_attention_heads": 32, "audio_attention_head_dim": 64, "num_connector_layers": 8, "apply_gated_attention": True},
+        "state_dict_converter": "diffsynth.utils.state_dict_converters.ltx2_text_encoder.LTX2TextEncoderPostModulesStateDictConverter",
+    },
 ]
-MODEL_CONFIGS = qwen_image_series + wan_series + flux_series + flux2_series + z_image_series + ltx2_series
+anima_series = [
+    {
+        # Example: ModelConfig(model_id="circlestone-labs/Anima", origin_file_pattern="split_files/vae/qwen_image_vae.safetensors")
+        "model_hash": "a9995952c2d8e63cf82e115005eb61b9",
+        "model_name": "z_image_text_encoder",
+        "model_class": "diffsynth.models.z_image_text_encoder.ZImageTextEncoder",
+        "extra_kwargs": {"model_size": "0.6B"},
+    },
+    {
+        # Example: ModelConfig(model_id="circlestone-labs/Anima", origin_file_pattern="split_files/diffusion_models/anima-preview.safetensors")
+        "model_hash": "417673936471e79e31ed4d186d7a3f4a",
+        "model_name": "anima_dit",
+        "model_class": "diffsynth.models.anima_dit.AnimaDiT",
+        "state_dict_converter": "diffsynth.utils.state_dict_converters.anima_dit.AnimaDiTStateDictConverter",
+    }
+]
+
+mova_series = [
+    # Example: ModelConfig(model_id="openmoss/MOVA-720p", origin_file_pattern="audio_dit/diffusion_pytorch_model.safetensors")
+    {
+        "model_hash": "8c57e12790e2c45a64817e0ce28cde2f",
+        "model_name": "mova_audio_dit",
+        "model_class": "diffsynth.models.mova_audio_dit.MovaAudioDit",
+        "extra_kwargs": {'has_image_input': False, 'patch_size': [1], 'in_dim': 128, 'dim': 1536, 'ffn_dim': 8960, 'freq_dim': 256, 'text_dim': 4096, 'out_dim': 128, 'num_heads': 12, 'num_layers': 30, 'eps': 1e-06}
+    },
+    # Example: ModelConfig(model_id="openmoss/MOVA-720p", origin_file_pattern="audio_vae/diffusion_pytorch_model.safetensors")
+    {
+        "model_hash": "418517fb2b4e919d2cac8f314fcf82ac",
+        "model_name": "mova_audio_vae",
+        "model_class": "diffsynth.models.mova_audio_vae.DacVAE",
+    },
+    # Example: ModelConfig(model_id="openmoss/MOVA-720p", origin_file_pattern="dual_tower_bridge/diffusion_pytorch_model.safetensors")
+    {
+        "model_hash": "d1139dbbc8b4ab53cf4b4243d57bbceb",
+        "model_name": "mova_dual_tower_bridge",
+        "model_class": "diffsynth.models.mova_dual_tower_bridge.DualTowerConditionalBridge",
+    },
+]
+MODEL_CONFIGS = qwen_image_series + wan_series + flux_series + flux2_series + z_image_series + ltx2_series + anima_series + mova_series

diffsynth/configs/vram_management_module_maps.py

@@ -243,4 +243,42 @@ VRAM_MANAGEMENT_MODULE_MAPS = {
         "transformers.models.gemma3.modeling_gemma3.Gemma3RMSNorm": "diffsynth.core.vram.layers.AutoWrappedModule",
         "transformers.models.gemma3.modeling_gemma3.Gemma3TextScaledWordEmbedding": "diffsynth.core.vram.layers.AutoWrappedModule",
     },
+    "diffsynth.models.anima_dit.AnimaDiT": {
+        "torch.nn.Linear": "diffsynth.core.vram.layers.AutoWrappedLinear",
+        "torch.nn.LayerNorm": "diffsynth.core.vram.layers.AutoWrappedModule",
+        "torch.nn.RMSNorm": "diffsynth.core.vram.layers.AutoWrappedModule",
+        "torch.nn.Embedding": "diffsynth.core.vram.layers.AutoWrappedModule",
+    },
+    "diffsynth.models.mova_audio_dit.MovaAudioDit": {
+        "diffsynth.models.wan_video_dit.DiTBlock": "diffsynth.core.vram.layers.AutoWrappedNonRecurseModule",
+        "diffsynth.models.wan_video_dit.Head": "diffsynth.core.vram.layers.AutoWrappedModule",
+        "torch.nn.Linear": "diffsynth.core.vram.layers.AutoWrappedLinear",
+        "torch.nn.Conv1d": "diffsynth.core.vram.layers.AutoWrappedModule",
+        "torch.nn.LayerNorm": "diffsynth.core.vram.layers.AutoWrappedModule",
+        "diffsynth.models.wan_video_dit.RMSNorm": "diffsynth.core.vram.layers.AutoWrappedModule",
+    },
+    "diffsynth.models.mova_dual_tower_bridge.DualTowerConditionalBridge": {
+        "torch.nn.Linear": "diffsynth.core.vram.layers.AutoWrappedLinear",
+        "torch.nn.LayerNorm": "diffsynth.core.vram.layers.AutoWrappedModule",
+        "diffsynth.models.wan_video_dit.RMSNorm": "diffsynth.core.vram.layers.AutoWrappedModule",
+    },
+    "diffsynth.models.mova_audio_vae.DacVAE": {
+        "diffsynth.models.mova_audio_vae.Snake1d": "diffsynth.core.vram.layers.AutoWrappedModule",
+        "torch.nn.Conv1d": "diffsynth.core.vram.layers.AutoWrappedModule",
+        "torch.nn.ConvTranspose1d": "diffsynth.core.vram.layers.AutoWrappedModule",
+    },
 }
+
+def QwenImageTextEncoder_Module_Map_Updater():
+    current = VRAM_MANAGEMENT_MODULE_MAPS["diffsynth.models.qwen_image_text_encoder.QwenImageTextEncoder"]
+    from packaging import version
+    import transformers
+    if version.parse(transformers.__version__) >= version.parse("5.2.0"):
+        # The Qwen2RMSNorm in transformers 5.2.0+ has been renamed to Qwen2_5_VLRMSNorm, so we need to update the module map accordingly
+        current.pop("transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2RMSNorm", None)
+        current["transformers.models.qwen2_5_vl.modeling_qwen2_5_vl.Qwen2_5_VLRMSNorm"] = "diffsynth.core.vram.layers.AutoWrappedModule"
+    return current
+
+VERSION_CHECKER_MAPS = {
+    "diffsynth.models.qwen_image_text_encoder.QwenImageTextEncoder": QwenImageTextEncoder_Module_Map_Updater,
+}

diffsynth/core/data/operators.py

@@ -1,6 +1,8 @@
+import math
 import torch, torchvision, imageio, os
 import imageio.v3 as iio
 from PIL import Image
+import torchaudio
 
 
 class DataProcessingPipeline:
@@ -105,27 +107,59 @@ class ToList(DataProcessingOperator):
         return [data]
     
 
-class LoadVideo(DataProcessingOperator):
-    def __init__(self, num_frames=81, time_division_factor=4, time_division_remainder=1, frame_processor=lambda x: x):
+class FrameSamplerByRateMixin:
+    def __init__(self, num_frames=81, time_division_factor=4, time_division_remainder=1, frame_rate=24, fix_frame_rate=False):
         self.num_frames = num_frames
         self.time_division_factor = time_division_factor
         self.time_division_remainder = time_division_remainder
-        # frame_processor is build in the video loader for high efficiency.
-        self.frame_processor = frame_processor
-        
+        self.frame_rate = frame_rate
+        self.fix_frame_rate = fix_frame_rate
+
+    def get_reader(self, data: str):
+        return imageio.get_reader(data)
+
+    def get_available_num_frames(self, reader):
+        if not self.fix_frame_rate:
+            return reader.count_frames()
+        meta_data = reader.get_meta_data()
+        total_original_frames = int(reader.count_frames())
+        duration = meta_data["duration"] if "duration" in meta_data else total_original_frames / meta_data['fps']
+        total_available_frames = math.floor(duration * self.frame_rate)
+        return int(total_available_frames)
+
     def get_num_frames(self, reader):
         num_frames = self.num_frames
-        if int(reader.count_frames()) < num_frames:
-            num_frames = int(reader.count_frames())
+        total_frames = self.get_available_num_frames(reader)
+        if int(total_frames) < num_frames:
+            num_frames = total_frames
             while num_frames > 1 and num_frames % self.time_division_factor != self.time_division_remainder:
                 num_frames -= 1
         return num_frames
-        
+
+    def map_single_frame_id(self, new_sequence_id: int, raw_frame_rate: float, total_raw_frames: int) -> int:
+        if not self.fix_frame_rate:
+            return new_sequence_id
+        target_time_in_seconds = new_sequence_id / self.frame_rate
+        raw_frame_index_float = target_time_in_seconds * raw_frame_rate
+        frame_id = int(round(raw_frame_index_float))        
+        frame_id = min(frame_id, total_raw_frames - 1)
+        return frame_id
+
+
+class LoadVideo(DataProcessingOperator, FrameSamplerByRateMixin):
+    def __init__(self, num_frames=81, time_division_factor=4, time_division_remainder=1, frame_processor=lambda x: x, frame_rate=24, fix_frame_rate=False):
+        FrameSamplerByRateMixin.__init__(self, num_frames, time_division_factor, time_division_remainder, frame_rate, fix_frame_rate)
+        # frame_processor is build in the video loader for high efficiency.
+        self.frame_processor = frame_processor
+
     def __call__(self, data: str):
-        reader = imageio.get_reader(data)
+        reader = self.get_reader(data)
+        raw_frame_rate = reader.get_meta_data()['fps']
         num_frames = self.get_num_frames(reader)
+        total_raw_frames = reader.count_frames()
         frames = []
         for frame_id in range(num_frames):
+            frame_id = self.map_single_frame_id(frame_id, raw_frame_rate, total_raw_frames)
             frame = reader.get_data(frame_id)
             frame = Image.fromarray(frame)
             frame = self.frame_processor(frame)
@@ -149,7 +183,7 @@ class LoadGIF(DataProcessingOperator):
         self.time_division_remainder = time_division_remainder
         # frame_processor is build in the video loader for high efficiency.
         self.frame_processor = frame_processor
-        
+
     def get_num_frames(self, path):
         num_frames = self.num_frames
         images = iio.imread(path, mode="RGB")
@@ -220,14 +254,17 @@ class LoadAudio(DataProcessingOperator):
         return input_audio
 
 
-class LoadAudioWithTorchaudio(DataProcessingOperator):
-    def __init__(self, duration=5):
-        self.duration = duration
+class LoadAudioWithTorchaudio(DataProcessingOperator, FrameSamplerByRateMixin):
+
+    def __init__(self, num_frames=121, time_division_factor=8, time_division_remainder=1, frame_rate=24, fix_frame_rate=True):
+        FrameSamplerByRateMixin.__init__(self, num_frames, time_division_factor, time_division_remainder, frame_rate, fix_frame_rate)
 
     def __call__(self, data: str):
-        import torchaudio
+        reader = self.get_reader(data)
+        num_frames = self.get_num_frames(reader)
+        duration = num_frames / self.frame_rate
         waveform, sample_rate = torchaudio.load(data)
-        target_samples = int(self.duration * sample_rate)
+        target_samples = int(duration * sample_rate)
         current_samples = waveform.shape[-1]
         if current_samples > target_samples:
             waveform = waveform[..., :target_samples]

diffsynth/core/data/unified_dataset.py

@@ -42,6 +42,7 @@ class UnifiedDataset(torch.utils.data.Dataset):
         max_pixels=1920*1080, height=None, width=None,
         height_division_factor=16, width_division_factor=16,
         num_frames=81, time_division_factor=4, time_division_remainder=1,
+        frame_rate=24, fix_frame_rate=False,
     ):
         return RouteByType(operator_map=[
             (str, ToAbsolutePath(base_path) >> RouteByExtensionName(operator_map=[
@@ -53,6 +54,7 @@ class UnifiedDataset(torch.utils.data.Dataset):
                 (("mp4", "avi", "mov", "wmv", "mkv", "flv", "webm"), LoadVideo(
                     num_frames, time_division_factor, time_division_remainder,
                     frame_processor=ImageCropAndResize(height, width, max_pixels, height_division_factor, width_division_factor),
+                    frame_rate=frame_rate, fix_frame_rate=fix_frame_rate,
                 )),
             ])),
         ])

diffsynth/core/gradient/gradient_checkpoint.py

@@ -1,12 +1,32 @@
 import torch
 
 
+try:
+    import deepspeed
+    _HAS_DEEPSPEED = True
+except ModuleNotFoundError:
+    _HAS_DEEPSPEED = False
+
+
 def create_custom_forward(module):
     def custom_forward(*inputs, **kwargs):
         return module(*inputs, **kwargs)
     return custom_forward
 
 
+def create_custom_forward_use_reentrant(module):
+    def custom_forward(*inputs):
+        return module(*inputs)
+    return custom_forward
+
+
+def judge_args_requires_grad(*args):
+    for arg in args:
+        if isinstance(arg, torch.Tensor) and arg.requires_grad:
+            return True
+    return False
+
+
 def gradient_checkpoint_forward(
     model,
     use_gradient_checkpointing,
@@ -14,6 +34,17 @@ def gradient_checkpoint_forward(
     *args,
     **kwargs,
 ):
+    if use_gradient_checkpointing and _HAS_DEEPSPEED and deepspeed.checkpointing.is_configured():
+        all_args = args + tuple(kwargs.values())
+        if not judge_args_requires_grad(*all_args):
+            # get the first grad_enabled tensor from un_checkpointed forward
+            model_output = model(*args, **kwargs)
+        else:
+            model_output = deepspeed.checkpointing.checkpoint(
+                create_custom_forward_use_reentrant(model),
+                *all_args,
+            )
+        return model_output
     if use_gradient_checkpointing_offload:
         with torch.autograd.graph.save_on_cpu():
             model_output = torch.utils.checkpoint.checkpoint(

diffsynth/core/vram/layers.py

@@ -417,7 +417,7 @@ class AutoWrappedLinear(torch.nn.Linear, AutoTorchModule):
     def lora_forward(self, x, out):
         if self.lora_merger is None:
             for lora_A, lora_B in zip(self.lora_A_weights, self.lora_B_weights):
-                out = out + x @ lora_A.T @ lora_B.T
+                out = out + x @ lora_A.T.to(device=x.device, dtype=x.dtype) @ lora_B.T.to(device=x.device, dtype=x.dtype)
         else:
             lora_output = []
             for lora_A, lora_B in zip(self.lora_A_weights, self.lora_B_weights):

diffsynth/diffusion/base_pipeline.py

@@ -94,20 +94,23 @@ class BasePipeline(torch.nn.Module):
         return self
 
 
-    def check_resize_height_width(self, height, width, num_frames=None):
+    def check_resize_height_width(self, height, width, num_frames=None, verbose=1):
         # Shape check
         if height % self.height_division_factor != 0:
             height = (height + self.height_division_factor - 1) // self.height_division_factor * self.height_division_factor
-            print(f"height % {self.height_division_factor} != 0. We round it up to {height}.")
+            if verbose > 0:
+                print(f"height % {self.height_division_factor} != 0. We round it up to {height}.")
         if width % self.width_division_factor != 0:
             width = (width + self.width_division_factor - 1) // self.width_division_factor * self.width_division_factor
-            print(f"width % {self.width_division_factor} != 0. We round it up to {width}.")
+            if verbose > 0:
+                print(f"width % {self.width_division_factor} != 0. We round it up to {width}.")
         if num_frames is None:
             return height, width
         else:
             if num_frames % self.time_division_factor != self.time_division_remainder:
                 num_frames = (num_frames + self.time_division_factor - 1) // self.time_division_factor * self.time_division_factor + self.time_division_remainder
-                print(f"num_frames % {self.time_division_factor} != {self.time_division_remainder}. We round it up to {num_frames}.")
+                if verbose > 0:
+                    print(f"num_frames % {self.time_division_factor} != {self.time_division_remainder}. We round it up to {num_frames}.")
             return height, width, num_frames
 
 
@@ -144,6 +147,12 @@ class BasePipeline(torch.nn.Module):
         video = [self.vae_output_to_image(image, pattern="H W C", min_value=min_value, max_value=max_value) for image in vae_output]
         return video
 
+    def output_audio_format_check(self, audio_output):
+        # output standard foramt: [C, T], output dtype: float()
+        # remove batch dim
+        if audio_output.ndim == 3:
+            audio_output = audio_output.squeeze(0)
+        return audio_output.float()
 
     def load_models_to_device(self, model_names):
         if self.vram_management_enabled:

diffsynth/diffusion/loss.py

@@ -121,7 +121,9 @@ class TrajectoryImitationLoss(torch.nn.Module):
                 progress_id_teacher = torch.argmin((timesteps_teacher - pipe.scheduler.timesteps[progress_id + 1]).abs())
                 latents_ = trajectory_teacher[progress_id_teacher]
             
-            target = (latents_ - inputs_shared["latents"]) / (sigma_ - sigma)
+            denom = sigma_ - sigma
+            denom = torch.sign(denom) * torch.clamp(denom.abs(), min=1e-6)
+            target = (latents_ - inputs_shared["latents"]) / denom
             loss = loss + torch.nn.functional.mse_loss(noise_pred.float(), target.float()) * pipe.scheduler.training_weight(timestep)
         return loss
     

diffsynth/diffusion/runner.py

@@ -29,7 +29,7 @@ def launch_training_task(
     dataloader = torch.utils.data.DataLoader(dataset, shuffle=True, collate_fn=lambda x: x[0], num_workers=num_workers)
     model.to(device=accelerator.device)
     model, optimizer, dataloader, scheduler = accelerator.prepare(model, optimizer, dataloader, scheduler)
-    
+    initialize_deepspeed_gradient_checkpointing(accelerator)
     for epoch_id in range(num_epochs):
         for data in tqdm(dataloader):
             with accelerator.accumulate(model):
@@ -70,3 +70,19 @@ def launch_data_process_task(
                 save_path = os.path.join(model_logger.output_path, str(accelerator.process_index), f"{data_id}.pth")
                 data = model(data)
                 torch.save(data, save_path)
+
+
+def initialize_deepspeed_gradient_checkpointing(accelerator: Accelerator):
+    if getattr(accelerator.state, "deepspeed_plugin", None) is not None:
+        ds_config = accelerator.state.deepspeed_plugin.deepspeed_config
+        if "activation_checkpointing" in ds_config:
+            import deepspeed
+            act_config = ds_config["activation_checkpointing"]
+            deepspeed.checkpointing.configure(
+                mpu_=None, 
+                partition_activations=act_config.get("partition_activations", False),
+                checkpoint_in_cpu=act_config.get("cpu_checkpointing", False),
+                contiguous_checkpointing=act_config.get("contiguous_memory_optimization", False)
+            )
+        else:
+            print("Do not find activation_checkpointing config in deepspeed config, skip initializing deepspeed gradient checkpointing.")

diffsynth/diffusion/training_module.py

@@ -1,9 +1,32 @@
-import torch, json, os
+import torch, json, os, inspect
 from ..core import ModelConfig, load_state_dict
 from ..utils.controlnet import ControlNetInput
+from .base_pipeline import PipelineUnit
 from peft import LoraConfig, inject_adapter_in_model
 
 
+class GeneralUnit_RemoveCache(PipelineUnit):
+    def __init__(self, required_params=tuple(), force_remove_params_shared=tuple(), force_remove_params_posi=tuple(), force_remove_params_nega=tuple()):
+        super().__init__(take_over=True)
+        self.required_params = required_params
+        self.force_remove_params_shared = force_remove_params_shared
+        self.force_remove_params_posi = force_remove_params_posi
+        self.force_remove_params_nega = force_remove_params_nega
+
+    def process_params(self, inputs, required_params, force_remove_params):
+        inputs_ = {}
+        for name, param in inputs.items():
+            if name in required_params and name not in force_remove_params:
+                inputs_[name] = param
+        return inputs_
+
+    def process(self, pipe, inputs_shared, inputs_posi, inputs_nega):
+        inputs_shared = self.process_params(inputs_shared, self.required_params, self.force_remove_params_shared)
+        inputs_posi = self.process_params(inputs_posi, self.required_params, self.force_remove_params_posi)
+        inputs_nega = self.process_params(inputs_nega, self.required_params, self.force_remove_params_nega)
+        return inputs_shared, inputs_posi, inputs_nega
+
+
 class DiffusionTrainingModule(torch.nn.Module):
     def __init__(self):
         super().__init__()
@@ -231,14 +254,30 @@ class DiffusionTrainingModule(torch.nn.Module):
             setattr(pipe, lora_base_model, model)
 
 
-    def split_pipeline_units(self, task, pipe, trainable_models=None, lora_base_model=None):
+    def split_pipeline_units(
+        self, task, pipe,
+        trainable_models=None, lora_base_model=None,
+        # TODO: set `remove_unnecessary_params` to `True` by default
+        remove_unnecessary_params=False,
+        # TODO: move `loss_required_params` to `loss.py`
+        loss_required_params=("input_latents", "max_timestep_boundary", "min_timestep_boundary", "first_frame_latents", "video_latents", "audio_input_latents", "num_inference_steps"),
+        force_remove_params_shared=tuple(),
+        force_remove_params_posi=tuple(),
+        force_remove_params_nega=tuple(),
+    ):
         models_require_backward = []
         if trainable_models is not None:
             models_require_backward += trainable_models.split(",")
         if lora_base_model is not None:
             models_require_backward += [lora_base_model]
         if task.endswith(":data_process"):
-            _, pipe.units = pipe.split_pipeline_units(models_require_backward)
+            other_units, pipe.units = pipe.split_pipeline_units(models_require_backward)
+            if remove_unnecessary_params:
+                required_params = list(loss_required_params) + [i for i in inspect.signature(self.pipe.model_fn).parameters]
+                for unit in other_units:
+                    required_params.extend(unit.fetch_input_params())
+                required_params = sorted(list(set(required_params)))
+                pipe.units.append(GeneralUnit_RemoveCache(required_params, force_remove_params_shared, force_remove_params_posi, force_remove_params_nega))
         elif task.endswith(":train"):
             pipe.units, _ = pipe.split_pipeline_units(models_require_backward)
         return pipe

diffsynth/models/ltx2_audio_vae.py

@@ -1279,9 +1279,268 @@ class LTX2AudioDecoder(torch.nn.Module):
         return torch.tanh(h) if self.tanh_out else h
 
 
+def get_padding(kernel_size: int, dilation: int = 1) -> int:
+    return int((kernel_size * dilation - dilation) / 2)
+
+
+# ---------------------------------------------------------------------------
+# Anti-aliased resampling helpers (kaiser-sinc filters) for BigVGAN v2
+# Adopted from https://github.com/NVIDIA/BigVGAN
+# ---------------------------------------------------------------------------
+
+
+def _sinc(x: torch.Tensor) -> torch.Tensor:
+    return torch.where(
+        x == 0,
+        torch.tensor(1.0, device=x.device, dtype=x.dtype),
+        torch.sin(math.pi * x) / math.pi / x,
+    )
+
+
+def kaiser_sinc_filter1d(cutoff: float, half_width: float, kernel_size: int) -> torch.Tensor:
+    even = kernel_size % 2 == 0
+    half_size = kernel_size // 2
+    delta_f = 4 * half_width
+    amplitude = 2.285 * (half_size - 1) * math.pi * delta_f + 7.95
+    if amplitude > 50.0:
+        beta = 0.1102 * (amplitude - 8.7)
+    elif amplitude >= 21.0:
+        beta = 0.5842 * (amplitude - 21) ** 0.4 + 0.07886 * (amplitude - 21.0)
+    else:
+        beta = 0.0
+    window = torch.kaiser_window(kernel_size, beta=beta, periodic=False)
+    time = torch.arange(-half_size, half_size) + 0.5 if even else torch.arange(kernel_size) - half_size
+    if cutoff == 0:
+        filter_ = torch.zeros_like(time)
+    else:
+        filter_ = 2 * cutoff * window * _sinc(2 * cutoff * time)
+        filter_ /= filter_.sum()
+    return filter_.view(1, 1, kernel_size)
+
+
+class LowPassFilter1d(nn.Module):
+    def __init__(
+        self,
+        cutoff: float = 0.5,
+        half_width: float = 0.6,
+        stride: int = 1,
+        padding: bool = True,
+        padding_mode: str = "replicate",
+        kernel_size: int = 12,
+    ) -> None:
+        super().__init__()
+        if cutoff < -0.0:
+            raise ValueError("Minimum cutoff must be larger than zero.")
+        if cutoff > 0.5:
+            raise ValueError("A cutoff above 0.5 does not make sense.")
+        self.kernel_size = kernel_size
+        self.even = kernel_size % 2 == 0
+        self.pad_left = kernel_size // 2 - int(self.even)
+        self.pad_right = kernel_size // 2
+        self.stride = stride
+        self.padding = padding
+        self.padding_mode = padding_mode
+        self.register_buffer("filter", kaiser_sinc_filter1d(cutoff, half_width, kernel_size))
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        _, n_channels, _ = x.shape
+        if self.padding:
+            x = F.pad(x, (self.pad_left, self.pad_right), mode=self.padding_mode)
+        return F.conv1d(x, self.filter.expand(n_channels, -1, -1), stride=self.stride, groups=n_channels)
+
+
+class UpSample1d(nn.Module):
+    def __init__(
+        self,
+        ratio: int = 2,
+        kernel_size: int | None = None,
+        persistent: bool = True,
+        window_type: str = "kaiser",
+    ) -> None:
+        super().__init__()
+        self.ratio = ratio
+        self.stride = ratio
+
+        if window_type == "hann":
+            # Hann-windowed sinc filter equivalent to torchaudio.functional.resample
+            rolloff = 0.99
+            lowpass_filter_width = 6
+            width = math.ceil(lowpass_filter_width / rolloff)
+            self.kernel_size = 2 * width * ratio + 1
+            self.pad = width
+            self.pad_left = 2 * width * ratio
+            self.pad_right = self.kernel_size - ratio
+            time_axis = (torch.arange(self.kernel_size) / ratio - width) * rolloff
+            time_clamped = time_axis.clamp(-lowpass_filter_width, lowpass_filter_width)
+            window = torch.cos(time_clamped * math.pi / lowpass_filter_width / 2) ** 2
+            sinc_filter = (torch.sinc(time_axis) * window * rolloff / ratio).view(1, 1, -1)
+        else:
+            # Kaiser-windowed sinc filter (BigVGAN default).
+            self.kernel_size = int(6 * ratio // 2) * 2 if kernel_size is None else kernel_size
+            self.pad = self.kernel_size // ratio - 1
+            self.pad_left = self.pad * self.stride + (self.kernel_size - self.stride) // 2
+            self.pad_right = self.pad * self.stride + (self.kernel_size - self.stride + 1) // 2
+            sinc_filter = kaiser_sinc_filter1d(
+                cutoff=0.5 / ratio,
+                half_width=0.6 / ratio,
+                kernel_size=self.kernel_size,
+            )
+
+        self.register_buffer("filter", sinc_filter, persistent=persistent)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        _, n_channels, _ = x.shape
+        x = F.pad(x, (self.pad, self.pad), mode="replicate")
+        filt = self.filter.to(dtype=x.dtype, device=x.device).expand(n_channels, -1, -1)
+        x = self.ratio * F.conv_transpose1d(x, filt, stride=self.stride, groups=n_channels)
+        return x[..., self.pad_left : -self.pad_right]
+
+
+class DownSample1d(nn.Module):
+    def __init__(self, ratio: int = 2, kernel_size: int | None = None) -> None:
+        super().__init__()
+        self.ratio = ratio
+        self.kernel_size = int(6 * ratio // 2) * 2 if kernel_size is None else kernel_size
+        self.lowpass = LowPassFilter1d(
+            cutoff=0.5 / ratio,
+            half_width=0.6 / ratio,
+            stride=ratio,
+            kernel_size=self.kernel_size,
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.lowpass(x)
+
+
+class Activation1d(nn.Module):
+    def __init__(
+        self,
+        activation: nn.Module,
+        up_ratio: int = 2,
+        down_ratio: int = 2,
+        up_kernel_size: int = 12,
+        down_kernel_size: int = 12,
+    ) -> None:
+        super().__init__()
+        self.act = activation
+        self.upsample = UpSample1d(up_ratio, up_kernel_size)
+        self.downsample = DownSample1d(down_ratio, down_kernel_size)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.upsample(x)
+        x = self.act(x)
+        return self.downsample(x)
+
+
+class Snake(nn.Module):
+    def __init__(
+        self,
+        in_features: int,
+        alpha: float = 1.0,
+        alpha_trainable: bool = True,
+        alpha_logscale: bool = True,
+    ) -> None:
+        super().__init__()
+        self.alpha_logscale = alpha_logscale
+        self.alpha = nn.Parameter(torch.zeros(in_features) if alpha_logscale else torch.ones(in_features) * alpha)
+        self.alpha.requires_grad = alpha_trainable
+        self.eps = 1e-9
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        alpha = self.alpha.unsqueeze(0).unsqueeze(-1)
+        if self.alpha_logscale:
+            alpha = torch.exp(alpha)
+        return x + (1.0 / (alpha + self.eps)) * torch.sin(x * alpha).pow(2)
+
+
+class SnakeBeta(nn.Module):
+    def __init__(
+        self,
+        in_features: int,
+        alpha: float = 1.0,
+        alpha_trainable: bool = True,
+        alpha_logscale: bool = True,
+    ) -> None:
+        super().__init__()
+        self.alpha_logscale = alpha_logscale
+        self.alpha = nn.Parameter(torch.zeros(in_features) if alpha_logscale else torch.ones(in_features) * alpha)
+        self.alpha.requires_grad = alpha_trainable
+        self.beta = nn.Parameter(torch.zeros(in_features) if alpha_logscale else torch.ones(in_features) * alpha)
+        self.beta.requires_grad = alpha_trainable
+        self.eps = 1e-9
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        alpha = self.alpha.unsqueeze(0).unsqueeze(-1)
+        beta = self.beta.unsqueeze(0).unsqueeze(-1)
+        if self.alpha_logscale:
+            alpha = torch.exp(alpha)
+            beta = torch.exp(beta)
+        return x + (1.0 / (beta + self.eps)) * torch.sin(x * alpha).pow(2)
+
+
+class AMPBlock1(nn.Module):
+    def __init__(
+        self,
+        channels: int,
+        kernel_size: int = 3,
+        dilation: tuple[int, int, int] = (1, 3, 5),
+        activation: str = "snake",
+    ) -> None:
+        super().__init__()
+        act_cls = SnakeBeta if activation == "snakebeta" else Snake
+        self.convs1 = nn.ModuleList(
+            [
+                nn.Conv1d(
+                    channels,
+                    channels,
+                    kernel_size,
+                    1,
+                    dilation=dilation[0],
+                    padding=get_padding(kernel_size, dilation[0]),
+                ),
+                nn.Conv1d(
+                    channels,
+                    channels,
+                    kernel_size,
+                    1,
+                    dilation=dilation[1],
+                    padding=get_padding(kernel_size, dilation[1]),
+                ),
+                nn.Conv1d(
+                    channels,
+                    channels,
+                    kernel_size,
+                    1,
+                    dilation=dilation[2],
+                    padding=get_padding(kernel_size, dilation[2]),
+                ),
+            ]
+        )
+
+        self.convs2 = nn.ModuleList(
+            [
+                nn.Conv1d(channels, channels, kernel_size, 1, dilation=1, padding=get_padding(kernel_size, 1)),
+                nn.Conv1d(channels, channels, kernel_size, 1, dilation=1, padding=get_padding(kernel_size, 1)),
+                nn.Conv1d(channels, channels, kernel_size, 1, dilation=1, padding=get_padding(kernel_size, 1)),
+            ]
+        )
+
+        self.acts1 = nn.ModuleList([Activation1d(act_cls(channels)) for _ in range(len(self.convs1))])
+        self.acts2 = nn.ModuleList([Activation1d(act_cls(channels)) for _ in range(len(self.convs2))])
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        for c1, c2, a1, a2 in zip(self.convs1, self.convs2, self.acts1, self.acts2, strict=True):
+            xt = a1(x)
+            xt = c1(xt)
+            xt = a2(xt)
+            xt = c2(xt)
+            x = x + xt
+        return x
+
+
 class LTX2Vocoder(torch.nn.Module):
     """
-    Vocoder model for synthesizing audio from Mel spectrograms.
+    LTX2Vocoder model for synthesizing audio from Mel spectrograms.
     Args:
         resblock_kernel_sizes: List of kernel sizes for the residual blocks.
                                This value is read from the checkpoint at `config.vocoder.resblock_kernel_sizes`.
@@ -1293,28 +1552,33 @@ class LTX2Vocoder(torch.nn.Module):
                                This value is read from the checkpoint at `config.vocoder.resblock_dilation_sizes`.
         upsample_initial_channel: Initial number of channels for the upsampling layers.
                                This value is read from the checkpoint at `config.vocoder.upsample_initial_channel`.
-        stereo: Whether to use stereo output.
-                               This value is read from the checkpoint at `config.vocoder.stereo`.
-        resblock: Type of residual block to use.
+        resblock: Type of residual block to use ("1", "2", or "AMP1").
                                 This value is read from the checkpoint at `config.vocoder.resblock`.
-        output_sample_rate: Waveform sample rate.
-                               This value is read from the checkpoint at `config.vocoder.output_sample_rate`.
+        output_sampling_rate: Waveform sample rate.
+                               This value is read from the checkpoint at `config.vocoder.output_sampling_rate`.
+        activation: Activation type for BigVGAN v2 ("snake" or "snakebeta"). Only used when resblock="AMP1".
+        use_tanh_at_final: Apply tanh at the output (when apply_final_activation=True).
+        apply_final_activation: Whether to apply the final tanh/clamp activation.
+        use_bias_at_final: Whether to use bias in the final conv layer.
     """
 
-    def __init__(
+    def __init__(  # noqa: PLR0913
         self,
         resblock_kernel_sizes: List[int] | None = [3, 7, 11],
         upsample_rates: List[int] | None = [6, 5, 2, 2, 2],
         upsample_kernel_sizes: List[int] | None = [16, 15, 8, 4, 4],
         resblock_dilation_sizes: List[List[int]] | None = [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
         upsample_initial_channel: int = 1024,
-        stereo: bool = True,
         resblock: str = "1",
-        output_sample_rate: int = 24000,
-    ):
+        output_sampling_rate: int = 24000,
+        activation: str = "snake",
+        use_tanh_at_final: bool = True,
+        apply_final_activation: bool = True,
+        use_bias_at_final: bool = True,
+    ) -> None:
         super().__init__()
 
-        # Initialize default values if not provided. Note that mutable default values are not supported.
+        # Mutable default values are not supported as default arguments.
         if resblock_kernel_sizes is None:
             resblock_kernel_sizes = [3, 7, 11]
         if upsample_rates is None:
@@ -1324,36 +1588,60 @@ class LTX2Vocoder(torch.nn.Module):
         if resblock_dilation_sizes is None:
             resblock_dilation_sizes = [[1, 3, 5], [1, 3, 5], [1, 3, 5]]
 
-        self.output_sample_rate = output_sample_rate
+        self.output_sampling_rate = output_sampling_rate
         self.num_kernels = len(resblock_kernel_sizes)
         self.num_upsamples = len(upsample_rates)
-        in_channels = 128 if stereo else 64
-        self.conv_pre = nn.Conv1d(in_channels, upsample_initial_channel, 7, 1, padding=3)
-        resblock_class = ResBlock1 if resblock == "1" else ResBlock2
+        self.use_tanh_at_final = use_tanh_at_final
+        self.apply_final_activation = apply_final_activation
+        self.is_amp = resblock == "AMP1"
 
-        self.ups = nn.ModuleList()
-        for i, (stride, kernel_size) in enumerate(zip(upsample_rates, upsample_kernel_sizes, strict=True)):
-            self.ups.append(
-                nn.ConvTranspose1d(
-                    upsample_initial_channel // (2**i),
-                    upsample_initial_channel // (2 ** (i + 1)),
-                    kernel_size,
-                    stride,
-                    padding=(kernel_size - stride) // 2,
-                )
+        # All production checkpoints are stereo: 128 input channels (2 stereo channels x 64 mel
+        # bins each), 2 output channels.
+        self.conv_pre = nn.Conv1d(
+            in_channels=128,
+            out_channels=upsample_initial_channel,
+            kernel_size=7,
+            stride=1,
+            padding=3,
+        )
+        resblock_cls = ResBlock1 if resblock == "1" else AMPBlock1
+
+        self.ups = nn.ModuleList(
+            nn.ConvTranspose1d(
+                upsample_initial_channel // (2**i),
+                upsample_initial_channel // (2 ** (i + 1)),
+                kernel_size,
+                stride,
+                padding=(kernel_size - stride) // 2,
             )
+            for i, (stride, kernel_size) in enumerate(zip(upsample_rates, upsample_kernel_sizes, strict=True))
+        )
 
+        final_channels = upsample_initial_channel // (2 ** len(upsample_rates))
         self.resblocks = nn.ModuleList()
-        for i, _ in enumerate(self.ups):
+
+        for i in range(len(upsample_rates)):
             ch = upsample_initial_channel // (2 ** (i + 1))
             for kernel_size, dilations in zip(resblock_kernel_sizes, resblock_dilation_sizes, strict=True):
-                self.resblocks.append(resblock_class(ch, kernel_size, dilations))
+                if self.is_amp:
+                    self.resblocks.append(resblock_cls(ch, kernel_size, dilations, activation=activation))
+                else:
+                    self.resblocks.append(resblock_cls(ch, kernel_size, dilations))
 
-        out_channels = 2 if stereo else 1
-        final_channels = upsample_initial_channel // (2**self.num_upsamples)
-        self.conv_post = nn.Conv1d(final_channels, out_channels, 7, 1, padding=3)
+        if self.is_amp:
+            self.act_post: nn.Module = Activation1d(SnakeBeta(final_channels))
+        else:
+            self.act_post = nn.LeakyReLU()
 
-        self.upsample_factor = math.prod(layer.stride[0] for layer in self.ups)
+        # All production checkpoints are stereo: this final conv maps `final_channels` to 2 output channels (stereo).
+        self.conv_post = nn.Conv1d(
+            in_channels=final_channels,
+            out_channels=2,
+            kernel_size=7,
+            stride=1,
+            padding=3,
+            bias=use_bias_at_final,
+        )
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         """
@@ -1374,7 +1662,8 @@ class LTX2Vocoder(torch.nn.Module):
         x = self.conv_pre(x)
 
         for i in range(self.num_upsamples):
-            x = F.leaky_relu(x, LRELU_SLOPE)
+            if not self.is_amp:
+                x = F.leaky_relu(x, LRELU_SLOPE)
             x = self.ups[i](x)
             start = i * self.num_kernels
             end = start + self.num_kernels
@@ -1386,23 +1675,198 @@ class LTX2Vocoder(torch.nn.Module):
                 [self.resblocks[idx](x) for idx in range(start, end)],
                 dim=0,
             )
-
             x = block_outputs.mean(dim=0)
 
-        x = self.conv_post(F.leaky_relu(x))
-        return torch.tanh(x)
+        x = self.act_post(x)
+        x = self.conv_post(x)
+
+        if self.apply_final_activation:
+            x = torch.tanh(x) if self.use_tanh_at_final else torch.clamp(x, -1, 1)
+
+        return x
 
 
-def decode_audio(latent: torch.Tensor, audio_decoder: "LTX2AudioDecoder", vocoder: "LTX2Vocoder") -> torch.Tensor:
+class _STFTFn(nn.Module):
+    """Implements STFT as a convolution with precomputed DFT x Hann-window bases.
+    The DFT basis rows (real and imaginary parts interleaved) multiplied by the causal
+    Hann window are stored as buffers and loaded from the checkpoint. Using the exact
+    bfloat16 bases from training ensures the mel values fed to the BWE generator are
+    bit-identical to what it was trained on.
     """
-    Decode an audio latent representation using the provided audio decoder and vocoder.
-    Args:
-        latent: Input audio latent tensor.
-        audio_decoder: Model to decode the latent to waveform features.
-        vocoder: Model to convert decoded features to audio waveform.
-    Returns:
-        Decoded audio as a float tensor.
+
+    def __init__(self, filter_length: int, hop_length: int, win_length: int) -> None:
+        super().__init__()
+        self.hop_length = hop_length
+        self.win_length = win_length
+        n_freqs = filter_length // 2 + 1
+        self.register_buffer("forward_basis", torch.zeros(n_freqs * 2, 1, filter_length))
+        self.register_buffer("inverse_basis", torch.zeros(n_freqs * 2, 1, filter_length))
+
+    def forward(self, y: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+        """Compute magnitude and phase spectrogram from a batch of waveforms.
+        Applies causal (left-only) padding of win_length - hop_length samples so that
+        each output frame depends only on past and present input — no lookahead.
+        Args:
+            y: Waveform tensor of shape (B, T).
+        Returns:
+            magnitude: Linear amplitude spectrogram, shape (B, n_freqs, T_frames).
+            phase:     Phase spectrogram in radians, shape (B, n_freqs, T_frames).
+        """
+        if y.dim() == 2:
+            y = y.unsqueeze(1)  # (B, 1, T)
+        left_pad = max(0, self.win_length - self.hop_length)  # causal: left-only
+        y = F.pad(y, (left_pad, 0))
+        spec = F.conv1d(y, self.forward_basis, stride=self.hop_length, padding=0)
+        n_freqs = spec.shape[1] // 2
+        real, imag = spec[:, :n_freqs], spec[:, n_freqs:]
+        magnitude = torch.sqrt(real**2 + imag**2)
+        phase = torch.atan2(imag.float(), real.float()).to(real.dtype)
+        return magnitude, phase
+
+
+class MelSTFT(nn.Module):
+    """Causal log-mel spectrogram module whose buffers are loaded from the checkpoint.
+    Computes a log-mel spectrogram by running the causal STFT (_STFTFn) on the input
+    waveform and projecting the linear magnitude spectrum onto the mel filterbank.
+    The module's state dict layout matches the 'mel_stft.*' keys stored in the checkpoint
+    (mel_basis, stft_fn.forward_basis, stft_fn.inverse_basis).
     """
-    decoded_audio = audio_decoder(latent)
-    decoded_audio = vocoder(decoded_audio).squeeze(0).float()
-    return decoded_audio
+
+    def __init__(
+        self,
+        filter_length: int,
+        hop_length: int,
+        win_length: int,
+        n_mel_channels: int,
+    ) -> None:
+        super().__init__()
+        self.stft_fn = _STFTFn(filter_length, hop_length, win_length)
+
+        # Initialized to zeros; load_state_dict overwrites with the checkpoint's
+        # exact bfloat16 filterbank (vocoder.mel_stft.mel_basis, shape [n_mels, n_freqs]).
+        n_freqs = filter_length // 2 + 1
+        self.register_buffer("mel_basis", torch.zeros(n_mel_channels, n_freqs))
+
+    def mel_spectrogram(self, y: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Compute log-mel spectrogram and auxiliary spectral quantities.
+        Args:
+            y: Waveform tensor of shape (B, T).
+        Returns:
+            log_mel:   Log-compressed mel spectrogram, shape (B, n_mel_channels, T_frames).
+            magnitude: Linear amplitude spectrogram, shape (B, n_freqs, T_frames).
+            phase:     Phase spectrogram in radians, shape (B, n_freqs, T_frames).
+            energy:    Per-frame energy (L2 norm over frequency), shape (B, T_frames).
+        """
+        magnitude, phase = self.stft_fn(y)
+        energy = torch.norm(magnitude, dim=1)
+        mel = torch.matmul(self.mel_basis.to(magnitude.dtype), magnitude)
+        log_mel = torch.log(torch.clamp(mel, min=1e-5))
+        return log_mel, magnitude, phase, energy
+
+
+class LTX2VocoderWithBWE(nn.Module):
+    """LTX2Vocoder with bandwidth extension (BWE) upsampling.
+    Chains a mel-to-wav vocoder with a BWE module that upsamples the output
+    to a higher sample rate. The BWE computes a mel spectrogram from the
+    vocoder output, runs it through a second generator to predict a residual,
+    and adds it to a sinc-resampled skip connection.
+    """
+
+    def __init__(
+        self,
+        input_sampling_rate: int = 16000,
+        output_sampling_rate: int = 48000,
+        hop_length: int = 80,
+    ) -> None:
+        super().__init__()
+        self.vocoder = LTX2Vocoder(
+            resblock_kernel_sizes=[3, 7, 11],
+            upsample_rates=[5, 2, 2, 2, 2, 2],
+            upsample_kernel_sizes=[11, 4, 4, 4, 4, 4],
+            resblock_dilation_sizes=[[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+            upsample_initial_channel=1536,
+            resblock="AMP1",
+            activation="snakebeta",
+            use_tanh_at_final=False,
+            apply_final_activation=True,
+            use_bias_at_final=False,
+            output_sampling_rate=input_sampling_rate,
+        )
+        self.bwe_generator = LTX2Vocoder(
+            resblock_kernel_sizes=[3, 7, 11],
+            upsample_rates=[6, 5, 2, 2, 2],
+            upsample_kernel_sizes=[12, 11, 4, 4, 4],
+            resblock_dilation_sizes=[[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+            upsample_initial_channel=512,
+            resblock="AMP1",
+            activation="snakebeta",
+            use_tanh_at_final=False,
+            apply_final_activation=False,
+            use_bias_at_final=False,
+            output_sampling_rate=output_sampling_rate,
+        )
+        
+        self.mel_stft = MelSTFT(
+            filter_length=512,
+            hop_length=hop_length,
+            win_length=512,
+            n_mel_channels=64,
+        )
+        self.input_sampling_rate = input_sampling_rate
+        self.output_sampling_rate = output_sampling_rate
+        self.hop_length = hop_length
+        # Compute the resampler on CPU so the sinc filter is materialized even when
+        # the model is constructed on meta device (SingleGPUModelBuilder pattern).
+        # The filter is not stored in the checkpoint (persistent=False).
+        with torch.device("cpu"):
+            self.resampler = UpSample1d(
+                ratio=output_sampling_rate // input_sampling_rate, persistent=False, window_type="hann"
+            )
+
+    @property
+    def conv_pre(self) -> nn.Conv1d:
+        return self.vocoder.conv_pre
+
+    @property
+    def conv_post(self) -> nn.Conv1d:
+        return self.vocoder.conv_post
+
+    def _compute_mel(self, audio: torch.Tensor) -> torch.Tensor:
+        """Compute log-mel spectrogram from waveform using causal STFT bases.
+        Args:
+            audio: Waveform tensor of shape (B, C, T).
+        Returns:
+            mel: Log-mel spectrogram of shape (B, C, n_mels, T_frames).
+        """
+        batch, n_channels, _ = audio.shape
+        flat = audio.reshape(batch * n_channels, -1)  # (B*C, T)
+        mel, _, _, _ = self.mel_stft.mel_spectrogram(flat)  # (B*C, n_mels, T_frames)
+        return mel.reshape(batch, n_channels, mel.shape[1], mel.shape[2])  # (B, C, n_mels, T_frames)
+
+    def forward(self, mel_spec: torch.Tensor) -> torch.Tensor:
+        """Run the full vocoder + BWE forward pass.
+        Args:
+            mel_spec: Mel spectrogram of shape (B, 2, T, mel_bins) for stereo
+                      or (B, T, mel_bins) for mono. Same format as LTX2Vocoder.forward.
+        Returns:
+            Waveform tensor of shape (B, out_channels, T_out) clipped to [-1, 1].
+        """
+        x = self.vocoder(mel_spec)
+        _, _, length_low_rate = x.shape
+        output_length = length_low_rate * self.output_sampling_rate // self.input_sampling_rate
+
+        # Pad to multiple of hop_length for exact mel frame count
+        remainder = length_low_rate % self.hop_length
+        if remainder != 0:
+            x = F.pad(x, (0, self.hop_length - remainder))
+
+        # Compute mel spectrogram from vocoder output: (B, C, n_mels, T_frames)
+        mel = self._compute_mel(x)
+
+        # LTX2Vocoder.forward expects (B, C, T, mel_bins) — transpose before calling bwe_generator
+        mel_for_bwe = mel.transpose(2, 3)  # (B, C, T_frames, mel_bins)
+        residual = self.bwe_generator(mel_for_bwe)
+        skip = self.resampler(x)
+        assert residual.shape == skip.shape, f"residual {residual.shape} != skip {skip.shape}"
+
+        return torch.clamp(residual + skip, -1, 1)[..., :output_length]

diffsynth/models/ltx2_common.py

@@ -251,11 +251,27 @@ class Modality:
     Input data for a single modality (video or audio) in the transformer.
     Bundles the latent tokens, timestep embeddings, positional information,
     and text conditioning context for processing by the diffusion transformer.
+    Attributes:
+        latent: Patchified latent tokens, shape ``(B, T, D)`` where *B* is
+            the batch size, *T* is the total number of tokens (noisy +
+            conditioning), and *D* is the input dimension.
+        timesteps: Per-token timestep embeddings, shape ``(B, T)``.
+        positions: Positional coordinates, shape ``(B, 3, T)`` for video
+            (time, height, width) or ``(B, 1, T)`` for audio.
+        context: Text conditioning embeddings from the prompt encoder.
+        enabled: Whether this modality is active in the current forward pass.
+        context_mask: Optional mask for the text context tokens.
+        attention_mask: Optional 2-D self-attention mask, shape ``(B, T, T)``.
+            Values in ``[0, 1]`` where ``1`` = full attention and ``0`` = no
+            attention. ``None`` means unrestricted (full) attention between
+            all tokens. Built incrementally by conditioning items; see
+            :class:`~ltx_core.conditioning.types.attention_strength_wrapper.ConditioningItemAttentionStrengthWrapper`.
     """
 
     latent: (
         torch.Tensor
     )  # Shape: (B, T, D) where B is the batch size, T is the number of tokens, and D is input dimension
+    sigma: torch.Tensor  # Shape: (B,). Current sigma value, used for cross-attention timestep calculation.
     timesteps: torch.Tensor  # Shape: (B, T) where T is the number of timesteps
     positions: (
         torch.Tensor
@@ -263,6 +279,7 @@ class Modality:
     context: torch.Tensor
     enabled: bool = True
     context_mask: torch.Tensor | None = None
+    attention_mask: torch.Tensor | None = None
 
 
 def to_denoised(

diffsynth/models/ltx2_dit.py

@@ -8,6 +8,7 @@ import torch
 from einops import rearrange
 from .ltx2_common import rms_norm, Modality
 from ..core.attention.attention import attention_forward
+from ..core import gradient_checkpoint_forward
 
 
 def get_timestep_embedding(
@@ -224,6 +225,17 @@ class BatchedPerturbationConfig:
         return BatchedPerturbationConfig([PerturbationConfig.empty() for _ in range(batch_size)])
 
 
+
+ADALN_NUM_BASE_PARAMS = 6
+# Cross-attention AdaLN adds 3 more (scale, shift, gate) for the CA norm.
+ADALN_NUM_CROSS_ATTN_PARAMS = 3
+
+
+def adaln_embedding_coefficient(cross_attention_adaln: bool) -> int:
+    """Total number of AdaLN parameters per block."""
+    return ADALN_NUM_BASE_PARAMS + (ADALN_NUM_CROSS_ATTN_PARAMS if cross_attention_adaln else 0)
+
+
 class AdaLayerNormSingle(torch.nn.Module):
     r"""
     Norm layer adaptive layer norm single (adaLN-single).
@@ -459,6 +471,7 @@ class Attention(torch.nn.Module):
         dim_head: int = 64,
         norm_eps: float = 1e-6,
         rope_type: LTXRopeType = LTXRopeType.INTERLEAVED,
+        apply_gated_attention: bool = False,
     ) -> None:
         super().__init__()
         self.rope_type = rope_type
@@ -476,6 +489,12 @@ class Attention(torch.nn.Module):
         self.to_k = torch.nn.Linear(context_dim, inner_dim, bias=True)
         self.to_v = torch.nn.Linear(context_dim, inner_dim, bias=True)
 
+        # Optional per-head gating
+        if apply_gated_attention:
+            self.to_gate_logits = torch.nn.Linear(query_dim, heads, bias=True)
+        else:
+            self.to_gate_logits = None
+
         self.to_out = torch.nn.Sequential(torch.nn.Linear(inner_dim, query_dim, bias=True), torch.nn.Identity())
 
     def forward(
@@ -485,6 +504,8 @@ class Attention(torch.nn.Module):
         mask: torch.Tensor | None = None,
         pe: torch.Tensor | None = None,
         k_pe: torch.Tensor | None = None,
+        perturbation_mask: torch.Tensor | None = None,
+        all_perturbed: bool = False,
     ) -> torch.Tensor:
         q = self.to_q(x)
         context = x if context is None else context
@@ -516,6 +537,19 @@ class Attention(torch.nn.Module):
 
         # Reshape back to original format
         out = out.flatten(2, 3)
+
+        # Apply per-head gating if enabled
+        if self.to_gate_logits is not None:
+            gate_logits = self.to_gate_logits(x)  # (B, T, H)
+            b, t, _ = out.shape
+            # Reshape to (B, T, H, D) for per-head gating
+            out = out.view(b, t, self.heads, self.dim_head)
+            # Apply gating: 2 * sigmoid(x) so that zero-init gives identity (2 * 0.5 = 1.0)
+            gates = 2.0 * torch.sigmoid(gate_logits)  # (B, T, H)
+            out = out * gates.unsqueeze(-1)  # (B, T, H, D) * (B, T, H, 1)
+            # Reshape back to (B, T, H*D)
+            out = out.view(b, t, self.heads * self.dim_head)
+
         return self.to_out(out)
 
 
@@ -544,7 +578,6 @@ class PixArtAlphaTextProjection(torch.nn.Module):
         hidden_states = self.linear_2(hidden_states)
         return hidden_states
 
-
 @dataclass(frozen=True)
 class TransformerArgs:
     x: torch.Tensor
@@ -557,7 +590,10 @@ class TransformerArgs:
     cross_scale_shift_timestep: torch.Tensor | None
     cross_gate_timestep: torch.Tensor | None
     enabled: bool
-
+    prompt_timestep: torch.Tensor | None = None
+    self_attention_mask: torch.Tensor | None = (
+        None  # Additive log-space self-attention bias (B, 1, T, T), None = full attention
+    )
 
 
 class TransformerArgsPreprocessor:
@@ -565,7 +601,6 @@ class TransformerArgsPreprocessor:
         self,
         patchify_proj: torch.nn.Linear,
         adaln: AdaLayerNormSingle,
-        caption_projection: PixArtAlphaTextProjection,
         inner_dim: int,
         max_pos: list[int],
         num_attention_heads: int,
@@ -574,10 +609,11 @@ class TransformerArgsPreprocessor:
         double_precision_rope: bool,
         positional_embedding_theta: float,
         rope_type: LTXRopeType,
+        caption_projection: torch.nn.Module | None = None,
+        prompt_adaln: AdaLayerNormSingle | None = None,
     ) -> None:
         self.patchify_proj = patchify_proj
         self.adaln = adaln
-        self.caption_projection = caption_projection
         self.inner_dim = inner_dim
         self.max_pos = max_pos
         self.num_attention_heads = num_attention_heads
@@ -586,18 +622,18 @@ class TransformerArgsPreprocessor:
         self.double_precision_rope = double_precision_rope
         self.positional_embedding_theta = positional_embedding_theta
         self.rope_type = rope_type
+        self.caption_projection = caption_projection
+        self.prompt_adaln = prompt_adaln
 
     def _prepare_timestep(
-        self, timestep: torch.Tensor, batch_size: int, hidden_dtype: torch.dtype
+        self, timestep: torch.Tensor, adaln: AdaLayerNormSingle, batch_size: int, hidden_dtype: torch.dtype
     ) -> tuple[torch.Tensor, torch.Tensor]:
         """Prepare timestep embeddings."""
-
-        timestep = timestep * self.timestep_scale_multiplier
-        timestep, embedded_timestep = self.adaln(
-            timestep.flatten(),
+        timestep_scaled = timestep * self.timestep_scale_multiplier
+        timestep, embedded_timestep = adaln(
+            timestep_scaled.flatten(),
             hidden_dtype=hidden_dtype,
         )
-
         # Second dimension is 1 or number of tokens (if timestep_per_token)
         timestep = timestep.view(batch_size, -1, timestep.shape[-1])
         embedded_timestep = embedded_timestep.view(batch_size, -1, embedded_timestep.shape[-1])
@@ -607,14 +643,12 @@ class TransformerArgsPreprocessor:
         self,
         context: torch.Tensor,
         x: torch.Tensor,
-        attention_mask: torch.Tensor | None = None,
-    ) -> tuple[torch.Tensor, torch.Tensor | None]:
+    ) -> torch.Tensor:
         """Prepare context for transformer blocks."""
+        if self.caption_projection is not None:
+            context = self.caption_projection(context)
         batch_size = x.shape[0]
-        context = self.caption_projection(context)
-        context = context.view(batch_size, -1, x.shape[-1])
-
-        return context, attention_mask
+        return context.view(batch_size, -1, x.shape[-1])
 
     def _prepare_attention_mask(self, attention_mask: torch.Tensor | None, x_dtype: torch.dtype) -> torch.Tensor | None:
         """Prepare attention mask."""
@@ -625,6 +659,34 @@ class TransformerArgsPreprocessor:
             (attention_mask.shape[0], 1, -1, attention_mask.shape[-1])
         ) * torch.finfo(x_dtype).max
 
+    def _prepare_self_attention_mask(
+        self, attention_mask: torch.Tensor | None, x_dtype: torch.dtype
+    ) -> torch.Tensor | None:
+        """Prepare self-attention mask by converting [0,1] values to additive log-space bias.
+        Input shape: (B, T, T) with values in [0, 1].
+        Output shape: (B, 1, T, T) with 0.0 for full attention and a large negative value
+        for masked positions.
+        Positions with attention_mask <= 0 are fully masked (mapped to the dtype's minimum
+        representable value). Strictly positive entries are converted via log-space for
+        smooth attenuation, with small values clamped for numerical stability.
+        Returns None if input is None (no masking).
+        """
+        if attention_mask is None:
+            return None
+
+        # Convert [0, 1] attention mask to additive log-space bias:
+        #   1.0 -> log(1.0) = 0.0  (no bias, full attention)
+        #   0.0 -> finfo.min        (fully masked)
+        finfo = torch.finfo(x_dtype)
+        eps = finfo.tiny
+
+        bias = torch.full_like(attention_mask, finfo.min, dtype=x_dtype)
+        positive = attention_mask > 0
+        if positive.any():
+            bias[positive] = torch.log(attention_mask[positive].clamp(min=eps)).to(x_dtype)
+
+        return bias.unsqueeze(1)  # (B, 1, T, T) for head broadcast
+
     def _prepare_positional_embeddings(
         self,
         positions: torch.Tensor,
@@ -652,11 +714,20 @@ class TransformerArgsPreprocessor:
     def prepare(
         self,
         modality: Modality,
+        cross_modality: Modality | None = None,  # noqa: ARG002
     ) -> TransformerArgs:
         x = self.patchify_proj(modality.latent)
-        timestep, embedded_timestep = self._prepare_timestep(modality.timesteps, x.shape[0], modality.latent.dtype)
-        context, attention_mask = self._prepare_context(modality.context, x, modality.context_mask)
-        attention_mask = self._prepare_attention_mask(attention_mask, modality.latent.dtype)
+        batch_size = x.shape[0]
+        timestep, embedded_timestep = self._prepare_timestep(
+            modality.timesteps, self.adaln, batch_size, modality.latent.dtype
+        )
+        prompt_timestep = None
+        if self.prompt_adaln is not None:
+            prompt_timestep, _ = self._prepare_timestep(
+                modality.sigma, self.prompt_adaln, batch_size, modality.latent.dtype
+            )
+        context = self._prepare_context(modality.context, x)
+        attention_mask = self._prepare_attention_mask(modality.context_mask, modality.latent.dtype)
         pe = self._prepare_positional_embeddings(
             positions=modality.positions,
             inner_dim=self.inner_dim,
@@ -665,6 +736,7 @@ class TransformerArgsPreprocessor:
             num_attention_heads=self.num_attention_heads,
             x_dtype=modality.latent.dtype,
         )
+        self_attention_mask = self._prepare_self_attention_mask(modality.attention_mask, modality.latent.dtype)
         return TransformerArgs(
             x=x,
             context=context,
@@ -676,6 +748,8 @@ class TransformerArgsPreprocessor:
             cross_scale_shift_timestep=None,
             cross_gate_timestep=None,
             enabled=modality.enabled,
+            prompt_timestep=prompt_timestep,
+            self_attention_mask=self_attention_mask,
         )
 
 
@@ -684,7 +758,6 @@ class MultiModalTransformerArgsPreprocessor:
         self,
         patchify_proj: torch.nn.Linear,
         adaln: AdaLayerNormSingle,
-        caption_projection: PixArtAlphaTextProjection,
         cross_scale_shift_adaln: AdaLayerNormSingle,
         cross_gate_adaln: AdaLayerNormSingle,
         inner_dim: int,
@@ -698,11 +771,12 @@ class MultiModalTransformerArgsPreprocessor:
         positional_embedding_theta: float,
         rope_type: LTXRopeType,
         av_ca_timestep_scale_multiplier: int,
+        caption_projection: torch.nn.Module | None = None,
+        prompt_adaln: AdaLayerNormSingle | None = None,
     ) -> None:
         self.simple_preprocessor = TransformerArgsPreprocessor(
             patchify_proj=patchify_proj,
             adaln=adaln,
-            caption_projection=caption_projection,
             inner_dim=inner_dim,
             max_pos=max_pos,
             num_attention_heads=num_attention_heads,
@@ -711,6 +785,8 @@ class MultiModalTransformerArgsPreprocessor:
             double_precision_rope=double_precision_rope,
             positional_embedding_theta=positional_embedding_theta,
             rope_type=rope_type,
+            caption_projection=caption_projection,
+            prompt_adaln=prompt_adaln,
         )
         self.cross_scale_shift_adaln = cross_scale_shift_adaln
         self.cross_gate_adaln = cross_gate_adaln
@@ -721,8 +797,22 @@ class MultiModalTransformerArgsPreprocessor:
     def prepare(
         self,
         modality: Modality,
+        cross_modality: Modality | None = None,
     ) -> TransformerArgs:
         transformer_args = self.simple_preprocessor.prepare(modality)
+        if cross_modality is None:
+            return transformer_args
+
+        if cross_modality.sigma.numel() > 1:
+            if cross_modality.sigma.shape[0] != modality.timesteps.shape[0]:
+                raise ValueError("Cross modality sigma must have the same batch size as the modality")
+            if cross_modality.sigma.ndim != 1:
+                raise ValueError("Cross modality sigma must be a 1D tensor")
+
+        cross_timestep = cross_modality.sigma.view(
+            modality.timesteps.shape[0], 1, *[1] * len(modality.timesteps.shape[2:])
+        )
+
         cross_pe = self.simple_preprocessor._prepare_positional_embeddings(
             positions=modality.positions[:, 0:1, :],
             inner_dim=self.audio_cross_attention_dim,
@@ -733,7 +823,7 @@ class MultiModalTransformerArgsPreprocessor:
         )
 
         cross_scale_shift_timestep, cross_gate_timestep = self._prepare_cross_attention_timestep(
-            timestep=modality.timesteps,
+            timestep=cross_timestep,
             timestep_scale_multiplier=self.simple_preprocessor.timestep_scale_multiplier,
             batch_size=transformer_args.x.shape[0],
             hidden_dtype=modality.latent.dtype,
@@ -748,7 +838,7 @@ class MultiModalTransformerArgsPreprocessor:
 
     def _prepare_cross_attention_timestep(
         self,
-        timestep: torch.Tensor,
+        timestep: torch.Tensor | None,
         timestep_scale_multiplier: int,
         batch_size: int,
         hidden_dtype: torch.dtype,
@@ -778,6 +868,8 @@ class TransformerConfig:
     heads: int
     d_head: int
     context_dim: int
+    apply_gated_attention: bool = False
+    cross_attention_adaln: bool = False
 
 
 class BasicAVTransformerBlock(torch.nn.Module):
@@ -800,6 +892,7 @@ class BasicAVTransformerBlock(torch.nn.Module):
                 context_dim=None,
                 rope_type=rope_type,
                 norm_eps=norm_eps,
+                apply_gated_attention=video.apply_gated_attention,
             )
             self.attn2 = Attention(
                 query_dim=video.dim,
@@ -808,9 +901,11 @@ class BasicAVTransformerBlock(torch.nn.Module):
                 dim_head=video.d_head,
                 rope_type=rope_type,
                 norm_eps=norm_eps,
+                apply_gated_attention=video.apply_gated_attention,
             )
             self.ff = FeedForward(video.dim, dim_out=video.dim)
-            self.scale_shift_table = torch.nn.Parameter(torch.empty(6, video.dim))
+            video_sst_size = adaln_embedding_coefficient(video.cross_attention_adaln)
+            self.scale_shift_table = torch.nn.Parameter(torch.empty(video_sst_size, video.dim))
 
         if audio is not None:
             self.audio_attn1 = Attention(
@@ -820,6 +915,7 @@ class BasicAVTransformerBlock(torch.nn.Module):
                 context_dim=None,
                 rope_type=rope_type,
                 norm_eps=norm_eps,
+                apply_gated_attention=audio.apply_gated_attention,
             )
             self.audio_attn2 = Attention(
                 query_dim=audio.dim,
@@ -828,9 +924,11 @@ class BasicAVTransformerBlock(torch.nn.Module):
                 dim_head=audio.d_head,
                 rope_type=rope_type,
                 norm_eps=norm_eps,
+                apply_gated_attention=audio.apply_gated_attention,
             )
             self.audio_ff = FeedForward(audio.dim, dim_out=audio.dim)
-            self.audio_scale_shift_table = torch.nn.Parameter(torch.empty(6, audio.dim))
+            audio_sst_size = adaln_embedding_coefficient(audio.cross_attention_adaln)
+            self.audio_scale_shift_table = torch.nn.Parameter(torch.empty(audio_sst_size, audio.dim))
 
         if audio is not None and video is not None:
             # Q: Video, K,V: Audio
@@ -841,6 +939,7 @@ class BasicAVTransformerBlock(torch.nn.Module):
                 dim_head=audio.d_head,
                 rope_type=rope_type,
                 norm_eps=norm_eps,
+                apply_gated_attention=video.apply_gated_attention,
             )
 
             # Q: Audio, K,V: Video
@@ -851,11 +950,21 @@ class BasicAVTransformerBlock(torch.nn.Module):
                 dim_head=audio.d_head,
                 rope_type=rope_type,
                 norm_eps=norm_eps,
+                apply_gated_attention=audio.apply_gated_attention,
             )
 
             self.scale_shift_table_a2v_ca_audio = torch.nn.Parameter(torch.empty(5, audio.dim))
             self.scale_shift_table_a2v_ca_video = torch.nn.Parameter(torch.empty(5, video.dim))
 
+        self.cross_attention_adaln = (video is not None and video.cross_attention_adaln) or (
+            audio is not None and audio.cross_attention_adaln
+        )
+
+        if self.cross_attention_adaln and video is not None:
+            self.prompt_scale_shift_table = torch.nn.Parameter(torch.empty(2, video.dim))
+        if self.cross_attention_adaln and audio is not None:
+            self.audio_prompt_scale_shift_table = torch.nn.Parameter(torch.empty(2, audio.dim))
+
         self.norm_eps = norm_eps
 
     def get_ada_values(
@@ -875,19 +984,49 @@ class BasicAVTransformerBlock(torch.nn.Module):
         batch_size: int,
         scale_shift_timestep: torch.Tensor,
         gate_timestep: torch.Tensor,
+        scale_shift_indices: slice,
         num_scale_shift_values: int = 4,
-    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
         scale_shift_ada_values = self.get_ada_values(
-            scale_shift_table[:num_scale_shift_values, :], batch_size, scale_shift_timestep, slice(None, None)
+            scale_shift_table[:num_scale_shift_values, :], batch_size, scale_shift_timestep, scale_shift_indices
         )
         gate_ada_values = self.get_ada_values(
             scale_shift_table[num_scale_shift_values:, :], batch_size, gate_timestep, slice(None, None)
         )
 
-        scale_shift_chunks = [t.squeeze(2) for t in scale_shift_ada_values]
-        gate_ada_values = [t.squeeze(2) for t in gate_ada_values]
+        scale, shift = (t.squeeze(2) for t in scale_shift_ada_values)
+        (gate,) = (t.squeeze(2) for t in gate_ada_values)
 
-        return (*scale_shift_chunks, *gate_ada_values)
+        return scale, shift, gate
+
+    def _apply_text_cross_attention(
+        self,
+        x: torch.Tensor,
+        context: torch.Tensor,
+        attn: Attention,
+        scale_shift_table: torch.Tensor,
+        prompt_scale_shift_table: torch.Tensor | None,
+        timestep: torch.Tensor,
+        prompt_timestep: torch.Tensor | None,
+        context_mask: torch.Tensor | None,
+        cross_attention_adaln: bool = False,
+    ) -> torch.Tensor:
+        """Apply text cross-attention, with optional AdaLN modulation."""
+        if cross_attention_adaln:
+            shift_q, scale_q, gate = self.get_ada_values(scale_shift_table, x.shape[0], timestep, slice(6, 9))
+            return apply_cross_attention_adaln(
+                x,
+                context,
+                attn,
+                shift_q,
+                scale_q,
+                gate,
+                prompt_scale_shift_table,
+                prompt_timestep,
+                context_mask,
+                self.norm_eps,
+            )
+        return attn(rms_norm(x, eps=self.norm_eps), context=context, mask=context_mask)
 
     def forward(  # noqa: PLR0915
         self,
@@ -895,7 +1034,11 @@ class BasicAVTransformerBlock(torch.nn.Module):
         audio: TransformerArgs | None,
         perturbations: BatchedPerturbationConfig | None = None,
     ) -> tuple[TransformerArgs | None, TransformerArgs | None]:
-        batch_size = video.x.shape[0]
+        if video is None and audio is None:
+            raise ValueError("At least one of video or audio must be provided")
+
+        batch_size = (video or audio).x.shape[0]
+
         if perturbations is None:
             perturbations = BatchedPerturbationConfig.empty(batch_size)
 
@@ -912,63 +1055,103 @@ class BasicAVTransformerBlock(torch.nn.Module):
             vshift_msa, vscale_msa, vgate_msa = self.get_ada_values(
                 self.scale_shift_table, vx.shape[0], video.timesteps, slice(0, 3)
             )
-            if not perturbations.all_in_batch(PerturbationType.SKIP_VIDEO_SELF_ATTN, self.idx):
-                norm_vx = rms_norm(vx, eps=self.norm_eps) * (1 + vscale_msa) + vshift_msa
-                v_mask = perturbations.mask_like(PerturbationType.SKIP_VIDEO_SELF_ATTN, self.idx, vx)
-                vx = vx + self.attn1(norm_vx, pe=video.positional_embeddings) * vgate_msa * v_mask
+            norm_vx = rms_norm(vx, eps=self.norm_eps) * (1 + vscale_msa) + vshift_msa
+            del vshift_msa, vscale_msa
 
-            vx = vx + self.attn2(rms_norm(vx, eps=self.norm_eps), context=video.context, mask=video.context_mask)
-
-            del vshift_msa, vscale_msa, vgate_msa
+            all_perturbed = perturbations.all_in_batch(PerturbationType.SKIP_VIDEO_SELF_ATTN, self.idx)
+            none_perturbed = not perturbations.any_in_batch(PerturbationType.SKIP_VIDEO_SELF_ATTN, self.idx)
+            v_mask = (
+                perturbations.mask_like(PerturbationType.SKIP_VIDEO_SELF_ATTN, self.idx, vx)
+                if not all_perturbed and not none_perturbed
+                else None
+            )
+            vx = (
+                vx
+                + self.attn1(
+                    norm_vx,
+                    pe=video.positional_embeddings,
+                    mask=video.self_attention_mask,
+                    perturbation_mask=v_mask,
+                    all_perturbed=all_perturbed,
+                )
+                * vgate_msa
+            )
+            del vgate_msa, norm_vx, v_mask
+            vx = vx + self._apply_text_cross_attention(
+                vx,
+                video.context,
+                self.attn2,
+                self.scale_shift_table,
+                getattr(self, "prompt_scale_shift_table", None),
+                video.timesteps,
+                video.prompt_timestep,
+                video.context_mask,
+                cross_attention_adaln=self.cross_attention_adaln,
+            )
 
         if run_ax:
             ashift_msa, ascale_msa, agate_msa = self.get_ada_values(
                 self.audio_scale_shift_table, ax.shape[0], audio.timesteps, slice(0, 3)
             )
 
-            if not perturbations.all_in_batch(PerturbationType.SKIP_AUDIO_SELF_ATTN, self.idx):
-                norm_ax = rms_norm(ax, eps=self.norm_eps) * (1 + ascale_msa) + ashift_msa
-                a_mask = perturbations.mask_like(PerturbationType.SKIP_AUDIO_SELF_ATTN, self.idx, ax)
-                ax = ax + self.audio_attn1(norm_ax, pe=audio.positional_embeddings) * agate_msa * a_mask
-
-            ax = ax + self.audio_attn2(rms_norm(ax, eps=self.norm_eps), context=audio.context, mask=audio.context_mask)
-
-            del ashift_msa, ascale_msa, agate_msa
+            norm_ax = rms_norm(ax, eps=self.norm_eps) * (1 + ascale_msa) + ashift_msa
+            del ashift_msa, ascale_msa
+            all_perturbed = perturbations.all_in_batch(PerturbationType.SKIP_AUDIO_SELF_ATTN, self.idx)
+            none_perturbed = not perturbations.any_in_batch(PerturbationType.SKIP_AUDIO_SELF_ATTN, self.idx)
+            a_mask = (
+                perturbations.mask_like(PerturbationType.SKIP_AUDIO_SELF_ATTN, self.idx, ax)
+                if not all_perturbed and not none_perturbed
+                else None
+            )
+            ax = (
+                ax
+                + self.audio_attn1(
+                    norm_ax,
+                    pe=audio.positional_embeddings,
+                    mask=audio.self_attention_mask,
+                    perturbation_mask=a_mask,
+                    all_perturbed=all_perturbed,
+                )
+                * agate_msa
+            )
+            del agate_msa, norm_ax, a_mask
+            ax = ax + self._apply_text_cross_attention(
+                ax,
+                audio.context,
+                self.audio_attn2,
+                self.audio_scale_shift_table,
+                getattr(self, "audio_prompt_scale_shift_table", None),
+                audio.timesteps,
+                audio.prompt_timestep,
+                audio.context_mask,
+                cross_attention_adaln=self.cross_attention_adaln,
+            )
 
         # Audio - Video cross attention.
         if run_a2v or run_v2a:
             vx_norm3 = rms_norm(vx, eps=self.norm_eps)
             ax_norm3 = rms_norm(ax, eps=self.norm_eps)
 
-            (
-                scale_ca_audio_hidden_states_a2v,
-                shift_ca_audio_hidden_states_a2v,
-                scale_ca_audio_hidden_states_v2a,
-                shift_ca_audio_hidden_states_v2a,
-                gate_out_v2a,
-            ) = self.get_av_ca_ada_values(
-                self.scale_shift_table_a2v_ca_audio,
-                ax.shape[0],
-                audio.cross_scale_shift_timestep,
-                audio.cross_gate_timestep,
-            )
+            if run_a2v and not perturbations.all_in_batch(PerturbationType.SKIP_A2V_CROSS_ATTN, self.idx):
+                scale_ca_video_a2v, shift_ca_video_a2v, gate_out_a2v = self.get_av_ca_ada_values(
+                    self.scale_shift_table_a2v_ca_video,
+                    vx.shape[0],
+                    video.cross_scale_shift_timestep,
+                    video.cross_gate_timestep,
+                    slice(0, 2),
+                )
+                vx_scaled = vx_norm3 * (1 + scale_ca_video_a2v) + shift_ca_video_a2v
+                del scale_ca_video_a2v, shift_ca_video_a2v
 
-            (
-                scale_ca_video_hidden_states_a2v,
-                shift_ca_video_hidden_states_a2v,
-                scale_ca_video_hidden_states_v2a,
-                shift_ca_video_hidden_states_v2a,
-                gate_out_a2v,
-            ) = self.get_av_ca_ada_values(
-                self.scale_shift_table_a2v_ca_video,
-                vx.shape[0],
-                video.cross_scale_shift_timestep,
-                video.cross_gate_timestep,
-            )
-
-            if run_a2v:
-                vx_scaled = vx_norm3 * (1 + scale_ca_video_hidden_states_a2v) + shift_ca_video_hidden_states_a2v
-                ax_scaled = ax_norm3 * (1 + scale_ca_audio_hidden_states_a2v) + shift_ca_audio_hidden_states_a2v
+                scale_ca_audio_a2v, shift_ca_audio_a2v, _ = self.get_av_ca_ada_values(
+                    self.scale_shift_table_a2v_ca_audio,
+                    ax.shape[0],
+                    audio.cross_scale_shift_timestep,
+                    audio.cross_gate_timestep,
+                    slice(0, 2),
+                )
+                ax_scaled = ax_norm3 * (1 + scale_ca_audio_a2v) + shift_ca_audio_a2v
+                del scale_ca_audio_a2v, shift_ca_audio_a2v
                 a2v_mask = perturbations.mask_like(PerturbationType.SKIP_A2V_CROSS_ATTN, self.idx, vx)
                 vx = vx + (
                     self.audio_to_video_attn(
@@ -980,10 +1163,27 @@ class BasicAVTransformerBlock(torch.nn.Module):
                     * gate_out_a2v
                     * a2v_mask
                 )
+                del gate_out_a2v, a2v_mask, vx_scaled, ax_scaled
 
-            if run_v2a:
-                ax_scaled = ax_norm3 * (1 + scale_ca_audio_hidden_states_v2a) + shift_ca_audio_hidden_states_v2a
-                vx_scaled = vx_norm3 * (1 + scale_ca_video_hidden_states_v2a) + shift_ca_video_hidden_states_v2a
+            if run_v2a and not perturbations.all_in_batch(PerturbationType.SKIP_V2A_CROSS_ATTN, self.idx):
+                scale_ca_audio_v2a, shift_ca_audio_v2a, gate_out_v2a = self.get_av_ca_ada_values(
+                    self.scale_shift_table_a2v_ca_audio,
+                    ax.shape[0],
+                    audio.cross_scale_shift_timestep,
+                    audio.cross_gate_timestep,
+                    slice(2, 4),
+                )
+                ax_scaled = ax_norm3 * (1 + scale_ca_audio_v2a) + shift_ca_audio_v2a
+                del scale_ca_audio_v2a, shift_ca_audio_v2a
+                scale_ca_video_v2a, shift_ca_video_v2a, _ = self.get_av_ca_ada_values(
+                    self.scale_shift_table_a2v_ca_video,
+                    vx.shape[0],
+                    video.cross_scale_shift_timestep,
+                    video.cross_gate_timestep,
+                    slice(2, 4),
+                )
+                vx_scaled = vx_norm3 * (1 + scale_ca_video_v2a) + shift_ca_video_v2a
+                del scale_ca_video_v2a, shift_ca_video_v2a
                 v2a_mask = perturbations.mask_like(PerturbationType.SKIP_V2A_CROSS_ATTN, self.idx, ax)
                 ax = ax + (
                     self.video_to_audio_attn(
@@ -995,40 +1195,53 @@ class BasicAVTransformerBlock(torch.nn.Module):
                     * gate_out_v2a
                     * v2a_mask
                 )
+                del gate_out_v2a, v2a_mask, ax_scaled, vx_scaled
 
-            del gate_out_a2v, gate_out_v2a
-            del (
-                scale_ca_video_hidden_states_a2v,
-                shift_ca_video_hidden_states_a2v,
-                scale_ca_audio_hidden_states_a2v,
-                shift_ca_audio_hidden_states_a2v,
-                scale_ca_video_hidden_states_v2a,
-                shift_ca_video_hidden_states_v2a,
-                scale_ca_audio_hidden_states_v2a,
-                shift_ca_audio_hidden_states_v2a,
-            )
+            del vx_norm3, ax_norm3
 
         if run_vx:
             vshift_mlp, vscale_mlp, vgate_mlp = self.get_ada_values(
-                self.scale_shift_table, vx.shape[0], video.timesteps, slice(3, None)
+                self.scale_shift_table, vx.shape[0], video.timesteps, slice(3, 6)
             )
             vx_scaled = rms_norm(vx, eps=self.norm_eps) * (1 + vscale_mlp) + vshift_mlp
             vx = vx + self.ff(vx_scaled) * vgate_mlp
 
-            del vshift_mlp, vscale_mlp, vgate_mlp
+            del vshift_mlp, vscale_mlp, vgate_mlp, vx_scaled
 
         if run_ax:
             ashift_mlp, ascale_mlp, agate_mlp = self.get_ada_values(
-                self.audio_scale_shift_table, ax.shape[0], audio.timesteps, slice(3, None)
+                self.audio_scale_shift_table, ax.shape[0], audio.timesteps, slice(3, 6)
             )
             ax_scaled = rms_norm(ax, eps=self.norm_eps) * (1 + ascale_mlp) + ashift_mlp
             ax = ax + self.audio_ff(ax_scaled) * agate_mlp
 
-            del ashift_mlp, ascale_mlp, agate_mlp
+            del ashift_mlp, ascale_mlp, agate_mlp, ax_scaled
 
         return replace(video, x=vx) if video is not None else None, replace(audio, x=ax) if audio is not None else None
 
 
+def apply_cross_attention_adaln(
+    x: torch.Tensor,
+    context: torch.Tensor,
+    attn: Attention,
+    q_shift: torch.Tensor,
+    q_scale: torch.Tensor,
+    q_gate: torch.Tensor,
+    prompt_scale_shift_table: torch.Tensor,
+    prompt_timestep: torch.Tensor,
+    context_mask: torch.Tensor | None = None,
+    norm_eps: float = 1e-6,
+) -> torch.Tensor:
+    batch_size = x.shape[0]
+    shift_kv, scale_kv = (
+        prompt_scale_shift_table[None, None].to(device=x.device, dtype=x.dtype)
+        + prompt_timestep.reshape(batch_size, prompt_timestep.shape[1], 2, -1)
+    ).unbind(dim=2)
+    attn_input = rms_norm(x, eps=norm_eps) * (1 + q_scale) + q_shift
+    encoder_hidden_states = context * (1 + scale_kv) + shift_kv
+    return attn(attn_input, context=encoder_hidden_states, mask=context_mask) * q_gate
+
+
 class GELUApprox(torch.nn.Module):
     def __init__(self, dim_in: int, dim_out: int) -> None:
         super().__init__()
@@ -1093,6 +1306,8 @@ class LTXModel(torch.nn.Module):
         av_ca_timestep_scale_multiplier: int = 1000,
         rope_type: LTXRopeType = LTXRopeType.SPLIT,
         double_precision_rope: bool = True,
+        apply_gated_attention: bool = False,
+        cross_attention_adaln: bool = False,
     ):
         super().__init__()
         self._enable_gradient_checkpointing = False
@@ -1102,6 +1317,7 @@ class LTXModel(torch.nn.Module):
         self.timestep_scale_multiplier = timestep_scale_multiplier
         self.positional_embedding_theta = positional_embedding_theta
         self.model_type = model_type
+        self.cross_attention_adaln = cross_attention_adaln
         cross_pe_max_pos = None
         if model_type.is_video_enabled():
             if positional_embedding_max_pos is None:
@@ -1144,8 +1360,13 @@ class LTXModel(torch.nn.Module):
             audio_attention_head_dim=audio_attention_head_dim if model_type.is_audio_enabled() else 0,
             audio_cross_attention_dim=audio_cross_attention_dim,
             norm_eps=norm_eps,
+            apply_gated_attention=apply_gated_attention,
         )
 
+    @property
+    def _adaln_embedding_coefficient(self) -> int:
+        return adaln_embedding_coefficient(self.cross_attention_adaln)
+
     def _init_video(
         self,
         in_channels: int,
@@ -1156,14 +1377,15 @@ class LTXModel(torch.nn.Module):
         """Initialize video-specific components."""
         # Video input components
         self.patchify_proj = torch.nn.Linear(in_channels, self.inner_dim, bias=True)
-
-        self.adaln_single = AdaLayerNormSingle(self.inner_dim)
+        self.adaln_single = AdaLayerNormSingle(self.inner_dim, embedding_coefficient=self._adaln_embedding_coefficient)
+        self.prompt_adaln_single = AdaLayerNormSingle(self.inner_dim, embedding_coefficient=2) if self.cross_attention_adaln else None
 
         # Video caption projection
-        self.caption_projection = PixArtAlphaTextProjection(
-            in_features=caption_channels,
-            hidden_size=self.inner_dim,
-        )
+        if caption_channels is not None:
+            self.caption_projection = PixArtAlphaTextProjection(
+                in_features=caption_channels,
+                hidden_size=self.inner_dim,
+            )
 
         # Video output components
         self.scale_shift_table = torch.nn.Parameter(torch.empty(2, self.inner_dim))
@@ -1182,15 +1404,15 @@ class LTXModel(torch.nn.Module):
         # Audio input components
         self.audio_patchify_proj = torch.nn.Linear(in_channels, self.audio_inner_dim, bias=True)
 
-        self.audio_adaln_single = AdaLayerNormSingle(
-            self.audio_inner_dim,
-        )
+        self.audio_adaln_single = AdaLayerNormSingle(self.audio_inner_dim, embedding_coefficient=self._adaln_embedding_coefficient)
+        self.audio_prompt_adaln_single = AdaLayerNormSingle(self.audio_inner_dim, embedding_coefficient=2) if self.cross_attention_adaln else None
 
         # Audio caption projection
-        self.audio_caption_projection = PixArtAlphaTextProjection(
-            in_features=caption_channels,
-            hidden_size=self.audio_inner_dim,
-        )
+        if caption_channels is not None:
+            self.audio_caption_projection = PixArtAlphaTextProjection(
+                in_features=caption_channels,
+                hidden_size=self.audio_inner_dim,
+            )
 
         # Audio output components
         self.audio_scale_shift_table = torch.nn.Parameter(torch.empty(2, self.audio_inner_dim))
@@ -1232,7 +1454,6 @@ class LTXModel(torch.nn.Module):
             self.video_args_preprocessor = MultiModalTransformerArgsPreprocessor(
                 patchify_proj=self.patchify_proj,
                 adaln=self.adaln_single,
-                caption_projection=self.caption_projection,
                 cross_scale_shift_adaln=self.av_ca_video_scale_shift_adaln_single,
                 cross_gate_adaln=self.av_ca_a2v_gate_adaln_single,
                 inner_dim=self.inner_dim,
@@ -1246,11 +1467,12 @@ class LTXModel(torch.nn.Module):
                 positional_embedding_theta=self.positional_embedding_theta,
                 rope_type=self.rope_type,
                 av_ca_timestep_scale_multiplier=self.av_ca_timestep_scale_multiplier,
+                caption_projection=getattr(self, "caption_projection", None),
+                prompt_adaln=getattr(self, "prompt_adaln_single", None),
             )
             self.audio_args_preprocessor = MultiModalTransformerArgsPreprocessor(
                 patchify_proj=self.audio_patchify_proj,
                 adaln=self.audio_adaln_single,
-                caption_projection=self.audio_caption_projection,
                 cross_scale_shift_adaln=self.av_ca_audio_scale_shift_adaln_single,
                 cross_gate_adaln=self.av_ca_v2a_gate_adaln_single,
                 inner_dim=self.audio_inner_dim,
@@ -1264,12 +1486,13 @@ class LTXModel(torch.nn.Module):
                 positional_embedding_theta=self.positional_embedding_theta,
                 rope_type=self.rope_type,
                 av_ca_timestep_scale_multiplier=self.av_ca_timestep_scale_multiplier,
+                caption_projection=getattr(self, "audio_caption_projection", None),
+                prompt_adaln=getattr(self, "audio_prompt_adaln_single", None),
             )
         elif self.model_type.is_video_enabled():
             self.video_args_preprocessor = TransformerArgsPreprocessor(
                 patchify_proj=self.patchify_proj,
                 adaln=self.adaln_single,
-                caption_projection=self.caption_projection,
                 inner_dim=self.inner_dim,
                 max_pos=self.positional_embedding_max_pos,
                 num_attention_heads=self.num_attention_heads,
@@ -1278,12 +1501,13 @@ class LTXModel(torch.nn.Module):
                 double_precision_rope=self.double_precision_rope,
                 positional_embedding_theta=self.positional_embedding_theta,
                 rope_type=self.rope_type,
+                caption_projection=getattr(self, "caption_projection", None),
+                prompt_adaln=getattr(self, "prompt_adaln_single", None),
             )
         elif self.model_type.is_audio_enabled():
             self.audio_args_preprocessor = TransformerArgsPreprocessor(
                 patchify_proj=self.audio_patchify_proj,
                 adaln=self.audio_adaln_single,
-                caption_projection=self.audio_caption_projection,
                 inner_dim=self.audio_inner_dim,
                 max_pos=self.audio_positional_embedding_max_pos,
                 num_attention_heads=self.audio_num_attention_heads,
@@ -1292,6 +1516,8 @@ class LTXModel(torch.nn.Module):
                 double_precision_rope=self.double_precision_rope,
                 positional_embedding_theta=self.positional_embedding_theta,
                 rope_type=self.rope_type,
+                caption_projection=getattr(self, "audio_caption_projection", None),
+                prompt_adaln=getattr(self, "audio_prompt_adaln_single", None),
             )
 
     def _init_transformer_blocks(
@@ -1302,6 +1528,7 @@ class LTXModel(torch.nn.Module):
         audio_attention_head_dim: int,
         audio_cross_attention_dim: int,
         norm_eps: float,
+        apply_gated_attention: bool,
     ) -> None:
         """Initialize transformer blocks for LTX."""
         video_config = (
@@ -1310,6 +1537,8 @@ class LTXModel(torch.nn.Module):
                 heads=self.num_attention_heads,
                 d_head=attention_head_dim,
                 context_dim=cross_attention_dim,
+                apply_gated_attention=apply_gated_attention,
+                cross_attention_adaln=self.cross_attention_adaln,
             )
             if self.model_type.is_video_enabled()
             else None
@@ -1320,6 +1549,8 @@ class LTXModel(torch.nn.Module):
                 heads=self.audio_num_attention_heads,
                 d_head=audio_attention_head_dim,
                 context_dim=audio_cross_attention_dim,
+                apply_gated_attention=apply_gated_attention,
+                cross_attention_adaln=self.cross_attention_adaln,
             )
             if self.model_type.is_audio_enabled()
             else None
@@ -1352,28 +1583,21 @@ class LTXModel(torch.nn.Module):
         video: TransformerArgs | None,
         audio: TransformerArgs | None,
         perturbations: BatchedPerturbationConfig,
+        use_gradient_checkpointing: bool = False,
+        use_gradient_checkpointing_offload: bool = False,
     ) -> tuple[TransformerArgs, TransformerArgs]:
         """Process transformer blocks for LTXAV."""
 
         # Process transformer blocks
         for block in self.transformer_blocks:
-            if self._enable_gradient_checkpointing and self.training:
-                # Use gradient checkpointing to save memory during training.
-                # With use_reentrant=False, we can pass dataclasses directly -
-                # PyTorch will track all tensor leaves in the computation graph.
-                video, audio = torch.utils.checkpoint.checkpoint(
-                    block,
-                    video,
-                    audio,
-                    perturbations,
-                    use_reentrant=False,
-                )
-            else:
-                video, audio = block(
-                    video=video,
-                    audio=audio,
-                    perturbations=perturbations,
-                )
+            video, audio = gradient_checkpoint_forward(
+                block,
+                use_gradient_checkpointing,
+                use_gradient_checkpointing_offload,
+                video=video,
+                audio=audio,
+                perturbations=perturbations,
+            )
 
         return video, audio
 
@@ -1398,7 +1622,12 @@ class LTXModel(torch.nn.Module):
         return x
 
     def _forward(
-        self, video: Modality | None, audio: Modality | None, perturbations: BatchedPerturbationConfig
+        self,
+        video: Modality | None,
+        audio: Modality | None,
+        perturbations: BatchedPerturbationConfig,
+        use_gradient_checkpointing: bool = False,
+        use_gradient_checkpointing_offload: bool = False,
     ) -> tuple[torch.Tensor, torch.Tensor]:
         """
         Forward pass for LTX models.
@@ -1410,13 +1639,15 @@ class LTXModel(torch.nn.Module):
         if not self.model_type.is_audio_enabled() and audio is not None:
             raise ValueError("Audio is not enabled for this model")
 
-        video_args = self.video_args_preprocessor.prepare(video) if video is not None else None
-        audio_args = self.audio_args_preprocessor.prepare(audio) if audio is not None else None
+        video_args = self.video_args_preprocessor.prepare(video, audio) if video is not None else None
+        audio_args = self.audio_args_preprocessor.prepare(audio, video) if audio is not None else None
         # Process transformer blocks
         video_out, audio_out = self._process_transformer_blocks(
             video=video_args,
             audio=audio_args,
             perturbations=perturbations,
+            use_gradient_checkpointing=use_gradient_checkpointing,
+            use_gradient_checkpointing_offload=use_gradient_checkpointing_offload,
         )
 
         # Process output
@@ -1440,12 +1671,12 @@ class LTXModel(torch.nn.Module):
         )
         return vx, ax
 
-    def forward(self, video_latents, video_positions, video_context, video_timesteps, audio_latents, audio_positions, audio_context, audio_timesteps):
+    def forward(self, video_latents, video_positions, video_context, video_timesteps, audio_latents, audio_positions, audio_context, audio_timesteps, sigma, use_gradient_checkpointing=False, use_gradient_checkpointing_offload=False):
         cross_pe_max_pos = None
         if self.model_type.is_video_enabled() and self.model_type.is_audio_enabled():
             cross_pe_max_pos = max(self.positional_embedding_max_pos[0], self.audio_positional_embedding_max_pos[0])
         self._init_preprocessors(cross_pe_max_pos)
-        video = Modality(video_latents, video_timesteps, video_positions, video_context)
-        audio = Modality(audio_latents, audio_timesteps, audio_positions, audio_context) if audio_latents is not None else None
-        vx, ax = self._forward(video=video, audio=audio, perturbations=None)
+        video = Modality(video_latents, sigma, video_timesteps, video_positions, video_context)
+        audio = Modality(audio_latents, sigma, audio_timesteps, audio_positions, audio_context) if audio_latents is not None else None
+        vx, ax = self._forward(video=video, audio=audio, perturbations=None, use_gradient_checkpointing=use_gradient_checkpointing, use_gradient_checkpointing_offload=use_gradient_checkpointing_offload)
         return vx, ax

diffsynth/models/ltx2_text_encoder.py

@@ -1,4 +1,7 @@
+import math
 import torch
+import torch.nn as nn
+from einops import rearrange
 from transformers import Gemma3ForConditionalGeneration, Gemma3Config, AutoTokenizer
 from .ltx2_dit import (LTXRopeType, generate_freq_grid_np, generate_freq_grid_pytorch, precompute_freqs_cis, Attention,
                        FeedForward)
@@ -147,14 +150,14 @@ class LTXVGemmaTokenizer:
         return out
 
 
-class GemmaFeaturesExtractorProjLinear(torch.nn.Module):
+class GemmaFeaturesExtractorProjLinear(nn.Module):
     """
     Feature extractor module for Gemma models.
     This module applies a single linear projection to the input tensor.
     It expects a flattened feature tensor of shape (batch_size, 3840*49).
     The linear layer maps this to a (batch_size, 3840) embedding.
     Attributes:
-        aggregate_embed (torch.nn.Linear): Linear projection layer.
+        aggregate_embed (nn.Linear): Linear projection layer.
     """
 
     def __init__(self) -> None:
@@ -163,26 +166,65 @@ class GemmaFeaturesExtractorProjLinear(torch.nn.Module):
         The input dimension is expected to be 3840 * 49, and the output is 3840.
         """
         super().__init__()
-        self.aggregate_embed = torch.nn.Linear(3840 * 49, 3840, bias=False)
+        self.aggregate_embed = nn.Linear(3840 * 49, 3840, bias=False)
 
-    def forward(self, x: torch.Tensor) -> torch.Tensor:
-        """
-        Forward pass for the feature extractor.
-        Args:
-            x (torch.Tensor): Input tensor of shape (batch_size, 3840 * 49).
-        Returns:
-            torch.Tensor: Output tensor of shape (batch_size, 3840).
-        """
-        return self.aggregate_embed(x)
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+        padding_side: str = "left",
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
+        encoded = torch.stack(hidden_states, dim=-1) if isinstance(hidden_states, (list, tuple)) else hidden_states
+        dtype = encoded.dtype
+        sequence_lengths = attention_mask.sum(dim=-1)
+        normed = _norm_and_concat_padded_batch(encoded, sequence_lengths, padding_side)
+        features = self.aggregate_embed(normed.to(dtype))
+        return features, features
 
 
-class _BasicTransformerBlock1D(torch.nn.Module):
+class GemmaSeperatedFeaturesExtractorProjLinear(nn.Module):
+    """22B: per-token RMS norm → rescale → dual aggregate embeds"""
+
+    def __init__(
+        self,
+        num_layers: int,
+        embedding_dim: int,
+        video_inner_dim: int,
+        audio_inner_dim: int,
+    ):
+        super().__init__()
+        in_dim = embedding_dim * num_layers
+        self.video_aggregate_embed = torch.nn.Linear(in_dim, video_inner_dim, bias=True)
+        self.audio_aggregate_embed = torch.nn.Linear(in_dim, audio_inner_dim, bias=True)
+        self.embedding_dim = embedding_dim
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+        padding_side: str = "left",  # noqa: ARG002
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
+        encoded = torch.stack(hidden_states, dim=-1) if isinstance(hidden_states, (list, tuple)) else hidden_states
+        normed = norm_and_concat_per_token_rms(encoded, attention_mask)
+        normed = normed.to(encoded.dtype)
+        v_dim = self.video_aggregate_embed.out_features
+        video = self.video_aggregate_embed(_rescale_norm(normed, v_dim, self.embedding_dim))
+        audio = None
+        if self.audio_aggregate_embed is not None:
+            a_dim = self.audio_aggregate_embed.out_features
+            audio = self.audio_aggregate_embed(_rescale_norm(normed, a_dim, self.embedding_dim))
+        return video, audio
+
+
+
+class _BasicTransformerBlock1D(nn.Module):
     def __init__(
         self,
         dim: int,
         heads: int,
         dim_head: int,
         rope_type: LTXRopeType = LTXRopeType.INTERLEAVED,
+        apply_gated_attention: bool = False,
     ):
         super().__init__()
 
@@ -191,6 +233,7 @@ class _BasicTransformerBlock1D(torch.nn.Module):
             heads=heads,
             dim_head=dim_head,
             rope_type=rope_type,
+            apply_gated_attention=apply_gated_attention,
         )
 
         self.ff = FeedForward(
@@ -231,7 +274,7 @@ class _BasicTransformerBlock1D(torch.nn.Module):
         return hidden_states
 
 
-class Embeddings1DConnector(torch.nn.Module):
+class Embeddings1DConnector(nn.Module):
     """
     Embeddings1DConnector applies a 1D transformer-based processing to sequential embeddings (e.g., for video, audio, or
     other modalities). It supports rotary positional encoding (rope), optional causal temporal positioning, and can
@@ -263,6 +306,7 @@ class Embeddings1DConnector(torch.nn.Module):
         num_learnable_registers: int | None = 128,
         rope_type: LTXRopeType = LTXRopeType.SPLIT,
         double_precision_rope: bool = True,
+        apply_gated_attention: bool = False,
     ):
         super().__init__()
         self.num_attention_heads = num_attention_heads
@@ -274,13 +318,14 @@ class Embeddings1DConnector(torch.nn.Module):
         )
         self.rope_type = rope_type
         self.double_precision_rope = double_precision_rope
-        self.transformer_1d_blocks = torch.nn.ModuleList(
+        self.transformer_1d_blocks = nn.ModuleList(
             [
                 _BasicTransformerBlock1D(
                     dim=self.inner_dim,
                     heads=num_attention_heads,
                     dim_head=attention_head_dim,
                     rope_type=rope_type,
+                    apply_gated_attention=apply_gated_attention,
                 )
                 for _ in range(num_layers)
             ]
@@ -288,7 +333,7 @@ class Embeddings1DConnector(torch.nn.Module):
 
         self.num_learnable_registers = num_learnable_registers
         if self.num_learnable_registers:
-            self.learnable_registers = torch.nn.Parameter(
+            self.learnable_registers = nn.Parameter(
                 torch.rand(self.num_learnable_registers, self.inner_dim, dtype=torch.bfloat16) * 2.0 - 1.0
             )
 
@@ -307,7 +352,7 @@ class Embeddings1DConnector(torch.nn.Module):
         non_zero_hidden_states = hidden_states[:, attention_mask_binary.squeeze().bool(), :]
         non_zero_nums = non_zero_hidden_states.shape[1]
         pad_length = hidden_states.shape[1] - non_zero_nums
-        adjusted_hidden_states = torch.nn.functional.pad(non_zero_hidden_states, pad=(0, 0, 0, pad_length), value=0)
+        adjusted_hidden_states = nn.functional.pad(non_zero_hidden_states, pad=(0, 0, 0, pad_length), value=0)
         flipped_mask = torch.flip(attention_mask_binary, dims=[1])
         hidden_states = flipped_mask * adjusted_hidden_states + (1 - flipped_mask) * learnable_registers
 
@@ -358,9 +403,147 @@ class Embeddings1DConnector(torch.nn.Module):
         return hidden_states, attention_mask
 
 
-class LTX2TextEncoderPostModules(torch.nn.Module):
-    def __init__(self,):
+class LTX2TextEncoderPostModules(nn.Module):
+    def __init__(
+        self,
+        separated_audio_video: bool = False,
+        embedding_dim_gemma: int = 3840,
+        num_layers_gemma: int = 49,
+        video_attention_heads: int = 32,
+        video_attention_head_dim: int = 128,
+        audio_attention_heads: int = 32,
+        audio_attention_head_dim: int = 64,
+        num_connector_layers: int = 2,
+        apply_gated_attention: bool = False,
+    ):
         super().__init__()
-        self.feature_extractor_linear = GemmaFeaturesExtractorProjLinear()
-        self.embeddings_connector = Embeddings1DConnector()
-        self.audio_embeddings_connector = Embeddings1DConnector()
+        if not separated_audio_video:
+            self.feature_extractor_linear = GemmaFeaturesExtractorProjLinear()
+            self.embeddings_connector = Embeddings1DConnector()
+            self.audio_embeddings_connector = Embeddings1DConnector()
+        else:
+            # LTX-2.3
+            self.feature_extractor_linear = GemmaSeperatedFeaturesExtractorProjLinear(
+                num_layers_gemma, embedding_dim_gemma, video_attention_heads * video_attention_head_dim,
+                audio_attention_heads * audio_attention_head_dim)
+            self.embeddings_connector = Embeddings1DConnector(
+                attention_head_dim=video_attention_head_dim,
+                num_attention_heads=video_attention_heads,
+                num_layers=num_connector_layers,
+                apply_gated_attention=apply_gated_attention,
+            )
+            self.audio_embeddings_connector = Embeddings1DConnector(
+                attention_head_dim=audio_attention_head_dim,
+                num_attention_heads=audio_attention_heads,
+                num_layers=num_connector_layers,
+                apply_gated_attention=apply_gated_attention,
+            )
+
+    def create_embeddings(
+        self,
+        video_features: torch.Tensor,
+        audio_features: torch.Tensor | None,
+        additive_attention_mask: torch.Tensor,
+    ) -> tuple[torch.Tensor, torch.Tensor | None, torch.Tensor]:
+        video_encoded, video_mask = self.embeddings_connector(video_features, additive_attention_mask)
+        video_encoded, binary_mask = _to_binary_mask(video_encoded, video_mask)
+        audio_encoded, _ = self.audio_embeddings_connector(audio_features, additive_attention_mask)
+
+        return video_encoded, audio_encoded, binary_mask
+
+    def process_hidden_states(
+        self,
+        hidden_states: tuple[torch.Tensor, ...],
+        attention_mask: torch.Tensor,
+        padding_side: str = "left",
+    ):
+        video_feats, audio_feats = self.feature_extractor_linear(hidden_states, attention_mask, padding_side)
+        additive_mask = _convert_to_additive_mask(attention_mask, video_feats.dtype)
+        video_enc, audio_enc, binary_mask = self.create_embeddings(video_feats, audio_feats, additive_mask)
+        return video_enc, audio_enc, binary_mask
+
+
+def _norm_and_concat_padded_batch(
+    encoded_text: torch.Tensor,
+    sequence_lengths: torch.Tensor,
+    padding_side: str = "right",
+) -> torch.Tensor:
+    """Normalize and flatten multi-layer hidden states, respecting padding.
+    Performs per-batch, per-layer normalization using masked mean and range,
+    then concatenates across the layer dimension.
+    Args:
+        encoded_text: Hidden states of shape [batch, seq_len, hidden_dim, num_layers].
+        sequence_lengths: Number of valid (non-padded) tokens per batch item.
+        padding_side: Whether padding is on "left" or "right".
+    Returns:
+        Normalized tensor of shape [batch, seq_len, hidden_dim * num_layers],
+        with padded positions zeroed out.
+    """
+    b, t, d, l = encoded_text.shape  # noqa: E741
+    device = encoded_text.device
+    # Build mask: [B, T, 1, 1]
+    token_indices = torch.arange(t, device=device)[None, :]  # [1, T]
+    if padding_side == "right":
+        # For right padding, valid tokens are from 0 to sequence_length-1
+        mask = token_indices < sequence_lengths[:, None]  # [B, T]
+    elif padding_side == "left":
+        # For left padding, valid tokens are from (T - sequence_length) to T-1
+        start_indices = t - sequence_lengths[:, None]  # [B, 1]
+        mask = token_indices >= start_indices  # [B, T]
+    else:
+        raise ValueError(f"padding_side must be 'left' or 'right', got {padding_side}")
+    mask = rearrange(mask, "b t -> b t 1 1")
+    eps = 1e-6
+    # Compute masked mean: [B, 1, 1, L]
+    masked = encoded_text.masked_fill(~mask, 0.0)
+    denom = (sequence_lengths * d).view(b, 1, 1, 1)
+    mean = masked.sum(dim=(1, 2), keepdim=True) / (denom + eps)
+    # Compute masked min/max: [B, 1, 1, L]
+    x_min = encoded_text.masked_fill(~mask, float("inf")).amin(dim=(1, 2), keepdim=True)
+    x_max = encoded_text.masked_fill(~mask, float("-inf")).amax(dim=(1, 2), keepdim=True)
+    range_ = x_max - x_min
+    # Normalize only the valid tokens
+    normed = 8 * (encoded_text - mean) / (range_ + eps)
+    # concat to be [Batch, T,  D * L] - this preserves the original structure
+    normed = normed.reshape(b, t, -1)  # [B, T, D * L]
+    # Apply mask to preserve original padding (set padded positions to 0)
+    mask_flattened = rearrange(mask, "b t 1 1 -> b t 1").expand(-1, -1, d * l)
+    normed = normed.masked_fill(~mask_flattened, 0.0)
+
+    return normed
+
+
+def _convert_to_additive_mask(attention_mask: torch.Tensor, dtype: torch.dtype) -> torch.Tensor:
+    return (attention_mask - 1).to(dtype).reshape(
+        (attention_mask.shape[0], 1, -1, attention_mask.shape[-1])) * torch.finfo(dtype).max
+
+def _to_binary_mask(encoded: torch.Tensor, encoded_mask: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+    """Convert connector output mask to binary mask and apply to encoded tensor."""
+    binary_mask = (encoded_mask < 0.000001).to(torch.int64)
+    binary_mask = binary_mask.reshape([encoded.shape[0], encoded.shape[1], 1])
+    encoded = encoded * binary_mask
+    return encoded, binary_mask
+
+
+def norm_and_concat_per_token_rms(
+    encoded_text: torch.Tensor,
+    attention_mask: torch.Tensor,
+) -> torch.Tensor:
+    """Per-token RMSNorm normalization for V2 models.
+    Args:
+        encoded_text: [B, T, D, L]
+        attention_mask: [B, T] binary mask
+    Returns:
+        [B, T, D*L] normalized tensor with padding zeroed out.
+    """
+    B, T, D, L = encoded_text.shape  # noqa: N806
+    variance = torch.mean(encoded_text**2, dim=2, keepdim=True)  # [B,T,1,L]
+    normed = encoded_text * torch.rsqrt(variance + 1e-6)
+    normed = normed.reshape(B, T, D * L)
+    mask_3d = attention_mask.bool().unsqueeze(-1)  # [B, T, 1]
+    return torch.where(mask_3d, normed, torch.zeros_like(normed))
+
+
+def _rescale_norm(x: torch.Tensor, target_dim: int, source_dim: int) -> torch.Tensor:
+    """Rescale normalization: x * sqrt(target_dim / source_dim)."""
+    return x * math.sqrt(target_dim / source_dim)

diffsynth/models/ltx2_video_vae.py

@@ -555,9 +555,6 @@ class PerChannelStatistics(nn.Module):
         super().__init__()
         self.register_buffer("std-of-means", torch.empty(latent_channels))
         self.register_buffer("mean-of-means", torch.empty(latent_channels))
-        self.register_buffer("mean-of-stds", torch.empty(latent_channels))
-        self.register_buffer("mean-of-stds_over_std-of-means", torch.empty(latent_channels))
-        self.register_buffer("channel", torch.empty(latent_channels))
 
     def un_normalize(self, x: torch.Tensor) -> torch.Tensor:
         return (x * self.get_buffer("std-of-means").view(1, -1, 1, 1, 1).to(x)) + self.get_buffer("mean-of-means").view(
@@ -1335,27 +1332,34 @@ class LTX2VideoEncoder(nn.Module):
         norm_layer: NormLayerType = NormLayerType.PIXEL_NORM,
         latent_log_var: LogVarianceType = LogVarianceType.UNIFORM,
         encoder_spatial_padding_mode: PaddingModeType = PaddingModeType.ZEROS,
+        encoder_version: str = "ltx-2",
     ):
         super().__init__()
-        encoder_blocks = [['res_x', {
-            'num_layers': 4
-        }], ['compress_space_res', {
-            'multiplier': 2
-        }], ['res_x', {
-            'num_layers': 6
-        }], ['compress_time_res', {
-            'multiplier': 2
-        }], ['res_x', {
-            'num_layers': 6
-        }], ['compress_all_res', {
-            'multiplier': 2
-        }], ['res_x', {
-            'num_layers': 2
-        }], ['compress_all_res', {
-            'multiplier': 2
-        }], ['res_x', {
-            'num_layers': 2
-        }]]
+        if encoder_version == "ltx-2":
+            encoder_blocks = [
+                ['res_x', {'num_layers': 4}],
+                ['compress_space_res', {'multiplier': 2}],
+                ['res_x', {'num_layers': 6}],
+                ['compress_time_res', {'multiplier': 2}],
+                ['res_x', {'num_layers': 6}],
+                ['compress_all_res', {'multiplier': 2}],
+                ['res_x', {'num_layers': 2}],
+                ['compress_all_res', {'multiplier': 2}],
+                ['res_x', {'num_layers': 2}]
+            ]
+        else:
+            # LTX-2.3
+            encoder_blocks = [
+                ["res_x", {"num_layers": 4}],
+                ["compress_space_res", {"multiplier": 2}],
+                ["res_x", {"num_layers": 6}],
+                ["compress_time_res", {"multiplier": 2}],
+                ["res_x", {"num_layers": 4}],
+                ["compress_all_res", {"multiplier": 2}],
+                ["res_x", {"num_layers": 2}],
+                ["compress_all_res", {"multiplier": 1}],
+                ["res_x", {"num_layers": 2}]
+            ]
         self.patch_size = patch_size
         self.norm_layer = norm_layer
         self.latent_channels = out_channels
@@ -1435,8 +1439,8 @@ class LTX2VideoEncoder(nn.Module):
         # Validate frame count
         frames_count = sample.shape[2]
         if ((frames_count - 1) % 8) != 0:
-            raise ValueError("Invalid number of frames: Encode input must have 1 + 8 * x frames "
-                             "(e.g., 1, 9, 17, ...). Please check your input.")
+            frames_to_crop = (frames_count - 1) % 8
+            sample = sample[:, :, :-frames_to_crop, ...]
 
         # Initial spatial compression: trade spatial resolution for channel depth
         # This reduces H,W by patch_size and increases channels, making convolutions more efficient
@@ -1712,17 +1716,21 @@ def _make_decoder_block(
             spatial_padding_mode=spatial_padding_mode,
         )
     elif block_name == "compress_time":
+        out_channels = in_channels // block_config.get("multiplier", 1)
         block = DepthToSpaceUpsample(
             dims=convolution_dimensions,
             in_channels=in_channels,
             stride=(2, 1, 1),
+            out_channels_reduction_factor=block_config.get("multiplier", 1),
             spatial_padding_mode=spatial_padding_mode,
         )
     elif block_name == "compress_space":
+        out_channels = in_channels // block_config.get("multiplier", 1)
         block = DepthToSpaceUpsample(
             dims=convolution_dimensions,
             in_channels=in_channels,
             stride=(1, 2, 2),
+            out_channels_reduction_factor=block_config.get("multiplier", 1),
             spatial_padding_mode=spatial_padding_mode,
         )
     elif block_name == "compress_all":
@@ -1782,6 +1790,8 @@ class LTX2VideoDecoder(nn.Module):
         causal: bool = False,
         timestep_conditioning: bool = False,
         decoder_spatial_padding_mode: PaddingModeType = PaddingModeType.REFLECT,
+        decoder_version: str = "ltx-2",
+        base_channels: int = 128,
     ):
         super().__init__()
 
@@ -1790,28 +1800,29 @@ class LTX2VideoDecoder(nn.Module):
         # video inputs by a factor of 8 in the temporal dimension and 32 in
         # each spatial dimension (height and width). This parameter determines how
         # many video frames and pixels correspond to a single latent cell.
-        decoder_blocks = [['res_x', {
-            'num_layers': 5,
-            'inject_noise': False
-        }], ['compress_all', {
-            'residual': True,
-            'multiplier': 2
-        }], ['res_x', {
-            'num_layers': 5,
-            'inject_noise': False
-        }], ['compress_all', {
-            'residual': True,
-            'multiplier': 2
-        }], ['res_x', {
-            'num_layers': 5,
-            'inject_noise': False
-        }], ['compress_all', {
-            'residual': True,
-            'multiplier': 2
-        }], ['res_x', {
-            'num_layers': 5,
-            'inject_noise': False
-        }]]
+        if decoder_version == "ltx-2":
+            decoder_blocks = [
+                ['res_x', {'num_layers': 5, 'inject_noise': False}],
+                ['compress_all', {'residual': True, 'multiplier': 2}],
+                ['res_x', {'num_layers': 5, 'inject_noise': False}],
+                ['compress_all', {'residual': True, 'multiplier': 2}],
+                ['res_x', {'num_layers': 5, 'inject_noise': False}],
+                ['compress_all', {'residual': True, 'multiplier': 2}],
+                ['res_x', {'num_layers': 5, 'inject_noise': False}]
+            ]
+        else:
+            # LTX-2.3
+            decoder_blocks = [
+                ["res_x", {"num_layers": 4}],
+                ["compress_space", {"multiplier": 2}],
+                ["res_x", {"num_layers": 6}],
+                ["compress_time", {"multiplier": 2}],
+                ["res_x", {"num_layers": 4}],
+                ["compress_all", {"multiplier": 1}],
+                ["res_x", {"num_layers": 2}],
+                ["compress_all", {"multiplier": 2}],
+                ["res_x", {"num_layers": 2}]
+            ]
         self.video_downscale_factors = SpatioTemporalScaleFactors(
             time=8,
             width=32,
@@ -1831,15 +1842,9 @@ class LTX2VideoDecoder(nn.Module):
         self.decode_noise_scale = 0.025
         self.decode_timestep = 0.05
 
-        # Compute initial feature_channels by going through blocks in reverse
-        # This determines the channel width at the start of the decoder
-        feature_channels = in_channels
-        for block_name, block_params in list(reversed(decoder_blocks)):
-            block_config = block_params if isinstance(block_params, dict) else {}
-            if block_name == "res_x_y":
-                feature_channels = feature_channels * block_config.get("multiplier", 2)
-            if block_name == "compress_all":
-                feature_channels = feature_channels * block_config.get("multiplier", 1)
+        # LTX VAE decoder architecture uses 3 upsampler blocks with multiplier equals to 2.
+        # Hence the total feature_channels is multiplied by 8 (2^3).
+        feature_channels = base_channels * 8
 
         self.conv_in = make_conv_nd(
             dims=convolution_dimensions,

diffsynth/models/model_loader.py

@@ -1,6 +1,6 @@
 from ..core.loader import load_model, hash_model_file
 from ..core.vram import AutoWrappedModule
-from ..configs import MODEL_CONFIGS, VRAM_MANAGEMENT_MODULE_MAPS
+from ..configs import MODEL_CONFIGS, VRAM_MANAGEMENT_MODULE_MAPS, VERSION_CHECKER_MAPS
 import importlib, json, torch
 
 
@@ -22,7 +22,8 @@ class ModelPool:
     def fetch_module_map(self, model_class, vram_config):
         if self.need_to_enable_vram_management(vram_config):
             if model_class in VRAM_MANAGEMENT_MODULE_MAPS:
-                module_map = {self.import_model_class(source): self.import_model_class(target) for source, target in VRAM_MANAGEMENT_MODULE_MAPS[model_class].items()}
+                vram_module_map = VRAM_MANAGEMENT_MODULE_MAPS[model_class] if model_class not in VERSION_CHECKER_MAPS else VERSION_CHECKER_MAPS[model_class]()
+                module_map = {self.import_model_class(source): self.import_model_class(target) for source, target in vram_module_map.items()}
             else:
                 module_map = {self.import_model_class(model_class): AutoWrappedModule}
         else:

diffsynth/models/mova_audio_dit.py

@@ -0,0 +1,57 @@
+import torch
+import torch.nn as nn
+from .wan_video_dit import WanModel, precompute_freqs_cis, sinusoidal_embedding_1d
+from einops import rearrange
+from ..core import gradient_checkpoint_forward
+
+def precompute_freqs_cis_1d(dim: int, end: int = 16384, theta: float = 10000.0):
+    f_freqs_cis = precompute_freqs_cis(dim, end, theta)
+    return f_freqs_cis.chunk(3, dim=-1)
+
+class MovaAudioDit(WanModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        head_dim = kwargs.get("dim", 1536) // kwargs.get("num_heads", 12)
+        self.freqs = precompute_freqs_cis_1d(head_dim)
+        self.patch_embedding = nn.Conv1d(
+            kwargs.get("in_dim", 128), kwargs.get("dim", 1536), kernel_size=[1], stride=[1]
+        )
+
+    def precompute_freqs_cis(self, dim: int, end: int = 16384, theta: float = 10000.0):
+        self.f_freqs_cis = precompute_freqs_cis_1d(dim, end, theta)
+
+    def forward(self,
+                x: torch.Tensor,
+                timestep: torch.Tensor,
+                context: torch.Tensor,
+                use_gradient_checkpointing: bool = False,
+                use_gradient_checkpointing_offload: bool = False,
+                **kwargs,
+                ):
+        t = self.time_embedding(sinusoidal_embedding_1d(self.freq_dim, timestep))
+        t_mod = self.time_projection(t).unflatten(1, (6, self.dim))
+        context = self.text_embedding(context)
+        x, (f, ) = self.patchify(x)
+        freqs = torch.cat([
+            self.freqs[0][:f].view(f, -1).expand(f, -1),
+            self.freqs[1][:f].view(f, -1).expand(f, -1),
+            self.freqs[2][:f].view(f, -1).expand(f, -1),
+        ], dim=-1).reshape(f, 1, -1).to(x.device)
+
+        for block in self.blocks:
+            x = gradient_checkpoint_forward(
+                block,
+                use_gradient_checkpointing,
+                use_gradient_checkpointing_offload,
+                x, context, t_mod, freqs,
+            )
+        x = self.head(x, t)
+        x = self.unpatchify(x, (f, ))
+        return x
+
+    def unpatchify(self, x: torch.Tensor, grid_size: torch.Tensor):
+        return rearrange(
+            x, 'b f (p c) -> b c (f p)',
+            f=grid_size[0],
+            p=self.patch_size[0]
+        )

diffsynth/models/mova_audio_vae.py

@@ -0,0 +1,796 @@
+import math
+from typing import List, Union
+import numpy as np
+import torch
+from torch import nn
+from torch.nn.utils import weight_norm
+import torch.nn.functional as F
+from einops import rearrange
+
+def WNConv1d(*args, **kwargs):
+    return weight_norm(nn.Conv1d(*args, **kwargs))
+
+
+def WNConvTranspose1d(*args, **kwargs):
+    return weight_norm(nn.ConvTranspose1d(*args, **kwargs))
+
+
+# Scripting this brings model speed up 1.4x
+@torch.jit.script
+def snake(x, alpha):
+    shape = x.shape
+    x = x.reshape(shape[0], shape[1], -1)
+    x = x + (alpha + 1e-9).reciprocal() * torch.sin(alpha * x).pow(2)
+    x = x.reshape(shape)
+    return x
+
+
+class Snake1d(nn.Module):
+    def __init__(self, channels):
+        super().__init__()
+        self.alpha = nn.Parameter(torch.ones(1, channels, 1))
+
+    def forward(self, x):
+        return snake(x, self.alpha)
+
+
+class VectorQuantize(nn.Module):
+    """
+    Implementation of VQ similar to Karpathy's repo:
+    https://github.com/karpathy/deep-vector-quantization
+    Additionally uses following tricks from Improved VQGAN
+    (https://arxiv.org/pdf/2110.04627.pdf):
+        1. Factorized codes: Perform nearest neighbor lookup in low-dimensional space
+            for improved codebook usage
+        2. l2-normalized codes: Converts euclidean distance to cosine similarity which
+            improves training stability
+    """
+
+    def __init__(self, input_dim: int, codebook_size: int, codebook_dim: int):
+        super().__init__()
+        self.codebook_size = codebook_size
+        self.codebook_dim = codebook_dim
+
+        self.in_proj = WNConv1d(input_dim, codebook_dim, kernel_size=1)
+        self.out_proj = WNConv1d(codebook_dim, input_dim, kernel_size=1)
+        self.codebook = nn.Embedding(codebook_size, codebook_dim)
+
+    def forward(self, z):
+        """Quantized the input tensor using a fixed codebook and returns
+        the corresponding codebook vectors
+
+        Parameters
+        ----------
+        z : Tensor[B x D x T]
+
+        Returns
+        -------
+        Tensor[B x D x T]
+            Quantized continuous representation of input
+        Tensor[1]
+            Commitment loss to train encoder to predict vectors closer to codebook
+            entries
+        Tensor[1]
+            Codebook loss to update the codebook
+        Tensor[B x T]
+            Codebook indices (quantized discrete representation of input)
+        Tensor[B x D x T]
+            Projected latents (continuous representation of input before quantization)
+        """
+
+        # Factorized codes (ViT-VQGAN) Project input into low-dimensional space
+        z_e = self.in_proj(z)  # z_e : (B x D x T)
+        z_q, indices = self.decode_latents(z_e)
+
+        commitment_loss = F.mse_loss(z_e, z_q.detach(), reduction="none").mean([1, 2])
+        codebook_loss = F.mse_loss(z_q, z_e.detach(), reduction="none").mean([1, 2])
+
+        z_q = (
+            z_e + (z_q - z_e).detach()
+        )  # noop in forward pass, straight-through gradient estimator in backward pass
+
+        z_q = self.out_proj(z_q)
+
+        return z_q, commitment_loss, codebook_loss, indices, z_e
+
+    def embed_code(self, embed_id):
+        return F.embedding(embed_id, self.codebook.weight)
+
+    def decode_code(self, embed_id):
+        return self.embed_code(embed_id).transpose(1, 2)
+
+    def decode_latents(self, latents):
+        encodings = rearrange(latents, "b d t -> (b t) d")
+        codebook = self.codebook.weight  # codebook: (N x D)
+
+        # L2 normalize encodings and codebook (ViT-VQGAN)
+        encodings = F.normalize(encodings)
+        codebook = F.normalize(codebook)
+
+        # Compute euclidean distance with codebook
+        dist = (
+            encodings.pow(2).sum(1, keepdim=True)
+            - 2 * encodings @ codebook.t()
+            + codebook.pow(2).sum(1, keepdim=True).t()
+        )
+        indices = rearrange((-dist).max(1)[1], "(b t) -> b t", b=latents.size(0))
+        z_q = self.decode_code(indices)
+        return z_q, indices
+
+
+class ResidualVectorQuantize(nn.Module):
+    """
+    Introduced in SoundStream: An end2end neural audio codec
+    https://arxiv.org/abs/2107.03312
+    """
+
+    def __init__(
+        self,
+        input_dim: int = 512,
+        n_codebooks: int = 9,
+        codebook_size: int = 1024,
+        codebook_dim: Union[int, list] = 8,
+        quantizer_dropout: float = 0.0,
+    ):
+        super().__init__()
+        if isinstance(codebook_dim, int):
+            codebook_dim = [codebook_dim for _ in range(n_codebooks)]
+
+        self.n_codebooks = n_codebooks
+        self.codebook_dim = codebook_dim
+        self.codebook_size = codebook_size
+
+        self.quantizers = nn.ModuleList(
+            [
+                VectorQuantize(input_dim, codebook_size, codebook_dim[i])
+                for i in range(n_codebooks)
+            ]
+        )
+        self.quantizer_dropout = quantizer_dropout
+
+    def forward(self, z, n_quantizers: int = None):
+        """Quantized the input tensor using a fixed set of `n` codebooks and returns
+        the corresponding codebook vectors
+        Parameters
+        ----------
+        z : Tensor[B x D x T]
+        n_quantizers : int, optional
+            No. of quantizers to use
+            (n_quantizers < self.n_codebooks ex: for quantizer dropout)
+            Note: if `self.quantizer_dropout` is True, this argument is ignored
+                when in training mode, and a random number of quantizers is used.
+        Returns
+        -------
+        dict
+            A dictionary with the following keys:
+
+            "z" : Tensor[B x D x T]
+                Quantized continuous representation of input
+            "codes" : Tensor[B x N x T]
+                Codebook indices for each codebook
+                (quantized discrete representation of input)
+            "latents" : Tensor[B x N*D x T]
+                Projected latents (continuous representation of input before quantization)
+            "vq/commitment_loss" : Tensor[1]
+                Commitment loss to train encoder to predict vectors closer to codebook
+                entries
+            "vq/codebook_loss" : Tensor[1]
+                Codebook loss to update the codebook
+        """
+        z_q = 0
+        residual = z
+        commitment_loss = 0
+        codebook_loss = 0
+
+        codebook_indices = []
+        latents = []
+
+        if n_quantizers is None:
+            n_quantizers = self.n_codebooks
+        if self.training:
+            n_quantizers = torch.ones((z.shape[0],)) * self.n_codebooks + 1
+            dropout = torch.randint(1, self.n_codebooks + 1, (z.shape[0],))
+            n_dropout = int(z.shape[0] * self.quantizer_dropout)
+            n_quantizers[:n_dropout] = dropout[:n_dropout]
+            n_quantizers = n_quantizers.to(z.device)
+
+        for i, quantizer in enumerate(self.quantizers):
+            if self.training is False and i >= n_quantizers:
+                break
+
+            z_q_i, commitment_loss_i, codebook_loss_i, indices_i, z_e_i = quantizer(
+                residual
+            )
+
+            # Create mask to apply quantizer dropout
+            mask = (
+                torch.full((z.shape[0],), fill_value=i, device=z.device) < n_quantizers
+            )
+            z_q = z_q + z_q_i * mask[:, None, None]
+            residual = residual - z_q_i
+
+            # Sum losses
+            commitment_loss += (commitment_loss_i * mask).mean()
+            codebook_loss += (codebook_loss_i * mask).mean()
+
+            codebook_indices.append(indices_i)
+            latents.append(z_e_i)
+
+        codes = torch.stack(codebook_indices, dim=1)
+        latents = torch.cat(latents, dim=1)
+
+        return z_q, codes, latents, commitment_loss, codebook_loss
+
+    def from_codes(self, codes: torch.Tensor):
+        """Given the quantized codes, reconstruct the continuous representation
+        Parameters
+        ----------
+        codes : Tensor[B x N x T]
+            Quantized discrete representation of input
+        Returns
+        -------
+        Tensor[B x D x T]
+            Quantized continuous representation of input
+        """
+        z_q = 0.0
+        z_p = []
+        n_codebooks = codes.shape[1]
+        for i in range(n_codebooks):
+            z_p_i = self.quantizers[i].decode_code(codes[:, i, :])
+            z_p.append(z_p_i)
+
+            z_q_i = self.quantizers[i].out_proj(z_p_i)
+            z_q = z_q + z_q_i
+        return z_q, torch.cat(z_p, dim=1), codes
+
+    def from_latents(self, latents: torch.Tensor):
+        """Given the unquantized latents, reconstruct the
+        continuous representation after quantization.
+
+        Parameters
+        ----------
+        latents : Tensor[B x N x T]
+            Continuous representation of input after projection
+
+        Returns
+        -------
+        Tensor[B x D x T]
+            Quantized representation of full-projected space
+        Tensor[B x D x T]
+            Quantized representation of latent space
+        """
+        z_q = 0
+        z_p = []
+        codes = []
+        dims = np.cumsum([0] + [q.codebook_dim for q in self.quantizers])
+
+        n_codebooks = np.where(dims <= latents.shape[1])[0].max(axis=0, keepdims=True)[
+            0
+        ]
+        for i in range(n_codebooks):
+            j, k = dims[i], dims[i + 1]
+            z_p_i, codes_i = self.quantizers[i].decode_latents(latents[:, j:k, :])
+            z_p.append(z_p_i)
+            codes.append(codes_i)
+
+            z_q_i = self.quantizers[i].out_proj(z_p_i)
+            z_q = z_q + z_q_i
+
+        return z_q, torch.cat(z_p, dim=1), torch.stack(codes, dim=1)
+
+
+class AbstractDistribution:
+    def sample(self):
+        raise NotImplementedError()
+
+    def mode(self):
+        raise NotImplementedError()
+
+
+class DiracDistribution(AbstractDistribution):
+    def __init__(self, value):
+        self.value = value
+
+    def sample(self):
+        return self.value
+
+    def mode(self):
+        return self.value
+
+
+class DiagonalGaussianDistribution(object):
+    def __init__(self, parameters, deterministic=False):
+        self.parameters = parameters
+        self.mean, self.logvar = torch.chunk(parameters, 2, dim=1)
+        self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
+        self.deterministic = deterministic
+        self.std = torch.exp(0.5 * self.logvar)
+        self.var = torch.exp(self.logvar)
+        if self.deterministic:
+            self.var = self.std = torch.zeros_like(self.mean).to(device=self.parameters.device)
+
+    def sample(self):
+        x = self.mean + self.std * torch.randn(self.mean.shape).to(device=self.parameters.device)
+        return x
+
+    def kl(self, other=None):
+        if self.deterministic:
+            return torch.Tensor([0.0])
+        else:
+            if other is None:
+                return 0.5 * torch.mean(
+                    torch.pow(self.mean, 2) + self.var - 1.0 - self.logvar,
+                    dim=[1, 2],
+                )
+            else:
+                return 0.5 * torch.mean(
+                    torch.pow(self.mean - other.mean, 2) / other.var
+                    + self.var / other.var
+                    - 1.0
+                    - self.logvar
+                    + other.logvar,
+                    dim=[1, 2],
+                )
+
+    def nll(self, sample, dims=[1, 2]):
+        if self.deterministic:
+            return torch.Tensor([0.0])
+        logtwopi = np.log(2.0 * np.pi)
+        return 0.5 * torch.sum(
+            logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var,
+            dim=dims,
+        )
+
+    def mode(self):
+        return self.mean
+
+
+def normal_kl(mean1, logvar1, mean2, logvar2):
+    """
+    source: https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/losses.py#L12
+    Compute the KL divergence between two gaussians.
+    Shapes are automatically broadcasted, so batches can be compared to
+    scalars, among other use cases.
+    """
+    tensor = None
+    for obj in (mean1, logvar1, mean2, logvar2):
+        if isinstance(obj, torch.Tensor):
+            tensor = obj
+            break
+    assert tensor is not None, "at least one argument must be a Tensor"
+
+    # Force variances to be Tensors. Broadcasting helps convert scalars to
+    # Tensors, but it does not work for torch.exp().
+    logvar1, logvar2 = [x if isinstance(x, torch.Tensor) else torch.tensor(x).to(tensor) for x in (logvar1, logvar2)]
+
+    return 0.5 * (
+        -1.0 + logvar2 - logvar1 + torch.exp(logvar1 - logvar2) + ((mean1 - mean2) ** 2) * torch.exp(-logvar2)
+    )
+
+
+def init_weights(m):
+    if isinstance(m, nn.Conv1d):
+        nn.init.trunc_normal_(m.weight, std=0.02)
+        nn.init.constant_(m.bias, 0)
+
+
+class ResidualUnit(nn.Module):
+    def __init__(self, dim: int = 16, dilation: int = 1):
+        super().__init__()
+        pad = ((7 - 1) * dilation) // 2
+        self.block = nn.Sequential(
+            Snake1d(dim),
+            WNConv1d(dim, dim, kernel_size=7, dilation=dilation, padding=pad),
+            Snake1d(dim),
+            WNConv1d(dim, dim, kernel_size=1),
+        )
+
+    def forward(self, x):
+        y = self.block(x)
+        pad = (x.shape[-1] - y.shape[-1]) // 2
+        if pad > 0:
+            x = x[..., pad:-pad]
+        return x + y
+
+
+class EncoderBlock(nn.Module):
+    def __init__(self, dim: int = 16, stride: int = 1):
+        super().__init__()
+        self.block = nn.Sequential(
+            ResidualUnit(dim // 2, dilation=1),
+            ResidualUnit(dim // 2, dilation=3),
+            ResidualUnit(dim // 2, dilation=9),
+            Snake1d(dim // 2),
+            WNConv1d(
+                dim // 2,
+                dim,
+                kernel_size=2 * stride,
+                stride=stride,
+                padding=math.ceil(stride / 2),
+            ),
+        )
+
+    def forward(self, x):
+        return self.block(x)
+
+
+class Encoder(nn.Module):
+    def __init__(
+        self,
+        d_model: int = 64,
+        strides: list = [2, 4, 8, 8],
+        d_latent: int = 64,
+    ):
+        super().__init__()
+        # Create first convolution
+        self.block = [WNConv1d(1, d_model, kernel_size=7, padding=3)]
+
+        # Create EncoderBlocks that double channels as they downsample by `stride`
+        for stride in strides:
+            d_model *= 2
+            self.block += [EncoderBlock(d_model, stride=stride)]
+
+        # Create last convolution
+        self.block += [
+            Snake1d(d_model),
+            WNConv1d(d_model, d_latent, kernel_size=3, padding=1),
+        ]
+
+        # Wrap black into nn.Sequential
+        self.block = nn.Sequential(*self.block)
+        self.enc_dim = d_model
+
+    def forward(self, x):
+        return self.block(x)
+
+
+class DecoderBlock(nn.Module):
+    def __init__(self, input_dim: int = 16, output_dim: int = 8, stride: int = 1):
+        super().__init__()
+        self.block = nn.Sequential(
+            Snake1d(input_dim),
+            WNConvTranspose1d(
+                input_dim,
+                output_dim,
+                kernel_size=2 * stride,
+                stride=stride,
+                padding=math.ceil(stride / 2),
+                output_padding=stride % 2,
+            ),
+            ResidualUnit(output_dim, dilation=1),
+            ResidualUnit(output_dim, dilation=3),
+            ResidualUnit(output_dim, dilation=9),
+        )
+
+    def forward(self, x):
+        return self.block(x)
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self,
+        input_channel,
+        channels,
+        rates,
+        d_out: int = 1,
+    ):
+        super().__init__()
+
+        # Add first conv layer
+        layers = [WNConv1d(input_channel, channels, kernel_size=7, padding=3)]
+
+        # Add upsampling + MRF blocks
+        for i, stride in enumerate(rates):
+            input_dim = channels // 2**i
+            output_dim = channels // 2 ** (i + 1)
+            layers += [DecoderBlock(input_dim, output_dim, stride)]
+
+        # Add final conv layer
+        layers += [
+            Snake1d(output_dim),
+            WNConv1d(output_dim, d_out, kernel_size=7, padding=3),
+            nn.Tanh(),
+        ]
+
+        self.model = nn.Sequential(*layers)
+
+    def forward(self, x):
+        return self.model(x)
+
+
+class DacVAE(nn.Module):
+
+    def __init__(
+        self,
+        encoder_dim: int = 128,
+        encoder_rates: List[int] = [2, 3, 4, 5, 8],
+        latent_dim: int = 128,
+        decoder_dim: int = 2048,
+        decoder_rates: List[int] = [8, 5, 4, 3, 2],
+        n_codebooks: int = 9,
+        codebook_size: int = 1024,
+        codebook_dim: Union[int, list] = 8,
+        quantizer_dropout: bool = False,
+        sample_rate: int = 48000,
+        continuous: bool = True,
+        use_weight_norm: bool = False,
+    ):
+        super().__init__()
+
+        self.encoder_dim = encoder_dim
+        self.encoder_rates = encoder_rates
+        self.decoder_dim = decoder_dim
+        self.decoder_rates = decoder_rates
+        self.sample_rate = sample_rate
+        self.continuous = continuous
+        self.use_weight_norm = use_weight_norm
+
+        if latent_dim is None:
+            latent_dim = encoder_dim * (2 ** len(encoder_rates))
+
+        self.latent_dim = latent_dim
+
+        self.hop_length = np.prod(encoder_rates)
+        self.encoder = Encoder(encoder_dim, encoder_rates, latent_dim)
+
+        if not continuous:
+            self.n_codebooks = n_codebooks
+            self.codebook_size = codebook_size
+            self.codebook_dim = codebook_dim
+            self.quantizer = ResidualVectorQuantize(
+                input_dim=latent_dim,
+                n_codebooks=n_codebooks,
+                codebook_size=codebook_size,
+                codebook_dim=codebook_dim,
+                quantizer_dropout=quantizer_dropout,
+            )
+        else:
+            self.quant_conv = torch.nn.Conv1d(latent_dim, 2 * latent_dim, 1)
+            self.post_quant_conv = torch.nn.Conv1d(latent_dim, latent_dim, 1)
+
+        self.decoder = Decoder(
+            latent_dim,
+            decoder_dim,
+            decoder_rates,
+        )
+        self.sample_rate = sample_rate
+        self.apply(init_weights)
+
+        self.delay = self.get_delay()
+
+        if not self.use_weight_norm:
+            self.remove_weight_norm()
+
+    def get_delay(self):
+        # Any number works here, delay is invariant to input length
+        l_out = self.get_output_length(0)
+        L = l_out
+
+        layers = []
+        for layer in self.modules():
+            if isinstance(layer, (nn.Conv1d, nn.ConvTranspose1d)):
+                layers.append(layer)
+
+        for layer in reversed(layers):
+            d = layer.dilation[0]
+            k = layer.kernel_size[0]
+            s = layer.stride[0]
+
+            if isinstance(layer, nn.ConvTranspose1d):
+                L = ((L - d * (k - 1) - 1) / s) + 1
+            elif isinstance(layer, nn.Conv1d):
+                L = (L - 1) * s + d * (k - 1) + 1
+
+            L = math.ceil(L)
+
+        l_in = L
+
+        return (l_in - l_out) // 2
+
+    def get_output_length(self, input_length):
+        L = input_length
+        # Calculate output length
+        for layer in self.modules():
+            if isinstance(layer, (nn.Conv1d, nn.ConvTranspose1d)):
+                d = layer.dilation[0]
+                k = layer.kernel_size[0]
+                s = layer.stride[0]
+
+                if isinstance(layer, nn.Conv1d):
+                    L = ((L - d * (k - 1) - 1) / s) + 1
+                elif isinstance(layer, nn.ConvTranspose1d):
+                    L = (L - 1) * s + d * (k - 1) + 1
+
+                L = math.floor(L)
+        return L
+
+    @property
+    def dtype(self):
+        """Get the dtype of the model parameters."""
+        # Return the dtype of the first parameter found
+        for param in self.parameters():
+            return param.dtype
+        return torch.float32  # fallback
+
+    @property
+    def device(self):
+        """Get the device of the model parameters."""
+        # Return the device of the first parameter found
+        for param in self.parameters():
+            return param.device
+        return torch.device('cpu')  # fallback
+
+    def preprocess(self, audio_data, sample_rate):
+        if sample_rate is None:
+            sample_rate = self.sample_rate
+        assert sample_rate == self.sample_rate
+
+        length = audio_data.shape[-1]
+        right_pad = math.ceil(length / self.hop_length) * self.hop_length - length
+        audio_data = nn.functional.pad(audio_data, (0, right_pad))
+
+        return audio_data
+
+    def encode(
+        self,
+        audio_data: torch.Tensor,
+        n_quantizers: int = None,
+    ):
+        """Encode given audio data and return quantized latent codes
+
+        Parameters
+        ----------
+        audio_data : Tensor[B x 1 x T]
+            Audio data to encode
+        n_quantizers : int, optional
+            Number of quantizers to use, by default None
+            If None, all quantizers are used.
+
+        Returns
+        -------
+        dict
+            A dictionary with the following keys:
+            "z" : Tensor[B x D x T]
+                Quantized continuous representation of input
+            "codes" : Tensor[B x N x T]
+                Codebook indices for each codebook
+                (quantized discrete representation of input)
+            "latents" : Tensor[B x N*D x T]
+                Projected latents (continuous representation of input before quantization)
+            "vq/commitment_loss" : Tensor[1]
+                Commitment loss to train encoder to predict vectors closer to codebook
+                entries
+            "vq/codebook_loss" : Tensor[1]
+                Codebook loss to update the codebook
+            "length" : int
+                Number of samples in input audio
+        """
+        z = self.encoder(audio_data)  # [B x D x T]
+        if not self.continuous:
+            z, codes, latents, commitment_loss, codebook_loss = self.quantizer(z, n_quantizers)
+        else:
+            z = self.quant_conv(z)  # [B x 2D x T]
+            z = DiagonalGaussianDistribution(z)
+            codes, latents, commitment_loss, codebook_loss = None, None, 0, 0
+
+        return z, codes, latents, commitment_loss, codebook_loss
+
+    def decode(self, z: torch.Tensor):
+        """Decode given latent codes and return audio data
+
+        Parameters
+        ----------
+        z : Tensor[B x D x T]
+            Quantized continuous representation of input
+        length : int, optional
+            Number of samples in output audio, by default None
+
+        Returns
+        -------
+        dict
+            A dictionary with the following keys:
+            "audio" : Tensor[B x 1 x length]
+                Decoded audio data.
+        """
+        if not self.continuous:
+            audio = self.decoder(z)
+        else:
+            z = self.post_quant_conv(z)
+            audio = self.decoder(z)
+
+        return audio
+
+    def forward(
+        self,
+        audio_data: torch.Tensor,
+        sample_rate: int = None,
+        n_quantizers: int = None,
+    ):
+        """Model forward pass
+
+        Parameters
+        ----------
+        audio_data : Tensor[B x 1 x T]
+            Audio data to encode
+        sample_rate : int, optional
+            Sample rate of audio data in Hz, by default None
+            If None, defaults to `self.sample_rate`
+        n_quantizers : int, optional
+            Number of quantizers to use, by default None.
+            If None, all quantizers are used.
+
+        Returns
+        -------
+        dict
+            A dictionary with the following keys:
+            "z" : Tensor[B x D x T]
+                Quantized continuous representation of input
+            "codes" : Tensor[B x N x T]
+                Codebook indices for each codebook
+                (quantized discrete representation of input)
+            "latents" : Tensor[B x N*D x T]
+                Projected latents (continuous representation of input before quantization)
+            "vq/commitment_loss" : Tensor[1]
+                Commitment loss to train encoder to predict vectors closer to codebook
+                entries
+            "vq/codebook_loss" : Tensor[1]
+                Codebook loss to update the codebook
+            "length" : int
+                Number of samples in input audio
+            "audio" : Tensor[B x 1 x length]
+                Decoded audio data.
+        """
+        length = audio_data.shape[-1]
+        audio_data = self.preprocess(audio_data, sample_rate)
+        if not self.continuous:
+            z, codes, latents, commitment_loss, codebook_loss = self.encode(audio_data, n_quantizers)
+
+            x = self.decode(z)
+            return {
+                "audio": x[..., :length],
+                "z": z,
+                "codes": codes,
+                "latents": latents,
+                "vq/commitment_loss": commitment_loss,
+                "vq/codebook_loss": codebook_loss,
+            }
+        else:
+            posterior, _, _, _, _ = self.encode(audio_data, n_quantizers)
+            z = posterior.sample()
+            x = self.decode(z)
+
+            kl_loss = posterior.kl()
+            kl_loss = kl_loss.mean()
+
+            return {
+                "audio": x[..., :length],
+                "z": z,
+                "kl_loss": kl_loss,
+            }
+
+    def remove_weight_norm(self):
+        """
+        Remove weight_norm from all modules in the model.
+        This fuses the weight_g and weight_v parameters into a single weight parameter.
+        Should be called before inference for better performance.
+        Returns:
+            self: The model with weight_norm removed
+        """
+        from torch.nn.utils import remove_weight_norm
+        num_removed = 0
+        for name, module in list(self.named_modules()):
+            if hasattr(module, "_forward_pre_hooks"):
+                for hook_id, hook in list(module._forward_pre_hooks.items()):
+                    if "WeightNorm" in str(type(hook)):
+                        try:
+                            remove_weight_norm(module)
+                            num_removed += 1
+                            # print(f"Removed weight_norm from: {name}")
+                        except ValueError as e:
+                            print(f"Failed to remove weight_norm from {name}: {e}")
+        if num_removed > 0:
+            # print(f"Successfully removed weight_norm from {num_removed} modules")
+            self.use_weight_norm = False
+        else:
+            print("No weight_norm found in the model")
+        return self

diffsynth/models/mova_dual_tower_bridge.py

@@ -0,0 +1,595 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from typing import Dict, List, Tuple, Optional
+from einops import rearrange
+from .wan_video_dit import AttentionModule, RMSNorm
+from ..core import gradient_checkpoint_forward
+
+class RotaryEmbedding(nn.Module):
+    inv_freq: torch.Tensor  # fix linting for `register_buffer`
+
+    def __init__(self, base: float, dim: int, device=None):
+        super().__init__()
+        self.base = base
+        self.dim = dim
+        self.attention_scaling = 1.0
+
+        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.int64).to(device=device, dtype=torch.float) / dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+    @torch.no_grad()
+    def forward(self, x, position_ids):
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
+        position_ids_expanded = position_ids[:, None, :].float()
+
+        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos() * self.attention_scaling
+            sin = emb.sin() * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+@torch.compile(fullgraph=True)
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+class PerFrameAttentionPooling(nn.Module):
+    """
+    Per-frame multi-head attention pooling.
+
+    Given a flattened token sequence [B, L, D] and grid size (T, H, W), perform a
+    single-query attention pooling over the H*W tokens for each time frame, producing
+    [B, T, D].
+
+    Inspired by SigLIP's Multihead Attention Pooling head (without MLP/residual stack).
+    """
+
+    def __init__(self, dim: int, num_heads: int, eps: float = 1e-6):
+        super().__init__()
+        assert dim % num_heads == 0, "dim must be divisible by num_heads"
+        self.dim = dim
+        self.num_heads = num_heads
+
+        self.probe = nn.Parameter(torch.randn(1, 1, dim))
+        nn.init.normal_(self.probe, std=0.02)
+
+        self.attention = nn.MultiheadAttention(embed_dim=dim, num_heads=num_heads, batch_first=True)
+        self.layernorm = nn.LayerNorm(dim, eps=eps)
+
+    def forward(self, x: torch.Tensor, grid_size: Tuple[int, int, int]) -> torch.Tensor:
+        """
+        Args:
+            x: [B, L, D], where L = T*H*W
+            grid_size: (T, H, W)
+        Returns:
+            pooled: [B, T, D]
+        """
+        B, L, D = x.shape
+        T, H, W = grid_size
+        assert D == self.dim, f"Channel dimension mismatch: D={D} vs dim={self.dim}"
+        assert L == T * H * W, f"Flattened length mismatch: L={L} vs T*H*W={T*H*W}"
+
+        S = H * W
+        # Re-arrange tokens grouped by frame.
+        x_bt_s_d = x.view(B, T, S, D).contiguous().view(B * T, S, D)  # [B*T, S, D]
+
+        # A learnable probe as the query (one query per frame).
+        probe = self.probe.expand(B * T, -1, -1)  # [B*T, 1, D]
+
+        # Attention pooling: query=probe, key/value=H*W tokens within the frame.
+        pooled_bt_1_d = self.attention(probe, x_bt_s_d, x_bt_s_d, need_weights=False)[0]  # [B*T, 1, D]
+        pooled_bt_d = pooled_bt_1_d.squeeze(1)  # [B*T, D]
+
+        # Restore to [B, T, D].
+        pooled = pooled_bt_d.view(B, T, D)
+        pooled = self.layernorm(pooled)
+        return pooled
+
+
+class CrossModalInteractionController:
+    """
+    Strategy class that controls interactions between two towers.
+    Manages the interaction mapping between visual DiT (e.g. 30 layers) and audio DiT (e.g. 30 layers).
+    """
+
+    def __init__(self, visual_layers: int = 30, audio_layers: int = 30):
+        self.visual_layers = visual_layers
+        self.audio_layers = audio_layers
+        self.min_layers = min(visual_layers, audio_layers)
+
+    def get_interaction_layers(self, strategy: str = "shallow_focus") -> Dict[str, List[Tuple[int, int]]]:
+        """
+        Get interaction layer mappings.
+
+        Args:
+            strategy: interaction strategy
+                - "shallow_focus": emphasize shallow layers to avoid deep-layer asymmetry
+                - "distributed": distributed interactions across the network
+                - "progressive": dense shallow interactions, sparse deeper interactions
+                - "custom": custom interaction layers
+
+        Returns:
+            A dict containing mappings for 'v2a' (visual -> audio) and 'a2v' (audio -> visual).
+        """
+
+        if strategy == "shallow_focus":
+            # Emphasize the first ~1/3 layers to avoid deep-layer asymmetry.
+            num_interact = min(10, self.min_layers // 3)
+            interact_layers = list(range(0, num_interact))
+
+        elif strategy == "distributed":
+            # Distribute interactions across the network (every few layers).
+            step = 3
+            interact_layers = list(range(0, self.min_layers, step))
+
+        elif strategy == "progressive":
+            # Progressive: dense shallow interactions, sparse deeper interactions.
+            shallow = list(range(0, min(8, self.min_layers)))  # Dense for the first 8 layers.
+            if self.min_layers > 8:
+                deep = list(range(8, self.min_layers, 3))  # Every 3 layers afterwards.
+                interact_layers = shallow + deep
+            else:
+                interact_layers = shallow
+
+        elif strategy == "custom":
+            # Custom strategy: adjust as needed.
+            interact_layers = [0, 2, 4, 6, 8, 12, 16, 20]  # Explicit layer indices.
+            interact_layers = [i for i in interact_layers if i < self.min_layers]
+
+        elif strategy == "full":
+            interact_layers = list(range(0, self.min_layers))
+
+        else:
+            raise ValueError(f"Unknown interaction strategy: {strategy}")
+
+        # Build bidirectional mapping.
+        mapping = {
+            'v2a': [(i, i) for i in interact_layers],  # visual layer i -> audio layer i
+            'a2v': [(i, i) for i in interact_layers]   # audio layer i -> visual layer i
+        }
+
+        return mapping
+
+    def should_interact(self, layer_idx: int, direction: str, interaction_mapping: Dict) -> bool:
+        """
+        Check whether a given layer should interact.
+
+        Args:
+            layer_idx: current layer index
+            direction: interaction direction ('v2a' or 'a2v')
+            interaction_mapping: interaction mapping table
+
+        Returns:
+            bool: whether to interact
+        """
+        if direction not in interaction_mapping:
+            return False
+
+        return any(src == layer_idx for src, _ in interaction_mapping[direction])
+
+
+class ConditionalCrossAttention(nn.Module):
+    def __init__(self, dim: int, kv_dim: int, num_heads: int, eps: float = 1e-6):
+        super().__init__()
+        self.q_dim = dim
+        self.kv_dim = kv_dim
+        self.num_heads = num_heads
+        self.head_dim = self.q_dim // num_heads
+
+        self.q = nn.Linear(dim, dim)
+        self.k = nn.Linear(kv_dim, dim)
+        self.v = nn.Linear(kv_dim, dim)
+        self.o = nn.Linear(dim, dim)
+        self.norm_q = RMSNorm(dim, eps=eps)
+        self.norm_k = RMSNorm(dim, eps=eps)
+
+        self.attn = AttentionModule(self.num_heads)
+
+    def forward(self, x: torch.Tensor, y: torch.Tensor, x_freqs: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, y_freqs: Optional[Tuple[torch.Tensor, torch.Tensor]] = None):
+        ctx = y
+        q = self.norm_q(self.q(x))
+        k = self.norm_k(self.k(ctx))
+        v = self.v(ctx)
+        if x_freqs is not None:
+            x_cos, x_sin = x_freqs
+            B, L, _ = q.shape
+            q_view = rearrange(q, 'b l (h d) -> b l h d', d=self.head_dim)
+            x_cos = x_cos.to(q_view.dtype).to(q_view.device)
+            x_sin = x_sin.to(q_view.dtype).to(q_view.device)
+            # Expect x_cos/x_sin shape: [B or 1, L, head_dim]
+            q_view, _ = apply_rotary_pos_emb(q_view, q_view, x_cos, x_sin, unsqueeze_dim=2)
+            q = rearrange(q_view, 'b l h d -> b l (h d)')
+        if y_freqs is not None:
+            y_cos, y_sin = y_freqs
+            Bc, Lc, _ = k.shape
+            k_view = rearrange(k, 'b l (h d) -> b l h d', d=self.head_dim)
+            y_cos = y_cos.to(k_view.dtype).to(k_view.device)
+            y_sin = y_sin.to(k_view.dtype).to(k_view.device)
+            # Expect y_cos/y_sin shape: [B or 1, L, head_dim]
+            _, k_view = apply_rotary_pos_emb(k_view, k_view, y_cos, y_sin, unsqueeze_dim=2)
+            k = rearrange(k_view, 'b l h d -> b l (h d)')
+        x = self.attn(q, k, v)
+        return self.o(x)
+
+
+# from diffusers.models.attention import AdaLayerNorm
+class AdaLayerNorm(nn.Module):
+    r"""
+    Norm layer modified to incorporate timestep embeddings.
+
+    Parameters:
+        embedding_dim (`int`): The size of each embedding vector.
+        num_embeddings (`int`, *optional*): The size of the embeddings dictionary.
+        output_dim (`int`, *optional*):
+        norm_elementwise_affine (`bool`, defaults to `False):
+        norm_eps (`bool`, defaults to `False`):
+        chunk_dim (`int`, defaults to `0`):
+    """
+
+    def __init__(
+        self,
+        embedding_dim: int,
+        num_embeddings: Optional[int] = None,
+        output_dim: Optional[int] = None,
+        norm_elementwise_affine: bool = False,
+        norm_eps: float = 1e-5,
+        chunk_dim: int = 0,
+    ):
+        super().__init__()
+
+        self.chunk_dim = chunk_dim
+        output_dim = output_dim or embedding_dim * 2
+
+        if num_embeddings is not None:
+            self.emb = nn.Embedding(num_embeddings, embedding_dim)
+        else:
+            self.emb = None
+
+        self.silu = nn.SiLU()
+        self.linear = nn.Linear(embedding_dim, output_dim)
+        self.norm = nn.LayerNorm(output_dim // 2, norm_eps, norm_elementwise_affine)
+
+    def forward(
+        self, x: torch.Tensor, timestep: Optional[torch.Tensor] = None, temb: Optional[torch.Tensor] = None
+    ) -> torch.Tensor:
+        if self.emb is not None:
+            temb = self.emb(timestep)
+
+        temb = self.linear(self.silu(temb))
+
+        if self.chunk_dim == 2:
+            scale, shift = temb.chunk(2, dim=2)
+            # print(f"{x.shape = }, {scale.shape = }, {shift.shape = }")
+        elif self.chunk_dim == 1:
+            # This is a bit weird why we have the order of "shift, scale" here and "scale, shift" in the
+            # other if-branch. This branch is specific to CogVideoX and OmniGen for now.
+            shift, scale = temb.chunk(2, dim=1)
+            shift = shift[:, None, :]
+            scale = scale[:, None, :]
+        else:
+            scale, shift = temb.chunk(2, dim=0)
+
+        x = self.norm(x) * (1 + scale) + shift
+        return x
+
+
+class ConditionalCrossAttentionBlock(nn.Module):
+    """
+    A thin wrapper around ConditionalCrossAttention.
+    Applies LayerNorm to the conditioning input `y` before cross-attention.
+    """
+    def __init__(self, dim: int, kv_dim: int, num_heads: int, eps: float = 1e-6, pooled_adaln: bool = False):
+        super().__init__()
+        self.y_norm = nn.LayerNorm(kv_dim, eps=eps)
+        self.inner = ConditionalCrossAttention(dim=dim, kv_dim=kv_dim, num_heads=num_heads, eps=eps)
+        self.pooled_adaln = pooled_adaln
+        if pooled_adaln:
+            self.per_frame_pooling = PerFrameAttentionPooling(kv_dim, num_heads=num_heads, eps=eps)
+            self.adaln = AdaLayerNorm(kv_dim, output_dim=dim*2, chunk_dim=2)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        y: torch.Tensor,
+        x_freqs: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        y_freqs: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        video_grid_size: Optional[Tuple[int, int, int]] = None,
+    ) -> torch.Tensor:
+        if self.pooled_adaln:
+            assert video_grid_size is not None, "video_grid_size must not be None"
+            pooled_y = self.per_frame_pooling(y, video_grid_size)
+            # Interpolate pooled_y along its temporal dimension to match x's sequence length.
+            if pooled_y.shape[1] != x.shape[1]:
+                pooled_y = F.interpolate(
+                    pooled_y.permute(0, 2, 1),  # [B, C, T]
+                    size=x.shape[1],
+                    mode='linear',
+                    align_corners=False,
+                ).permute(0, 2, 1)  # [B, T, C]
+            x = self.adaln(x, temb=pooled_y)
+        y = self.y_norm(y)
+        return self.inner(x=x, y=y, x_freqs=x_freqs, y_freqs=y_freqs)
+
+
+class DualTowerConditionalBridge(nn.Module):
+    """
+    Dual-tower conditional bridge.
+    """
+    def __init__(self,
+                 visual_layers: int = 40,
+                 audio_layers: int = 30,
+                 visual_hidden_dim: int = 5120,    # visual DiT hidden state dimension
+                 audio_hidden_dim: int = 1536,     # audio DiT hidden state dimension
+                 audio_fps: float = 50.0,
+                 head_dim: int = 128,              # attention head dimension
+                 interaction_strategy: str = "full",
+                 apply_cross_rope: bool = True,   # whether to apply RoPE in cross-attention
+                 apply_first_frame_bias_in_rope: bool = False,  # whether to account for 1/video_fps bias for the first frame in RoPE alignment
+                 trainable_condition_scale: bool = False,
+                 pooled_adaln: bool = False,
+                 ):
+        super().__init__()
+
+        self.visual_hidden_dim = visual_hidden_dim
+        self.audio_hidden_dim = audio_hidden_dim
+        self.audio_fps = audio_fps
+        self.head_dim = head_dim
+        self.apply_cross_rope = apply_cross_rope
+        self.apply_first_frame_bias_in_rope = apply_first_frame_bias_in_rope
+        self.trainable_condition_scale = trainable_condition_scale
+        self.pooled_adaln = pooled_adaln
+        if self.trainable_condition_scale:
+            self.condition_scale = nn.Parameter(torch.tensor([1.0], dtype=torch.float32))
+        else:
+            self.condition_scale = 1.0
+
+        self.controller = CrossModalInteractionController(visual_layers, audio_layers)
+        self.interaction_mapping = self.controller.get_interaction_layers(interaction_strategy)
+
+        # Conditional cross-attention modules operating at the DiT hidden-state level.
+        self.audio_to_video_conditioners = nn.ModuleDict()  # audio hidden states -> visual DiT conditioning
+        self.video_to_audio_conditioners = nn.ModuleDict()  # visual hidden states -> audio DiT conditioning
+
+        # Build conditioners for layers that should interact.
+        # audio hidden states condition the visual DiT
+        self.rotary = RotaryEmbedding(base=10000.0, dim=head_dim)
+        for v_layer, _ in self.interaction_mapping['a2v']:
+            self.audio_to_video_conditioners[str(v_layer)] = ConditionalCrossAttentionBlock(
+                dim=visual_hidden_dim,     # 3072 (visual DiT hidden states)
+                kv_dim=audio_hidden_dim,    # 1536 (audio DiT hidden states)
+                num_heads=visual_hidden_dim // head_dim, # derive number of heads from hidden dim
+                pooled_adaln=False # a2v typically does not need pooled AdaLN
+            )
+
+        # visual hidden states condition the audio DiT
+        for a_layer, _ in self.interaction_mapping['v2a']:
+            self.video_to_audio_conditioners[str(a_layer)] = ConditionalCrossAttentionBlock(
+                dim=audio_hidden_dim,      # 1536 (audio DiT hidden states)
+                kv_dim=visual_hidden_dim,   # 3072 (visual DiT hidden states)
+                num_heads=audio_hidden_dim // head_dim, # safe head count derivation
+                pooled_adaln=self.pooled_adaln
+            )
+
+    @torch.no_grad()
+    def build_aligned_freqs(self,
+                            video_fps: float,
+                            grid_size: Tuple[int, int, int],
+                            audio_steps: int,
+                            device: Optional[torch.device] = None,
+                            dtype: Optional[torch.dtype] = None) -> Tuple[Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor, torch.Tensor]]:
+        """
+        Build aligned RoPE (cos, sin) pairs based on video fps, video grid size (f_v, h, w),
+        and audio sequence length `audio_steps` (with fixed audio fps = 44100/2048).
+
+        Returns:
+            visual_freqs: (cos_v, sin_v), shape [1, f_v*h*w, head_dim]
+            audio_freqs:  (cos_a, sin_a), shape [1, audio_steps, head_dim]
+        """
+        f_v, h, w = grid_size
+        L_v = f_v * h * w
+        L_a = int(audio_steps)
+
+        device = device or next(self.parameters()).device
+        dtype = dtype or torch.float32
+
+        # Audio positions: 0,1,2,...,L_a-1 (audio as reference).
+        audio_pos = torch.arange(L_a, device=device, dtype=torch.float32).unsqueeze(0)
+
+        # Video positions: align video frames to audio-step units.
+        # FIXME(dhyu): hard-coded VAE temporal stride = 4
+        if self.apply_first_frame_bias_in_rope:
+            # Account for the "first frame lasts 1/video_fps" bias.
+            video_effective_fps = float(video_fps) / 4.0
+            if f_v > 0:
+                t_starts = torch.zeros((f_v,), device=device, dtype=torch.float32)
+                if f_v > 1:
+                    t_starts[1:] = (1.0 / float(video_fps)) + torch.arange(f_v - 1, device=device, dtype=torch.float32) * (1.0 / video_effective_fps)
+            else:
+                t_starts = torch.zeros((0,), device=device, dtype=torch.float32)
+            # Convert to audio-step units.
+            video_pos_per_frame = t_starts * float(self.audio_fps)
+        else:
+            # No first-frame bias: uniform alignment.
+            scale = float(self.audio_fps) / float(video_fps / 4.0)
+            video_pos_per_frame = torch.arange(f_v, device=device, dtype=torch.float32) * scale
+        # Flatten to f*h*w; tokens within the same frame share the same time position.
+        video_pos = video_pos_per_frame.repeat_interleave(h * w).unsqueeze(0)
+
+        # print(f"video fps: {video_fps}, audio fps: {self.audio_fps}, scale: {scale}")
+        # print(f"video pos: {video_pos.shape}, audio pos: {audio_pos.shape}")
+
+        # Build dummy x to produce cos/sin, dim=head_dim.
+        dummy_v = torch.zeros((1, L_v, self.head_dim), device=device, dtype=dtype)
+        dummy_a = torch.zeros((1, L_a, self.head_dim), device=device, dtype=dtype)
+
+        cos_v, sin_v = self.rotary(dummy_v, position_ids=video_pos)
+        cos_a, sin_a = self.rotary(dummy_a, position_ids=audio_pos)
+
+        return (cos_v, sin_v), (cos_a, sin_a)
+
+    def should_interact(self, layer_idx: int, direction: str) -> bool:
+        return self.controller.should_interact(layer_idx, direction, self.interaction_mapping)
+
+    def apply_conditional_control(
+        self,
+        layer_idx: int,
+        direction: str,
+        primary_hidden_states: torch.Tensor,
+        condition_hidden_states: torch.Tensor,
+        x_freqs: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        y_freqs: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        condition_scale: Optional[float] = None,
+        video_grid_size: Optional[Tuple[int, int, int]] = None,
+        use_gradient_checkpointing: Optional[bool] = False,
+        use_gradient_checkpointing_offload: Optional[bool] = False,
+    ) -> torch.Tensor:
+        """
+        Apply conditional control (at the DiT hidden-state level).
+
+        Args:
+            layer_idx: current layer index
+            direction: conditioning direction
+                - 'a2v': audio hidden states -> visual DiT
+                - 'v2a': visual hidden states -> audio DiT
+            primary_hidden_states: primary DiT hidden states [B, L, hidden_dim]
+            condition_hidden_states: condition DiT hidden states [B, L, hidden_dim]
+            condition_scale: conditioning strength (similar to CFG scale)
+
+        Returns:
+            Conditioned primary DiT hidden states [B, L, hidden_dim]
+        """
+
+        if not self.controller.should_interact(layer_idx, direction, self.interaction_mapping):
+            return primary_hidden_states
+
+        if direction == 'a2v':
+            # audio hidden states condition the visual DiT
+            conditioner = self.audio_to_video_conditioners[str(layer_idx)]
+
+        elif direction == 'v2a':
+            # visual hidden states condition the audio DiT
+            conditioner = self.video_to_audio_conditioners[str(layer_idx)]
+        else:
+            raise ValueError(f"Invalid direction: {direction}")
+
+        conditioned_features = gradient_checkpoint_forward(
+            conditioner,
+            use_gradient_checkpointing,
+            use_gradient_checkpointing_offload,
+            x=primary_hidden_states,
+            y=condition_hidden_states,
+            x_freqs=x_freqs,
+            y_freqs=y_freqs,
+            video_grid_size=video_grid_size,
+        )
+
+        if self.trainable_condition_scale and condition_scale is not None:
+            print(
+                "[WARN] This model has a trainable condition_scale, but an external "
+                f"condition_scale={condition_scale} was provided. The trainable condition_scale "
+                "will be ignored in favor of the external value."
+            )
+
+        scale = condition_scale if condition_scale is not None else self.condition_scale
+
+        primary_hidden_states = primary_hidden_states + conditioned_features * scale
+
+        return primary_hidden_states
+
+    def forward(
+        self,
+        layer_idx: int,
+        visual_hidden_states: torch.Tensor,
+        audio_hidden_states: torch.Tensor,
+        *,
+        x_freqs: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        y_freqs: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        a2v_condition_scale: Optional[float] = None,
+        v2a_condition_scale: Optional[float] = None,
+        condition_scale: Optional[float] = None,
+        video_grid_size: Optional[Tuple[int, int, int]] = None,
+        use_gradient_checkpointing: Optional[bool] = False,
+        use_gradient_checkpointing_offload: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Apply bidirectional conditional control to both visual/audio towers.
+
+        Args:
+            layer_idx: current layer index
+            visual_hidden_states: visual DiT hidden states
+            audio_hidden_states: audio DiT hidden states
+            x_freqs / y_freqs: cross-modal RoPE (cos, sin) pairs.
+                If provided, x_freqs is assumed to correspond to the primary tower and y_freqs
+                to the conditioning tower.
+            a2v_condition_scale: audio->visual conditioning strength (overrides global condition_scale)
+            v2a_condition_scale: visual->audio conditioning strength (overrides global condition_scale)
+            condition_scale: fallback conditioning strength when per-direction scale is None
+            video_grid_size: (F, H, W), used on the audio side when pooled_adaln is enabled
+
+        Returns:
+            (visual_hidden_states, audio_hidden_states), both conditioned in their respective directions.
+        """
+
+        visual_conditioned = self.apply_conditional_control(
+            layer_idx=layer_idx,
+            direction="a2v",
+            primary_hidden_states=visual_hidden_states,
+            condition_hidden_states=audio_hidden_states,
+            x_freqs=x_freqs,
+            y_freqs=y_freqs,
+            condition_scale=a2v_condition_scale if a2v_condition_scale is not None else condition_scale,
+            video_grid_size=video_grid_size,
+            use_gradient_checkpointing=use_gradient_checkpointing,
+            use_gradient_checkpointing_offload=use_gradient_checkpointing_offload,
+        )
+
+        audio_conditioned = self.apply_conditional_control(
+            layer_idx=layer_idx,
+            direction="v2a",
+            primary_hidden_states=audio_hidden_states,
+            condition_hidden_states=visual_hidden_states,
+            x_freqs=y_freqs,
+            y_freqs=x_freqs,
+            condition_scale=v2a_condition_scale if v2a_condition_scale is not None else condition_scale,
+            video_grid_size=video_grid_size,
+            use_gradient_checkpointing=use_gradient_checkpointing,
+            use_gradient_checkpointing_offload=use_gradient_checkpointing_offload,
+        )
+
+        return visual_conditioned, audio_conditioned

diffsynth/models/wan_video_dit.py

@@ -99,18 +99,30 @@ def rope_apply(x, freqs, num_heads):
     return x_out.to(x.dtype)
 
 
+def set_to_torch_norm(models):
+    for model in models:
+        for module in model.modules():
+            if isinstance(module, RMSNorm):
+                module.use_torch_norm = True
+
+
 class RMSNorm(nn.Module):
     def __init__(self, dim, eps=1e-5):
         super().__init__()
         self.eps = eps
         self.weight = nn.Parameter(torch.ones(dim))
+        self.use_torch_norm = False
+        self.normalized_shape = (dim,)
 
     def norm(self, x):
         return x * torch.rsqrt(x.pow(2).mean(dim=-1, keepdim=True) + self.eps)
 
     def forward(self, x):
         dtype = x.dtype
-        return self.norm(x.float()).to(dtype) * self.weight
+        if self.use_torch_norm:
+            return F.rms_norm(x, self.normalized_shape, self.weight, self.eps)
+        else:        
+            return self.norm(x.float()).to(dtype) * self.weight
 
 
 class AttentionModule(nn.Module):

diffsynth/models/wan_video_vae.py

@@ -469,7 +469,7 @@ class Down_ResidualBlock(nn.Module):
     def forward(self, x, feat_cache=None, feat_idx=[0]):
         x_copy = x.clone()
         for module in self.downsamples:
-            x = module(x, feat_cache, feat_idx)
+            x, feat_cache, feat_idx = module(x, feat_cache, feat_idx)
 
         return x + self.avg_shortcut(x_copy), feat_cache, feat_idx
 
@@ -506,10 +506,10 @@ class Up_ResidualBlock(nn.Module):
     def forward(self, x, feat_cache=None, feat_idx=[0], first_chunk=False):
         x_main = x.clone()
         for module in self.upsamples:
-            x_main = module(x_main, feat_cache, feat_idx)
+            x_main, feat_cache, feat_idx = module(x_main, feat_cache, feat_idx)
         if self.avg_shortcut is not None:
             x_shortcut = self.avg_shortcut(x, first_chunk)
-            return x_main + x_shortcut
+            return x_main + x_shortcut, feat_cache, feat_idx
         else:
             return x_main, feat_cache, feat_idx
 

diffsynth/pipelines/anima_image.py

@@ -0,0 +1,263 @@
+import torch, math
+from PIL import Image
+from typing import Union
+from tqdm import tqdm
+from einops import rearrange
+import numpy as np
+from math import prod
+from transformers import AutoTokenizer
+
+from ..core.device.npu_compatible_device import get_device_type
+from ..diffusion import FlowMatchScheduler
+from ..core import ModelConfig, gradient_checkpoint_forward
+from ..diffusion.base_pipeline import BasePipeline, PipelineUnit, ControlNetInput
+from ..utils.lora.merge import merge_lora
+
+from ..models.anima_dit import AnimaDiT
+from ..models.z_image_text_encoder import ZImageTextEncoder
+from ..models.wan_video_vae import WanVideoVAE
+
+
+class AnimaImagePipeline(BasePipeline):
+
+    def __init__(self, device=get_device_type(), torch_dtype=torch.bfloat16):
+        super().__init__(
+            device=device, torch_dtype=torch_dtype,
+            height_division_factor=16, width_division_factor=16,
+        )
+        self.scheduler = FlowMatchScheduler("Z-Image")
+        self.text_encoder: ZImageTextEncoder = None
+        self.dit: AnimaDiT = None
+        self.vae: WanVideoVAE = None
+        self.tokenizer: AutoTokenizer = None
+        self.tokenizer_t5xxl: AutoTokenizer = None
+        self.in_iteration_models = ("dit",)
+        self.units = [
+            AnimaUnit_ShapeChecker(),
+            AnimaUnit_NoiseInitializer(),
+            AnimaUnit_InputImageEmbedder(),
+            AnimaUnit_PromptEmbedder(),
+        ]
+        self.model_fn = model_fn_anima
+    
+    
+    @staticmethod
+    def from_pretrained(
+        torch_dtype: torch.dtype = torch.bfloat16,
+        device: Union[str, torch.device] = get_device_type(),
+        model_configs: list[ModelConfig] = [],
+        tokenizer_config: ModelConfig = ModelConfig(model_id="Qwen/Qwen3-0.6B", origin_file_pattern="./"),
+        tokenizer_t5xxl_config: ModelConfig = ModelConfig(model_id="stabilityai/stable-diffusion-3.5-large", origin_file_pattern="tokenizer_3/"),
+        vram_limit: float = None,
+    ):
+        # Initialize pipeline
+        pipe = AnimaImagePipeline(device=device, torch_dtype=torch_dtype)
+        model_pool = pipe.download_and_load_models(model_configs, vram_limit)
+        
+        # Fetch models
+        pipe.text_encoder = model_pool.fetch_model("z_image_text_encoder")
+        pipe.dit = model_pool.fetch_model("anima_dit")
+        pipe.vae = model_pool.fetch_model("wan_video_vae")
+        if tokenizer_config is not None:
+            tokenizer_config.download_if_necessary()
+            pipe.tokenizer = AutoTokenizer.from_pretrained(tokenizer_config.path)
+        if tokenizer_t5xxl_config is not None:
+            tokenizer_t5xxl_config.download_if_necessary()
+            pipe.tokenizer_t5xxl = AutoTokenizer.from_pretrained(tokenizer_t5xxl_config.path)
+        # VRAM Management
+        pipe.vram_management_enabled = pipe.check_vram_management_state()
+        return pipe
+    
+    
+    @torch.no_grad()
+    def __call__(
+        self,
+        # Prompt
+        prompt: str,
+        negative_prompt: str = "",
+        cfg_scale: float = 4.0,
+        # Image
+        input_image: Image.Image = None,
+        denoising_strength: float = 1.0,
+        # Shape
+        height: int = 1024,
+        width: int = 1024,
+        # Randomness
+        seed: int = None,
+        rand_device: str = "cpu",
+        # Steps
+        num_inference_steps: int = 30,
+        sigma_shift: float = None,
+        # Progress bar
+        progress_bar_cmd = tqdm,
+    ):
+        # Scheduler
+        self.scheduler.set_timesteps(num_inference_steps, denoising_strength=denoising_strength, shift=sigma_shift)
+        
+        # Parameters
+        inputs_posi = {
+            "prompt": prompt,
+        }
+        inputs_nega = {
+            "negative_prompt": negative_prompt,
+        }
+        inputs_shared = {
+            "cfg_scale": cfg_scale,
+            "input_image": input_image, "denoising_strength": denoising_strength,
+            "height": height, "width": width,
+            "seed": seed, "rand_device": rand_device,
+            "num_inference_steps": num_inference_steps,
+        }
+        for unit in self.units:
+            inputs_shared, inputs_posi, inputs_nega = self.unit_runner(unit, self, inputs_shared, inputs_posi, inputs_nega)
+
+        # Denoise
+        self.load_models_to_device(self.in_iteration_models)
+        models = {name: getattr(self, name) for name in self.in_iteration_models}
+        for progress_id, timestep in enumerate(progress_bar_cmd(self.scheduler.timesteps)):
+            timestep = timestep.unsqueeze(0).to(dtype=self.torch_dtype, device=self.device)
+            noise_pred = self.cfg_guided_model_fn(
+                self.model_fn, cfg_scale,
+                inputs_shared, inputs_posi, inputs_nega,
+                **models, timestep=timestep, progress_id=progress_id
+            )
+            inputs_shared["latents"] = self.step(self.scheduler, progress_id=progress_id, noise_pred=noise_pred, **inputs_shared)
+        
+        # Decode
+        self.load_models_to_device(['vae'])
+        image = self.vae.decode(inputs_shared["latents"].unsqueeze(2), device=self.device).squeeze(2)
+        image = self.vae_output_to_image(image)
+        self.load_models_to_device([])
+
+        return image
+
+
+class AnimaUnit_ShapeChecker(PipelineUnit):
+    def __init__(self):
+        super().__init__(
+            input_params=("height", "width"),
+            output_params=("height", "width"),
+        )
+
+    def process(self, pipe: AnimaImagePipeline, height, width):
+        height, width = pipe.check_resize_height_width(height, width)
+        return {"height": height, "width": width}
+
+
+
+class AnimaUnit_NoiseInitializer(PipelineUnit):
+    def __init__(self):
+        super().__init__(
+            input_params=("height", "width", "seed", "rand_device"),
+            output_params=("noise",),
+        )
+
+    def process(self, pipe: AnimaImagePipeline, height, width, seed, rand_device):
+        noise = pipe.generate_noise((1, 16, height//8, width//8), seed=seed, rand_device=rand_device, rand_torch_dtype=pipe.torch_dtype)
+        return {"noise": noise}
+
+
+
+class AnimaUnit_InputImageEmbedder(PipelineUnit):
+    def __init__(self):
+        super().__init__(
+            input_params=("input_image", "noise"),
+            output_params=("latents", "input_latents"),
+            onload_model_names=("vae",)
+        )
+
+    def process(self, pipe: AnimaImagePipeline, input_image, noise):
+        if input_image is None:
+            return {"latents": noise, "input_latents": None}
+        pipe.load_models_to_device(['vae'])
+        if isinstance(input_image, list):
+            input_latents = []
+            for image in input_image:
+                image = pipe.preprocess_image(image).to(device=pipe.device, dtype=pipe.torch_dtype)
+                input_latents.append(pipe.vae.encode(image))
+            input_latents = torch.concat(input_latents, dim=0)
+        else:
+            image = pipe.preprocess_image(input_image).to(device=pipe.device, dtype=pipe.torch_dtype)
+            input_latents = pipe.vae.encode(image.unsqueeze(2), device=pipe.device).squeeze(2)
+        if pipe.scheduler.training:
+            return {"latents": noise, "input_latents": input_latents}
+        else:
+            latents = pipe.scheduler.add_noise(input_latents, noise, timestep=pipe.scheduler.timesteps[0])
+            return {"latents": latents, "input_latents": input_latents}
+
+
+class AnimaUnit_PromptEmbedder(PipelineUnit):
+    def __init__(self):
+        super().__init__(
+            seperate_cfg=True,
+            input_params_posi={"prompt": "prompt"},
+            input_params_nega={"prompt": "negative_prompt"},
+            output_params=("prompt_emb",),
+            onload_model_names=("text_encoder",)
+        )
+
+    def encode_prompt(
+        self,
+        pipe: AnimaImagePipeline,
+        prompt,
+        device = None,
+        max_sequence_length: int = 512,
+    ):
+        if isinstance(prompt, str):
+            prompt = [prompt]
+
+        text_inputs = pipe.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+
+        text_input_ids = text_inputs.input_ids.to(device)
+        prompt_masks = text_inputs.attention_mask.to(device).bool()
+
+        prompt_embeds = pipe.text_encoder(
+            input_ids=text_input_ids,
+            attention_mask=prompt_masks,
+            output_hidden_states=True,
+        ).hidden_states[-1]
+        
+        t5xxl_text_inputs = pipe.tokenizer_t5xxl(
+            prompt,
+            max_length=max_sequence_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+        t5xxl_ids = t5xxl_text_inputs.input_ids.to(device)
+
+        return prompt_embeds.to(pipe.torch_dtype), t5xxl_ids
+
+    def process(self, pipe: AnimaImagePipeline, prompt):
+        pipe.load_models_to_device(self.onload_model_names)
+        prompt_embeds, t5xxl_ids = self.encode_prompt(pipe, prompt, pipe.device)
+        return {"prompt_emb": prompt_embeds, "t5xxl_ids": t5xxl_ids}
+
+
+def model_fn_anima(
+    dit: AnimaDiT = None,
+    latents=None,
+    timestep=None,
+    prompt_emb=None,
+    t5xxl_ids=None,
+    use_gradient_checkpointing=False,
+    use_gradient_checkpointing_offload=False,
+    **kwargs
+):
+    latents = latents.unsqueeze(2)
+    timestep = timestep / 1000
+    model_output = dit(
+        x=latents,
+        timesteps=timestep,
+        context=prompt_emb,
+        t5xxl_ids=t5xxl_ids,
+        use_gradient_checkpointing=use_gradient_checkpointing,
+        use_gradient_checkpointing_offload=use_gradient_checkpointing_offload,
+    )
+    model_output = model_output.squeeze(2)
+    return model_output

diffsynth/pipelines/flux2_image.py

@@ -90,6 +90,7 @@ class Flux2ImagePipeline(BasePipeline):
         # Randomness
         seed: int = None,
         rand_device: str = "cpu",
+        initial_noise: torch.Tensor = None,
         # Steps
         num_inference_steps: int = 30,
         # Progress bar
@@ -109,7 +110,7 @@ class Flux2ImagePipeline(BasePipeline):
             "input_image": input_image, "denoising_strength": denoising_strength,
             "edit_image": edit_image, "edit_image_auto_resize": edit_image_auto_resize,
             "height": height, "width": width,
-            "seed": seed, "rand_device": rand_device,
+            "seed": seed, "rand_device": rand_device, "initial_noise": initial_noise,
             "num_inference_steps": num_inference_steps,
         }
         for unit in self.units:
@@ -429,12 +430,15 @@ class Flux2Unit_Qwen3PromptEmbedder(PipelineUnit):
 class Flux2Unit_NoiseInitializer(PipelineUnit):
     def __init__(self):
         super().__init__(
-            input_params=("height", "width", "seed", "rand_device"),
+            input_params=("height", "width", "seed", "rand_device", "initial_noise"),
             output_params=("noise",),
         )
 
-    def process(self, pipe: Flux2ImagePipeline, height, width, seed, rand_device):
-        noise = pipe.generate_noise((1, 128, height//16, width//16), seed=seed, rand_device=rand_device, rand_torch_dtype=pipe.torch_dtype)
+    def process(self, pipe: Flux2ImagePipeline, height, width, seed, rand_device, initial_noise):
+        if initial_noise is not None:
+            noise = initial_noise.clone()
+        else:
+            noise = pipe.generate_noise((1, 128, height//16, width//16), seed=seed, rand_device=rand_device, rand_torch_dtype=pipe.torch_dtype)
         noise = noise.reshape(1, 128, height//16 * width//16).permute(0, 2, 1)
         return {"noise": noise}
 

diffsynth/pipelines/ltx2_audio_video.py

@@ -12,7 +12,7 @@ from transformers import AutoImageProcessor, Gemma3Processor
 
 from ..core.device.npu_compatible_device import get_device_type
 from ..diffusion import FlowMatchScheduler
-from ..core import ModelConfig, gradient_checkpoint_forward
+from ..core import ModelConfig
 from ..diffusion.base_pipeline import BasePipeline, PipelineUnit
 
 from ..models.ltx2_text_encoder import LTX2TextEncoder, LTX2TextEncoderPostModules, LTXVGemmaTokenizer
@@ -22,6 +22,7 @@ from ..models.ltx2_audio_vae import LTX2AudioEncoder, LTX2AudioDecoder, LTX2Voco
 from ..models.ltx2_upsampler import LTX2LatentUpsampler
 from ..models.ltx2_common import VideoLatentShape, AudioLatentShape, VideoPixelShape, get_pixel_coords, VIDEO_SCALE_FACTORS
 from ..utils.data.media_io_ltx2 import ltx2_preprocess
+from ..utils.data.audio import convert_to_stereo
 
 
 class LTX2AudioVideoPipeline(BasePipeline):
@@ -58,12 +59,53 @@ class LTX2AudioVideoPipeline(BasePipeline):
             LTX2AudioVideoUnit_ShapeChecker(),
             LTX2AudioVideoUnit_PromptEmbedder(),
             LTX2AudioVideoUnit_NoiseInitializer(),
+            LTX2AudioVideoUnit_VideoRetakeEmbedder(),
+            LTX2AudioVideoUnit_AudioRetakeEmbedder(),
             LTX2AudioVideoUnit_InputAudioEmbedder(),
             LTX2AudioVideoUnit_InputVideoEmbedder(),
             LTX2AudioVideoUnit_InputImagesEmbedder(),
+            LTX2AudioVideoUnit_InContextVideoEmbedder(),
+        ]
+        self.stage2_units = [
+            LTX2AudioVideoUnit_SwitchStage2(),
+            LTX2AudioVideoUnit_NoiseInitializer(),
+            LTX2AudioVideoUnit_LatentsUpsampler(),
+            LTX2AudioVideoUnit_VideoRetakeEmbedder(),
+            LTX2AudioVideoUnit_AudioRetakeEmbedder(),
+            LTX2AudioVideoUnit_InputImagesEmbedder(),
+            LTX2AudioVideoUnit_SetScheduleStage2(),
         ]
         self.model_fn = model_fn_ltx2
 
+        self.default_negative_prompt = {
+            "LTX-2": (
+                "blurry, out of focus, overexposed, underexposed, low contrast, washed out colors, excessive noise, "
+                "grainy texture, poor lighting, flickering, motion blur, distorted proportions, unnatural skin tones, "
+                "deformed facial features, asymmetrical face, missing facial features, extra limbs, disfigured hands, "
+                "wrong hand count, artifacts around text, inconsistent perspective, camera shake, incorrect depth of "
+                "field, background too sharp, background clutter, distracting reflections, harsh shadows, inconsistent "
+                "lighting direction, color banding, cartoonish rendering, 3D CGI look, unrealistic materials, uncanny "
+                "valley effect, incorrect ethnicity, wrong gender, exaggerated expressions, wrong gaze direction, "
+                "mismatched lip sync, silent or muted audio, distorted voice, robotic voice, echo, background noise, "
+                "off-sync audio, incorrect dialogue, added dialogue, repetitive speech, jittery movement, awkward "
+                "pauses, incorrect timing, unnatural transitions, inconsistent framing, tilted camera, flat lighting, "
+                "inconsistent tone, cinematic oversaturation, stylized filters, or AI artifacts."
+            ),
+            "LTX-2.3": (
+                "blurry, out of focus, overexposed, underexposed, low contrast, washed out colors, excessive noise, "
+                "grainy texture, poor lighting, flickering, motion blur, distorted proportions, unnatural skin tones, "
+                "deformed facial features, asymmetrical face, missing facial features, extra limbs, disfigured hands, "
+                "wrong hand count, artifacts around text, inconsistent perspective, camera shake, incorrect depth of "
+                "field, background too sharp, background clutter, distracting reflections, harsh shadows, inconsistent "
+                "lighting direction, color banding, cartoonish rendering, 3D CGI look, unrealistic materials, uncanny "
+                "valley effect, incorrect ethnicity, wrong gender, exaggerated expressions, wrong gaze direction, "
+                "mismatched lip sync, silent or muted audio, distorted voice, robotic voice, echo, background noise, "
+                "off-sync audio, incorrect dialogue, added dialogue, repetitive speech, jittery movement, awkward "
+                "pauses, incorrect timing, unnatural transitions, inconsistent framing, tilted camera, flat lighting, "
+                "inconsistent tone, cinematic oversaturation, stylized filters, or AI artifacts."
+            ),
+        }
+
     @staticmethod
     def from_pretrained(
         torch_dtype: torch.dtype = torch.bfloat16,
@@ -71,6 +113,7 @@ class LTX2AudioVideoPipeline(BasePipeline):
         model_configs: list[ModelConfig] = [],
         tokenizer_config: ModelConfig = ModelConfig(model_id="google/gemma-3-12b-it-qat-q4_0-unquantized"),
         stage2_lora_config: Optional[ModelConfig] = None,
+        stage2_lora_strength: float = 0.8,
         vram_limit: float = None,
     ):
         # Initialize pipeline
@@ -91,112 +134,22 @@ class LTX2AudioVideoPipeline(BasePipeline):
         pipe.audio_vae_decoder = model_pool.fetch_model("ltx2_audio_vae_decoder")
         pipe.audio_vocoder = model_pool.fetch_model("ltx2_audio_vocoder")
         pipe.upsampler = model_pool.fetch_model("ltx2_latent_upsampler")
+        pipe.audio_vae_encoder = model_pool.fetch_model("ltx2_audio_vae_encoder")
 
         # Stage 2
         if stage2_lora_config is not None:
-            stage2_lora_config.download_if_necessary()
-            pipe.stage2_lora_path = stage2_lora_config.path
-        # Optional, currently not used
-        pipe.audio_vae_encoder = model_pool.fetch_model("ltx2_audio_vae_encoder")
+            pipe.stage2_lora_config = stage2_lora_config
+            pipe.stage2_lora_strength = stage2_lora_strength
 
         # VRAM Management
         pipe.vram_management_enabled = pipe.check_vram_management_state()
         return pipe
 
-    def stage2_denoise(self, inputs_shared, inputs_posi, inputs_nega, progress_bar_cmd=tqdm):
-        if inputs_shared["use_two_stage_pipeline"]:
-            latent = self.video_vae_encoder.per_channel_statistics.un_normalize(inputs_shared["video_latents"])
-            self.load_models_to_device('upsampler',)
-            latent = self.upsampler(latent)
-            latent = self.video_vae_encoder.per_channel_statistics.normalize(latent)
-            self.scheduler.set_timesteps(special_case="stage2")
-            inputs_shared.update({k.replace("stage2_", ""): v for k, v in inputs_shared.items() if k.startswith("stage2_")})
-            denoise_mask_video = 1.0
-            if inputs_shared.get("input_images", None) is not None:
-                latent, denoise_mask_video, initial_latents = self.apply_input_images_to_latents(
-                    latent, inputs_shared.pop("input_latents"), inputs_shared["input_images_indexes"],
-                    inputs_shared["input_images_strength"], latent.clone())
-                inputs_shared.update({"input_latents_video": initial_latents, "denoise_mask_video": denoise_mask_video})
-            inputs_shared["video_latents"] = self.scheduler.sigmas[0] * denoise_mask_video * inputs_shared[
-                "video_noise"] + (1 - self.scheduler.sigmas[0] * denoise_mask_video) * latent
-            inputs_shared["audio_latents"] = self.scheduler.sigmas[0] * inputs_shared["audio_noise"] + (
-                1 - self.scheduler.sigmas[0]) * inputs_shared["audio_latents"]
-
-            self.load_models_to_device(self.in_iteration_models)
-            if not inputs_shared["use_distilled_pipeline"]:
-                self.load_lora(self.dit, self.stage2_lora_path, alpha=0.8)
-            models = {name: getattr(self, name) for name in self.in_iteration_models}
-            for progress_id, timestep in enumerate(progress_bar_cmd(self.scheduler.timesteps)):
-                timestep = timestep.unsqueeze(0).to(dtype=self.torch_dtype, device=self.device)
-                noise_pred_video, noise_pred_audio = self.cfg_guided_model_fn(
-                    self.model_fn, 1.0, inputs_shared, inputs_posi, inputs_nega,
-                    **models, timestep=timestep, progress_id=progress_id
-                )
-                inputs_shared["video_latents"] = self.step(self.scheduler, inputs_shared["video_latents"], progress_id=progress_id,
-                                                           noise_pred=noise_pred_video, inpaint_mask=inputs_shared.get("denoise_mask_video", None),
-                                                           input_latents=inputs_shared.get("input_latents_video", None), **inputs_shared)
-                inputs_shared["audio_latents"] = self.step(self.scheduler, inputs_shared["audio_latents"], progress_id=progress_id,
-                                                           noise_pred=noise_pred_audio, **inputs_shared)
-        return inputs_shared
-
-    @torch.no_grad()
-    def __call__(
-        self,
-        # Prompt
-        prompt: str,
-        negative_prompt: Optional[str] = "",
-        # Image-to-video
-        denoising_strength: float = 1.0,
-        input_images: Optional[list[Image.Image]] = None,
-        input_images_indexes: Optional[list[int]] = None,
-        input_images_strength: Optional[float] = 1.0,
-        # Randomness
-        seed: Optional[int] = None,
-        rand_device: Optional[str] = "cpu",
-        # Shape
-        height: Optional[int] = 512,
-        width: Optional[int] = 768,
-        num_frames=121,
-        # Classifier-free guidance
-        cfg_scale: Optional[float] = 3.0,
-        # Scheduler
-        num_inference_steps: Optional[int] = 40,
-        # VAE tiling
-        tiled: Optional[bool] = True,
-        tile_size_in_pixels: Optional[int] = 512,
-        tile_overlap_in_pixels: Optional[int] = 128,
-        tile_size_in_frames: Optional[int] = 128,
-        tile_overlap_in_frames: Optional[int] = 24,
-        # Special Pipelines
-        use_two_stage_pipeline: Optional[bool] = False,
-        use_distilled_pipeline: Optional[bool] = False,
-        # progress_bar
-        progress_bar_cmd=tqdm,
-    ):
-        # Scheduler
-        self.scheduler.set_timesteps(num_inference_steps, denoising_strength=denoising_strength,
-                                     special_case="ditilled_stage1" if use_distilled_pipeline else None)
-        # Inputs
-        inputs_posi = {
-            "prompt": prompt,
-        }
-        inputs_nega = {
-            "negative_prompt": negative_prompt,
-        }
-        inputs_shared = {
-            "input_images": input_images, "input_images_indexes": input_images_indexes, "input_images_strength": input_images_strength,
-            "seed": seed, "rand_device": rand_device,
-            "height": height, "width": width, "num_frames": num_frames,
-            "cfg_scale": cfg_scale,
-            "tiled": tiled, "tile_size_in_pixels": tile_size_in_pixels, "tile_overlap_in_pixels": tile_overlap_in_pixels,
-            "tile_size_in_frames": tile_size_in_frames, "tile_overlap_in_frames": tile_overlap_in_frames,
-            "use_two_stage_pipeline": use_two_stage_pipeline, "use_distilled_pipeline": use_distilled_pipeline,
-            "video_patchifier": self.video_patchifier, "audio_patchifier": self.audio_patchifier,
-        }
-        for unit in self.units:
+    def denoise_stage(self, inputs_shared, inputs_posi, inputs_nega, units, cfg_scale=1.0, progress_bar_cmd=tqdm, skip_stage=False):
+        if skip_stage:
+            return inputs_shared, inputs_posi, inputs_nega
+        for unit in units:
             inputs_shared, inputs_posi, inputs_nega = self.unit_runner(unit, self, inputs_shared, inputs_posi, inputs_nega)
-
-        # Denoise Stage 1
         self.load_models_to_device(self.in_iteration_models)
         models = {name: getattr(self, name) for name in self.in_iteration_models}
         for progress_id, timestep in enumerate(progress_bar_cmd(self.scheduler.timesteps)):
@@ -207,34 +160,93 @@ class LTX2AudioVideoPipeline(BasePipeline):
             )
             inputs_shared["video_latents"] = self.step(self.scheduler, inputs_shared["video_latents"], progress_id=progress_id, noise_pred=noise_pred_video,
                                                        inpaint_mask=inputs_shared.get("denoise_mask_video", None), input_latents=inputs_shared.get("input_latents_video", None), **inputs_shared)
-            inputs_shared["audio_latents"] = self.step(self.scheduler, inputs_shared["audio_latents"], progress_id=progress_id,
-                                                       noise_pred=noise_pred_audio, **inputs_shared)
-
-        # Denoise Stage 2
-        inputs_shared = self.stage2_denoise(inputs_shared, inputs_posi, inputs_nega, progress_bar_cmd)
+            inputs_shared["audio_latents"] = self.step(self.scheduler, inputs_shared["audio_latents"], progress_id=progress_id, noise_pred=noise_pred_audio,
+                                                       inpaint_mask=inputs_shared.get("denoise_mask_audio", None), input_latents=inputs_shared.get("input_latents_audio", None), **inputs_shared)
+        return inputs_shared, inputs_posi, inputs_nega
 
+    @torch.no_grad()
+    def __call__(
+        self,
+        # Prompt
+        prompt: str,
+        negative_prompt: Optional[str] = "",
+        denoising_strength: float = 1.0,
+        # Image-to-video
+        input_images: Optional[list[Image.Image]] = None,
+        input_images_indexes: Optional[list[int]] = [0],
+        input_images_strength: Optional[float] = 1.0,
+        # In-Context Video Control
+        in_context_videos: Optional[list[list[Image.Image]]] = None,
+        in_context_downsample_factor: Optional[int] = 2,
+        # Video-to-video
+        retake_video: Optional[list[Image.Image]] = None,
+        retake_video_regions: Optional[list[tuple[float, float]]] = None,
+        # Audio-to-video
+        retake_audio: Optional[torch.Tensor] = None,
+        audio_sample_rate: Optional[int] = 48000,
+        retake_audio_regions: Optional[list[tuple[float, float]]] = None,
+        # Randomness
+        seed: Optional[int] = None,
+        rand_device: Optional[str] = "cpu",
+        # Shape
+        height: Optional[int] = 512,
+        width: Optional[int] = 768,
+        num_frames: Optional[int] = 121,
+        frame_rate: Optional[int] = 24,
+        # Classifier-free guidance
+        cfg_scale: Optional[float] = 3.0,
+        # Scheduler
+        num_inference_steps: Optional[int] = 30,
+        # VAE tiling
+        tiled: Optional[bool] = True,
+        tile_size_in_pixels: Optional[int] = 512,
+        tile_overlap_in_pixels: Optional[int] = 128,
+        tile_size_in_frames: Optional[int] = 128,
+        tile_overlap_in_frames: Optional[int] = 24,
+        # Special Pipelines
+        use_two_stage_pipeline: Optional[bool] = False,
+        stage2_spatial_upsample_factor: Optional[int] = 2,
+        clear_lora_before_state_two: Optional[bool] = False,
+        use_distilled_pipeline: Optional[bool] = False,
+        # progress_bar
+        progress_bar_cmd=tqdm,
+    ):
+        # Scheduler
+        self.scheduler.set_timesteps(num_inference_steps, denoising_strength=denoising_strength, special_case="ditilled_stage1" if use_distilled_pipeline else None)
+        # Inputs
+        inputs_posi = {
+            "prompt": prompt,
+        }
+        inputs_nega = {
+            "negative_prompt": negative_prompt,
+        }
+        inputs_shared = {
+            "input_images": input_images, "input_images_indexes": input_images_indexes, "input_images_strength": input_images_strength,
+            "retake_video": retake_video, "retake_video_regions": retake_video_regions,
+            "retake_audio": (retake_audio, audio_sample_rate) if retake_audio is not None else None, "retake_audio_regions": retake_audio_regions,
+            "in_context_videos": in_context_videos, "in_context_downsample_factor": in_context_downsample_factor,
+            "seed": seed, "rand_device": rand_device,
+            "height": height, "width": width, "num_frames": num_frames, "frame_rate": frame_rate,
+            "cfg_scale": cfg_scale,
+            "tiled": tiled, "tile_size_in_pixels": tile_size_in_pixels, "tile_overlap_in_pixels": tile_overlap_in_pixels,
+            "tile_size_in_frames": tile_size_in_frames, "tile_overlap_in_frames": tile_overlap_in_frames,
+            "use_two_stage_pipeline": use_two_stage_pipeline, "use_distilled_pipeline": use_distilled_pipeline, "clear_lora_before_state_two": clear_lora_before_state_two, "stage2_spatial_upsample_factor": stage2_spatial_upsample_factor,
+            "video_patchifier": self.video_patchifier, "audio_patchifier": self.audio_patchifier,
+        }
+        # Stage 1
+        inputs_shared, inputs_posi, inputs_nega = self.denoise_stage(inputs_shared, inputs_posi, inputs_nega, self.units, cfg_scale, progress_bar_cmd)
+        # Stage 2
+        inputs_shared, inputs_posi, inputs_nega = self.denoise_stage(inputs_shared, inputs_posi, inputs_nega, self.stage2_units, 1.0, progress_bar_cmd, not inputs_shared["use_two_stage_pipeline"])
         # Decode
         self.load_models_to_device(['video_vae_decoder'])
-        video = self.video_vae_decoder.decode(inputs_shared["video_latents"], tiled, tile_size_in_pixels,
-                                              tile_overlap_in_pixels, tile_size_in_frames, tile_overlap_in_frames)
+        video = self.video_vae_decoder.decode(inputs_shared["video_latents"], tiled, tile_size_in_pixels, tile_overlap_in_pixels, tile_size_in_frames, tile_overlap_in_frames)
         video = self.vae_output_to_video(video)
         self.load_models_to_device(['audio_vae_decoder', 'audio_vocoder'])
         decoded_audio = self.audio_vae_decoder(inputs_shared["audio_latents"])
-        decoded_audio = self.audio_vocoder(decoded_audio).squeeze(0).float()
+        decoded_audio = self.audio_vocoder(decoded_audio)
+        decoded_audio = self.output_audio_format_check(decoded_audio)
         return video, decoded_audio
 
-    def apply_input_images_to_latents(self, latents, input_latents, input_indexes, input_strength, initial_latents=None, num_frames=121):
-        b, _, f, h, w = latents.shape
-        denoise_mask = torch.ones((b, 1, f, h, w), dtype=latents.dtype, device=latents.device)
-        initial_latents = torch.zeros_like(latents) if initial_latents is None else initial_latents
-        for idx, input_latent in zip(input_indexes, input_latents):
-            idx = min(max(1 + (idx-1) // 8, 0), f - 1)
-            input_latent = input_latent.to(dtype=latents.dtype, device=latents.device)
-            initial_latents[:, :, idx:idx + input_latent.shape[2], :, :] = input_latent
-            denoise_mask[:, :, idx:idx + input_latent.shape[2], :, :] = 1.0 - input_strength
-        latents = latents * denoise_mask + initial_latents * (1.0 - denoise_mask)
-        return latents, denoise_mask, initial_latents
-
 
 class LTX2AudioVideoUnit_PipelineChecker(PipelineUnit):
     def __init__(self):
@@ -252,8 +264,8 @@ class LTX2AudioVideoUnit_PipelineChecker(PipelineUnit):
         if inputs_shared.get("use_two_stage_pipeline", False):
             # distill pipeline also uses two-stage, but it does not needs lora
             if not inputs_shared.get("use_distilled_pipeline", False):
-                if not (hasattr(pipe, "stage2_lora_path") and pipe.stage2_lora_path is not None):
-                    raise ValueError("Two-stage pipeline requested, but stage2_lora_path is not set in the pipeline.")
+                if not (hasattr(pipe, "stage2_lora_config") and pipe.stage2_lora_config is not None):
+                    raise ValueError("Two-stage pipeline requested, but stage2_lora_config is not set in the pipeline.")
             if not (hasattr(pipe, "upsampler") and pipe.upsampler is not None):
                 raise ValueError("Two-stage pipeline requested, but upsampler model is not loaded in the pipeline.")
         return inputs_shared, inputs_posi, inputs_nega
@@ -263,22 +275,23 @@ class LTX2AudioVideoUnit_ShapeChecker(PipelineUnit):
     """
     For two-stage pipelines, the resolution must be divisible by 64.
     For one-stage pipelines, the resolution must be divisible by 32.
+    This unit set height and width to stage 1 resolution, and stage_2_width and stage_2_height.
     """
     def __init__(self):
         super().__init__(
-            input_params=("height", "width", "num_frames"),
-            output_params=("height", "width", "num_frames"),
+            input_params=("height", "width", "num_frames", "use_two_stage_pipeline", "stage2_spatial_upsample_factor"),
+            output_params=("height", "width", "num_frames", "stage_2_height", "stage_2_width"),
         )
 
-    def process(self, pipe: LTX2AudioVideoPipeline, height, width, num_frames, use_two_stage_pipeline=False):
+    def process(self, pipe: LTX2AudioVideoPipeline, height, width, num_frames, use_two_stage_pipeline=False, stage2_spatial_upsample_factor=2):
         if use_two_stage_pipeline:
-            self.width_division_factor = 64
-            self.height_division_factor = 64
-        height, width, num_frames = pipe.check_resize_height_width(height, width, num_frames)
-        if use_two_stage_pipeline:
-            self.width_division_factor = 32
-            self.height_division_factor = 32
-        return {"height": height, "width": width, "num_frames": num_frames}
+            height, width = height // stage2_spatial_upsample_factor, width // stage2_spatial_upsample_factor
+            height, width, num_frames = pipe.check_resize_height_width(height, width, num_frames)
+            stage_2_height, stage_2_width = int(height * stage2_spatial_upsample_factor), int(width * stage2_spatial_upsample_factor)
+        else:
+            stage_2_height, stage_2_width = None, None
+            height, width, num_frames = pipe.check_resize_height_width(height, width, num_frames)
+        return {"height": height, "width": width, "num_frames": num_frames, "stage_2_height": stage_2_height, "stage_2_width": stage_2_width}
 
 
 class LTX2AudioVideoUnit_PromptEmbedder(PipelineUnit):
@@ -291,121 +304,20 @@ class LTX2AudioVideoUnit_PromptEmbedder(PipelineUnit):
             output_params=("video_context", "audio_context"),
             onload_model_names=("text_encoder", "text_encoder_post_modules"),
         )
-
-    def _convert_to_additive_mask(self, attention_mask: torch.Tensor, dtype: torch.dtype) -> torch.Tensor:
-        return (attention_mask - 1).to(dtype).reshape(
-            (attention_mask.shape[0], 1, -1, attention_mask.shape[-1])) * torch.finfo(dtype).max
-
-    def _run_connectors(self, pipe, encoded_input: torch.Tensor,
-                        attention_mask: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
-        connector_attention_mask = self._convert_to_additive_mask(attention_mask, encoded_input.dtype)
-
-        encoded, encoded_connector_attention_mask = pipe.text_encoder_post_modules.embeddings_connector(
-            encoded_input,
-            connector_attention_mask,
-        )
-
-        # restore the mask values to int64
-        attention_mask = (encoded_connector_attention_mask < 0.000001).to(torch.int64)
-        attention_mask = attention_mask.reshape([encoded.shape[0], encoded.shape[1], 1])
-        encoded = encoded * attention_mask
-
-        encoded_for_audio, _ = pipe.text_encoder_post_modules.audio_embeddings_connector(
-            encoded_input, connector_attention_mask)
-
-        return encoded, encoded_for_audio, attention_mask.squeeze(-1)
-
-    def _norm_and_concat_padded_batch(
-        self,
-        encoded_text: torch.Tensor,
-        sequence_lengths: torch.Tensor,
-        padding_side: str = "right",
-    ) -> torch.Tensor:
-        """Normalize and flatten multi-layer hidden states, respecting padding.
-        Performs per-batch, per-layer normalization using masked mean and range,
-        then concatenates across the layer dimension.
-        Args:
-            encoded_text: Hidden states of shape [batch, seq_len, hidden_dim, num_layers].
-            sequence_lengths: Number of valid (non-padded) tokens per batch item.
-            padding_side: Whether padding is on "left" or "right".
-        Returns:
-            Normalized tensor of shape [batch, seq_len, hidden_dim * num_layers],
-            with padded positions zeroed out.
-        """
-        b, t, d, l = encoded_text.shape  # noqa: E741
-        device = encoded_text.device
-        # Build mask: [B, T, 1, 1]
-        token_indices = torch.arange(t, device=device)[None, :]  # [1, T]
-        if padding_side == "right":
-            # For right padding, valid tokens are from 0 to sequence_length-1
-            mask = token_indices < sequence_lengths[:, None]  # [B, T]
-        elif padding_side == "left":
-            # For left padding, valid tokens are from (T - sequence_length) to T-1
-            start_indices = t - sequence_lengths[:, None]  # [B, 1]
-            mask = token_indices >= start_indices  # [B, T]
-        else:
-            raise ValueError(f"padding_side must be 'left' or 'right', got {padding_side}")
-        mask = rearrange(mask, "b t -> b t 1 1")
-        eps = 1e-6
-        # Compute masked mean: [B, 1, 1, L]
-        masked = encoded_text.masked_fill(~mask, 0.0)
-        denom = (sequence_lengths * d).view(b, 1, 1, 1)
-        mean = masked.sum(dim=(1, 2), keepdim=True) / (denom + eps)
-        # Compute masked min/max: [B, 1, 1, L]
-        x_min = encoded_text.masked_fill(~mask, float("inf")).amin(dim=(1, 2), keepdim=True)
-        x_max = encoded_text.masked_fill(~mask, float("-inf")).amax(dim=(1, 2), keepdim=True)
-        range_ = x_max - x_min
-        # Normalize only the valid tokens
-        normed = 8 * (encoded_text - mean) / (range_ + eps)
-        # concat to be [Batch, T,  D * L] - this preserves the original structure
-        normed = normed.reshape(b, t, -1)  # [B, T, D * L]
-        # Apply mask to preserve original padding (set padded positions to 0)
-        mask_flattened = rearrange(mask, "b t 1 1 -> b t 1").expand(-1, -1, d * l)
-        normed = normed.masked_fill(~mask_flattened, 0.0)
-
-        return normed
-
-    def _run_feature_extractor(self,
-                               pipe,
-                               hidden_states: torch.Tensor,
-                               attention_mask: torch.Tensor,
-                               padding_side: str = "right") -> torch.Tensor:
-        encoded_text_features = torch.stack(hidden_states, dim=-1)
-        encoded_text_features_dtype = encoded_text_features.dtype
-        sequence_lengths = attention_mask.sum(dim=-1)
-        normed_concated_encoded_text_features = self._norm_and_concat_padded_batch(encoded_text_features,
-                                                                                   sequence_lengths,
-                                                                                   padding_side=padding_side)
-
-        return pipe.text_encoder_post_modules.feature_extractor_linear(
-            normed_concated_encoded_text_features.to(encoded_text_features_dtype))
-
     def _preprocess_text(
         self,
         pipe,
         text: str,
-        padding_side: str = "left",
     ) -> tuple[torch.Tensor, dict[str, torch.Tensor]]:
-        """
-        Encode a given string into feature tensors suitable for downstream tasks.
-        Args:
-            text (str): Input string to encode.
-        Returns:
-            tuple[torch.Tensor, dict[str, torch.Tensor]]: Encoded features and a dictionary with attention mask.
-        """
         token_pairs = pipe.tokenizer.tokenize_with_weights(text)["gemma"]
         input_ids = torch.tensor([[t[0] for t in token_pairs]], device=pipe.device)
         attention_mask = torch.tensor([[w[1] for w in token_pairs]], device=pipe.device)
         outputs = pipe.text_encoder(input_ids=input_ids, attention_mask=attention_mask, output_hidden_states=True)
-        projected = self._run_feature_extractor(pipe,
-                                                hidden_states=outputs.hidden_states,
-                                                attention_mask=attention_mask,
-                                                padding_side=padding_side)
-        return projected, attention_mask
-
+        return outputs.hidden_states, attention_mask
     def encode_prompt(self, pipe, text, padding_side="left"):
-        encoded_inputs, attention_mask = self._preprocess_text(pipe, text, padding_side)
-        video_encoding, audio_encoding, attention_mask = self._run_connectors(pipe, encoded_inputs, attention_mask)
+        hidden_states, attention_mask = self._preprocess_text(pipe, text)
+        video_encoding, audio_encoding, attention_mask = pipe.text_encoder_post_modules.process_hidden_states(
+            hidden_states, attention_mask, padding_side)
         return video_encoding, audio_encoding, attention_mask
 
     def process(self, pipe: LTX2AudioVideoPipeline, prompt: str):
@@ -417,8 +329,8 @@ class LTX2AudioVideoUnit_PromptEmbedder(PipelineUnit):
 class LTX2AudioVideoUnit_NoiseInitializer(PipelineUnit):
     def __init__(self):
         super().__init__(
-            input_params=("height", "width", "num_frames", "seed", "rand_device", "use_two_stage_pipeline"),
-            output_params=("video_noise", "audio_noise",),
+            input_params=("height", "width", "num_frames", "seed", "rand_device", "frame_rate"),
+            output_params=("video_noise", "audio_noise", "video_positions", "audio_positions", "video_latent_shape", "audio_latent_shape")
         )
 
     def process_stage(self, pipe: LTX2AudioVideoPipeline, height, width, num_frames, seed, rand_device, frame_rate=24.0):
@@ -443,15 +355,9 @@ class LTX2AudioVideoUnit_NoiseInitializer(PipelineUnit):
             "audio_latent_shape": audio_latent_shape
         }
 
-    def process(self, pipe: LTX2AudioVideoPipeline, height, width, num_frames, seed, rand_device, frame_rate=24.0, use_two_stage_pipeline=False):
-        if use_two_stage_pipeline:
-            stage1_dict = self.process_stage(pipe, height // 2, width // 2, num_frames, seed, rand_device, frame_rate)
-            stage2_dict = self.process_stage(pipe, height, width, num_frames, seed, rand_device, frame_rate)
-            initial_dict = stage1_dict
-            initial_dict.update({"stage2_" + k: v for k, v in stage2_dict.items()})
-            return initial_dict
-        else:
-            return self.process_stage(pipe, height, width, num_frames, seed, rand_device, frame_rate)
+    def process(self, pipe: LTX2AudioVideoPipeline, height, width, num_frames, seed, rand_device, frame_rate=24.0):
+        return self.process_stage(pipe, height, width, num_frames, seed, rand_device, frame_rate)
+
 
 class LTX2AudioVideoUnit_InputVideoEmbedder(PipelineUnit):
     def __init__(self):
@@ -462,17 +368,13 @@ class LTX2AudioVideoUnit_InputVideoEmbedder(PipelineUnit):
         )
 
     def process(self, pipe: LTX2AudioVideoPipeline, input_video, video_noise, tiled, tile_size_in_pixels, tile_overlap_in_pixels):
-        if input_video is None:
+        if input_video is None or not pipe.scheduler.training:
             return {"video_latents": video_noise}
         else:
             pipe.load_models_to_device(self.onload_model_names)
             input_video = pipe.preprocess_video(input_video)
             input_latents = pipe.video_vae_encoder.encode(input_video, tiled, tile_size_in_pixels, tile_overlap_in_pixels).to(dtype=pipe.torch_dtype, device=pipe.device)
-            if pipe.scheduler.training:
-                return {"video_latents": input_latents, "input_latents": input_latents}
-            else:
-                # TODO: implement video-to-video
-                raise NotImplementedError("Video-to-video not implemented yet.")
+            return {"video_latents": input_latents, "input_latents": input_latents}
 
 class LTX2AudioVideoUnit_InputAudioEmbedder(PipelineUnit):
     def __init__(self):
@@ -483,26 +385,95 @@ class LTX2AudioVideoUnit_InputAudioEmbedder(PipelineUnit):
         )
 
     def process(self, pipe: LTX2AudioVideoPipeline, input_audio, audio_noise):
-        if input_audio is None:
+        if input_audio is None or not pipe.scheduler.training:
             return {"audio_latents": audio_noise}
         else:
             input_audio, sample_rate = input_audio
+            input_audio = convert_to_stereo(input_audio)
             pipe.load_models_to_device(self.onload_model_names)
             input_audio = pipe.audio_processor.waveform_to_mel(input_audio.unsqueeze(0), waveform_sample_rate=sample_rate).to(dtype=pipe.torch_dtype)
             audio_input_latents = pipe.audio_vae_encoder(input_audio)
             audio_latent_shape = AudioLatentShape.from_torch_shape(audio_input_latents.shape)
             audio_positions = pipe.audio_patchifier.get_patch_grid_bounds(audio_latent_shape, device=pipe.device)
-            if pipe.scheduler.training:
-                return {"audio_latents": audio_input_latents, "audio_input_latents": audio_input_latents, "audio_positions": audio_positions, "audio_latent_shape": audio_latent_shape}
-            else:
-                # TODO: implement video-to-video
-                raise NotImplementedError("Video-to-video not implemented yet.")
+            return {"audio_latents": audio_input_latents, "audio_input_latents": audio_input_latents, "audio_positions": audio_positions, "audio_latent_shape": audio_latent_shape}
+
+
+class LTX2AudioVideoUnit_VideoRetakeEmbedder(PipelineUnit):
+    def __init__(self):
+        super().__init__(
+            input_params=("retake_video", "height", "width", "tiled", "tile_size_in_pixels", "tile_overlap_in_pixels", "video_positions", "retake_video_regions"),
+            output_params=("input_latents_video", "denoise_mask_video"),
+            onload_model_names=("video_vae_encoder")
+        )
+
+    def process(self, pipe: LTX2AudioVideoPipeline, retake_video, height, width, tiled, tile_size_in_pixels, tile_overlap_in_pixels, video_positions, retake_video_regions=None):
+        if retake_video is None:
+            return {}
+        pipe.load_models_to_device(self.onload_model_names)
+        resized_video = [frame.resize((width, height)) for frame in retake_video]
+        input_video = pipe.preprocess_video(resized_video)
+        input_latents_video = pipe.video_vae_encoder.encode(input_video, tiled, tile_size_in_pixels, tile_overlap_in_pixels).to(dtype=pipe.torch_dtype, device=pipe.device)
+
+        b, c, f, h, w = input_latents_video.shape
+        denoise_mask_video = torch.zeros((b, 1, f, h, w), device=input_latents_video.device, dtype=input_latents_video.dtype)
+        if retake_video_regions is not None and len(retake_video_regions) > 0:
+            for start_time, end_time in retake_video_regions:
+                t_start, t_end = video_positions[0, 0].unbind(dim=-1)
+                in_region = (t_end >= start_time) & (t_start <= end_time)
+                in_region = pipe.video_patchifier.unpatchify_video(in_region.unsqueeze(0).unsqueeze(-1), f, h, w)
+                denoise_mask_video = torch.where(in_region, torch.ones_like(denoise_mask_video), denoise_mask_video)
+
+        return {"input_latents_video": input_latents_video, "denoise_mask_video": denoise_mask_video}
+
+
+class LTX2AudioVideoUnit_AudioRetakeEmbedder(PipelineUnit):
+    """
+    Functionality of audio2video, audio retaking.
+    """
+    def __init__(self):
+        super().__init__(
+            input_params=("retake_audio", "seed", "rand_device", "retake_audio_regions"),
+            output_params=("input_latents_audio", "audio_noise", "audio_positions", "audio_latent_shape", "denoise_mask_audio"),
+            onload_model_names=("audio_vae_encoder",)
+        )
+
+    def process(self, pipe: LTX2AudioVideoPipeline, retake_audio, seed, rand_device, retake_audio_regions=None):
+        if retake_audio is None:
+            return {}
+        else:
+            input_audio, sample_rate = retake_audio
+            input_audio = convert_to_stereo(input_audio)
+            pipe.load_models_to_device(self.onload_model_names)
+            input_audio = pipe.audio_processor.waveform_to_mel(input_audio.unsqueeze(0), waveform_sample_rate=sample_rate).to(dtype=pipe.torch_dtype, device=pipe.device)
+            input_latents_audio = pipe.audio_vae_encoder(input_audio)
+            audio_latent_shape = AudioLatentShape.from_torch_shape(input_latents_audio.shape)
+            audio_positions = pipe.audio_patchifier.get_patch_grid_bounds(audio_latent_shape, device=pipe.device)
+            # Regenerate noise for the new shape if retake_audio is provided, to avoid shape mismatch.
+            audio_noise = pipe.generate_noise(input_latents_audio.shape, seed=seed, rand_device=rand_device)
+
+            b, c, t, f = input_latents_audio.shape
+            denoise_mask_audio = torch.zeros((b, 1, t, 1), device=input_latents_audio.device, dtype=input_latents_audio.dtype)
+            if retake_audio_regions is not None and len(retake_audio_regions) > 0:
+                for start_time, end_time in retake_audio_regions:
+                    t_start, t_end = audio_positions[:, 0, :, 0], audio_positions[:, 0, :, 1]
+                    in_region = (t_end >= start_time) & (t_start <= end_time)
+                    in_region = pipe.audio_patchifier.unpatchify_audio(in_region.unsqueeze(-1), 1, 1)
+                    denoise_mask_audio = torch.where(in_region, torch.ones_like(denoise_mask_audio), denoise_mask_audio)
+
+            return {
+                "input_latents_audio": input_latents_audio,
+                "denoise_mask_audio": denoise_mask_audio,
+                "audio_noise": audio_noise,
+                "audio_positions": audio_positions,
+                "audio_latent_shape": audio_latent_shape,
+            }
+
 
 class LTX2AudioVideoUnit_InputImagesEmbedder(PipelineUnit):
     def __init__(self):
         super().__init__(
-            input_params=("input_images", "input_images_indexes", "input_images_strength", "video_latents", "height", "width", "num_frames", "tiled", "tile_size_in_pixels", "tile_overlap_in_pixels", "use_two_stage_pipeline"),
-            output_params=("video_latents"),
+            input_params=("input_images", "input_images_indexes", "input_images_strength", "video_latents", "height", "width", "frame_rate", "tiled", "tile_size_in_pixels", "tile_overlap_in_pixels", "input_latents_video", "denoise_mask_video"),
+            output_params=("denoise_mask_video", "input_latents_video", "ref_frames_latents", "ref_frames_positions"),
             onload_model_names=("video_vae_encoder")
         )
 
@@ -511,30 +482,166 @@ class LTX2AudioVideoUnit_InputImagesEmbedder(PipelineUnit):
         image = torch.Tensor(np.array(image, dtype=np.float32)).to(dtype=pipe.torch_dtype, device=pipe.device)
         image = image / 127.5 - 1.0
         image = repeat(image, f"H W C -> B C F H W", B=1, F=1)
-        latent = pipe.video_vae_encoder.encode(image, tiled, tile_size_in_pixels, tile_overlap_in_pixels).to(pipe.device)
-        return latent
+        latents = pipe.video_vae_encoder.encode(image, tiled, tile_size_in_pixels, tile_overlap_in_pixels).to(pipe.device)
+        return latents
 
-    def process(self, pipe: LTX2AudioVideoPipeline, input_images, input_images_indexes, input_images_strength, video_latents, height, width, num_frames, tiled, tile_size_in_pixels, tile_overlap_in_pixels, use_two_stage_pipeline=False):
+    def apply_input_images_to_latents(self, latents, input_latents, input_indexes, input_strength=1.0, input_latents_video=None, denoise_mask_video=None):
+        b, _, f, h, w = latents.shape
+        denoise_mask = torch.ones((b, 1, f, h, w), dtype=latents.dtype, device=latents.device) if denoise_mask_video is None else denoise_mask_video
+        input_latents_video = torch.zeros_like(latents) if input_latents_video is None else input_latents_video
+        for idx, input_latent in zip(input_indexes, input_latents):
+            idx = min(max(1 + (idx-1) // 8, 0), f - 1)
+            input_latent = input_latent.to(dtype=latents.dtype, device=latents.device)
+            input_latents_video[:, :, idx:idx + input_latent.shape[2], :, :] = input_latent
+            denoise_mask[:, :, idx:idx + input_latent.shape[2], :, :] = 1.0 - input_strength
+        return input_latents_video, denoise_mask
+
+    def process(
+        self,
+        pipe: LTX2AudioVideoPipeline,
+        video_latents,
+        input_images,
+        height,
+        width,
+        frame_rate,
+        tiled,
+        tile_size_in_pixels,
+        tile_overlap_in_pixels,
+        input_images_indexes=[0],
+        input_images_strength=1.0,
+        input_latents_video=None,
+        denoise_mask_video=None,
+    ):
         if input_images is None or len(input_images) == 0:
-            return {"video_latents": video_latents}
+            return {}
+        else:
+            if len(input_images_indexes) != len(set(input_images_indexes)):
+                raise ValueError("Input images must have unique indexes.")
+            pipe.load_models_to_device(self.onload_model_names)
+            frame_conditions = {"input_latents_video": None, "denoise_mask_video": None, "ref_frames_latents": [], "ref_frames_positions": []}
+            for img, index in zip(input_images, input_images_indexes):
+                latents = self.get_image_latent(pipe, img, height, width, tiled, tile_size_in_pixels, tile_overlap_in_pixels)
+                # first_frame by replacing latents
+                if index == 0:
+                    input_latents_video, denoise_mask_video = self.apply_input_images_to_latents(
+                        video_latents, [latents], [0], input_images_strength, input_latents_video, denoise_mask_video)
+                    frame_conditions.update({"input_latents_video": input_latents_video, "denoise_mask_video": denoise_mask_video})
+                # other frames by adding reference latents
+                else:
+                    latent_coords = pipe.video_patchifier.get_patch_grid_bounds(output_shape=VideoLatentShape.from_torch_shape(latents.shape), device=pipe.device)
+                    video_positions = get_pixel_coords(latent_coords, VIDEO_SCALE_FACTORS, False).float()
+                    video_positions[:, 0, ...] = (video_positions[:, 0, ...] + index) / frame_rate
+                    video_positions = video_positions.to(pipe.torch_dtype)
+                    frame_conditions["ref_frames_latents"].append(latents)
+                    frame_conditions["ref_frames_positions"].append(video_positions)
+            if len(frame_conditions["ref_frames_latents"]) == 0:
+                frame_conditions.update({"ref_frames_latents": None, "ref_frames_positions": None})
+            return frame_conditions
+
+
+class LTX2AudioVideoUnit_InContextVideoEmbedder(PipelineUnit):
+    def __init__(self):
+        super().__init__(
+            input_params=("in_context_videos", "height", "width", "num_frames", "frame_rate", "in_context_downsample_factor", "tiled", "tile_size_in_pixels", "tile_overlap_in_pixels"),
+            output_params=("in_context_video_latents", "in_context_video_positions"),
+            onload_model_names=("video_vae_encoder")
+        )
+
+    def check_in_context_video(self, pipe, in_context_video, height, width, num_frames, in_context_downsample_factor):
+        if in_context_video is None or len(in_context_video) == 0:
+            raise ValueError("In-context video is None or empty.")
+        in_context_video = in_context_video[:num_frames]
+        expected_height = height // in_context_downsample_factor
+        expected_width = width // in_context_downsample_factor
+        current_h, current_w, current_f = in_context_video[0].size[1], in_context_video[0].size[0], len(in_context_video)
+        h, w, f = pipe.check_resize_height_width(expected_height, expected_width, current_f, verbose=0)
+        if current_h != h or current_w != w:
+            in_context_video = [img.resize((w, h)) for img in in_context_video]
+        if current_f != f:
+            # pad black frames at the end
+            in_context_video = in_context_video + [Image.new("RGB", (w, h), (0, 0, 0))] * (f - current_f)
+        return in_context_video
+
+    def process(self, pipe: LTX2AudioVideoPipeline, in_context_videos, height, width, num_frames, frame_rate, in_context_downsample_factor, tiled, tile_size_in_pixels, tile_overlap_in_pixels):
+        if in_context_videos is None or len(in_context_videos) == 0:
+            return {}
         else:
             pipe.load_models_to_device(self.onload_model_names)
-            output_dicts = {}
-            stage1_height = height // 2 if use_two_stage_pipeline else height
-            stage1_width = width // 2 if use_two_stage_pipeline else width
-            stage1_latents = [
-                self.get_image_latent(pipe, img, stage1_height, stage1_width, tiled, tile_size_in_pixels,
-                                      tile_overlap_in_pixels) for img in input_images
-            ]
-            video_latents, denoise_mask_video, initial_latents = pipe.apply_input_images_to_latents(video_latents, stage1_latents, input_images_indexes, input_images_strength, num_frames=num_frames)
-            output_dicts.update({"video_latents": video_latents, "denoise_mask_video": denoise_mask_video, "input_latents_video": initial_latents})
-            if use_two_stage_pipeline:
-                stage2_latents = [
-                    self.get_image_latent(pipe, img, height, width, tiled, tile_size_in_pixels,
-                                          tile_overlap_in_pixels) for img in input_images
-                ]
-                output_dicts.update({"stage2_input_latents": stage2_latents})
-            return output_dicts
+            latents, positions = [], []
+            for in_context_video in in_context_videos:
+                in_context_video = self.check_in_context_video(pipe, in_context_video, height, width, num_frames, in_context_downsample_factor)
+                in_context_video = pipe.preprocess_video(in_context_video)
+                in_context_latents = pipe.video_vae_encoder.encode(in_context_video, tiled, tile_size_in_pixels, tile_overlap_in_pixels).to(dtype=pipe.torch_dtype, device=pipe.device)
+
+                latent_coords = pipe.video_patchifier.get_patch_grid_bounds(output_shape=VideoLatentShape.from_torch_shape(in_context_latents.shape), device=pipe.device)
+                video_positions = get_pixel_coords(latent_coords, VIDEO_SCALE_FACTORS, True).float()
+                video_positions[:, 0, ...] = video_positions[:, 0, ...] / frame_rate
+                video_positions[:, 1, ...] *= in_context_downsample_factor  # height axis
+                video_positions[:, 2, ...] *= in_context_downsample_factor  # width axis
+                video_positions = video_positions.to(pipe.torch_dtype)
+
+                latents.append(in_context_latents)
+                positions.append(video_positions)
+            latents = torch.cat(latents, dim=1)
+            positions = torch.cat(positions, dim=1)
+            return {"in_context_video_latents": latents, "in_context_video_positions": positions}
+
+
+class LTX2AudioVideoUnit_SwitchStage2(PipelineUnit):
+    """
+    1. switch height and width to stage 2 resolution
+    2. clear in_context_video_latents and in_context_video_positions
+    3. switch stage 2 lora model
+    """
+    def __init__(self):
+        super().__init__(
+            input_params=("stage_2_height", "stage_2_width", "clear_lora_before_state_two", "use_distilled_pipeline"),
+            output_params=("height", "width", "in_context_video_latents", "in_context_video_positions"),
+        )
+
+    def process(self, pipe: LTX2AudioVideoPipeline, stage_2_height, stage_2_width, clear_lora_before_state_two, use_distilled_pipeline):
+        stage2_params = {}
+        stage2_params.update({"height": stage_2_height, "width": stage_2_width})
+        stage2_params.update({"in_context_video_latents": None, "in_context_video_positions": None})
+        stage2_params.update({"input_latents_video": None, "denoise_mask_video": None})
+        if clear_lora_before_state_two:
+            pipe.clear_lora()
+        if not use_distilled_pipeline:
+            pipe.load_lora(pipe.dit, pipe.stage2_lora_config, alpha=pipe.stage2_lora_strength, state_dict=pipe.stage2_lora_config.state_dict)
+        return stage2_params
+
+
+class LTX2AudioVideoUnit_SetScheduleStage2(PipelineUnit):
+    def __init__(self):
+        super().__init__(
+            input_params=("video_latents", "video_noise", "audio_latents", "audio_noise"),
+            output_params=("video_latents", "audio_latents"),
+        )
+
+    def process(self, pipe: LTX2AudioVideoPipeline, video_latents, video_noise, audio_latents, audio_noise):
+        pipe.scheduler.set_timesteps(special_case="stage2")
+        video_latents = pipe.scheduler.add_noise(video_latents, video_noise, pipe.scheduler.timesteps[0])
+        audio_latents = pipe.scheduler.add_noise(audio_latents, audio_noise, pipe.scheduler.timesteps[0])
+        return {"video_latents": video_latents, "audio_latents": audio_latents}
+
+
+class LTX2AudioVideoUnit_LatentsUpsampler(PipelineUnit):
+    def __init__(self):
+        super().__init__(
+            input_params=("video_latents",),
+            output_params=("video_latents",),
+            onload_model_names=("upsampler",),
+        )
+
+    def process(self, pipe: LTX2AudioVideoPipeline, video_latents):
+        if video_latents is None or pipe.upsampler is None:
+            raise ValueError("No upsampler or no video latents before stage 2.")
+        else:
+            pipe.load_models_to_device(self.onload_model_names)
+            video_latents = pipe.video_vae_encoder.per_channel_statistics.un_normalize(video_latents)
+            video_latents = pipe.upsampler(video_latents)
+            video_latents = pipe.video_vae_encoder.per_channel_statistics.normalize(video_latents)
+            return {"video_latents": video_latents}
 
 
 def model_fn_ltx2(
@@ -548,7 +655,19 @@ def model_fn_ltx2(
     audio_positions=None,
     audio_patchifier=None,
     timestep=None,
+    # First Frame Conditioning
+    input_latents_video=None,
     denoise_mask_video=None,
+    # Other Frames Conditioning
+    ref_frames_latents=None,
+    ref_frames_positions=None,
+    # In-Context Conditioning
+    in_context_video_latents=None,
+    in_context_video_positions=None,
+    # Audio Inputs
+    input_latents_audio=None,
+    denoise_mask_audio=None,
+    # Gradient Checkpointing
     use_gradient_checkpointing=False,
     use_gradient_checkpointing_offload=False,
     **kwargs,
@@ -558,16 +677,38 @@ def model_fn_ltx2(
     # patchify
     b, c_v, f, h, w = video_latents.shape
     video_latents = video_patchifier.patchify(video_latents)
+    seq_len_video = video_latents.shape[1]
     video_timesteps = timestep.repeat(1, video_latents.shape[1], 1)
-    if denoise_mask_video is not None:
-        video_timesteps = video_patchifier.patchify(denoise_mask_video) * video_timesteps
+    # Frist frame conditioning by replacing the video latents
+    if input_latents_video is not None:
+        denoise_mask_video = video_patchifier.patchify(denoise_mask_video)
+        video_latents = video_latents * denoise_mask_video + video_patchifier.patchify(input_latents_video) * (1.0 - denoise_mask_video)
+        video_timesteps = denoise_mask_video * video_timesteps
+
+    # Reference conditioning by appending the reference video or frame latents
+    total_ref_latents = ref_frames_latents if ref_frames_latents is not None else []
+    total_ref_positions = ref_frames_positions if ref_frames_positions is not None else []
+    total_ref_latents += [in_context_video_latents] if in_context_video_latents is not None else []
+    total_ref_positions += [in_context_video_positions] if in_context_video_positions is not None else []
+    if len(total_ref_latents) > 0:
+        for ref_frames_latent, ref_frames_position in zip(total_ref_latents, total_ref_positions):
+            ref_frames_latent = video_patchifier.patchify(ref_frames_latent)
+            ref_frames_timestep = timestep.repeat(1, ref_frames_latent.shape[1], 1) * 0.
+            video_latents = torch.cat([video_latents, ref_frames_latent], dim=1)
+            video_positions = torch.cat([video_positions, ref_frames_position], dim=2)
+            video_timesteps = torch.cat([video_timesteps, ref_frames_timestep], dim=1)
+
     if audio_latents is not None:
         _, c_a, _, mel_bins  = audio_latents.shape
         audio_latents = audio_patchifier.patchify(audio_latents)
         audio_timesteps = timestep.repeat(1, audio_latents.shape[1], 1)
     else:
         audio_timesteps = None
-    #TODO: support gradient checkpointing in training
+    if input_latents_audio is not None:
+        denoise_mask_audio = audio_patchifier.patchify(denoise_mask_audio)
+        audio_latents = audio_latents * denoise_mask_audio + audio_patchifier.patchify(input_latents_audio) * (1.0 - denoise_mask_audio)
+        audio_timesteps = denoise_mask_audio * audio_timesteps
+
     vx, ax = dit(
         video_latents=video_latents,
         video_positions=video_positions,
@@ -577,7 +718,12 @@ def model_fn_ltx2(
         audio_positions=audio_positions,
         audio_context=audio_context,
         audio_timesteps=audio_timesteps,
+        sigma=timestep,
+        use_gradient_checkpointing=use_gradient_checkpointing,
+        use_gradient_checkpointing_offload=use_gradient_checkpointing_offload,
     )
+
+    vx = vx[:, :seq_len_video, ...]
     # unpatchify
     vx = video_patchifier.unpatchify_video(vx, f, h, w)
     ax = audio_patchifier.unpatchify_audio(ax, c_a, mel_bins) if ax is not None else None

diffsynth/pipelines/mova_audio_video.py

@@ -0,0 +1,460 @@
+import sys
+import torch, types
+from PIL import Image
+from typing import Optional, Union
+from einops import rearrange
+import numpy as np
+from PIL import Image
+from tqdm import tqdm
+from typing import Optional
+
+from ..core.device.npu_compatible_device import get_device_type
+from ..diffusion import FlowMatchScheduler
+from ..core import ModelConfig, gradient_checkpoint_forward
+from ..diffusion.base_pipeline import BasePipeline, PipelineUnit
+
+from ..models.wan_video_dit import WanModel, sinusoidal_embedding_1d, set_to_torch_norm
+from ..models.wan_video_text_encoder import WanTextEncoder, HuggingfaceTokenizer
+from ..models.wan_video_vae import WanVideoVAE
+from ..models.mova_audio_dit import MovaAudioDit
+from ..models.mova_audio_vae import DacVAE
+from ..models.mova_dual_tower_bridge import DualTowerConditionalBridge
+from ..utils.data.audio import convert_to_mono, resample_waveform
+
+
+class MovaAudioVideoPipeline(BasePipeline):
+
+    def __init__(self, device=get_device_type(), torch_dtype=torch.bfloat16):
+        super().__init__(
+            device=device, torch_dtype=torch_dtype,
+            height_division_factor=16, width_division_factor=16, time_division_factor=4, time_division_remainder=1
+        )
+        self.scheduler = FlowMatchScheduler("Wan")
+        self.tokenizer: HuggingfaceTokenizer = None
+        self.text_encoder: WanTextEncoder = None
+        self.video_dit: WanModel = None # high noise model
+        self.video_dit2: WanModel = None # low noise model
+        self.audio_dit: MovaAudioDit = None
+        self.dual_tower_bridge: DualTowerConditionalBridge = None
+        self.video_vae: WanVideoVAE = None
+        self.audio_vae: DacVAE = None
+
+        self.in_iteration_models = ("video_dit", "audio_dit", "dual_tower_bridge")
+        self.in_iteration_models_2 = ("video_dit2", "audio_dit", "dual_tower_bridge")
+
+        self.units = [
+            MovaAudioVideoUnit_ShapeChecker(),
+            MovaAudioVideoUnit_NoiseInitializer(),
+            MovaAudioVideoUnit_InputVideoEmbedder(),
+            MovaAudioVideoUnit_InputAudioEmbedder(),
+            MovaAudioVideoUnit_PromptEmbedder(),
+            MovaAudioVideoUnit_ImageEmbedderVAE(),
+            MovaAudioVideoUnit_UnifiedSequenceParallel(),
+        ]
+        self.model_fn = model_fn_mova_audio_video
+
+    def enable_usp(self):
+        from ..utils.xfuser import get_sequence_parallel_world_size, usp_attn_forward
+        for block in self.video_dit.blocks + self.audio_dit.blocks + self.video_dit2.blocks:
+            block.self_attn.forward = types.MethodType(usp_attn_forward, block.self_attn)
+        self.sp_size = get_sequence_parallel_world_size()
+        self.use_unified_sequence_parallel = True
+
+    @staticmethod
+    def from_pretrained(
+        torch_dtype: torch.dtype = torch.bfloat16,
+        device: Union[str, torch.device] = get_device_type(),
+        model_configs: list[ModelConfig] = [],
+        tokenizer_config: ModelConfig = ModelConfig(model_id="openmoss/MOVA-720p", origin_file_pattern="tokenizer/"),
+        use_usp: bool = False,
+        vram_limit: float = None,
+    ):
+        if use_usp:
+            from ..utils.xfuser import initialize_usp
+            initialize_usp(device)
+            import torch.distributed as dist
+            from ..core.device.npu_compatible_device import get_device_name
+            if dist.is_available() and dist.is_initialized():
+                device = get_device_name()
+        # Initialize pipeline
+        pipe = MovaAudioVideoPipeline(device=device, torch_dtype=torch_dtype)
+        model_pool = pipe.download_and_load_models(model_configs, vram_limit)
+
+        # Fetch models
+        pipe.text_encoder = model_pool.fetch_model("wan_video_text_encoder")
+        dit = model_pool.fetch_model("wan_video_dit", index=2)
+        if isinstance(dit, list):
+            pipe.video_dit, pipe.video_dit2 = dit
+        else:
+            pipe.video_dit = dit
+        pipe.audio_dit = model_pool.fetch_model("mova_audio_dit")
+        pipe.dual_tower_bridge = model_pool.fetch_model("mova_dual_tower_bridge")
+        pipe.video_vae = model_pool.fetch_model("wan_video_vae")
+        pipe.audio_vae = model_pool.fetch_model("mova_audio_vae")
+        set_to_torch_norm([pipe.video_dit, pipe.audio_dit, pipe.dual_tower_bridge] + ([pipe.video_dit2] if pipe.video_dit2 is not None else []))
+
+        # Size division factor
+        if pipe.video_vae is not None:
+            pipe.height_division_factor = pipe.video_vae.upsampling_factor * 2
+            pipe.width_division_factor = pipe.video_vae.upsampling_factor * 2
+
+        # Initialize tokenizer and processor
+        if tokenizer_config is not None:
+            tokenizer_config.download_if_necessary()
+            pipe.tokenizer = HuggingfaceTokenizer(name=tokenizer_config.path, seq_len=512, clean='whitespace')
+
+        # Unified Sequence Parallel
+        if use_usp: pipe.enable_usp()
+
+        # VRAM Management
+        pipe.vram_management_enabled = pipe.check_vram_management_state()
+        return pipe
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        # Prompt
+        prompt: str,
+        negative_prompt: Optional[str] = "",
+        # Image-to-video
+        input_image: Optional[Image.Image] = None,
+        # First-last-frame-to-video
+        end_image: Optional[Image.Image] = None,
+        # Video-to-video
+        denoising_strength: Optional[float] = 1.0,
+        # Randomness
+        seed: Optional[int] = None,
+        rand_device: Optional[str] = "cpu",
+        # Shape
+        height: Optional[int] = 352,
+        width: Optional[int] = 640,
+        num_frames: Optional[int] = 81,
+        frame_rate: Optional[int] = 24,
+        # Classifier-free guidance
+        cfg_scale: Optional[float] = 5.0,
+        # Boundary
+        switch_DiT_boundary: Optional[float] = 0.9,
+        # Scheduler
+        num_inference_steps: Optional[int] = 50,
+        sigma_shift: Optional[float] = 5.0,
+        # VAE tiling
+        tiled: Optional[bool] = True,
+        tile_size: Optional[tuple[int, int]] = (30, 52),
+        tile_stride: Optional[tuple[int, int]] = (15, 26),
+        # progress_bar
+        progress_bar_cmd=tqdm,
+    ):
+        # Scheduler
+        self.scheduler.set_timesteps(num_inference_steps, denoising_strength=denoising_strength, shift=sigma_shift)
+
+        # Inputs
+        inputs_posi = {
+            "prompt": prompt,
+        }
+        inputs_nega = {
+            "negative_prompt": negative_prompt,
+        }
+        inputs_shared = {
+            "input_image": input_image,
+            "end_image": end_image,
+            "denoising_strength": denoising_strength,
+            "seed": seed, "rand_device": rand_device,
+            "height": height, "width": width, "num_frames": num_frames, "frame_rate": frame_rate,
+            "cfg_scale": cfg_scale,
+            "sigma_shift": sigma_shift,
+            "tiled": tiled, "tile_size": tile_size, "tile_stride": tile_stride,
+        }
+        for unit in self.units:
+            inputs_shared, inputs_posi, inputs_nega = self.unit_runner(unit, self, inputs_shared, inputs_posi, inputs_nega)
+
+        # Denoise
+        self.load_models_to_device(self.in_iteration_models)
+        models = {name: getattr(self, name) for name in self.in_iteration_models}
+        for progress_id, timestep in enumerate(progress_bar_cmd(self.scheduler.timesteps)):
+            # Switch DiT if necessary
+            if timestep.item() < switch_DiT_boundary * 1000 and self.video_dit2 is not None and not models["video_dit"] is self.video_dit2:
+                self.load_models_to_device(self.in_iteration_models_2)
+                models["video_dit"] = self.video_dit2
+            # Timestep
+            timestep = timestep.unsqueeze(0).to(dtype=self.torch_dtype, device=self.device)
+            noise_pred_video, noise_pred_audio = self.cfg_guided_model_fn(
+                self.model_fn, cfg_scale, inputs_shared, inputs_posi, inputs_nega,
+                **models, timestep=timestep, progress_id=progress_id
+            )
+            # Scheduler
+            inputs_shared["video_latents"] = self.step(self.scheduler, inputs_shared["video_latents"], progress_id=progress_id, noise_pred=noise_pred_video, **inputs_shared)
+            inputs_shared["audio_latents"] = self.step(self.scheduler, inputs_shared["audio_latents"], progress_id=progress_id, noise_pred=noise_pred_audio, **inputs_shared)
+
+        # Decode
+        self.load_models_to_device(['video_vae'])
+        video = self.video_vae.decode(inputs_shared["video_latents"], device=self.device, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride)
+        video = self.vae_output_to_video(video)
+        self.load_models_to_device(["audio_vae"])
+        audio = self.audio_vae.decode(inputs_shared["audio_latents"])
+        audio = self.output_audio_format_check(audio)
+        self.load_models_to_device([])
+        return video, audio
+
+
+class MovaAudioVideoUnit_ShapeChecker(PipelineUnit):
+    def __init__(self):
+        super().__init__(
+            input_params=("height", "width", "num_frames"),
+            output_params=("height", "width", "num_frames"),
+        )
+
+    def process(self, pipe: MovaAudioVideoPipeline, height, width, num_frames):
+        height, width, num_frames = pipe.check_resize_height_width(height, width, num_frames)
+        return {"height": height, "width": width, "num_frames": num_frames}
+
+
+class MovaAudioVideoUnit_NoiseInitializer(PipelineUnit):
+    def __init__(self):
+        super().__init__(
+            input_params=("height", "width", "num_frames", "seed", "rand_device", "frame_rate"),
+            output_params=("video_noise", "audio_noise")
+        )
+
+    def process(self, pipe: MovaAudioVideoPipeline, height, width, num_frames, seed, rand_device, frame_rate):
+        length = (num_frames - 1) // 4 + 1
+        video_shape = (1, pipe.video_vae.model.z_dim, length, height // pipe.video_vae.upsampling_factor, width // pipe.video_vae.upsampling_factor)
+        video_noise = pipe.generate_noise(video_shape, seed=seed, rand_device=rand_device)
+
+        audio_num_samples = (int(pipe.audio_vae.sample_rate * num_frames / frame_rate) - 1) // int(pipe.audio_vae.hop_length) + 1
+        audio_shape = (1, pipe.audio_vae.latent_dim, audio_num_samples)
+        audio_noise = pipe.generate_noise(audio_shape, seed=seed, rand_device=rand_device)
+        return {"video_noise": video_noise, "audio_noise": audio_noise}
+
+
+class MovaAudioVideoUnit_InputVideoEmbedder(PipelineUnit):
+    def __init__(self):
+        super().__init__(
+            input_params=("input_video", "video_noise", "tiled", "tile_size", "tile_stride"),
+            output_params=("video_latents", "input_latents"),
+            onload_model_names=("video_vae",)
+        )
+
+    def process(self, pipe: MovaAudioVideoPipeline, input_video, video_noise, tiled, tile_size, tile_stride):
+        if input_video is None or not pipe.scheduler.training:
+            return {"video_latents": video_noise}
+        else:
+            pipe.load_models_to_device(self.onload_model_names)
+            input_video = pipe.preprocess_video(input_video)
+            input_latents = pipe.video_vae.encode(input_video, device=pipe.device, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride).to(dtype=pipe.torch_dtype, device=pipe.device)
+            return {"input_latents": input_latents}
+
+
+class MovaAudioVideoUnit_InputAudioEmbedder(PipelineUnit):
+    def __init__(self):
+        super().__init__(
+            input_params=("input_audio", "audio_noise"),
+            output_params=("audio_latents", "audio_input_latents"),
+            onload_model_names=("audio_vae",)
+        )
+
+    def process(self, pipe: MovaAudioVideoPipeline, input_audio, audio_noise):
+        if input_audio is None or not pipe.scheduler.training:
+            return {"audio_latents": audio_noise}
+        else:
+            pipe.load_models_to_device(self.onload_model_names)
+            input_audio, sample_rate = input_audio
+            input_audio = convert_to_mono(input_audio)
+            input_audio = resample_waveform(input_audio, sample_rate, pipe.audio_vae.sample_rate)
+            input_audio = pipe.audio_vae.preprocess(input_audio.unsqueeze(0), pipe.audio_vae.sample_rate)
+            z, _, _, _, _ = pipe.audio_vae.encode(input_audio)
+            return {"audio_input_latents": z.mode()}
+
+
+class MovaAudioVideoUnit_PromptEmbedder(PipelineUnit):
+    def __init__(self):
+        super().__init__(
+            seperate_cfg=True,
+            input_params_posi={"prompt": "prompt"},
+            input_params_nega={"prompt": "negative_prompt"},
+            output_params=("context",),
+            onload_model_names=("text_encoder",)
+        )
+
+    def encode_prompt(self, pipe: MovaAudioVideoPipeline, prompt):
+        ids, mask = pipe.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=512,
+            truncation=True,
+            add_special_tokens=True,
+            return_mask=True,
+            return_tensors="pt",
+        )
+        ids = ids.to(pipe.device)
+        mask = mask.to(pipe.device)
+        seq_lens = mask.gt(0).sum(dim=1).long()
+        prompt_emb = pipe.text_encoder(ids, mask)
+        for i, v in enumerate(seq_lens):
+            prompt_emb[:, v:] = 0
+        return prompt_emb
+
+    def process(self, pipe: MovaAudioVideoPipeline, prompt) -> dict:
+        pipe.load_models_to_device(self.onload_model_names)
+        prompt_emb = self.encode_prompt(pipe, prompt)
+        return {"context": prompt_emb}
+
+
+class MovaAudioVideoUnit_ImageEmbedderVAE(PipelineUnit):
+    def __init__(self):
+        super().__init__(
+            input_params=("input_image", "end_image", "num_frames", "height", "width", "tiled", "tile_size", "tile_stride"),
+            output_params=("y",),
+            onload_model_names=("video_vae",)
+        )
+
+    def process(self, pipe: MovaAudioVideoPipeline, input_image, end_image, num_frames, height, width, tiled, tile_size, tile_stride):
+        if input_image is None or not pipe.video_dit.require_vae_embedding:
+            return {}
+        pipe.load_models_to_device(self.onload_model_names)
+
+        image = pipe.preprocess_image(input_image.resize((width, height))).to(pipe.device)
+        msk = torch.ones(1, num_frames, height//8, width//8, device=pipe.device)
+        msk[:, 1:] = 0
+        if end_image is not None:
+            end_image = pipe.preprocess_image(end_image.resize((width, height))).to(pipe.device)
+            vae_input = torch.concat([image.transpose(0,1), torch.zeros(3, num_frames-2, height, width).to(image.device), end_image.transpose(0,1)],dim=1)
+            msk[:, -1:] = 1
+        else:
+            vae_input = torch.concat([image.transpose(0, 1), torch.zeros(3, num_frames-1, height, width).to(image.device)], dim=1)
+
+        msk = torch.concat([torch.repeat_interleave(msk[:, 0:1], repeats=4, dim=1), msk[:, 1:]], dim=1)
+        msk = msk.view(1, msk.shape[1] // 4, 4, height//8, width//8)
+        msk = msk.transpose(1, 2)[0]
+
+        y = pipe.video_vae.encode([vae_input.to(dtype=pipe.torch_dtype, device=pipe.device)], device=pipe.device, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride)[0]
+        y = y.to(dtype=pipe.torch_dtype, device=pipe.device)
+        y = torch.concat([msk, y])
+        y = y.unsqueeze(0)
+        y = y.to(dtype=pipe.torch_dtype, device=pipe.device)
+        return {"y": y}
+
+
+class MovaAudioVideoUnit_UnifiedSequenceParallel(PipelineUnit):
+    def __init__(self):
+        super().__init__(input_params=(), output_params=("use_unified_sequence_parallel",))
+
+    def process(self, pipe: MovaAudioVideoPipeline):
+        if hasattr(pipe, "use_unified_sequence_parallel") and pipe.use_unified_sequence_parallel:
+            return {"use_unified_sequence_parallel": True}
+        return {"use_unified_sequence_parallel": False}
+
+
+def model_fn_mova_audio_video(
+    video_dit: WanModel,
+    audio_dit: MovaAudioDit,
+    dual_tower_bridge: DualTowerConditionalBridge,
+    video_latents: torch.Tensor = None,
+    audio_latents: torch.Tensor = None,
+    timestep: torch.Tensor = None,
+    context: torch.Tensor = None,
+    y: Optional[torch.Tensor] = None,
+    frame_rate: Optional[int] = 24,
+    use_unified_sequence_parallel: bool = False,
+    use_gradient_checkpointing: bool = False,
+    use_gradient_checkpointing_offload: bool = False,
+    **kwargs,
+):
+    video_x, audio_x = video_latents, audio_latents
+    # First-Last Frame
+    if y is not None:
+        video_x = torch.cat([video_x, y], dim=1)
+
+    # Timestep
+    video_t = video_dit.time_embedding(sinusoidal_embedding_1d(video_dit.freq_dim, timestep))
+    video_t_mod = video_dit.time_projection(video_t).unflatten(1, (6, video_dit.dim))
+    audio_t = audio_dit.time_embedding(sinusoidal_embedding_1d(audio_dit.freq_dim, timestep))
+    audio_t_mod = audio_dit.time_projection(audio_t).unflatten(1, (6, audio_dit.dim))
+
+    # Context
+    video_context = video_dit.text_embedding(context)
+    audio_context = audio_dit.text_embedding(context)
+
+    # Patchify
+    video_x = video_dit.patch_embedding(video_x)
+    f_v, h, w = video_x.shape[2:]
+    video_x = rearrange(video_x, 'b c f h w -> b (f h w) c').contiguous()
+    seq_len_video = video_x.shape[1]
+
+    audio_x = audio_dit.patch_embedding(audio_x)
+    f_a = audio_x.shape[2]
+    audio_x = rearrange(audio_x, 'b c f -> b f c').contiguous()
+    seq_len_audio = audio_x.shape[1]
+
+    # Freqs
+    video_freqs = torch.cat([
+        video_dit.freqs[0][:f_v].view(f_v, 1, 1, -1).expand(f_v, h, w, -1),
+        video_dit.freqs[1][:h].view(1, h, 1, -1).expand(f_v, h, w, -1),
+        video_dit.freqs[2][:w].view(1, 1, w, -1).expand(f_v, h, w, -1)
+    ], dim=-1).reshape(f_v * h * w, 1, -1).to(video_x.device)
+    audio_freqs = torch.cat([
+        audio_dit.freqs[0][:f_a].view(f_a, -1).expand(f_a, -1),
+        audio_dit.freqs[1][:f_a].view(f_a, -1).expand(f_a, -1),
+        audio_dit.freqs[2][:f_a].view(f_a, -1).expand(f_a, -1),
+    ], dim=-1).reshape(f_a, 1, -1).to(audio_x.device)
+
+    video_rope, audio_rope = dual_tower_bridge.build_aligned_freqs(
+        video_fps=frame_rate,
+        grid_size=(f_v, h, w),
+        audio_steps=audio_x.shape[1],
+        device=video_x.device,
+        dtype=video_x.dtype,
+    )
+    # usp func
+    if use_unified_sequence_parallel:
+        from ..utils.xfuser import get_current_chunk, gather_all_chunks
+    else:
+        get_current_chunk = lambda x, dim=1: x
+        gather_all_chunks = lambda x, seq_len, dim=1: x
+    # Forward blocks
+    for block_id in range(len(audio_dit.blocks)):
+        if dual_tower_bridge.should_interact(block_id, "a2v"):
+            video_x, audio_x = dual_tower_bridge(
+                block_id,
+                video_x,
+                audio_x,
+                x_freqs=video_rope,
+                y_freqs=audio_rope,
+                condition_scale=1.0,
+                video_grid_size=(f_v, h, w),
+                use_gradient_checkpointing=use_gradient_checkpointing,
+                use_gradient_checkpointing_offload=use_gradient_checkpointing_offload,
+            )
+        video_x = get_current_chunk(video_x, dim=1)
+        video_x = gradient_checkpoint_forward(
+            video_dit.blocks[block_id],
+            use_gradient_checkpointing,
+            use_gradient_checkpointing_offload,
+            video_x, video_context, video_t_mod, video_freqs
+        )
+        video_x = gather_all_chunks(video_x, seq_len=seq_len_video, dim=1)
+        audio_x = get_current_chunk(audio_x, dim=1)
+        audio_x = gradient_checkpoint_forward(
+            audio_dit.blocks[block_id],
+            use_gradient_checkpointing,
+            use_gradient_checkpointing_offload,
+            audio_x, audio_context, audio_t_mod, audio_freqs
+        )
+        audio_x = gather_all_chunks(audio_x, seq_len=seq_len_audio, dim=1)
+
+    video_x = get_current_chunk(video_x, dim=1)
+    for block_id in range(len(audio_dit.blocks), len(video_dit.blocks)):
+        video_x = gradient_checkpoint_forward(
+            video_dit.blocks[block_id],
+            use_gradient_checkpointing,
+            use_gradient_checkpointing_offload,
+            video_x, video_context, video_t_mod, video_freqs
+        )
+    video_x = gather_all_chunks(video_x, seq_len=seq_len_video, dim=1)
+
+    # Head
+    video_x = video_dit.head(video_x, video_t)
+    video_x = video_dit.unpatchify(video_x, (f_v, h, w))
+
+    audio_x = audio_dit.head(audio_x, audio_t)
+    audio_x = audio_dit.unpatchify(audio_x, (f_a,))
+    return video_x, audio_x

diffsynth/pipelines/qwen_image.py

@@ -682,14 +682,16 @@ class QwenImageUnit_Image2LoRADecode(PipelineUnit):
 class QwenImageUnit_ContextImageEmbedder(PipelineUnit):
     def __init__(self):
         super().__init__(
-            input_params=("context_image", "height", "width", "tiled", "tile_size", "tile_stride"),
+            input_params=("context_image", "height", "width", "tiled", "tile_size", "tile_stride", "layer_input_image"),
             output_params=("context_latents",),
             onload_model_names=("vae",)
         )
 
-    def process(self, pipe: QwenImagePipeline, context_image, height, width, tiled, tile_size, tile_stride):
+    def process(self, pipe: QwenImagePipeline, context_image, height, width, tiled, tile_size, tile_stride, layer_input_image=None):
         if context_image is None:
             return {}
+        if layer_input_image is not None:
+            context_image = context_image.convert("RGBA")
         pipe.load_models_to_device(self.onload_model_names)
         context_image = pipe.preprocess_image(context_image.resize((width, height))).to(device=pipe.device, dtype=pipe.torch_dtype)
         context_latents = pipe.vae.encode(context_image, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride)

diffsynth/pipelines/z_image.py

@@ -299,7 +299,7 @@ class ZImageUnit_PromptEmbedder(PipelineUnit):
 
     def process(self, pipe: ZImagePipeline, prompt, edit_image):
         pipe.load_models_to_device(self.onload_model_names)
-        if hasattr(pipe, "dit") and pipe.dit.siglip_embedder is not None:
+        if hasattr(pipe, "dit") and pipe.dit is not None and pipe.dit.siglip_embedder is not None:
             # Z-Image-Turbo and Z-Image-Omni-Base use different prompt encoding methods.
             # We determine which encoding method to use based on the model architecture.
             # If you are using two-stage split training,

diffsynth/utils/data/__init__.py

@@ -116,7 +116,7 @@ class VideoData:
         if self.height is not None and self.width is not None:
             return self.height, self.width
         else:
-            height, width, _ = self.__getitem__(0).shape
+            width, height = self.__getitem__(0).size
             return height, width
 
     def __getitem__(self, item):

diffsynth/utils/data/audio.py

@@ -0,0 +1,108 @@
+import torch
+import torchaudio
+
+
+def convert_to_mono(audio_tensor: torch.Tensor) -> torch.Tensor:
+    """
+    Convert audio to mono by averaging channels.
+    Supports [C, T] or [B, C, T]. Output shape: [1, T] or [B, 1, T].
+    """
+    return audio_tensor.mean(dim=-2, keepdim=True)
+
+
+def convert_to_stereo(audio_tensor: torch.Tensor) -> torch.Tensor:
+    """
+    Convert audio to stereo.
+    Supports [C, T] or [B, C, T]. Duplicate mono, keep stereo.
+    """
+    if audio_tensor.size(-2) == 1:
+        return audio_tensor.repeat(1, 2, 1) if audio_tensor.dim() == 3 else audio_tensor.repeat(2, 1)
+    return audio_tensor
+
+
+def resample_waveform(waveform: torch.Tensor, source_rate: int, target_rate: int) -> torch.Tensor:
+    """Resample waveform to target sample rate if needed."""
+    if source_rate == target_rate:
+        return waveform
+    resampled = torchaudio.functional.resample(waveform, source_rate, target_rate)
+    return resampled.to(dtype=waveform.dtype)
+
+
+def read_audio_with_torchcodec(
+    path: str,
+    start_time: float = 0,
+    duration: float | None = None,
+) -> tuple[torch.Tensor, int]:
+    """
+    Read audio from file natively using torchcodec, with optional start time and duration.
+    
+    Args:
+        path (str): The file path to the audio file.
+        start_time (float, optional): The start time in seconds to read from. Defaults to 0.
+        duration (float | None, optional): The duration in seconds to read. If None, reads until the end. Defaults to None.
+        
+    Returns:
+        tuple[torch.Tensor, int]: A tuple containing the audio tensor and the sample rate.
+            The audio tensor shape is [C, T] where C is the number of channels and T is the number of audio frames.
+    """
+    from torchcodec.decoders import AudioDecoder
+    decoder = AudioDecoder(path)
+    stop_seconds = None if duration is None else start_time + duration
+    waveform = decoder.get_samples_played_in_range(start_seconds=start_time, stop_seconds=stop_seconds).data
+    return waveform, decoder.metadata.sample_rate
+
+
+def read_audio(
+    path: str,
+    start_time: float = 0,
+    duration: float | None = None,
+    resample: bool = False,
+    resample_rate: int = 48000,
+    backend: str = "torchcodec",
+) -> tuple[torch.Tensor, int]:
+    """
+    Read audio from file, with optional start time, duration, and resampling.
+    
+    Args:
+        path (str): The file path to the audio file.
+        start_time (float, optional): The start time in seconds to read from. Defaults to 0.
+        duration (float | None, optional): The duration in seconds to read. If None, reads until the end. Defaults to None.
+        resample (bool, optional): Whether to resample the audio to a different sample rate. Defaults to False.
+        resample_rate (int, optional): The target sample rate for resampling if resample is True. Defaults to 48000.
+        backend (str, optional): The audio backend to use for reading. Defaults to "torchcodec".
+        
+    Returns:
+        tuple[torch.Tensor, int]: A tuple containing the audio tensor and the sample rate.
+            The audio tensor shape is [C, T] where C is the number of channels and T is the number of audio frames.
+    """
+    if backend == "torchcodec":
+        waveform, sample_rate = read_audio_with_torchcodec(path, start_time, duration)
+    else:
+        raise ValueError(f"Unsupported audio backend: {backend}")
+
+    if resample:
+        waveform = resample_waveform(waveform, sample_rate, resample_rate)
+        sample_rate = resample_rate
+
+    return waveform, sample_rate
+
+
+def save_audio(waveform: torch.Tensor, sample_rate: int, save_path: str, backend: str = "torchcodec"):
+    """
+    Save audio tensor to file.
+    
+    Args:
+        waveform (torch.Tensor): The audio tensor to save. Shape can be [C, T] or [B, C, T].
+        sample_rate (int): The sample rate of the audio.
+        save_path (str): The file path to save the audio to.
+        backend (str, optional): The audio backend to use for saving. Defaults to "torchcodec".
+    """
+    if waveform.dim() == 3:
+        waveform = waveform[0]
+
+    if backend == "torchcodec":
+        from torchcodec.encoders import AudioEncoder
+        encoder = AudioEncoder(waveform, sample_rate=sample_rate)
+        encoder.to_file(dest=save_path)
+    else:
+        raise ValueError(f"Unsupported audio backend: {backend}")

diffsynth/utils/data/audio_video.py

@@ -0,0 +1,134 @@
+import av
+from fractions import Fraction
+import torch
+from PIL import Image
+from tqdm import tqdm
+from .audio import convert_to_stereo
+
+
+def _resample_audio(
+    container: av.container.Container, audio_stream: av.audio.AudioStream, frame_in: av.AudioFrame
+) -> None:
+    cc = audio_stream.codec_context
+
+    # Use the encoder's format/layout/rate as the *target*
+    target_format = cc.format or "fltp"  # AAC → usually fltp
+    target_layout = cc.layout or "stereo"
+    target_rate = cc.sample_rate or frame_in.sample_rate
+
+    audio_resampler = av.audio.resampler.AudioResampler(
+        format=target_format,
+        layout=target_layout,
+        rate=target_rate,
+    )
+
+    audio_next_pts = 0
+    for rframe in audio_resampler.resample(frame_in):
+        if rframe.pts is None:
+            rframe.pts = audio_next_pts
+        audio_next_pts += rframe.samples
+        rframe.sample_rate = frame_in.sample_rate
+        container.mux(audio_stream.encode(rframe))
+
+    # flush audio encoder
+    for packet in audio_stream.encode():
+        container.mux(packet)
+
+
+def _write_audio(
+    container: av.container.Container, audio_stream: av.audio.AudioStream, samples: torch.Tensor, audio_sample_rate: int
+) -> None:
+    if samples.ndim == 1:
+        samples = samples.unsqueeze(0)
+    samples = convert_to_stereo(samples)
+    assert samples.ndim == 2 and samples.shape[0] == 2, "audio samples must be [C, S] or [S], C must be 1 or 2"
+    samples = samples.T
+    # Convert to int16 packed for ingestion; resampler converts to encoder fmt.
+    if samples.dtype != torch.int16:
+        samples = torch.clip(samples, -1.0, 1.0)
+        samples = (samples * 32767.0).to(torch.int16)
+
+    frame_in = av.AudioFrame.from_ndarray(
+        samples.contiguous().reshape(1, -1).cpu().numpy(),
+        format="s16",
+        layout="stereo",
+    )
+    frame_in.sample_rate = audio_sample_rate
+
+    _resample_audio(container, audio_stream, frame_in)
+
+
+def _prepare_audio_stream(container: av.container.Container, audio_sample_rate: int) -> av.audio.AudioStream:
+    """
+    Prepare the audio stream for writing.
+    """
+    audio_stream = container.add_stream("aac")
+    supported_sample_rates = audio_stream.codec_context.codec.audio_rates
+    if supported_sample_rates:
+        best_rate = min(supported_sample_rates, key=lambda x: abs(x - audio_sample_rate))
+        if best_rate != audio_sample_rate:
+            print(f"Using closest supported audio sample rate: {best_rate}")
+    else:
+        best_rate = audio_sample_rate
+    audio_stream.codec_context.sample_rate = best_rate
+    audio_stream.codec_context.layout = "stereo"
+    audio_stream.codec_context.time_base = Fraction(1, best_rate)
+    return audio_stream
+
+
+def write_video_audio(
+    video: list[Image.Image],
+    audio: torch.Tensor | None,
+    output_path: str,
+    fps: int = 24,
+    audio_sample_rate: int | None = None,
+) -> None:
+    """
+    Writes a sequence of images and an audio tensor to a video file.
+
+    This function utilizes PyAV (or a similar multimedia library) to encode a list of PIL images into a video stream
+    and multiplex a PyTorch tensor as the audio stream into the output container.
+
+    Args:
+        video (list[Image.Image]): A list of PIL Image objects representing the video frames. 
+            The length of this list determines the total duration of the video based on the FPS.
+        audio (torch.Tensor | None): The audio data as a PyTorch tensor.
+            The shape is typically (channels, samples). If no audio is required, pass None.
+            channels can be 1 or 2. 1 for mono, 2 for stereo.
+        output_path (str): The file path (including extension) where the output video will be saved.
+        fps (int, optional): The frame rate (frames per second) for the video. Defaults to 24.
+        audio_sample_rate (int | None, optional): The sample rate (e.g., 44100, 48000) for the audio.
+            If the audio tensor is provided and this is None, the function attempts to infer the rate 
+            based on the audio tensor's length and the video duration.
+    Raises:
+        ValueError: If an audio tensor is provided but the sample rate cannot be determined.
+    """
+    duration = len(video) / fps
+    if audio_sample_rate is None:
+        audio_sample_rate = int(audio.shape[-1] / duration)
+
+    width, height = video[0].size
+    container = av.open(output_path, mode="w")
+    stream = container.add_stream("libx264", rate=int(fps))
+    stream.width = width
+    stream.height = height
+    stream.pix_fmt = "yuv420p"
+
+    if audio is not None:
+        if audio_sample_rate is None:
+            raise ValueError("audio_sample_rate is required when audio is provided")
+        audio_stream = _prepare_audio_stream(container, audio_sample_rate)
+
+    for frame in tqdm(video, total=len(video)):
+        frame = av.VideoFrame.from_image(frame)
+        for packet in stream.encode(frame):
+            container.mux(packet)
+
+    # Flush encoder
+    for packet in stream.encode():
+        container.mux(packet)
+
+    if audio is not None:
+        _write_audio(container, audio_stream, audio, audio_sample_rate)
+
+    container.close()

diffsynth/utils/data/media_io_ltx2.py

@@ -1,113 +1,7 @@
-
-from fractions import Fraction
-import torch
 import av
-from tqdm import tqdm
-from PIL import Image
 import numpy as np
 from io import BytesIO
-from collections.abc import Generator, Iterator
-
-
-def _resample_audio(
-    container: av.container.Container, audio_stream: av.audio.AudioStream, frame_in: av.AudioFrame
-) -> None:
-    cc = audio_stream.codec_context
-
-    # Use the encoder's format/layout/rate as the *target*
-    target_format = cc.format or "fltp"  # AAC → usually fltp
-    target_layout = cc.layout or "stereo"
-    target_rate = cc.sample_rate or frame_in.sample_rate
-
-    audio_resampler = av.audio.resampler.AudioResampler(
-        format=target_format,
-        layout=target_layout,
-        rate=target_rate,
-    )
-
-    audio_next_pts = 0
-    for rframe in audio_resampler.resample(frame_in):
-        if rframe.pts is None:
-            rframe.pts = audio_next_pts
-        audio_next_pts += rframe.samples
-        rframe.sample_rate = frame_in.sample_rate
-        container.mux(audio_stream.encode(rframe))
-
-    # flush audio encoder
-    for packet in audio_stream.encode():
-        container.mux(packet)
-
-
-def _write_audio(
-    container: av.container.Container, audio_stream: av.audio.AudioStream, samples: torch.Tensor, audio_sample_rate: int
-) -> None:
-    if samples.ndim == 1:
-        samples = samples[:, None]
-
-    if samples.shape[1] != 2 and samples.shape[0] == 2:
-        samples = samples.T
-
-    if samples.shape[1] != 2:
-        raise ValueError(f"Expected samples with 2 channels; got shape {samples.shape}.")
-
-    # Convert to int16 packed for ingestion; resampler converts to encoder fmt.
-    if samples.dtype != torch.int16:
-        samples = torch.clip(samples, -1.0, 1.0)
-        samples = (samples * 32767.0).to(torch.int16)
-
-    frame_in = av.AudioFrame.from_ndarray(
-        samples.contiguous().reshape(1, -1).cpu().numpy(),
-        format="s16",
-        layout="stereo",
-    )
-    frame_in.sample_rate = audio_sample_rate
-
-    _resample_audio(container, audio_stream, frame_in)
-
-
-def _prepare_audio_stream(container: av.container.Container, audio_sample_rate: int) -> av.audio.AudioStream:
-    """
-    Prepare the audio stream for writing.
-    """
-    audio_stream = container.add_stream("aac", rate=audio_sample_rate)
-    audio_stream.codec_context.sample_rate = audio_sample_rate
-    audio_stream.codec_context.layout = "stereo"
-    audio_stream.codec_context.time_base = Fraction(1, audio_sample_rate)
-    return audio_stream
-
-def write_video_audio_ltx2(
-    video: list[Image.Image],
-    audio: torch.Tensor | None,
-    output_path: str,
-    fps: int = 24,
-    audio_sample_rate: int | None = 24000,
-) -> None:
-
-    width, height = video[0].size
-    container = av.open(output_path, mode="w")
-    stream = container.add_stream("libx264", rate=int(fps))
-    stream.width = width
-    stream.height = height
-    stream.pix_fmt = "yuv420p"
-    
-    if audio is not None:
-        if audio_sample_rate is None:
-            raise ValueError("audio_sample_rate is required when audio is provided")
-        audio_stream = _prepare_audio_stream(container, audio_sample_rate)
-
-    for frame in tqdm(video, total=len(video)):
-        frame = av.VideoFrame.from_image(frame)
-        for packet in stream.encode(frame):
-            container.mux(packet)
-
-    # Flush encoder
-    for packet in stream.encode():
-        container.mux(packet)
-
-    if audio is not None:
-        _write_audio(container, audio_stream, audio, audio_sample_rate)
-
-    container.close()
+from .audio_video import write_video_audio as write_video_audio_ltx2
 
 
 def encode_single_frame(output_file: str, image_array: np.ndarray, crf: float) -> None:

diffsynth/utils/lora/general.py

@@ -1,4 +1,4 @@
-import torch
+import torch, warnings
 
 
 class GeneralLoRALoader:
@@ -26,7 +26,11 @@ class GeneralLoRALoader:
                 keys.pop(0)
             keys.pop(-1)
             target_name = ".".join(keys)
-            lora_name_dict[target_name] = (key, key.replace(lora_B_key, lora_A_key))
+            # Alpha: Deprecated but retained for compatibility.
+            key_alpha = key.replace(lora_B_key + ".weight", "alpha").replace(lora_B_key + ".default.weight", "alpha")
+            if key_alpha == key or key_alpha not in lora_state_dict:
+                key_alpha = None
+            lora_name_dict[target_name] = (key, key.replace(lora_B_key, lora_A_key), key_alpha)
         return lora_name_dict
     
     
@@ -36,6 +40,10 @@ class GeneralLoRALoader:
         for name in name_dict:
             weight_up = state_dict[name_dict[name][0]]
             weight_down = state_dict[name_dict[name][1]]
+            if name_dict[name][2] is not None:
+                warnings.warn("Alpha detected in the LoRA file. This may be a LoRA model not trained by DiffSynth-Studio. To ensure compatibility, the LoRA weights will be converted to weight * alpha / rank.")
+                alpha = state_dict[name_dict[name][2]] / weight_down.shape[0]
+                weight_down = weight_down * alpha
             state_dict_[name + f".lora_B{suffix}"] = weight_up
             state_dict_[name + f".lora_A{suffix}"] = weight_down
         return state_dict_

diffsynth/utils/ses/README.md

@@ -0,0 +1 @@
+Please see `docs/en/Research_Tutorial/inference_time_scaling.md` or `docs/zh/Research_Tutorial/inference_time_scaling.md` for more details.

diffsynth/utils/ses/__init__.py

@@ -0,0 +1 @@
+from .ses import ses_search

diffsynth/utils/ses/ses.py

@@ -0,0 +1,117 @@
+import torch
+import pywt
+import numpy as np
+from tqdm import tqdm
+
+
+def split_dwt(z_tensor_cpu, wavelet_name, dwt_level):
+    all_clow_np = []
+    all_chigh_list = []
+    z_tensor_cpu = z_tensor_cpu.cpu().float()
+    
+    for i in range(z_tensor_cpu.shape[0]): 
+        z_numpy_ch = z_tensor_cpu[i].numpy()
+        
+        coeffs_ch = pywt.wavedec2(z_numpy_ch, wavelet_name, level=dwt_level, mode='symmetric', axes=(-2, -1))
+        
+        clow_np = coeffs_ch[0]
+        chigh_list = coeffs_ch[1:]
+        
+        all_clow_np.append(clow_np)
+        all_chigh_list.append(chigh_list)
+        
+    all_clow_tensor = torch.from_numpy(np.stack(all_clow_np, axis=0))
+    return all_clow_tensor, all_chigh_list
+
+
+def reconstruct_dwt(c_low_tensor_cpu, c_high_coeffs, wavelet_name, original_shape):
+    H_high, W_high = original_shape
+    c_low_tensor_cpu = c_low_tensor_cpu.cpu().float()
+    
+    clow_np = c_low_tensor_cpu.numpy()
+    
+    if clow_np.ndim == 4 and clow_np.shape[0] == 1:
+        clow_np = clow_np[0]
+
+    coeffs_combined = [clow_np] + c_high_coeffs
+    z_recon_np = pywt.waverec2(coeffs_combined, wavelet_name, mode='symmetric', axes=(-2, -1))
+    if z_recon_np.shape[-2] != H_high or z_recon_np.shape[-1] != W_high:
+        z_recon_np = z_recon_np[..., :H_high, :W_high]
+    z_recon_tensor = torch.from_numpy(z_recon_np)
+    if z_recon_tensor.ndim == 3:
+        z_recon_tensor = z_recon_tensor.unsqueeze(0)
+    return z_recon_tensor
+
+
+def ses_search(
+    base_latents,
+    objective_reward_fn,
+    total_eval_budget=30,
+    popsize=10,
+    k_elites=5,
+    wavelet_name="db1",
+    dwt_level=4,
+):
+    latent_h, latent_w = base_latents.shape[-2], base_latents.shape[-1]
+    c_low_init, c_high_fixed_batch = split_dwt(base_latents, wavelet_name, dwt_level)
+    c_high_fixed = c_high_fixed_batch[0]    
+    c_low_shape = c_low_init.shape[1:]
+    mu = torch.zeros_like(c_low_init.view(-1).cpu()) 
+    sigma_sq = torch.ones_like(mu) * 1.0 
+    
+    best_overall = {"fitness": -float('inf'), "score": -float('inf'), "c_low": c_low_init[0]}
+    eval_count = 0
+    
+    elite_db = []    
+    n_generations = (total_eval_budget // popsize) + 5
+    pbar = tqdm(total=total_eval_budget, desc="[SES] Searching", unit="img")
+
+    for gen in range(n_generations):
+        if eval_count >= total_eval_budget: break
+        
+        std = torch.sqrt(torch.clamp(sigma_sq, min=1e-9))
+        z_noise = torch.randn(popsize, mu.shape[0])
+        samples_flat = mu + z_noise * std
+        samples_reshaped = samples_flat.view(popsize, *c_low_shape) 
+        
+        batch_results = []
+        
+        for i in range(popsize):
+            if eval_count >= total_eval_budget: break
+            
+            c_low_sample = samples_reshaped[i].unsqueeze(0) 
+            z_recon = reconstruct_dwt(c_low_sample, c_high_fixed, wavelet_name, (latent_h, latent_w))
+            z_recon = z_recon.to(base_latents.device, dtype=base_latents.dtype)  
+            # img = pipeline_callback(z_recon)
+
+            # score = scorer.get_score(img, prompt)
+            score = objective_reward_fn(z_recon)
+            res = {
+                "score": score, 
+                "c_low": c_low_sample.cpu()
+            }
+            batch_results.append(res)
+            if score > best_overall['score']:
+                best_overall = res
+                
+            eval_count += 1
+            pbar.update(1)
+            
+        if not batch_results: break
+        elite_db.extend(batch_results)        
+        elite_db.sort(key=lambda x: x['score'], reverse=True)        
+        elite_db = elite_db[:k_elites]        
+        elites_flat = torch.stack([x['c_low'].view(-1) for x in elite_db])
+        mu_new = torch.mean(elites_flat, dim=0)
+        
+        if len(elite_db) > 1:
+            sigma_sq_new = torch.var(elites_flat, dim=0, unbiased=True) + 1e-7
+        else:
+            sigma_sq_new = sigma_sq
+        mu = mu_new
+        sigma_sq = sigma_sq_new
+    pbar.close()
+    best_c_low = best_overall['c_low']
+    final_latents = reconstruct_dwt(best_c_low, c_high_fixed, wavelet_name, (latent_h, latent_w))
+    
+    return final_latents.to(base_latents.device, dtype=base_latents.dtype)

diffsynth/utils/state_dict_converters/anima_dit.py

@@ -0,0 +1,6 @@
+def AnimaDiTStateDictConverter(state_dict):
+    new_state_dict = {}
+    for key in state_dict:
+        value = state_dict[key]
+        new_state_dict[key.replace("net.", "")] = value
+    return new_state_dict

diffsynth/utils/state_dict_converters/ltx2_audio_vae.py

@@ -27,6 +27,6 @@ def LTX2VocoderStateDictConverter(state_dict):
     state_dict_ = {}
     for name in state_dict:
         if name.startswith("vocoder."):
-            new_name = name.replace("vocoder.", "")
+            new_name = name[len("vocoder."):]
             state_dict_[new_name] = state_dict[name]
     return state_dict_

diffsynth/utils/state_dict_converters/ltx2_video_vae.py

@@ -6,7 +6,8 @@ def LTX2VideoEncoderStateDictConverter(state_dict):
             state_dict_[new_name] = state_dict[name]
         elif name.startswith("vae.per_channel_statistics."):
             new_name = name.replace("vae.per_channel_statistics.", "per_channel_statistics.")
-            state_dict_[new_name] = state_dict[name]
+            if new_name not in ["per_channel_statistics.channel", "per_channel_statistics.mean-of-stds", "per_channel_statistics.mean-of-stds_over_std-of-means"]:
+                state_dict_[new_name] = state_dict[name]
     return state_dict_
 
 
@@ -18,5 +19,6 @@ def LTX2VideoDecoderStateDictConverter(state_dict):
             state_dict_[new_name] = state_dict[name]
         elif name.startswith("vae.per_channel_statistics."):
             new_name = name.replace("vae.per_channel_statistics.", "per_channel_statistics.")
-            state_dict_[new_name] = state_dict[name]
-    return state_dict_
+            if new_name not in ["per_channel_statistics.channel", "per_channel_statistics.mean-of-stds", "per_channel_statistics.mean-of-stds_over_std-of-means"]:
+                state_dict_[new_name] = state_dict[name]
+    return state_dict_

diffsynth/utils/xfuser/__init__.py

@@ -1 +1 @@
-from .xdit_context_parallel import usp_attn_forward, usp_dit_forward, get_sequence_parallel_world_size, initialize_usp
+from .xdit_context_parallel import usp_attn_forward, usp_dit_forward, get_sequence_parallel_world_size, initialize_usp, get_current_chunk, gather_all_chunks

diffsynth/utils/xfuser/xdit_context_parallel.py

@@ -143,4 +143,31 @@ def usp_attn_forward(self, x, freqs):
 
     del q, k, v
     getattr(torch, parse_device_type(x.device)).empty_cache()
-    return self.o(x)
+    return self.o(x)
+
+
+def get_current_chunk(x, dim=1):
+    chunks = torch.chunk(x, get_sequence_parallel_world_size(), dim=dim)
+    ndims = len(chunks[0].shape)
+    pad_list = [0] * (2 * ndims)
+    pad_end_index = 2 * (ndims - 1 - dim) + 1
+    max_size = chunks[0].size(dim)
+    chunks = [
+        torch.nn.functional.pad(
+            chunk, 
+            tuple(pad_list[:pad_end_index] + [max_size - chunk.size(dim)] + pad_list[pad_end_index+1:]), 
+            value=0
+        ) 
+        for chunk in chunks
+    ]
+    x = chunks[get_sequence_parallel_rank()]
+    return x
+
+
+def gather_all_chunks(x, seq_len=None, dim=1):
+    x = get_sp_group().all_gather(x, dim=dim)
+    if seq_len is not None:
+        slices = [slice(None)] * x.ndim
+        slices[dim] = slice(0, seq_len)
+        x = x[tuple(slices)]
+    return x

docs/en/Model_Details/Anima.md

@@ -0,0 +1,139 @@
+# Anima
+
+Anima is an image generation model trained and open-sourced by CircleStone Labs and Comfy Org.
+
+## Installation
+
+Before using this project for model inference and training, please install DiffSynth-Studio first.
+
+```shell
+git clone https://github.com/modelscope/DiffSynth-Studio.git
+cd DiffSynth-Studio
+pip install -e .
+```
+
+For more installation information, please refer to [Install Dependencies](../Pipeline_Usage/Setup.md).
+
+## Quick Start
+
+The following code demonstrates how to quickly load the [circlestone-labs/Anima](https://www.modelscope.cn/models/circlestone-labs/Anima) model for inference. VRAM management is enabled by default, allowing the framework to automatically control model parameter loading based on available VRAM. Minimum 8GB VRAM required.
+
+```python
+from diffsynth.pipelines.anima_image import AnimaImagePipeline, ModelConfig
+import torch
+
+vram_config = {
+    "offload_dtype": "disk",
+    "offload_device": "disk",
+    "onload_dtype": "disk",
+    "onload_device": "disk",
+    "preparing_dtype": torch.bfloat16,
+    "preparing_device": "cuda",
+    "computation_dtype": torch.bfloat16,
+    "computation_device": "cuda",
+}
+pipe = AnimaImagePipeline.from_pretrained(
+    torch_dtype=torch.bfloat16,
+    device="cuda",
+    model_configs=[
+        ModelConfig(model_id="circlestone-labs/Anima", origin_file_pattern="split_files/diffusion_models/anima-preview.safetensors", **vram_config),
+        ModelConfig(model_id="circlestone-labs/Anima", origin_file_pattern="split_files/text_encoders/qwen_3_06b_base.safetensors", **vram_config),
+        ModelConfig(model_id="circlestone-labs/Anima", origin_file_pattern="split_files/vae/qwen_image_vae.safetensors", **vram_config),
+    ],
+    tokenizer_config=ModelConfig(model_id="Qwen/Qwen3-0.6B", origin_file_pattern="./"),
+    tokenizer_t5xxl_config=ModelConfig(model_id="stabilityai/stable-diffusion-3.5-large", origin_file_pattern="tokenizer_3/"),
+    vram_limit=torch.cuda.mem_get_info("cuda")[1] / (1024 ** 3) - 0.5,
+)
+prompt = "Masterpiece, best quality, solo, long hair, wavy hair, silver hair, blue eyes, blue dress, medium breasts, dress, underwater, air bubble, floating hair, refraction, portrait."
+negative_prompt = "worst quality, low quality, monochrome, zombie, interlocked fingers, Aissist, cleavage, nsfw,"
+image = pipe(prompt, seed=0, num_inference_steps=50)
+image.save("image.jpg")
+```
+
+## Model Overview
+
+|Model ID|Inference|Low VRAM Inference|Full Training|Validation after Full Training|LoRA Training|Validation after LoRA Training|
+|-|-|-|-|-|-|-|
+|[circlestone-labs/Anima](https://www.modelscope.cn/models/circlestone-labs/Anima)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/anima/model_inference/anima-preview.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/anima/model_inference_low_vram/anima-preview.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/anima/model_training/full/anima-preview.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/anima/model_training/validate_full/anima-preview.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/anima/model_training/lora/anima-preview.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/anima/model_training/validate_lora/anima-preview.py)|
+
+Special training scripts:
+
+* Differential LoRA Training: [doc](../Training/Differential_LoRA.md)
+* FP8 Precision Training: [doc](../Training/FP8_Precision.md)
+* Two-Stage Split Training: [doc](../Training/Split_Training.md)
+* End-to-End Direct Distillation: [doc](../Training/Direct_Distill.md)
+
+## Model Inference
+
+Models are loaded through `AnimaImagePipeline.from_pretrained`, see [Model Inference](../Pipeline_Usage/Model_Inference.md#loading-models) for details.
+
+Input parameters for `AnimaImagePipeline` inference include:
+
+* `prompt`: Text description of the desired image content.
+* `negative_prompt`: Content to exclude from the generated image (default: `""`).
+* `cfg_scale`: Classifier-free guidance parameter (default: 4.0).
+* `input_image`: Input image for image-to-image generation (default: `None`).
+* `denoising_strength`: Controls similarity to input image (default: 1.0).
+* `height`: Image height (must be multiple of 16, default: 1024).
+* `width`: Image width (must be multiple of 16, default: 1024).
+* `seed`: Random seed (default: `None`).
+* `rand_device`: Device for random noise generation (default: `"cpu"`).
+* `num_inference_steps`: Inference steps (default: 30).
+* `sigma_shift`: Scheduler sigma offset (default: `None`).
+* `progress_bar_cmd`: Progress bar implementation (default: `tqdm.tqdm`).
+
+For VRAM constraints, enable [VRAM Management](../Pipeline_Usage/VRAM_management.md). Recommended low-VRAM configurations are provided in the "Model Overview" table above.
+
+## Model Training
+
+Anima models are trained through [`examples/anima/model_training/train.py`](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/anima/model_training/train.py) with parameters including:
+
+* General Training Parameters
+    * Dataset Configuration
+        * `--dataset_base_path`: Dataset root directory.
+        * `--dataset_metadata_path`: Metadata file path.
+        * `--dataset_repeat`: Dataset repetition per epoch.
+        * `--dataset_num_workers`: Dataloader worker count.
+        * `--data_file_keys`: Metadata fields to load (comma-separated).
+    * Model Loading
+        * `--model_paths`: Model paths (JSON format).
+        * `--model_id_with_origin_paths`: Model IDs with origin paths (e.g., `"anima-team/anima-1B:text_encoder/*.safetensors"`).
+        * `--extra_inputs`: Additional pipeline inputs (e.g., `controlnet_inputs` for ControlNet).
+        * `--fp8_models`: FP8-formatted models (same format as `--model_paths`).
+    * Training Configuration
+        * `--learning_rate`: Learning rate.
+        * `--num_epochs`: Training epochs.
+        * `--trainable_models`: Trainable components (e.g., `dit`, `vae`, `text_encoder`).
+        * `--find_unused_parameters`: Handle unused parameters in DDP training.
+        * `--weight_decay`: Weight decay value.
+        * `--task`: Training task (default: `sft`).
+    * Output Configuration
+        * `--output_path`: Model output directory.
+        * `--remove_prefix_in_ckpt`: Remove state dict prefixes.
+        * `--save_steps`: Model saving interval.
+    * LoRA Configuration
+        * `--lora_base_model`: Target model for LoRA.
+        * `--lora_target_modules`: Target modules for LoRA.
+        * `--lora_rank`: LoRA rank.
+        * `--lora_checkpoint`: LoRA checkpoint path.
+        * `--preset_lora_path`: Preloaded LoRA checkpoint path.
+        * `--preset_lora_model`: Model to merge LoRA with (e.g., `dit`).
+    * Gradient Configuration
+        * `--use_gradient_checkpointing`: Enable gradient checkpointing.
+        * `--use_gradient_checkpointing_offload`: Offload checkpointing to CPU.
+        * `--gradient_accumulation_steps`: Gradient accumulation steps.
+    * Image Resolution
+        * `--height`: Image height (empty for dynamic resolution).
+        * `--width`: Image width (empty for dynamic resolution).
+        * `--max_pixels`: Maximum pixel area for dynamic resolution.
+* Anima-Specific Parameters
+    * `--tokenizer_path`: Tokenizer path for text-to-image models.
+    * `--tokenizer_t5xxl_path`: T5-XXL tokenizer path.
+
+We provide a sample image dataset for testing:
+
+```shell
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --local_dir ./data/diffsynth_example_dataset
+```
+
+For training script details, refer to [Model Training](../Pipeline_Usage/Model_Training.md). For advanced training techniques, see [Training Framework Documentation](https://github.com/modelscope/DiffSynth-Studio/tree/main/docs/zh/Training/).

docs/en/Model_Details/FLUX.md

@@ -195,7 +195,7 @@ FLUX series models are uniformly trained through [`examples/flux/model_training/
 We have built a sample image dataset for your testing. You can download this dataset with the following command:
 
 ```shell
-modelscope download --dataset DiffSynth-Studio/example_image_dataset --local_dir ./data/example_image_dataset
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --local_dir ./data/diffsynth_example_dataset
 ```
 
 We have written recommended training scripts for each model, please refer to the table in the "Model Overview" section above. For how to write model training scripts, please refer to [Model Training](../Pipeline_Usage/Model_Training.md); for more advanced training algorithms, please refer to [Training Framework Detailed Explanation](https://github.com/modelscope/DiffSynth-Studio/tree/main/docs/en/Training/).

docs/en/Model_Details/FLUX2.md

@@ -145,7 +145,7 @@ FLUX.2 series models are uniformly trained through [`examples/flux2/model_traini
 We have built a sample image dataset for your testing. You can download this dataset with the following command:
 
 ```shell
-modelscope download --dataset DiffSynth-Studio/example_image_dataset --local_dir ./data/example_image_dataset
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --local_dir ./data/diffsynth_example_dataset
 ```
 
 We have written recommended training scripts for each model, please refer to the table in the "Model Overview" section above. For how to write model training scripts, please refer to [Model Training](../Pipeline_Usage/Model_Training.md); for more advanced training algorithms, please refer to [Training Framework Detailed Explanation](https://github.com/modelscope/DiffSynth-Studio/tree/main/docs/en/Training/).

docs/en/Model_Details/LTX-2.md

@@ -111,7 +111,19 @@ write_video_audio_ltx2(
 ## Model Overview
 |Model ID|Additional Parameters|Inference|Low VRAM Inference|Full Training|Validation After Full Training|LoRA Training|Validation After LoRA Training|
 |-|-|-|-|-|-|-|-|
+|[Lightricks/LTX-2.3: OneStagePipeline-I2AV](https://www.modelscope.cn/models/Lightricks/LTX-2.3)|`input_images`|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2.3-I2AV-OneStage.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2.3-I2AV-OneStage.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/full/LTX-2.3-I2AV-splited.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/validate_full/LTX-2.3-I2AV.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/lora/LTX-2.3-I2AV-splited.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/validate_lora/LTX-2.3-I2AV.py)|
+|[Lightricks/LTX-2.3: TwoStagePipeline-I2AV](https://www.modelscope.cn/models/Lightricks/LTX-2.3)|`input_images`|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2.3-I2AV-TwoStage.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2.3-I2AV-TwoStage.py)|-|-|-|-|
+|[Lightricks/LTX-2.3: DistilledPipeline-I2AV](https://www.modelscope.cn/models/Lightricks/LTX-2.3)|`input_images`|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2.3-I2AV-DistilledPipeline.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2.3-I2AV-DistilledPipeline.py)|-|-|-|-|
+|[Lightricks/LTX-2.3: OneStagePipeline-T2AV](https://www.modelscope.cn/models/Lightricks/LTX-2.3)||[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2.3-T2AV-OneStage.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2.3-T2AV-OneStage.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/full/LTX-2.3-T2AV-splited.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/validate_full/LTX-2.3-T2AV.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/lora/LTX-2.3-T2AV-splited.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/validate_lora/LTX-2.3-T2AV.py)|
+|[Lightricks/LTX-2.3: TwoStagePipeline-T2AV](https://www.modelscope.cn/models/Lightricks/LTX-2.3)||[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2.3-T2AV-TwoStage.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2.3-T2AV-TwoStage.py)|-|-|-|-|
+|[Lightricks/LTX-2.3: DistilledPipeline-T2AV](https://www.modelscope.cn/models/Lightricks/LTX-2.3)||[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2.3-T2AV-DistilledPipeline.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2.3-T2AV-DistilledPipeline.py)|-|-|-|-|
+|[Lightricks/LTX-2.3: A2V](https://www.modelscope.cn/models/Lightricks/LTX-2.3)|`retake_audio`,`audio_sample_rate`,`retake_audio_regions`|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2.3-A2V-TwoStage.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2.3-A2V-TwoStage.py)|-|-|-|-|
+|[Lightricks/LTX-2.3: Retake](https://www.modelscope.cn/models/Lightricks/LTX-2.3)|`retake_video`,`retake_video_regions`,`retake_audio`,`audio_sample_rate`,`retake_audio_regions`|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2.3-T2AV-TwoStage-Retake.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2.3-T2AV-TwoStage-Retake.py)|-|-|-|-|
+|[Lightricks/LTX-2.3-22b-IC-LoRA-Union-Control](https://www.modelscope.cn/models/Lightricks/LTX-2.3-22b-IC-LoRA-Union-Control)|`in_context_videos`,`in_context_downsample_factor`|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2.3-T2AV-IC-LoRA-Union-Control.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2.3-T2AV-IC-LoRA-Union-Control.py)|-|-|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/lora/LTX-2.3-T2AV-IC-LoRA-splited.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/validate_lora/LTX-2.3-T2AV-IC-LoRA.py)|
+|[Lightricks/LTX-2.3-22b-IC-LoRA-Motion-Track-Control](https://www.modelscope.cn/models/Lightricks/LTX-2.3-22b-IC-LoRA-Motion-Track-Control)|`in_context_videos`,`in_context_downsample_factor`|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2.3-T2AV-IC-LoRA-Motion-Track-Control.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2.3-T2AV-IC-LoRA-Motion-Track-Control.py)|-|-|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/lora/LTX-2.3-T2AV-IC-LoRA-splited.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/validate_lora/LTX-2.3-T2AV-IC-LoRA.py)|
 |[Lightricks/LTX-2: OneStagePipeline-T2AV](https://www.modelscope.cn/models/Lightricks/LTX-2)||[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2-T2AV-OneStage.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2-T2AV-OneStage.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/full/LTX-2-T2AV-splited.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/validate_full/LTX-2-T2AV.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/lora/LTX-2-T2AV-splited.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/validate_lora/LTX-2-T2AV.py)|
+|[Lightricks/LTX-2-19b-IC-LoRA-Union-Control](https://www.modelscope.cn/models/Lightricks/LTX-2-19b-IC-LoRA-Union-Control)|`in_context_videos`,`in_context_downsample_factor`|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2-T2AV-IC-LoRA-Union-Control.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2-T2AV-IC-LoRA-Union-Control.py)|-|-|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/lora/LTX-2-T2AV-IC-LoRA-splited.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/validate_lora/LTX-2-T2AV-IC-LoRA.py)|
+|[Lightricks/LTX-2-19b-IC-LoRA-Detailer](https://www.modelscope.cn/models/Lightricks/LTX-2-19b-IC-LoRA-Detailer)|`in_context_videos`,`in_context_downsample_factor`|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2-T2AV-IC-LoRA-Detailer.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2-T2AV-IC-LoRA-Detailer.py)|-|-|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/lora/LTX-2-T2AV-IC-LoRA-splited.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/validate_lora/LTX-2-T2AV-IC-LoRA.py)|
 |[Lightricks/LTX-2: TwoStagePipeline-T2AV](https://www.modelscope.cn/models/Lightricks/LTX-2)||[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2-T2AV-TwoStage.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2-T2AV-TwoStage.py)|-|-|-|-|
 |[Lightricks/LTX-2: DistilledPipeline-T2AV](https://www.modelscope.cn/models/Lightricks/LTX-2)||[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2-T2AV-DistilledPipeline.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2-T2AV-DistilledPipeline.py)|-|-|-|-|
 |[Lightricks/LTX-2: OneStagePipeline-I2AV](https://www.modelscope.cn/models/Lightricks/LTX-2)|`input_images`|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2-I2AV-OneStage.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2-I2AV-OneStage.py)|-|-|-|-|
@@ -205,7 +217,7 @@ LTX-2 series models are uniformly trained through [`examples/ltx2/model_training
 We have built a sample video dataset for your testing. You can download this dataset with the following command:
 
 ```shell
-modelscope download --dataset DiffSynth-Studio/example_video_dataset --local_dir ./data/example_video_dataset
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --local_dir ./data/diffsynth_example_dataset
 ```
 
 We have written recommended training scripts for each model, please refer to the table in the "Model Overview" section above. For how to write model training scripts, please refer to [Model Training](../Pipeline_Usage/Model_Training.md); for more advanced training algorithms, please refer to [Training Framework Detailed Explanation](https://github.com/modelscope/DiffSynth-Studio/tree/main/docs/en/Training/).

docs/en/Model_Details/Qwen-Image.md

@@ -86,9 +86,11 @@ graph LR;
 | [Qwen/Qwen-Image-Edit-2509](https://www.modelscope.cn/models/Qwen/Qwen-Image-Edit-2509) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference/Qwen-Image-Edit-2509.py) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference_low_vram/Qwen-Image-Edit-2509.py) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/full/Qwen-Image-Edit-2509.sh) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_full/Qwen-Image-Edit-2509.py) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/lora/Qwen-Image-Edit-2509.sh) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_lora/Qwen-Image-Edit-2509.py) |
 |[Qwen/Qwen-Image-Edit-2511](https://www.modelscope.cn/models/Qwen/Qwen-Image-Edit-2511)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference/Qwen-Image-Edit-2511.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference_low_vram/Qwen-Image-Edit-2511.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/full/Qwen-Image-Edit-2511.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_full/Qwen-Image-Edit-2511.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/lora/Qwen-Image-Edit-2511.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_lora/Qwen-Image-Edit-2511.py)|
 |[FireRedTeam/FireRed-Image-Edit-1.0](https://www.modelscope.cn/models/FireRedTeam/FireRed-Image-Edit-1.0)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference/FireRed-Image-Edit-1.0.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference_low_vram/FireRed-Image-Edit-1.0.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/full/FireRed-Image-Edit-1.0.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_full/FireRed-Image-Edit-1.0.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/lora/FireRed-Image-Edit-1.0.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_lora/FireRed-Image-Edit-1.0.py)|
+|[FireRedTeam/FireRed-Image-Edit-1.1](https://www.modelscope.cn/models/FireRedTeam/FireRed-Image-Edit-1.1)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference/FireRed-Image-Edit-1.1.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference_low_vram/FireRed-Image-Edit-1.1.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/full/FireRed-Image-Edit-1.1.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_full/FireRed-Image-Edit-1.1.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/lora/FireRed-Image-Edit-1.1.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_lora/FireRed-Image-Edit-1.1.py)|
 |[lightx2v/Qwen-Image-Edit-2511-Lightning](https://modelscope.cn/models/lightx2v/Qwen-Image-Edit-2511-Lightning)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference/Qwen-Image-Edit-2511-Lightning.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference_low_vram/Qwen-Image-Edit-2511-Lightning.py)|-|-|-|-|
 |[Qwen/Qwen-Image-Layered](https://www.modelscope.cn/models/Qwen/Qwen-Image-Layered)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference/Qwen-Image-Layered.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference_low_vram/Qwen-Image-Layered.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/full/Qwen-Image-Layered.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_full/Qwen-Image-Layered.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/lora/Qwen-Image-Layered.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_lora/Qwen-Image-Layered.py)|
 |[DiffSynth-Studio/Qwen-Image-Layered-Control](https://www.modelscope.cn/models/DiffSynth-Studio/Qwen-Image-Layered-Control)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference/Qwen-Image-Layered-Control.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference_low_vram/Qwen-Image-Layered-Control.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/full/Qwen-Image-Layered-Control.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_full/Qwen-Image-Layered-Control.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/lora/Qwen-Image-Layered-Control.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_lora/Qwen-Image-Layered-Control.py)|
+|[DiffSynth-Studio/Qwen-Image-Layered-Control-V2](https://www.modelscope.cn/models/DiffSynth-Studio/Qwen-Image-Layered-Control-V2)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference/Qwen-Image-Layered-Control-V2.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference_low_vram/Qwen-Image-Layered-Control-V2.py)|-|-|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/lora/Qwen-Image-Layered-Control-V2.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_lora/Qwen-Image-Layered-Control-V2.py)|
 | [DiffSynth-Studio/Qwen-Image-EliGen](https://www.modelscope.cn/models/DiffSynth-Studio/Qwen-Image-EliGen) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference/Qwen-Image-EliGen.py) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference_low_vram/Qwen-Image-EliGen.py) | - | - | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/lora/Qwen-Image-EliGen.sh) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_lora/Qwen-Image-EliGen.py) |
 | [DiffSynth-Studio/Qwen-Image-EliGen-V2](https://www.modelscope.cn/models/DiffSynth-Studio/Qwen-Image-EliGen-V2) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference/Qwen-Image-EliGen-V2.py) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference_low_vram/Qwen-Image-EliGen-V2.py) | - | - | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/lora/Qwen-Image-EliGen.sh) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_lora/Qwen-Image-EliGen.py) |
 | [DiffSynth-Studio/Qwen-Image-EliGen-Poster](https://www.modelscope.cn/models/DiffSynth-Studio/Qwen-Image-EliGen-Poster) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference/Qwen-Image-EliGen-Poster.py) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference_low_vram/Qwen-Image-EliGen-Poster.py) | - | - | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/lora/Qwen-Image-EliGen-Poster.sh) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_lora/Qwen-Image-EliGen-Poster.py) |
@@ -197,7 +199,7 @@ Qwen-Image series models are uniformly trained through [`examples/qwen_image/mod
 We have built a sample image dataset for your testing. You can download this dataset with the following command:
 
 ```shell
-modelscope download --dataset DiffSynth-Studio/example_image_dataset --local_dir ./data/example_image_dataset
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --local_dir ./data/diffsynth_example_dataset
 ```
 
 We have written recommended training scripts for each model, please refer to the table in the "Model Overview" section above. For how to write model training scripts, please refer to [Model Training](../Pipeline_Usage/Model_Training.md); for more advanced training algorithms, please refer to [Training Framework Detailed Explanation](https://github.com/modelscope/DiffSynth-Studio/tree/main/docs/en/Training/).

docs/en/Model_Details/Wan.md

@@ -137,6 +137,8 @@ graph LR;
 | [PAI/Wan2.2-Fun-A14B-InP](https://modelscope.cn/models/PAI/Wan2.2-Fun-A14B-InP) | `input_image`, `end_image` | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_inference/Wan2.2-Fun-A14B-InP.py) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_training/full/Wan2.2-Fun-A14B-InP.sh) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_training/validate_full/Wan2.2-Fun-A14B-InP.py) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_training/lora/Wan2.2-Fun-A14B-InP.sh) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_training/validate_lora/Wan2.2-Fun-A14B-InP.py) |
 | [PAI/Wan2.2-Fun-A14B-Control](https://modelscope.cn/models/PAI/Wan2.2-Fun-A14B-Control) | `control_video`, `reference_image` | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_inference/Wan2.2-Fun-A14B-Control.py) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_training/full/Wan2.2-Fun-A14B-Control.sh) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_training/validate_full/Wan2.2-Fun-A14B-Control.py) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_training/lora/Wan2.2-Fun-A14B-Control.sh) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_training/validate_lora/Wan2.2-Fun-A14B-Control.py) |
 | [PAI/Wan2.2-Fun-A14B-Control-Camera](https://modelscope.cn/models/PAI/Wan2.2-Fun-A14B-Control-Camera) | `control_camera_video`, `input_image` | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_inference/Wan2.2-Fun-A14B-Control-Camera.py) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_training/full/Wan2.2-Fun-A14B-Control-Camera.sh) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_training/validate_full/Wan2.2-Fun-A14B-Control-Camera.py) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_training/lora/Wan2.2-Fun-A14B-Control-Camera.sh) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_training/validate_lora/Wan2.2-Fun-A14B-Control-Camera.py) |
+| [openmoss/MOVA-360p](https://modelscope.cn/models/openmoss/MOVA-360p) | `input_image` | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/mova/model_inference/MOVA-360p-I2AV.py) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/mova/model_training/full/MOVA-360P-I2AV.sh) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/mova/model_training/validate_full/MOVA-360p-I2AV.py) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/mova/model_training/lora/MOVA-360P-I2AV.sh) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/mova/model_training/validate_lora/MOVA-360p-I2AV.py) |
+| [openmoss/MOVA-720p](https://modelscope.cn/models/openmoss/MOVA-720p) | `input_image` | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/mova/model_inference/MOVA-720p-I2AV.py) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/mova/model_training/full/MOVA-720P-I2AV.sh) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/mova/model_training/validate_full/MOVA-720p-I2AV.py) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/mova/model_training/lora/MOVA-720P-I2AV.sh) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/mova/model_training/validate_lora/MOVA-720p-I2AV.py) |
 
 * FP8 Precision Training: [doc](../Training/FP8_Precision.md), [code](https://github.com/modelscope/DiffSynth-Studio/tree/main/examples/wanvideo/model_training/special/fp8_training/)
 * Two-stage Split Training: [doc](../Training/Split_Training.md), [code](https://github.com/modelscope/DiffSynth-Studio/tree/main/examples/wanvideo/model_training/special/split_training/)
@@ -251,7 +253,7 @@ Wan series models are uniformly trained through [`examples/wanvideo/model_traini
 We have built a sample video dataset for your testing. You can download this dataset with the following command:
 
 ```shell
-modelscope download --dataset DiffSynth-Studio/example_video_dataset --local_dir ./data/example_video_dataset
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --local_dir ./data/diffsynth_example_dataset
 ```
 
 We have written recommended training scripts for each model, please refer to the table in the "Model Overview" section above. For how to write model training scripts, please refer to [Model Training](../Pipeline_Usage/Model_Training.md); for more advanced training algorithms, please refer to [Training Framework Detailed Explanation](https://github.com/modelscope/DiffSynth-Studio/tree/main/docs/en/Training/).

docs/en/Model_Details/Z-Image.md

@@ -134,7 +134,7 @@ Z-Image series models are uniformly trained through [`examples/z_image/model_tra
 We have built a sample image dataset for your testing. You can download this dataset with the following command:
 
 ```shell
-modelscope download --dataset DiffSynth-Studio/example_image_dataset --local_dir ./data/example_image_dataset
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --local_dir ./data/diffsynth_example_dataset
 ```
 
 We have written recommended training scripts for each model, please refer to the table in the "Model Overview" section above. For how to write model training scripts, please refer to [Model Training](../Pipeline_Usage/Model_Training.md); for more advanced training algorithms, please refer to [Training Framework Detailed Explanation](https://github.com/modelscope/DiffSynth-Studio/tree/main/docs/en/Training/).

docs/en/Pipeline_Usage/GPU_support.md

@@ -90,4 +90,5 @@ Set 0 or not set: indicates not enabling the binding function
 | Model          | Parameter                 | Note              |
 |----------------|---------------------------|-------------------|
 | Wan 14B series | --initialize_model_on_cpu | The 14B model needs to be initialized on the CPU |
-| Qwen-Image series | --initialize_model_on_cpu | The model needs to be initialized on the CPU |
+| Qwen-Image series | --initialize_model_on_cpu | The model needs to be initialized on the CPU |
+| Z-Image series | --enable_npu_patch | Using NPU fusion operator to replace the corresponding operator in Z-image model to improve the performance of the model on NPU |

docs/en/Pipeline_Usage/Model_Training.md

@@ -69,25 +69,11 @@ We have built sample datasets for your testing. To understand how the universal
 
 <details>
 
-<summary>Sample Image Dataset</summary>
+<summary>Sample Dataset</summary>
 
 > ```shell
-> modelscope download --dataset DiffSynth-Studio/example_image_dataset --local_dir ./data/example_image_dataset
+> modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --local_dir ./data/diffsynth_example_dataset
 > ```
-> 
-> Applicable to training of image generation models such as Qwen-Image and FLUX.
-
-</details>
-
-<details>
-
-<summary>Sample Video Dataset</summary>
-
-> ```shell
-> modelscope download --dataset DiffSynth-Studio/example_video_dataset --local_dir ./data/example_video_dataset
-> ```
-> 
-> Applicable to training of video generation models such as Wan.
 
 </details>
 
@@ -123,7 +109,6 @@ Similar to [model loading during inference](../Pipeline_Usage/Model_Inference.md
 
 <details>
 
-<details>
 
 <summary>Load models from local file paths</summary>
 
@@ -244,4 +229,119 @@ accelerate launch --config_file examples/qwen_image/model_training/full/accelera
 * The training framework does not support batch size > 1. The reasons are complex. See [Q&A: Why doesn't the training framework support batch size > 1?](../QA.md#why-doesnt-the-training-framework-support-batch-size--1)
 * Some models contain redundant parameters. For example, the text encoding part of the last layer of Qwen-Image's DiT part. When training these models, `--find_unused_parameters` needs to be set to avoid errors in multi-GPU training. For compatibility with community models, we do not intend to remove these redundant parameters.
 * The loss function value of Diffusion models has little relationship with actual effects. Therefore, we do not record loss function values during training. We recommend setting `--num_epochs` to a sufficiently large value, testing while training, and manually closing the training program after the effect converges.
-* `--use_gradient_checkpointing` is usually enabled unless GPU VRAM is sufficient; `--use_gradient_checkpointing_offload` is enabled as needed. See [`diffsynth.core.gradient`](../API_Reference/core/gradient.md) for details.
+* `--use_gradient_checkpointing` is usually enabled unless GPU VRAM is sufficient; `--use_gradient_checkpointing_offload` is enabled as needed. See [`diffsynth.core.gradient`](../API_Reference/core/gradient.md) for details.
+
+## Low VRAM Training
+
+If you want to complete LoRA model training on GPU with low vram, you can combine [Two-Stage Split Training](../Training/Split_Training.md) with `deepspeed_zero3_offload` training. First, split the preprocessing steps into the first stage and store the computed results onto the hard disk. Second, read these results from the disk and train the denoising model. By using `deepspeed_zero3_offload`, the training parameters and optimizer states are offloaded to the CPU or disk. We provide examples for some models, primarily by specifying the `deepspeed` configuration via `--config_file`.
+
+Please note that the `deepspeed_zero3_offload` mode is incompatible with PyTorch's native gradient checkpointing mechanism. To address this, we have adapted the `checkpointing` interface of `deepspeed`. Users need to fill the `activation_checkpointing` field in the `deepspeed` configuration to enable gradient checkpointing.
+
+Below is the script for low VRAM model training for the Qwen-Image model:
+
+```shell
+accelerate launch examples/qwen_image/model_training/train.py \
+  --dataset_base_path data/example_image_dataset \
+  --dataset_metadata_path data/example_image_dataset/metadata.csv \
+  --max_pixels 1048576 \
+  --dataset_repeat 1 \
+  --model_id_with_origin_paths "Qwen/Qwen-Image:text_encoder/model*.safetensors,Qwen/Qwen-Image:vae/diffusion_pytorch_model.safetensors" \
+  --learning_rate 1e-4 \
+  --num_epochs 5 \
+  --remove_prefix_in_ckpt "pipe.dit." \
+  --output_path "./models/train/Qwen-Image_lora-splited-cache" \
+  --lora_base_model "dit" \
+  --lora_target_modules "to_q,to_k,to_v,add_q_proj,add_k_proj,add_v_proj,to_out.0,to_add_out,img_mlp.net.2,img_mod.1,txt_mlp.net.2,txt_mod.1" \
+  --lora_rank 32 \
+  --task "sft:data_process" \
+  --use_gradient_checkpointing \
+  --dataset_num_workers 8 \
+  --find_unused_parameters
+
+accelerate launch --config_file examples/qwen_image/model_training/special/low_vram_training/deepspeed_zero3_cpuoffload.yaml examples/qwen_image/model_training/train.py \
+  --dataset_base_path "./models/train/Qwen-Image_lora-splited-cache" \
+  --max_pixels 1048576 \
+  --dataset_repeat 50 \
+  --model_id_with_origin_paths "Qwen/Qwen-Image:transformer/diffusion_pytorch_model*.safetensors" \
+  --learning_rate 1e-4 \
+  --num_epochs 5 \
+  --remove_prefix_in_ckpt "pipe.dit." \
+  --output_path "./models/train/Qwen-Image_lora" \
+  --lora_base_model "dit" \
+  --lora_target_modules "to_q,to_k,to_v,add_q_proj,add_k_proj,add_v_proj,to_out.0,to_add_out,img_mlp.net.2,img_mod.1,txt_mlp.net.2,txt_mod.1" \
+  --lora_rank 32 \
+  --task "sft:train" \
+  --use_gradient_checkpointing \
+  --dataset_num_workers 8 \
+  --find_unused_parameters \
+  --initialize_model_on_cpu
+```
+
+The configurations for `accelerate` and `deepspeed` are as follows:
+
+```yaml
+compute_environment: LOCAL_MACHINE
+debug: true
+deepspeed_config:
+  deepspeed_config_file: examples/qwen_image/model_training/special/low_vram_training/ds_z3_cpuoffload.json
+  zero3_init_flag: true
+distributed_type: DEEPSPEED
+downcast_bf16: 'no'
+enable_cpu_affinity: false
+machine_rank: 0
+main_training_function: main
+num_machines: 1
+num_processes: 1
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false
+```
+
+```json
+{
+    "fp16": {
+        "enabled": "auto",
+        "loss_scale": 0,
+        "loss_scale_window": 1000,
+        "initial_scale_power": 16,
+        "hysteresis": 2,
+        "min_loss_scale": 1
+    },
+    "bf16": {
+        "enabled": "auto"
+    },
+    "zero_optimization": {
+        "stage": 3,
+        "offload_optimizer": {
+            "device": "cpu",
+            "pin_memory": true
+        },
+        "offload_param": {
+            "device": "cpu",
+            "pin_memory": true
+        },
+        "overlap_comm": false,
+        "contiguous_gradients": true,
+        "sub_group_size": 1e9,
+        "reduce_bucket_size": 5e7,
+        "stage3_prefetch_bucket_size": 5e7,
+        "stage3_param_persistence_threshold": 1e5,
+        "stage3_max_live_parameters": 1e8,
+        "stage3_max_reuse_distance": 1e8,
+        "stage3_gather_16bit_weights_on_model_save": true
+    },
+    "activation_checkpointing": {
+        "partition_activations": false,
+        "cpu_checkpointing": false,
+        "contiguous_memory_optimization": false
+    },
+    "gradient_accumulation_steps": "auto",
+    "gradient_clipping": "auto",
+    "train_batch_size": "auto",
+    "train_micro_batch_size_per_gpu": "auto",
+    "wall_clock_breakdown": false
+}
+```

docs/en/Pipeline_Usage/Setup.md

@@ -37,9 +37,9 @@ pip install torch torchvision --index-url https://download.pytorch.org/whl/rocm6
    git clone https://github.com/modelscope/DiffSynth-Studio.git
    cd DiffSynth-Studio
    # aarch64/ARM
-   pip install -e .[npu_aarch64] --extra-index-url "https://download.pytorch.org/whl/cpu"
+   pip install -e .[npu_aarch64] 
    # x86
-   pip install -e .[npu]
+   pip install -e .[npu] --extra-index-url "https://download.pytorch.org/whl/cpu"
 
 When using Ascend NPU, please replace `"cuda"` with `"npu"` in your Python code. For details, see [NPU Support](../Pipeline_Usage/GPU_support.md#ascend-npu).
 

docs/en/README.md

@@ -42,6 +42,8 @@ This section introduces the Diffusion models supported by `DiffSynth-Studio`. So
 * [Qwen-Image](./Model_Details/Qwen-Image.md)
 * [FLUX.2](./Model_Details/FLUX2.md)
 * [Z-Image](./Model_Details/Z-Image.md)
+* [Anima](./Model_Details/Anima.md)
+* [LTX-2](./Model_Details/LTX-2.md)
 
 ## Section 3: Training Framework
 
@@ -78,7 +80,7 @@ This section introduces the independent core module `diffsynth.core` in `DiffSyn
 This section introduces how to use `DiffSynth-Studio` to train new models, helping researchers explore new model technologies.
 
 * [Training models from scratch](./Research_Tutorial/train_from_scratch.md)
-* Inference improvement techniques 【coming soon】
+* [Inference improvement techniques](./Research_Tutorial/inference_time_scaling.md)
 * Designing controllable generation models 【coming soon】
 * Creating new training paradigms 【coming soon】
 

docs/en/Research_Tutorial/inference_time_scaling.ipynb

@@ -0,0 +1,236 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "id": "8db54992",
+   "metadata": {},
+   "source": [
+    "# Inference Optimization Techniques\n",
+    "\n",
+    "DiffSynth-Studio aims to drive technological innovation through its foundational framework. This article demonstrates how to build a training-free image generation enhancement solution using DiffSynth-Studio, taking Inference-time scaling as an example."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "0911cad4",
+   "metadata": {},
+   "source": [
+    "## 1. Image Quality Quantification\n",
+    "\n",
+    "First, we need to find an indicator to quantify image quality from generation models. Manual scoring is the most straightforward solution but too costly for large-scale applications. However, after collecting manual scores, training an image classification model to predict human scoring is completely feasible. PickScore [[1]](https://arxiv.org/abs/2305.01569) is such a model. Running the following code will automatically download and load the [PickScore model](https://modelscope.cn/models/AI-ModelScope/PickScore_v1)."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "4faca4ca",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from modelscope import AutoProcessor, AutoModel\n",
+    "import torch\n",
+    "\n",
+    "class PickScore(torch.nn.Module):\n",
+    "    def __init__(self):\n",
+    "        super().__init__()\n",
+    "        self.processor = AutoProcessor.from_pretrained(\"laion/CLIP-ViT-H-14-laion2B-s32B-b79K\")\n",
+    "        self.model = AutoModel.from_pretrained(\"AI-ModelScope/PickScore_v1\").eval().to(\"cuda\")\n",
+    "\n",
+    "    def forward(self, image, prompt):\n",
+    "        image_inputs = self.processor(images=image, padding=True, truncation=True, max_length=77, return_tensors=\"pt\").to(\"cuda\")\n",
+    "        text_inputs = self.processor(text=prompt, padding=True, truncation=True, max_length=77, return_tensors=\"pt\").to(\"cuda\")\n",
+    "        with torch.inference_mode():\n",
+    "            image_embs = self.model.get_image_features(**image_inputs).pooler_output\n",
+    "            image_embs = image_embs / torch.norm(image_embs, dim=-1, keepdim=True)\n",
+    "            text_embs = self.model.get_text_features(**text_inputs).pooler_output\n",
+    "            text_embs = text_embs / torch.norm(text_embs, dim=-1, keepdim=True)\n",
+    "            score = (text_embs @ image_embs.T).flatten().item()\n",
+    "        return score\n",
+    "\n",
+    "reward_model = PickScore()"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "5f807cec",
+   "metadata": {},
+   "source": [
+    "## 2. Inference-time Scaling Techniques\n",
+    "\n",
+    "Inference-time Scaling [[2]](https://arxiv.org/abs/2504.00294) is an interesting technique aiming to improve generation quality by increasing computational costs during inference. For example, in language models, models like [Qwen/Qwen3.5-27B](https://modelscope.cn/models/Qwen/Qwen3.5-27B) and [deepseek-ai/DeepSeek-R1](deepseek-ai/DeepSeek-R1) use \"thinking mode\" to guide the model to spend more time considering results more carefully, producing more accurate answers. Next, we'll use the [black-forest-labs/FLUX.2-klein-4B](https://modelscope.cn/models/black-forest-labs/FLUX.2-klein-4B) model as an example to explore how to design Inference-time Scaling solutions for image generation models.\n",
+    "\n",
+    "> Before starting, we slightly modified the `Flux2ImagePipeline` code to allow initialization with specific Gaussian noise matrices for result reproducibility. See `Flux2Unit_NoiseInitializer` in [diffsynth/pipelines/flux2_image.py](https://github.com/modelscope/DiffSynth-Studio/blob/main/diffsynth/pipelines/flux2_image.py).\n",
+    "\n",
+    "Run the following code to load the [black-forest-labs/FLUX.2-klein-4B](https://modelscope.cn/models/black-forest-labs/FLUX.2-klein-4B) model."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "c5818a87",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from diffsynth.pipelines.flux2_image import Flux2ImagePipeline, ModelConfig\n",
+    "\n",
+    "pipe = Flux2ImagePipeline.from_pretrained(\n",
+    "    torch_dtype=torch.bfloat16,\n",
+    "    device=\"cuda\",\n",
+    "    model_configs=[\n",
+    "        ModelConfig(model_id=\"black-forest-labs/FLUX.2-klein-4B\", origin_file_pattern=\"text_encoder/*.safetensors\"),\n",
+    "        ModelConfig(model_id=\"black-forest-labs/FLUX.2-klein-4B\", origin_file_pattern=\"transformer/*.safetensors\"),\n",
+    "        ModelConfig(model_id=\"black-forest-labs/FLUX.2-klein-4B\", origin_file_pattern=\"vae/diffusion_pytorch_model.safetensors\"),\n",
+    "    ],\n",
+    "    tokenizer_config=ModelConfig(model_id=\"black-forest-labs/FLUX.2-klein-4B\", origin_file_pattern=\"tokenizer/\"),\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "f58e9945",
+   "metadata": {},
+   "source": [
+    "Generate a sketch cat image using the prompt `\"sketch, a cat\"` and score it with the PickScore model."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "6ea2d258",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def evaluate_noise(noise, pipe, reward_model, prompt):\n",
+    "    # Generate an image and compute the score.\n",
+    "    image = pipe(\n",
+    "        prompt=prompt,\n",
+    "        num_inference_steps=4,\n",
+    "        initial_noise=noise,\n",
+    "        progress_bar_cmd=lambda x: x,\n",
+    "    )\n",
+    "    score = reward_model(image, prompt)\n",
+    "    return score\n",
+    "\n",
+    "torch.manual_seed(1)\n",
+    "prompt = \"sketch, a cat\"\n",
+    "noise = pipe.generate_noise((1, 128, 64, 64), rand_device=\"cuda\", rand_torch_dtype=pipe.torch_dtype)\n",
+    "\n",
+    "image_1 = pipe(prompt, num_inference_steps=4, initial_noise=noise)\n",
+    "print(\"Score:\", reward_model(image_1, prompt))\n",
+    "image_1"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "5e11694e",
+   "metadata": {},
+   "source": [
+    "### 2.1 Best-of-N Random Search\n",
+    "\n",
+    "Model generation results have inherent randomness. Different random seeds produce different images - sometimes high quality, sometimes low. This leads to a simple Inference-time scaling solution: generate images using multiple random seeds, score them with PickScore, and retain only the highest-scoring image."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "241f10d2",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from tqdm import tqdm\n",
+    "\n",
+    "def random_search(base_latents, objective_reward_fn, total_eval_budget):\n",
+    "    # Search for the noise randomly.\n",
+    "    best_noise = base_latents\n",
+    "    best_score = objective_reward_fn(base_latents)\n",
+    "    for it in tqdm(range(total_eval_budget - 1)):\n",
+    "        noise = pipe.generate_noise((1, 128, 64, 64), seed=None)\n",
+    "        score = objective_reward_fn(noise)\n",
+    "        if score > best_score:\n",
+    "            best_score, best_noise = score, noise\n",
+    "    return best_noise\n",
+    "\n",
+    "best_noise = random_search(\n",
+    "    base_latents=noise,\n",
+    "    objective_reward_fn=lambda noise: evaluate_noise(noise, pipe, reward_model, prompt),\n",
+    "    total_eval_budget=50,\n",
+    ")\n",
+    "image_2 = pipe(prompt, num_inference_steps=4, initial_noise=best_noise)\n",
+    "print(\"Score:\", reward_model(image_2, prompt))\n",
+    "image_2"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "8e9bf966",
+   "metadata": {},
+   "source": [
+    "We can clearly see that after multiple random searches, the final selected cat image shows richer fur details and significantly improved PickScore. However, this brute-force random search is extremely inefficient - generation time multiplies while easily hitting quality limits. Therefore, we need a more efficient search method that achieves higher scores within the same computational budget."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "c9578349",
+   "metadata": {},
+   "source": [
+    "### 2.2 SES Search\n",
+    "\n",
+    "To overcome random search limitations, we introduce the Spectral Evolution Search (SES) algorithm [[3]](https://arxiv.org/abs/2602.03208). Detailed code is available at [diffsynth/utils/ses](https://github.com/modelscope/DiffSynth-Studio/blob/main/diffsynth/utils/ses).\n",
+    "\n",
+    "Image generation in diffusion models is largely determined by low-frequency components in the initial noise. The SES algorithm decomposes Gaussian noise through wavelet transforms, fixes high-frequency details, and applies an evolution search using the cross-entropy method specifically on low-frequency components to find optimal initial noise with higher efficiency.\n",
+    "\n",
+    "Run the following code to perform efficient best Gaussian noise matrix search using SES."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "adeed2aa",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from diffsynth.utils.ses import ses_search\n",
+    "\n",
+    "best_noise = ses_search(\n",
+    "    base_latents=noise,\n",
+    "    objective_reward_fn=lambda noise: evaluate_noise(noise, pipe, reward_model, prompt),\n",
+    "    total_eval_budget=50,\n",
+    ")\n",
+    "image_3 = pipe(prompt, num_inference_steps=4, initial_noise=best_noise)\n",
+    "print(\"Score:\", reward_model(image_3, prompt))\n",
+    "image_3"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "940a97f1",
+   "metadata": {},
+   "source": [
+    "Observing the results, under the same computational budget, SES achieves significantly higher PickScore compared to random search. The \"sketch cat\" demonstrates more refined overall composition and more layered contrast between light and shadow.\n",
+    "\n",
+    "Inference-time scaling can achieve higher image quality at the cost of longer inference time. The generated image data can then be used to train the model itself through methods like DPO [[4]](https://arxiv.org/abs/2311.12908) or differential training [[5]](https://arxiv.org/abs/2412.12888), opening another interesting research direction."
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "dzj8",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.19"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

docs/en/Research_Tutorial/inference_time_scaling.md

@@ -0,0 +1,140 @@
+# Inference Optimization Techniques
+
+DiffSynth-Studio aims to drive technological innovation through its foundational framework. This article demonstrates how to build a training-free image generation enhancement solution using DiffSynth-Studio, taking Inference-time scaling as an example.
+
+Notebook: https://github.com/modelscope/DiffSynth-Studio/blob/main/docs/en/Research_Tutorial/inference_time_scaling.ipynb
+
+## 1. Image Quality Quantification
+
+First, we need to find an indicator to quantify image quality from generation models. Manual scoring is the most straightforward solution but too costly for large-scale applications. However, after collecting manual scores, training an image classification model to predict human scoring is completely feasible. PickScore [[1]](https://arxiv.org/abs/2305.01569) is such a model. Running the following code will automatically download and load the [PickScore model](https://modelscope.cn/models/AI-ModelScope/PickScore_v1).
+
+```python
+from modelscope import AutoProcessor, AutoModel
+import torch
+
+class PickScore(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.processor = AutoProcessor.from_pretrained("laion/CLIP-ViT-H-14-laion2B-s32B-b79K")
+        self.model = AutoModel.from_pretrained("AI-ModelScope/PickScore_v1").eval().to("cuda")
+
+    def forward(self, image, prompt):
+        image_inputs = self.processor(images=image, padding=True, truncation=True, max_length=77, return_tensors="pt").to("cuda")
+        text_inputs = self.processor(text=prompt, padding=True, truncation=True, max_length=77, return_tensors="pt").to("cuda")
+        with torch.inference_mode():
+            image_embs = self.model.get_image_features(**image_inputs).pooler_output
+            image_embs = image_embs / torch.norm(image_embs, dim=-1, keepdim=True)
+            text_embs = self.model.get_text_features(**text_inputs).pooler_output
+            text_embs = text_embs / torch.norm(text_embs, dim=-1, keepdim=True)
+            score = (text_embs @ image_embs.T).flatten().item()
+        return score
+
+reward_model = PickScore()
+```
+
+## 2. Inference-time Scaling Techniques
+
+Inference-time Scaling [[2]](https://arxiv.org/abs/2504.00294) is an interesting technique aiming to improve generation quality by increasing computational costs during inference. For example, in language models, models like [Qwen/Qwen3.5-27B](https://modelscope.cn/models/Qwen/Qwen3.5-27B) and [deepseek-ai/DeepSeek-R1](deepseek-ai/DeepSeek-R1) use "thinking mode" to guide the model to spend more time considering results more carefully, producing more accurate answers. Next, we'll use the [black-forest-labs/FLUX.2-klein-4B](https://modelscope.cn/models/black-forest-labs/FLUX.2-klein-4B) model as an example to explore how to design Inference-time Scaling solutions for image generation models.
+
+> Before starting, we slightly modified the `Flux2ImagePipeline` code to allow initialization with specific Gaussian noise matrices for result reproducibility. See `Flux2Unit_NoiseInitializer` in [diffsynth/pipelines/flux2_image.py](https://github.com/modelscope/DiffSynth-Studio/blob/main/diffsynth/pipelines/flux2_image.py).
+
+Run the following code to load the [black-forest-labs/FLUX.2-klein-4B](https://modelscope.cn/models/black-forest-labs/FLUX.2-klein-4B) model.
+
+```python
+from diffsynth.pipelines.flux2_image import Flux2ImagePipeline, ModelConfig
+
+pipe = Flux2ImagePipeline.from_pretrained(
+    torch_dtype=torch.bfloat16,
+    device="cuda",
+    model_configs=[
+        ModelConfig(model_id="black-forest-labs/FLUX.2-klein-4B", origin_file_pattern="text_encoder/*.safetensors"),
+        ModelConfig(model_id="black-forest-labs/FLUX.2-klein-4B", origin_file_pattern="transformer/*.safetensors"),
+        ModelConfig(model_id="black-forest-labs/FLUX.2-klein-4B", origin_file_pattern="vae/diffusion_pytorch_model.safetensors"),
+    ],
+    tokenizer_config=ModelConfig(model_id="black-forest-labs/FLUX.2-klein-4B", origin_file_pattern="tokenizer/"),
+)
+```
+
+Generate a sketch cat image using the prompt `"sketch, a cat"` and score it with the PickScore model.
+
+```python
+def evaluate_noise(noise, pipe, reward_model, prompt):
+    # Generate an image and compute the score.
+    image = pipe(
+        prompt=prompt,
+        num_inference_steps=4,
+        initial_noise=noise,
+        progress_bar_cmd=lambda x: x,
+    )
+    score = reward_model(image, prompt)
+    return score
+
+torch.manual_seed(1)
+prompt = "sketch, a cat"
+noise = pipe.generate_noise((1, 128, 64, 64), rand_device="cuda", rand_torch_dtype=pipe.torch_dtype)
+
+image_1 = pipe(prompt, num_inference_steps=4, initial_noise=noise)
+print("Score:", reward_model(image_1, prompt))
+image_1
+```
+
+![Image](https://github.com/user-attachments/assets/b6546c6d-b368-4463-b703-d561a9134ba0)
+
+### 2.1 Best-of-N Random Search
+
+Model generation results have inherent randomness. Different random seeds produce different images - sometimes high quality, sometimes low. This leads to a simple Inference-time scaling solution: generate images using multiple random seeds, score them with PickScore, and retain only the highest-scoring image.
+
+```python
+from tqdm import tqdm
+
+def random_search(base_latents, objective_reward_fn, total_eval_budget):
+    # Search for the noise randomly.
+    best_noise = base_latents
+    best_score = objective_reward_fn(base_latents)
+    for it in tqdm(range(total_eval_budget - 1)):
+        noise = pipe.generate_noise((1, 128, 64, 64), seed=None)
+        score = objective_reward_fn(noise)
+        if score > best_score:
+            best_score, best_noise = score, noise
+    return best_noise
+
+best_noise = random_search(
+    base_latents=noise,
+    objective_reward_fn=lambda noise: evaluate_noise(noise, pipe, reward_model, prompt),
+    total_eval_budget=50,
+)
+image_2 = pipe(prompt, num_inference_steps=4, initial_noise=best_noise)
+print("Score:", reward_model(image_2, prompt))
+image_2
+```
+
+![Image](https://github.com/user-attachments/assets/b8dba70a-daa8-4368-8f32-a6c150daecb5)
+
+We can clearly see that after multiple random searches, the final selected cat image shows richer fur details and significantly improved PickScore. However, this brute-force random search is extremely inefficient - generation time multiplies while easily hitting quality limits. Therefore, we need a more efficient search method that achieves higher scores within the same computational budget.
+
+### 2.2 SES Search
+
+To overcome random search limitations, we introduce the Spectral Evolution Search (SES) algorithm [[3]](https://arxiv.org/abs/2602.03208). Detailed code is available at [diffsynth/utils/ses](https://github.com/modelscope/DiffSynth-Studio/blob/main/diffsynth/utils/ses).
+
+Image generation in diffusion models is largely determined by low-frequency components in the initial noise. The SES algorithm decomposes Gaussian noise through wavelet transforms, fixes high-frequency details, and applies an evolution search using the cross-entropy method specifically on low-frequency components to find optimal initial noise with higher efficiency.
+
+Run the following code to perform efficient best Gaussian noise matrix search using SES.
+
+```python
+from diffsynth.utils.ses import ses_search
+
+best_noise = ses_search(
+    base_latents=noise,
+    objective_reward_fn=lambda noise: evaluate_noise(noise, pipe, reward_model, prompt),
+    total_eval_budget=50,
+)
+image_3 = pipe(prompt, num_inference_steps=4, initial_noise=best_noise)
+print("Score:", reward_model(image_3, prompt))
+image_3
+```
+
+![Image](https://github.com/user-attachments/assets/9a3f7598-3812-46d2-b333-cd65e49886ab)
+
+Observing the results, under the same computational budget, SES achieves significantly higher PickScore compared to random search. The "sketch cat" demonstrates more refined overall composition and more layered contrast between light and shadow.
+
+Inference-time scaling can achieve higher image quality at the cost of longer inference time. The generated image data can then be used to train the model itself through methods like DPO [[4]](https://arxiv.org/abs/2311.12908) or differential training [[5]](https://arxiv.org/abs/2412.12888), opening another interesting research direction.

docs/en/Training/Direct_Distill.md

@@ -77,7 +77,7 @@ distill_qwen/image.jpg,"精致肖像，水下少女，蓝裙飘逸，发丝轻
 This sample dataset can be downloaded directly:
 
 ```shell
-modelscope download --dataset DiffSynth-Studio/example_image_dataset --local_dir ./data/example_image_dataset
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --local_dir ./data/diffsynth_example_dataset
 ```
 
 Then start LoRA distillation accelerated training:

docs/en/index.rst

@@ -27,6 +27,8 @@ Welcome to DiffSynth-Studio's Documentation
    Model_Details/Qwen-Image
    Model_Details/FLUX2
    Model_Details/Z-Image
+   Model_Details/Anima
+   Model_Details/LTX-2
 
 .. toctree::
    :maxdepth: 2
@@ -63,6 +65,7 @@ Welcome to DiffSynth-Studio's Documentation
    :caption: Research Guide
 
    Research_Tutorial/train_from_scratch
+   Research_Tutorial/inference_time_scaling
 
 .. toctree::
    :maxdepth: 2

docs/zh/Model_Details/Anima.md

@@ -0,0 +1,139 @@
+# Anima
+
+Anima 是由 CircleStone Labs 与 Comfy Org 训练并开源的图像生成模型。
+
+## 安装
+
+在使用本项目进行模型推理和训练前，请先安装 DiffSynth-Studio。
+
+```shell
+git clone https://github.com/modelscope/DiffSynth-Studio.git
+cd DiffSynth-Studio
+pip install -e .
+```
+
+更多关于安装的信息，请参考[安装依赖](../Pipeline_Usage/Setup.md)。
+
+## 快速开始
+
+运行以下代码可以快速加载 [circlestone-labs/Anima](https://www.modelscope.cn/models/circlestone-labs/Anima) 模型并进行推理。显存管理已启动，框架会自动根据剩余显存控制模型参数的加载，最低 8G 显存即可运行。
+
+```python
+from diffsynth.pipelines.anima_image import AnimaImagePipeline, ModelConfig
+import torch
+
+vram_config = {
+    "offload_dtype": "disk",
+    "offload_device": "disk",
+    "onload_dtype": "disk",
+    "onload_device": "disk",
+    "preparing_dtype": torch.bfloat16,
+    "preparing_device": "cuda",
+    "computation_dtype": torch.bfloat16,
+    "computation_device": "cuda",
+}
+pipe = AnimaImagePipeline.from_pretrained(
+    torch_dtype=torch.bfloat16,
+    device="cuda",
+    model_configs=[
+        ModelConfig(model_id="circlestone-labs/Anima", origin_file_pattern="split_files/diffusion_models/anima-preview.safetensors", **vram_config),
+        ModelConfig(model_id="circlestone-labs/Anima", origin_file_pattern="split_files/text_encoders/qwen_3_06b_base.safetensors", **vram_config),
+        ModelConfig(model_id="circlestone-labs/Anima", origin_file_pattern="split_files/vae/qwen_image_vae.safetensors", **vram_config),
+    ],
+    tokenizer_config=ModelConfig(model_id="Qwen/Qwen3-0.6B", origin_file_pattern="./"),
+    tokenizer_t5xxl_config=ModelConfig(model_id="stabilityai/stable-diffusion-3.5-large", origin_file_pattern="tokenizer_3/"),
+    vram_limit=torch.cuda.mem_get_info("cuda")[1] / (1024 ** 3) - 0.5,
+)
+prompt = "Masterpiece, best quality, solo, long hair, wavy hair, silver hair, blue eyes, blue dress, medium breasts, dress, underwater, air bubble, floating hair, refraction, portrait."
+negative_prompt = "worst quality, low quality, monochrome, zombie, interlocked fingers, Aissist, cleavage, nsfw,"
+image = pipe(prompt, seed=0, num_inference_steps=50)
+image.save("image.jpg")
+```
+
+## 模型总览
+
+|模型 ID|推理|低显存推理|全量训练|全量训练后验证|LoRA 训练|LoRA 训练后验证|
+|-|-|-|-|-|-|-|
+|[circlestone-labs/Anima](https://www.modelscope.cn/models/circlestone-labs/Anima)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/anima/model_inference/anima-preview.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/anima/model_inference_low_vram/anima-preview.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/anima/model_training/full/anima-preview.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/anima/model_training/validate_full/anima-preview.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/anima/model_training/lora/anima-preview.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/anima/model_training/validate_lora/anima-preview.py)|
+
+特殊训练脚本：
+
+* 差分 LoRA 训练：[doc](../Training/Differential_LoRA.md)
+* FP8 精度训练：[doc](../Training/FP8_Precision.md)
+* 两阶段拆分训练：[doc](../Training/Split_Training.md)
+* 端到端直接蒸馏：[doc](../Training/Direct_Distill.md)
+
+## 模型推理
+
+模型通过 `AnimaImagePipeline.from_pretrained` 加载，详见[加载模型](../Pipeline_Usage/Model_Inference.md#加载模型)。
+
+`AnimaImagePipeline` 推理的输入参数包括：
+
+* `prompt`: 提示词，描述画面中出现的内容。
+* `negative_prompt`: 负向提示词，描述画面中不应该出现的内容，默认值为 `""`。
+* `cfg_scale`: Classifier-free guidance 的参数，默认值为 4.0。
+* `input_image`: 输入图像，用于图像到图像的生成。默认为 `None`。
+* `denoising_strength`: 去噪强度，控制生成图像与输入图像的相似度，默认值为 1.0。
+* `height`: 图像高度，需保证高度为 16 的倍数，默认值为 1024。
+* `width`: 图像宽度，需保证宽度为 16 的倍数，默认值为 1024。
+* `seed`: 随机种子。默认为 `None`，即完全随机。
+* `rand_device`: 生成随机高斯噪声矩阵的计算设备，默认为 `"cpu"`。当设置为 `cuda` 时，在不同 GPU 上会导致不同的生成结果。
+* `num_inference_steps`: 推理次数，默认值为 30。
+* `sigma_shift`: 调度器的 sigma 偏移量，默认为 `None`。
+* `progress_bar_cmd`: 进度条，默认为 `tqdm.tqdm`。可通过设置为 `lambda x:x` 来屏蔽进度条。
+
+如果显存不足，请开启[显存管理](../Pipeline_Usage/VRAM_management.md)，我们在示例代码中提供了每个模型推荐的低显存配置，详见前文"模型总览"中的表格。
+
+## 模型训练
+
+Anima 系列模型统一通过 [`examples/anima/model_training/train.py`](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/anima/model_training/train.py) 进行训练，脚本的参数包括：
+
+* 通用训练参数
+    * 数据集基础配置
+        * `--dataset_base_path`: 数据集的根目录。
+        * `--dataset_metadata_path`: 数据集的元数据文件路径。
+        * `--dataset_repeat`: 每个 epoch 中数据集重复的次数。
+        * `--dataset_num_workers`: 每个 Dataloder 的进程数量。
+        * `--data_file_keys`: 元数据中需要加载的字段名称，通常是图像或视频文件的路径，以 `,` 分隔。
+    * 模型加载配置
+        * `--model_paths`: 要加载的模型路径。JSON 格式。
+        * `--model_id_with_origin_paths`: 带原始路径的模型 ID，例如 `"anima-team/anima-1B:text_encoder/*.safetensors"`。用逗号分隔。
+        * `--extra_inputs`: 模型 Pipeline 所需的额外输入参数，例如训练 ControlNet 模型时需要额外参数 `controlnet_inputs`，以 `,` 分隔。
+        * `--fp8_models`：以 FP8 格式加载的模型，格式与 `--model_paths` 或 `--model_id_with_origin_paths` 一致，目前仅支持参数不被梯度更新的模型（不需要梯度回传，或梯度仅更新其 LoRA）。
+    * 训练基础配置
+        * `--learning_rate`: 学习率。
+        * `--num_epochs`: 轮数（Epoch）。
+        * `--trainable_models`: 可训练的模型，例如 `dit`、`vae`、`text_encoder`。
+        * `--find_unused_parameters`: DDP 训练中是否存在未使用的参数，少数模型包含不参与梯度计算的冗余参数，需开启这一设置避免在多 GPU 训练中报错。
+        * `--weight_decay`：权重衰减大小，详见 [torch.optim.AdamW](https://docs.pytorch.org/docs/stable/generated/torch.optim.AdamW.html)。
+        * `--task`: 训练任务，默认为 `sft`，部分模型支持更多训练模式，请参考每个特定模型的文档。
+    * 输出配置
+        * `--output_path`: 模型保存路径。
+        * `--remove_prefix_in_ckpt`: 在模型文件的 state dict 中移除前缀。
+        * `--save_steps`: 保存模型的训练步数间隔，若此参数留空，则每个 epoch 保存一次。
+    * LoRA 配置
+        * `--lora_base_model`: LoRA 添加到哪个模型上。
+        * `--lora_target_modules`: LoRA 添加到哪些层上。
+        * `--lora_rank`: LoRA 的秩（Rank）。
+        * `--lora_checkpoint`: LoRA 检查点的路径。如果提供此路径，LoRA 将从此检查点加载。
+        * `--preset_lora_path`: 预置 LoRA 检查点路径，如果提供此路径，这一 LoRA 将会以融入基础模型的形式加载。此参数用于 LoRA 差分训练。
+        * `--preset_lora_model`: 预置 LoRA 融入的模型，例如 `dit`。
+    * 梯度配置
+        * `--use_gradient_checkpointing`: 是否启用 gradient checkpointing。
+        * `--use_gradient_checkpointing_offload`: 是否将 gradient checkpointing 卸载到内存中。
+        * `--gradient_accumulation_steps`: 梯度累积步数。
+    * 图像宽高配置（适用于图像生成模型和视频生成模型）
+        * `--height`: 图像或视频的高度。将 `height` 和 `width` 留空以启用动态分辨率。
+        * `--width`: 图像或视频的宽度。将 `height` 和 `width` 留空以启用动态分辨率。
+        * `--max_pixels`: 图像或视频帧的最大像素面积，当启用动态分辨率时，分辨率大于这个数值的图片都会被缩小，分辨率小于这个数值的图片保持不变。
+* Anima 专有参数
+    * `--tokenizer_path`: tokenizer 的路径，适用于文生图模型，留空则自动从远程下载。
+    * `--tokenizer_t5xxl_path`: T5-XXL tokenizer 的路径，适用于文生图模型，留空则自动从远程下载。
+
+我们构建了一个样例图像数据集，以方便您进行测试，通过以下命令可以下载这个数据集：
+
+```shell
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --local_dir ./data/diffsynth_example_dataset
+```
+
+我们为每个模型编写了推荐的训练脚本，请参考前文"模型总览"中的表格。关于如何编写模型训练脚本，请参考[模型训练](../Pipeline_Usage/Model_Training.md)；更多高阶训练算法，请参考[训练框架详解](https://github.com/modelscope/DiffSynth-Studio/tree/main/docs/zh/Training/)。

docs/zh/Model_Details/FLUX.md

@@ -195,7 +195,7 @@ FLUX 系列模型统一通过 [`examples/flux/model_training/train.py`](https://
 我们构建了一个样例图像数据集，以方便您进行测试，通过以下命令可以下载这个数据集：
 
 ```shell
-modelscope download --dataset DiffSynth-Studio/example_image_dataset --local_dir ./data/example_image_dataset
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --local_dir ./data/diffsynth_example_dataset
 ```
 
 我们为每个模型编写了推荐的训练脚本，请参考前文"模型总览"中的表格。关于如何编写模型训练脚本，请参考[模型训练](../Pipeline_Usage/Model_Training.md)；更多高阶训练算法，请参考[训练框架详解](https://github.com/modelscope/DiffSynth-Studio/tree/main/docs/zh/Training/)。

docs/zh/Model_Details/FLUX2.md

@@ -145,7 +145,7 @@ FLUX.2 系列模型统一通过 [`examples/flux2/model_training/train.py`](https
 我们构建了一个样例图像数据集，以方便您进行测试，通过以下命令可以下载这个数据集：
 
 ```shell
-modelscope download --dataset DiffSynth-Studio/example_image_dataset --local_dir ./data/example_image_dataset
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --local_dir ./data/diffsynth_example_dataset
 ```
 
 我们为每个模型编写了推荐的训练脚本，请参考前文"模型总览"中的表格。关于如何编写模型训练脚本，请参考[模型训练](../Pipeline_Usage/Model_Training.md)；更多高阶训练算法，请参考[训练框架详解](https://github.com/modelscope/DiffSynth-Studio/tree/main/docs/zh/Training/)。

docs/zh/Model_Details/LTX-2.md

@@ -111,7 +111,19 @@ write_video_audio_ltx2(
 ## 模型总览
 |模型 ID|额外参数|推理|低显存推理|全量训练|全量训练后验证|LoRA 训练|LoRA 训练后验证|
 |-|-|-|-|-|-|-|-|
+|[Lightricks/LTX-2.3: OneStagePipeline-I2AV](https://www.modelscope.cn/models/Lightricks/LTX-2.3)|`input_images`|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2.3-I2AV-OneStage.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2.3-I2AV-OneStage.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/full/LTX-2.3-I2AV-splited.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/validate_full/LTX-2.3-I2AV.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/lora/LTX-2.3-I2AV-splited.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/validate_lora/LTX-2.3-I2AV.py)|
+|[Lightricks/LTX-2.3: TwoStagePipeline-I2AV](https://www.modelscope.cn/models/Lightricks/LTX-2.3)|`input_images`|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2.3-I2AV-TwoStage.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2.3-I2AV-TwoStage.py)|-|-|-|-|
+|[Lightricks/LTX-2.3: DistilledPipeline-I2AV](https://www.modelscope.cn/models/Lightricks/LTX-2.3)|`input_images`|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2.3-I2AV-DistilledPipeline.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2.3-I2AV-DistilledPipeline.py)|-|-|-|-|
+|[Lightricks/LTX-2.3: OneStagePipeline-T2AV](https://www.modelscope.cn/models/Lightricks/LTX-2.3)||[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2.3-T2AV-OneStage.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2.3-T2AV-OneStage.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/full/LTX-2.3-T2AV-splited.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/validate_full/LTX-2.3-T2AV.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/lora/LTX-2.3-T2AV-splited.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/validate_lora/LTX-2.3-T2AV.py)|
+|[Lightricks/LTX-2.3: TwoStagePipeline-T2AV](https://www.modelscope.cn/models/Lightricks/LTX-2.3)||[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2.3-T2AV-TwoStage.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2.3-T2AV-TwoStage.py)|-|-|-|-|
+|[Lightricks/LTX-2.3: DistilledPipeline-T2AV](https://www.modelscope.cn/models/Lightricks/LTX-2.3)||[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2.3-T2AV-DistilledPipeline.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2.3-T2AV-DistilledPipeline.py)|-|-|-|-|
+|[Lightricks/LTX-2.3: A2V](https://www.modelscope.cn/models/Lightricks/LTX-2.3)|`retake_audio`,`audio_sample_rate`,`retake_audio_regions`|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2.3-A2V-TwoStage.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2.3-A2V-TwoStage.py)|-|-|-|-|
+|[Lightricks/LTX-2.3: Retake](https://www.modelscope.cn/models/Lightricks/LTX-2.3)|`retake_video`,`retake_video_regions`,`retake_audio`,`audio_sample_rate`,`retake_audio_regions`|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2.3-T2AV-TwoStage-Retake.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2.3-T2AV-TwoStage-Retake.py)|-|-|-|-|
+|[Lightricks/LTX-2.3-22b-IC-LoRA-Union-Control](https://www.modelscope.cn/models/Lightricks/LTX-2.3-22b-IC-LoRA-Union-Control)|`in_context_videos`,`in_context_downsample_factor`|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2.3-T2AV-IC-LoRA-Union-Control.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2.3-T2AV-IC-LoRA-Union-Control.py)|-|-|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/lora/LTX-2.3-T2AV-IC-LoRA-splited.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/validate_lora/LTX-2.3-T2AV-IC-LoRA.py)|
+|[Lightricks/LTX-2.3-22b-IC-LoRA-Motion-Track-Control](https://www.modelscope.cn/models/Lightricks/LTX-2.3-22b-IC-LoRA-Motion-Track-Control)|`in_context_videos`,`in_context_downsample_factor`|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2.3-T2AV-IC-LoRA-Motion-Track-Control.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2.3-T2AV-IC-LoRA-Motion-Track-Control.py)|-|-|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/lora/LTX-2.3-T2AV-IC-LoRA-splited.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/validate_lora/LTX-2.3-T2AV-IC-LoRA.py)|
 |[Lightricks/LTX-2: OneStagePipeline-T2AV](https://www.modelscope.cn/models/Lightricks/LTX-2)||[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2-T2AV-OneStage.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2-T2AV-OneStage.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/full/LTX-2-T2AV-splited.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/validate_full/LTX-2-T2AV.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/lora/LTX-2-T2AV-splited.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/validate_lora/LTX-2-T2AV.py)|
+|[Lightricks/LTX-2-19b-IC-LoRA-Union-Control](https://www.modelscope.cn/models/Lightricks/LTX-2-19b-IC-LoRA-Union-Control)|`in_context_videos`,`in_context_downsample_factor`|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2-T2AV-IC-LoRA-Union-Control.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2-T2AV-IC-LoRA-Union-Control.py)|-|-|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/lora/LTX-2-T2AV-IC-LoRA-splited.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/validate_lora/LTX-2-T2AV-IC-LoRA.py)|
+|[Lightricks/LTX-2-19b-IC-LoRA-Detailer](https://www.modelscope.cn/models/Lightricks/LTX-2-19b-IC-LoRA-Detailer)|`in_context_videos`,`in_context_downsample_factor`|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2-T2AV-IC-LoRA-Detailer.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2-T2AV-IC-LoRA-Detailer.py)|-|-|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/lora/LTX-2-T2AV-IC-LoRA-splited.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_training/validate_lora/LTX-2-T2AV-IC-LoRA.py)|
 |[Lightricks/LTX-2: TwoStagePipeline-T2AV](https://www.modelscope.cn/models/Lightricks/LTX-2)||[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2-T2AV-TwoStage.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2-T2AV-TwoStage.py)|-|-|-|-|
 |[Lightricks/LTX-2: DistilledPipeline-T2AV](https://www.modelscope.cn/models/Lightricks/LTX-2)||[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2-T2AV-DistilledPipeline.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2-T2AV-DistilledPipeline.py)|-|-|-|-|
 |[Lightricks/LTX-2: OneStagePipeline-I2AV](https://www.modelscope.cn/models/Lightricks/LTX-2)|`input_images`|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference/LTX-2-I2AV-OneStage.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/ltx2/model_inference_low_vram/LTX-2-I2AV-OneStage.py)|-|-|-|-|
@@ -205,7 +217,7 @@ LTX-2 系列模型统一通过 [`examples/ltx2/model_training/train.py`](https:/
 我们构建了一个样例视频数据集，以方便您进行测试，通过以下命令可以下载这个数据集：
 
 ```shell
-modelscope download --dataset DiffSynth-Studio/example_video_dataset --local_dir ./data/example_video_dataset
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --local_dir ./data/diffsynth_example_dataset
 ```
 
 我们为每个模型编写了推荐的训练脚本，请参考前文"模型总览"中的表格。关于如何编写模型训练脚本，请参考[模型训练](../Pipeline_Usage/Model_Training.md)；更多高阶训练算法，请参考[训练框架详解](https://github.com/modelscope/DiffSynth-Studio/tree/main/docs/zh/Training/)。

docs/zh/Model_Details/Qwen-Image.md

@@ -86,9 +86,11 @@ graph LR;
 |[Qwen/Qwen-Image-Edit-2509](https://www.modelscope.cn/models/Qwen/Qwen-Image-Edit-2509)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference/Qwen-Image-Edit-2509.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference_low_vram/Qwen-Image-Edit-2509.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/full/Qwen-Image-Edit-2509.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_full/Qwen-Image-Edit-2509.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/lora/Qwen-Image-Edit-2509.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_lora/Qwen-Image-Edit-2509.py)|
 |[Qwen/Qwen-Image-Edit-2511](https://www.modelscope.cn/models/Qwen/Qwen-Image-Edit-2511)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference/Qwen-Image-Edit-2511.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference_low_vram/Qwen-Image-Edit-2511.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/full/Qwen-Image-Edit-2511.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_full/Qwen-Image-Edit-2511.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/lora/Qwen-Image-Edit-2511.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_lora/Qwen-Image-Edit-2511.py)|
 |[FireRedTeam/FireRed-Image-Edit-1.0](https://www.modelscope.cn/models/FireRedTeam/FireRed-Image-Edit-1.0)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference/FireRed-Image-Edit-1.0.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference_low_vram/FireRed-Image-Edit-1.0.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/full/FireRed-Image-Edit-1.0.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_full/FireRed-Image-Edit-1.0.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/lora/FireRed-Image-Edit-1.0.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_lora/FireRed-Image-Edit-1.0.py)|
+|[FireRedTeam/FireRed-Image-Edit-1.1](https://www.modelscope.cn/models/FireRedTeam/FireRed-Image-Edit-1.1)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference/FireRed-Image-Edit-1.1.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference_low_vram/FireRed-Image-Edit-1.1.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/full/FireRed-Image-Edit-1.1.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_full/FireRed-Image-Edit-1.1.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/lora/FireRed-Image-Edit-1.1.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_lora/FireRed-Image-Edit-1.1.py)|
 |[lightx2v/Qwen-Image-Edit-2511-Lightning](https://modelscope.cn/models/lightx2v/Qwen-Image-Edit-2511-Lightning)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference/Qwen-Image-Edit-2511-Lightning.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference_low_vram/Qwen-Image-Edit-2511-Lightning.py)|-|-|-|-|
 |[Qwen/Qwen-Image-Layered](https://www.modelscope.cn/models/Qwen/Qwen-Image-Layered)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference/Qwen-Image-Layered.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference_low_vram/Qwen-Image-Layered.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/full/Qwen-Image-Layered.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_full/Qwen-Image-Layered.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/lora/Qwen-Image-Layered.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_lora/Qwen-Image-Layered.py)|
 |[DiffSynth-Studio/Qwen-Image-Layered-Control](https://www.modelscope.cn/models/DiffSynth-Studio/Qwen-Image-Layered-Control)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference/Qwen-Image-Layered-Control.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference_low_vram/Qwen-Image-Layered-Control.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/full/Qwen-Image-Layered-Control.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_full/Qwen-Image-Layered-Control.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/lora/Qwen-Image-Layered-Control.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_lora/Qwen-Image-Layered-Control.py)|
+|[DiffSynth-Studio/Qwen-Image-Layered-Control-V2](https://www.modelscope.cn/models/DiffSynth-Studio/Qwen-Image-Layered-Control-V2)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference/Qwen-Image-Layered-Control-V2.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference_low_vram/Qwen-Image-Layered-Control-V2.py)|-|-|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/lora/Qwen-Image-Layered-Control-V2.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_lora/Qwen-Image-Layered-Control-V2.py)|
 |[DiffSynth-Studio/Qwen-Image-EliGen](https://www.modelscope.cn/models/DiffSynth-Studio/Qwen-Image-EliGen)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference/Qwen-Image-EliGen.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference_low_vram/Qwen-Image-EliGen.py)|-|-|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/lora/Qwen-Image-EliGen.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_lora/Qwen-Image-EliGen.py)|
 |[DiffSynth-Studio/Qwen-Image-EliGen-V2](https://www.modelscope.cn/models/DiffSynth-Studio/Qwen-Image-EliGen-V2)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference/Qwen-Image-EliGen-V2.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference_low_vram/Qwen-Image-EliGen-V2.py)|-|-|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/lora/Qwen-Image-EliGen.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_lora/Qwen-Image-EliGen.py)|
 |[DiffSynth-Studio/Qwen-Image-EliGen-Poster](https://www.modelscope.cn/models/DiffSynth-Studio/Qwen-Image-EliGen-Poster)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference/Qwen-Image-EliGen-Poster.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_inference_low_vram/Qwen-Image-EliGen-Poster.py)|-|-|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/lora/Qwen-Image-EliGen-Poster.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/qwen_image/model_training/validate_lora/Qwen-Image-EliGen-Poster.py)|
@@ -197,7 +199,7 @@ Qwen-Image 系列模型统一通过 [`examples/qwen_image/model_training/train.p
 我们构建了一个样例图像数据集，以方便您进行测试，通过以下命令可以下载这个数据集：
 
 ```shell
-modelscope download --dataset DiffSynth-Studio/example_image_dataset --local_dir ./data/example_image_dataset
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --local_dir ./data/diffsynth_example_dataset
 ```
 
 我们为每个模型编写了推荐的训练脚本，请参考前文“模型总览”中的表格。关于如何编写模型训练脚本，请参考[模型训练](../Pipeline_Usage/Model_Training.md)；更多高阶训练算法，请参考[训练框架详解](https://github.com/modelscope/DiffSynth-Studio/tree/main/docs/zh/Training/)。

docs/zh/Model_Details/Wan.md

@@ -138,6 +138,8 @@ graph LR;
 |[PAI/Wan2.2-Fun-A14B-InP](https://modelscope.cn/models/PAI/Wan2.2-Fun-A14B-InP)|`input_image`, `end_image`|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_inference/Wan2.2-Fun-A14B-InP.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_training/full/Wan2.2-Fun-A14B-InP.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_training/validate_full/Wan2.2-Fun-A14B-InP.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_training/lora/Wan2.2-Fun-A14B-InP.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_training/validate_lora/Wan2.2-Fun-A14B-InP.py)|
 |[PAI/Wan2.2-Fun-A14B-Control](https://modelscope.cn/models/PAI/Wan2.2-Fun-A14B-Control)|`control_video`, `reference_image`|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_inference/Wan2.2-Fun-A14B-Control.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_training/full/Wan2.2-Fun-A14B-Control.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_training/validate_full/Wan2.2-Fun-A14B-Control.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_training/lora/Wan2.2-Fun-A14B-Control.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_training/validate_lora/Wan2.2-Fun-A14B-Control.py)|
 |[PAI/Wan2.2-Fun-A14B-Control-Camera](https://modelscope.cn/models/PAI/Wan2.2-Fun-A14B-Control-Camera)|`control_camera_video`, `input_image`|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_inference/Wan2.2-Fun-A14B-Control-Camera.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_training/full/Wan2.2-Fun-A14B-Control-Camera.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_training/validate_full/Wan2.2-Fun-A14B-Control-Camera.py)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_training/lora/Wan2.2-Fun-A14B-Control-Camera.sh)|[code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/wanvideo/model_training/validate_lora/Wan2.2-Fun-A14B-Control-Camera.py)|
+| [openmoss/MOVA-360p](https://modelscope.cn/models/openmoss/MOVA-360p) | `input_image` | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/mova/model_inference/MOVA-360p-I2AV.py) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/mova/model_training/full/MOVA-360P-I2AV.sh) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/mova/model_training/validate_full/MOVA-360p-I2AV.py) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/mova/model_training/lora/MOVA-360P-I2AV.sh) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/mova/model_training/validate_lora/MOVA-360p-I2AV.py) |
+| [openmoss/MOVA-720p](https://modelscope.cn/models/openmoss/MOVA-720p) | `input_image` | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/mova/model_inference/MOVA-720p-I2AV.py) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/mova/model_training/full/MOVA-720P-I2AV.sh) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/mova/model_training/validate_full/MOVA-720p-I2AV.py) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/mova/model_training/lora/MOVA-720P-I2AV.sh) | [code](https://github.com/modelscope/DiffSynth-Studio/blob/main/examples/mova/model_training/validate_lora/MOVA-720p-I2AV.py) |
 
 * FP8 精度训练：[doc](../Training/FP8_Precision.md)、[code](https://github.com/modelscope/DiffSynth-Studio/tree/main/examples/wanvideo/model_training/special/fp8_training/)
 * 两阶段拆分训练：[doc](../Training/Split_Training.md)、[code](https://github.com/modelscope/DiffSynth-Studio/tree/main/examples/wanvideo/model_training/special/split_training/)
@@ -252,7 +254,7 @@ Wan 系列模型统一通过 [`examples/wanvideo/model_training/train.py`](https
 我们构建了一个样例视频数据集，以方便您进行测试，通过以下命令可以下载这个数据集：
 
 ```shell
-modelscope download --dataset DiffSynth-Studio/example_video_dataset --local_dir ./data/example_video_dataset
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --local_dir ./data/diffsynth_example_dataset
 ```
 
 我们为每个模型编写了推荐的训练脚本，请参考前文"模型总览"中的表格。关于如何编写模型训练脚本，请参考[模型训练](../Pipeline_Usage/Model_Training.md)；更多高阶训练算法，请参考[训练框架详解](https://github.com/modelscope/DiffSynth-Studio/tree/main/docs/zh/Training/)。

docs/zh/Model_Details/Z-Image.md

@@ -134,7 +134,7 @@ Z-Image 系列模型统一通过 [`examples/z_image/model_training/train.py`](ht
 我们构建了一个样例图像数据集，以方便您进行测试，通过以下命令可以下载这个数据集：
 
 ```shell
-modelscope download --dataset DiffSynth-Studio/example_image_dataset --local_dir ./data/example_image_dataset
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --local_dir ./data/diffsynth_example_dataset
 ```
 
 我们为每个模型编写了推荐的训练脚本，请参考前文"模型总览"中的表格。关于如何编写模型训练脚本，请参考[模型训练](../Pipeline_Usage/Model_Training.md)；更多高阶训练算法，请参考[训练框架详解](https://github.com/modelscope/DiffSynth-Studio/tree/main/docs/zh/Training/)。

docs/zh/Pipeline_Usage/GPU_support.md

@@ -89,4 +89,5 @@ export CPU_AFFINITY_CONF=1
 | 模型        | 参数 | 备注                |
 |-----------|------|-------------------|
 | Wan 14B系列 | --initialize_model_on_cpu | 14B模型需要在cpu上进行初始化 |
-| Qwen-Image系列 | --initialize_model_on_cpu | 模型需要在cpu上进行初始化 |
+| Qwen-Image系列 | --initialize_model_on_cpu | 模型需要在cpu上进行初始化 |
+| Z-Image 系列 | --enable_npu_patch | 使用NPU融合算子来替换Z-image模型中的对应算子以提升模型在NPU上的性能 |

docs/zh/Pipeline_Usage/Model_Training.md

@@ -69,28 +69,16 @@ image_2.jpg,"a cat"
 
 <details>
 
-<summary>样例图像数据集</summary>
+<summary>样例数据集</summary>
 
 > ```shell
-> modelscope download --dataset DiffSynth-Studio/example_image_dataset --local_dir ./data/example_image_dataset
+> modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --local_dir ./data/diffsynth_example_dataset
 > ```
 > 
 > 适用于 Qwen-Image、FLUX 等图像生成模型的训练。
 
 </details>
 
-<details>
-
-<summary>样例视频数据集</summary>
-
-> ```shell
-> modelscope download --dataset DiffSynth-Studio/example_video_dataset --local_dir ./data/example_video_dataset
-> ```
-> 
-> 适用于 Wan 等视频生成模型的训练。
-
-</details>
-
 ## 加载模型
 
 类似于[推理时的模型加载](../Pipeline_Usage/Model_Inference.md#加载模型)，我们支持多种方式配置模型路径，两种方式是可以混用的。
@@ -243,3 +231,116 @@ accelerate launch --config_file examples/qwen_image/model_training/full/accelera
 * 少数模型包含冗余参数，例如 Qwen-Image 的 DiT 部分最后一层的文本编码部分，在训练这些模型时，需设置 `--find_unused_parameters` 避免在多 GPU 训练中报错。出于对开源社区模型兼容性的考虑，我们不打算删除这些冗余参数。
 * Diffusion 模型的损失函数值与实际效果的关系不大，因此我们在训练过程中不会记录损失函数值。我们建议把 `--num_epochs` 设置为足够大的数值，边训边测，直至效果收敛后手动关闭训练程序。
 * `--use_gradient_checkpointing` 通常是开启的，除非 GPU 显存足够；`--use_gradient_checkpointing_offload` 则按需开启，详见 [`diffsynth.core.gradient`](../API_Reference/core/gradient.md)。
+
+## 低显存训练
+如果想在低显存显卡上完成 LoRA 模型训练，可以同时采用 [两阶段拆分训练](../Training/Split_Training.md) 和 `deepspeed_zero3_offload` 训练。 首先，将前处理过程拆分到第一阶段，将计算结果存储到硬盘中。其次，在第二阶段从硬盘中读取这些结果并进行去噪模型的训练，训练通过采用 `deepspeed_zero3_offload`，将训练参数和优化器状态 offload 到 cpu 或者 disk 上。我们为部分模型提供了样例，主要是通过 `--config_file` 指定 `deepspeed` 配置。
+
+需要注意的是，`deepspeed_zero3_offload` 模式与 `pytorch` 原生的梯度检查点机制不兼容，我们为此对 `deepspeed` 的`checkpointing` 接口做了适配。用户需要在 `deepspeed` 配置中填写 `activation_checkpointing` 字段以启用梯度检查点。
+
+以下为 Qwen-Image 模型的低显存模型训练脚本：
+```shell
+accelerate launch examples/qwen_image/model_training/train.py \
+  --dataset_base_path data/example_image_dataset \
+  --dataset_metadata_path data/example_image_dataset/metadata.csv \
+  --max_pixels 1048576 \
+  --dataset_repeat 1 \
+  --model_id_with_origin_paths "Qwen/Qwen-Image:text_encoder/model*.safetensors,Qwen/Qwen-Image:vae/diffusion_pytorch_model.safetensors" \
+  --learning_rate 1e-4 \
+  --num_epochs 5 \
+  --remove_prefix_in_ckpt "pipe.dit." \
+  --output_path "./models/train/Qwen-Image_lora-splited-cache" \
+  --lora_base_model "dit" \
+  --lora_target_modules "to_q,to_k,to_v,add_q_proj,add_k_proj,add_v_proj,to_out.0,to_add_out,img_mlp.net.2,img_mod.1,txt_mlp.net.2,txt_mod.1" \
+  --lora_rank 32 \
+  --task "sft:data_process" \
+  --use_gradient_checkpointing \
+  --dataset_num_workers 8 \
+  --find_unused_parameters
+
+accelerate launch --config_file examples/qwen_image/model_training/special/low_vram_training/deepspeed_zero3_cpuoffload.yaml examples/qwen_image/model_training/train.py \
+  --dataset_base_path "./models/train/Qwen-Image_lora-splited-cache" \
+  --max_pixels 1048576 \
+  --dataset_repeat 50 \
+  --model_id_with_origin_paths "Qwen/Qwen-Image:transformer/diffusion_pytorch_model*.safetensors" \
+  --learning_rate 1e-4 \
+  --num_epochs 5 \
+  --remove_prefix_in_ckpt "pipe.dit." \
+  --output_path "./models/train/Qwen-Image_lora" \
+  --lora_base_model "dit" \
+  --lora_target_modules "to_q,to_k,to_v,add_q_proj,add_k_proj,add_v_proj,to_out.0,to_add_out,img_mlp.net.2,img_mod.1,txt_mlp.net.2,txt_mod.1" \
+  --lora_rank 32 \
+  --task "sft:train" \
+  --use_gradient_checkpointing \
+  --dataset_num_workers 8 \
+  --find_unused_parameters \
+  --initialize_model_on_cpu
+```
+
+其中，`accelerate` 和 `deepspeed` 的配置文件如下：
+
+```yaml
+compute_environment: LOCAL_MACHINE
+debug: true
+deepspeed_config:
+  deepspeed_config_file: examples/qwen_image/model_training/special/low_vram_training/ds_z3_cpuoffload.json
+  zero3_init_flag: true
+distributed_type: DEEPSPEED
+downcast_bf16: 'no'
+enable_cpu_affinity: false
+machine_rank: 0
+main_training_function: main
+num_machines: 1
+num_processes: 1
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false
+```
+
+```json
+{
+    "fp16": {
+        "enabled": "auto",
+        "loss_scale": 0,
+        "loss_scale_window": 1000,
+        "initial_scale_power": 16,
+        "hysteresis": 2,
+        "min_loss_scale": 1
+    },
+    "bf16": {
+        "enabled": "auto"
+    },
+    "zero_optimization": {
+        "stage": 3,
+        "offload_optimizer": {
+            "device": "cpu",
+            "pin_memory": true
+        },
+        "offload_param": {
+            "device": "cpu",
+            "pin_memory": true
+        },
+        "overlap_comm": false,
+        "contiguous_gradients": true,
+        "sub_group_size": 1e9,
+        "reduce_bucket_size": 5e7,
+        "stage3_prefetch_bucket_size": 5e7,
+        "stage3_param_persistence_threshold": 1e5,
+        "stage3_max_live_parameters": 1e8,
+        "stage3_max_reuse_distance": 1e8,
+        "stage3_gather_16bit_weights_on_model_save": true
+    },
+    "activation_checkpointing": {
+        "partition_activations": false,
+        "cpu_checkpointing": false,
+        "contiguous_memory_optimization": false
+    },
+    "gradient_accumulation_steps": "auto",
+    "gradient_clipping": "auto",
+    "train_batch_size": "auto",
+    "train_micro_batch_size_per_gpu": "auto",
+    "wall_clock_breakdown": false
+}
+```

docs/zh/Pipeline_Usage/Setup.md

@@ -37,9 +37,9 @@ pip install torch torchvision --index-url https://download.pytorch.org/whl/rocm6
    git clone https://github.com/modelscope/DiffSynth-Studio.git
    cd DiffSynth-Studio
    # aarch64/ARM
-   pip install -e .[npu_aarch64] --extra-index-url "https://download.pytorch.org/whl/cpu"
+   pip install -e .[npu_aarch64] 
    # x86
-   pip install -e .[npu]
+   pip install -e .[npu] --extra-index-url "https://download.pytorch.org/whl/cpu"
 
 使用 Ascend NPU 时，请将 Python 代码中的 `"cuda"` 改为 `"npu"`，详见[NPU 支持](../Pipeline_Usage/GPU_support.md#ascend-npu)。
 

docs/zh/README.md

@@ -42,6 +42,8 @@ graph LR;
 * [Qwen-Image](./Model_Details/Qwen-Image.md)
 * [FLUX.2](./Model_Details/FLUX2.md)
 * [Z-Image](./Model_Details/Z-Image.md)
+* [Anima](./Model_Details/Anima.md)
+* [LTX-2](./Model_Details/LTX-2.md)
 
 ## Section 3: 训练框架
 
@@ -78,7 +80,7 @@ graph LR;
 本节介绍如何利用 `DiffSynth-Studio` 训练新的模型，帮助科研工作者探索新的模型技术。
 
 * [从零开始训练模型](./Research_Tutorial/train_from_scratch.md)
-* 推理改进优化技术【coming soon】
+* [推理改进优化技术](./Research_Tutorial/inference_time_scaling.md)
 * 设计可控生成模型【coming soon】
 * 创建新的训练范式【coming soon】
 

docs/zh/Research_Tutorial/inference_time_scaling.ipynb

@@ -0,0 +1,236 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "id": "8db54992",
+   "metadata": {},
+   "source": [
+    "# 推理改进优化技术\n",
+    "\n",
+    "DiffSynth-Studio 旨在以基础框架驱动技术创新。本文以 Inference-time scaling 为例，展示如何基于 DiffSynth-Studio 构建免训练（Training-free）的图像生成增强方案。"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "0911cad4",
+   "metadata": {},
+   "source": [
+    "## 1. 图像质量量化\n",
+    "\n",
+    "首先，我们需要找到一个指标来量化图像生成模型生成的图像质量。最简单直接的方案是人工打分，但这样做的成本太高，无法大规模使用。不过，收集人工打分后，训练一个图像分类模型来预测人类的打分结果，是完全可行的。PickScore [[1]](https://arxiv.org/abs/2305.01569) 就是这样一个模型，运行下面的代码，将会自动下载并加载 [PickScore 模型](https://modelscope.cn/models/AI-ModelScope/PickScore_v1)。"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "4faca4ca",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from modelscope import AutoProcessor, AutoModel\n",
+    "import torch\n",
+    "\n",
+    "class PickScore(torch.nn.Module):\n",
+    "    def __init__(self):\n",
+    "        super().__init__()\n",
+    "        self.processor = AutoProcessor.from_pretrained(\"laion/CLIP-ViT-H-14-laion2B-s32B-b79K\")\n",
+    "        self.model = AutoModel.from_pretrained(\"AI-ModelScope/PickScore_v1\").eval().to(\"cuda\")\n",
+    "\n",
+    "    def forward(self, image, prompt):\n",
+    "        image_inputs = self.processor(images=image, padding=True, truncation=True, max_length=77, return_tensors=\"pt\").to(\"cuda\")\n",
+    "        text_inputs = self.processor(text=prompt, padding=True, truncation=True, max_length=77, return_tensors=\"pt\").to(\"cuda\")\n",
+    "        with torch.inference_mode():\n",
+    "            image_embs = self.model.get_image_features(**image_inputs).pooler_output\n",
+    "            image_embs = image_embs / torch.norm(image_embs, dim=-1, keepdim=True)\n",
+    "            text_embs = self.model.get_text_features(**text_inputs).pooler_output\n",
+    "            text_embs = text_embs / torch.norm(text_embs, dim=-1, keepdim=True)\n",
+    "            score = (text_embs @ image_embs.T).flatten().item()\n",
+    "        return score\n",
+    "\n",
+    "reward_model = PickScore()"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "5f807cec",
+   "metadata": {},
+   "source": [
+    "## 2. Inference-time Scaling 技术\n",
+    "\n",
+    "Inference-time Scaling [[2]](https://arxiv.org/abs/2504.00294) 是一类有趣的技术，旨在通过增加推理时的计算量来提升生成结果的质量。例如，在语言模型中，[Qwen/Qwen3.5-27B](https://modelscope.cn/models/Qwen/Qwen3.5-27B)、[deepseek-ai/DeepSeek-R1](deepseek-ai/DeepSeek-R1) 等模型通过“思考模式”引导模型花更多时间仔细思考，让回答结果更准确。接下来我们以模型 [black-forest-labs/FLUX.2-klein-4B](https://modelscope.cn/models/black-forest-labs/FLUX.2-klein-4B) 为例，探讨如何为图像生成模型设计 Inference-time Scaling 方案。\n",
+    "\n",
+    "> 在开始前，我们稍微改造了 `Flux2ImagePipeline` 的代码，使其能够根据输入的特定高斯噪声矩阵进行初始化，便于复现结果，详见 [diffsynth/pipelines/flux2_image.py](https://github.com/modelscope/DiffSynth-Studio/blob/main/diffsynth/pipelines/flux2_image.py) 中的 `Flux2Unit_NoiseInitializer`。\n",
+    "\n",
+    "运行以下代码，加载模型 [black-forest-labs/FLUX.2-klein-4B](https://modelscope.cn/models/black-forest-labs/FLUX.2-klein-4B)。"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "c5818a87",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from diffsynth.pipelines.flux2_image import Flux2ImagePipeline, ModelConfig\n",
+    "\n",
+    "pipe = Flux2ImagePipeline.from_pretrained(\n",
+    "    torch_dtype=torch.bfloat16,\n",
+    "    device=\"cuda\",\n",
+    "    model_configs=[\n",
+    "        ModelConfig(model_id=\"black-forest-labs/FLUX.2-klein-4B\", origin_file_pattern=\"text_encoder/*.safetensors\"),\n",
+    "        ModelConfig(model_id=\"black-forest-labs/FLUX.2-klein-4B\", origin_file_pattern=\"transformer/*.safetensors\"),\n",
+    "        ModelConfig(model_id=\"black-forest-labs/FLUX.2-klein-4B\", origin_file_pattern=\"vae/diffusion_pytorch_model.safetensors\"),\n",
+    "    ],\n",
+    "    tokenizer_config=ModelConfig(model_id=\"black-forest-labs/FLUX.2-klein-4B\", origin_file_pattern=\"tokenizer/\"),\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "f58e9945",
+   "metadata": {},
+   "source": [
+    "用提示词 `\"sketch, a cat\"` 生成一只素描猫猫，并用 PickScore 模型打分。"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "6ea2d258",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def evaluate_noise(noise, pipe, reward_model, prompt):\n",
+    "    # Generate an image and compute the score.\n",
+    "    image = pipe(\n",
+    "        prompt=prompt,\n",
+    "        num_inference_steps=4,\n",
+    "        initial_noise=noise,\n",
+    "        progress_bar_cmd=lambda x: x,\n",
+    "    )\n",
+    "    score = reward_model(image, prompt)\n",
+    "    return score\n",
+    "\n",
+    "torch.manual_seed(1)\n",
+    "prompt = \"sketch, a cat\"\n",
+    "noise = pipe.generate_noise((1, 128, 64, 64), rand_device=\"cuda\", rand_torch_dtype=pipe.torch_dtype)\n",
+    "\n",
+    "image_1 = pipe(prompt, num_inference_steps=4, initial_noise=noise)\n",
+    "print(\"Score:\", reward_model(image_1, prompt))\n",
+    "image_1"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "5e11694e",
+   "metadata": {},
+   "source": [
+    "### 2.1 Best-of-N 随机搜索\n",
+    "\n",
+    "模型的生成结果具有一定的随机性，如果用不同的随机种子，生成的图像结果也是不同的，有时图像质量高，有时图像质量低。那么，我们有一个简单的 Inference-time scaling 方案：使用多个不同的随机种子分别生成图像，然后利用 PickScore 进行打分，只保留分数最高的那一张。"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "241f10d2",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from tqdm import tqdm\n",
+    "\n",
+    "def random_search(base_latents, objective_reward_fn, total_eval_budget):\n",
+    "    # Search for the noise randomly.\n",
+    "    best_noise = base_latents\n",
+    "    best_score = objective_reward_fn(base_latents)\n",
+    "    for it in tqdm(range(total_eval_budget - 1)):\n",
+    "        noise = pipe.generate_noise((1, 128, 64, 64), seed=None)\n",
+    "        score = objective_reward_fn(noise)\n",
+    "        if score > best_score:\n",
+    "            best_score, best_noise = score, noise\n",
+    "    return best_noise\n",
+    "\n",
+    "best_noise = random_search(\n",
+    "    base_latents=noise,\n",
+    "    objective_reward_fn=lambda noise: evaluate_noise(noise, pipe, reward_model, prompt),\n",
+    "    total_eval_budget=50,\n",
+    ")\n",
+    "image_2 = pipe(prompt, num_inference_steps=4, initial_noise=best_noise)\n",
+    "print(\"Score:\", reward_model(image_2, prompt))\n",
+    "image_2"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "8e9bf966",
+   "metadata": {},
+   "source": [
+    "我们可以清晰地看到，经过多次随机搜索后，最终选出的猫猫毛发细节更加丰富，PickScore 分数也有明显提升。但这种暴力的随机搜索效率极低，生成时间成倍增长，且很容易触及质量上限。因此，我们希望能够找到一种更高效的搜索方法，在同等计算预算下达到更高的分数。"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "c9578349",
+   "metadata": {},
+   "source": [
+    "### 2.2 SES 搜索\n",
+    "\n",
+    "为了突破随机搜索的瓶颈，我们引入了 SES (Spectral Evolution Search) 算法 [[3]](https://arxiv.org/abs/2602.03208)，详细的代码位于 [diffsynth/utils/ses](https://github.com/modelscope/DiffSynth-Studio/blob/main/diffsynth/utils/ses)。\n",
+    "\n",
+    "扩散模型生成的图像，很大程度上由初始噪声的低频分量决定。SES 算法通过小波变换将高斯噪声分解，固定高频细节，专门针对低频部分使用交叉熵方法进行演化搜索，能以更高的效率找到优质的初始噪声。\n",
+    "\n",
+    "运行下面的代码，即可使用 SES 更高效地搜索最佳的高斯噪声矩阵。"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "adeed2aa",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from diffsynth.utils.ses import ses_search\n",
+    "\n",
+    "best_noise = ses_search(\n",
+    "    base_latents=noise,\n",
+    "    objective_reward_fn=lambda noise: evaluate_noise(noise, pipe, reward_model, prompt),\n",
+    "    total_eval_budget=50,\n",
+    ")\n",
+    "image_3 = pipe(prompt, num_inference_steps=4, initial_noise=best_noise)\n",
+    "print(\"Score:\", reward_model(image_3, prompt))\n",
+    "image_3"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "940a97f1",
+   "metadata": {},
+   "source": [
+    "可以观察到，在同样的计算预算下，相比于随机搜索，SES 的结果在 PickScore 得分上取得了显著的提升。“素描猫猫”展现出了更精致的整体构图以及更具层次感的明暗对比。\n",
+    "\n",
+    "Inference-time scaling 能够以更长推理时间为代价获得更高的图像质量，那么它生成的图像数据也可以用 DPO [[4]](https://arxiv.org/abs/2311.12908)、差分训练 [[5]](https://arxiv.org/abs/2412.12888) 等方式赋予模型自身，那就是另外一个有趣的探索方向了。"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "dzj8",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.19"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

docs/zh/Research_Tutorial/inference_time_scaling.md

@@ -0,0 +1,140 @@
+# 推理改进优化技术
+
+DiffSynth-Studio 旨在以基础框架驱动技术创新。本文以 Inference-time scaling 为例，展示如何基于 DiffSynth-Studio 构建免训练（Training-free）的图像生成增强方案。
+
+Notebook: https://github.com/modelscope/DiffSynth-Studio/blob/main/docs/zh/Research_Tutorial/inference_time_scaling.ipynb
+
+## 1. 图像质量量化
+
+首先，我们需要找到一个指标来量化图像生成模型生成的图像质量。最简单直接的方案是人工打分，但这样做的成本太高，无法大规模使用。不过，收集人工打分后，训练一个图像分类模型来预测人类的打分结果，是完全可行的。PickScore [[1]](https://arxiv.org/abs/2305.01569) 就是这样一个模型，运行下面的代码，将会自动下载并加载 [PickScore 模型](https://modelscope.cn/models/AI-ModelScope/PickScore_v1)。
+
+```python
+from modelscope import AutoProcessor, AutoModel
+import torch
+
+class PickScore(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.processor = AutoProcessor.from_pretrained("laion/CLIP-ViT-H-14-laion2B-s32B-b79K")
+        self.model = AutoModel.from_pretrained("AI-ModelScope/PickScore_v1").eval().to("cuda")
+
+    def forward(self, image, prompt):
+        image_inputs = self.processor(images=image, padding=True, truncation=True, max_length=77, return_tensors="pt").to("cuda")
+        text_inputs = self.processor(text=prompt, padding=True, truncation=True, max_length=77, return_tensors="pt").to("cuda")
+        with torch.inference_mode():
+            image_embs = self.model.get_image_features(**image_inputs).pooler_output
+            image_embs = image_embs / torch.norm(image_embs, dim=-1, keepdim=True)
+            text_embs = self.model.get_text_features(**text_inputs).pooler_output
+            text_embs = text_embs / torch.norm(text_embs, dim=-1, keepdim=True)
+            score = (text_embs @ image_embs.T).flatten().item()
+        return score
+
+reward_model = PickScore()
+```
+
+## 2. Inference-time Scaling 技术
+
+Inference-time Scaling [[2]](https://arxiv.org/abs/2504.00294) 是一类有趣的技术，旨在通过增加推理时的计算量来提升生成结果的质量。例如，在语言模型中，[Qwen/Qwen3.5-27B](https://modelscope.cn/models/Qwen/Qwen3.5-27B)、[deepseek-ai/DeepSeek-R1](deepseek-ai/DeepSeek-R1) 等模型通过“思考模式”引导模型花更多时间仔细思考，让回答结果更准确。接下来我们以模型 [black-forest-labs/FLUX.2-klein-4B](https://modelscope.cn/models/black-forest-labs/FLUX.2-klein-4B) 为例，探讨如何为图像生成模型设计 Inference-time Scaling 方案。
+
+> 在开始前，我们稍微改造了 `Flux2ImagePipeline` 的代码，使其能够根据输入的特定高斯噪声矩阵进行初始化，便于复现结果，详见 [diffsynth/pipelines/flux2_image.py](https://github.com/modelscope/DiffSynth-Studio/blob/main/diffsynth/pipelines/flux2_image.py) 中的 `Flux2Unit_NoiseInitializer`。
+
+运行以下代码，加载模型 [black-forest-labs/FLUX.2-klein-4B](https://modelscope.cn/models/black-forest-labs/FLUX.2-klein-4B)。
+
+```python
+from diffsynth.pipelines.flux2_image import Flux2ImagePipeline, ModelConfig
+
+pipe = Flux2ImagePipeline.from_pretrained(
+    torch_dtype=torch.bfloat16,
+    device="cuda",
+    model_configs=[
+        ModelConfig(model_id="black-forest-labs/FLUX.2-klein-4B", origin_file_pattern="text_encoder/*.safetensors"),
+        ModelConfig(model_id="black-forest-labs/FLUX.2-klein-4B", origin_file_pattern="transformer/*.safetensors"),
+        ModelConfig(model_id="black-forest-labs/FLUX.2-klein-4B", origin_file_pattern="vae/diffusion_pytorch_model.safetensors"),
+    ],
+    tokenizer_config=ModelConfig(model_id="black-forest-labs/FLUX.2-klein-4B", origin_file_pattern="tokenizer/"),
+)
+```
+
+用提示词 `"sketch, a cat"` 生成一只素描猫猫，并用 PickScore 模型打分。
+
+```python
+def evaluate_noise(noise, pipe, reward_model, prompt):
+    # Generate an image and compute the score.
+    image = pipe(
+        prompt=prompt,
+        num_inference_steps=4,
+        initial_noise=noise,
+        progress_bar_cmd=lambda x: x,
+    )
+    score = reward_model(image, prompt)
+    return score
+
+torch.manual_seed(1)
+prompt = "sketch, a cat"
+noise = pipe.generate_noise((1, 128, 64, 64), rand_device="cuda", rand_torch_dtype=pipe.torch_dtype)
+
+image_1 = pipe(prompt, num_inference_steps=4, initial_noise=noise)
+print("Score:", reward_model(image_1, prompt))
+image_1
+```
+
+![Image](https://github.com/user-attachments/assets/b6546c6d-b368-4463-b703-d561a9134ba0)
+
+### 2.1 Best-of-N 随机搜索
+
+模型的生成结果具有一定的随机性，如果用不同的随机种子，生成的图像结果也是不同的，有时图像质量高，有时图像质量低。那么，我们有一个简单的 Inference-time scaling 方案：使用多个不同的随机种子分别生成图像，然后利用 PickScore 进行打分，只保留分数最高的那一张。
+
+```python
+from tqdm import tqdm
+
+def random_search(base_latents, objective_reward_fn, total_eval_budget):
+    # Search for the noise randomly.
+    best_noise = base_latents
+    best_score = objective_reward_fn(base_latents)
+    for it in tqdm(range(total_eval_budget - 1)):
+        noise = pipe.generate_noise((1, 128, 64, 64), seed=None)
+        score = objective_reward_fn(noise)
+        if score > best_score:
+            best_score, best_noise = score, noise
+    return best_noise
+
+best_noise = random_search(
+    base_latents=noise,
+    objective_reward_fn=lambda noise: evaluate_noise(noise, pipe, reward_model, prompt),
+    total_eval_budget=50,
+)
+image_2 = pipe(prompt, num_inference_steps=4, initial_noise=best_noise)
+print("Score:", reward_model(image_2, prompt))
+image_2
+```
+
+![Image](https://github.com/user-attachments/assets/b8dba70a-daa8-4368-8f32-a6c150daecb5)
+
+我们可以清晰地看到，经过多次随机搜索后，最终选出的猫猫毛发细节更加丰富，PickScore 分数也有明显提升。但这种暴力的随机搜索效率极低，生成时间成倍增长，且很容易触及质量上限。因此，我们希望能够找到一种更高效的搜索方法，在同等计算预算下达到更高的分数。
+
+### 2.2 SES 搜索
+
+为了突破随机搜索的瓶颈，我们引入了 SES (Spectral Evolution Search) 算法 [[3]](https://arxiv.org/abs/2602.03208)，详细的代码位于 [diffsynth/utils/ses](https://github.com/modelscope/DiffSynth-Studio/blob/main/diffsynth/utils/ses)。
+
+扩散模型生成的图像，很大程度上由初始噪声的低频分量决定。SES 算法通过小波变换将高斯噪声分解，固定高频细节，专门针对低频部分使用交叉熵方法进行演化搜索，能以更高的效率找到优质的初始噪声。
+
+运行下面的代码，即可使用 SES 更高效地搜索最佳的高斯噪声矩阵。
+
+```python
+from diffsynth.utils.ses import ses_search
+
+best_noise = ses_search(
+    base_latents=noise,
+    objective_reward_fn=lambda noise: evaluate_noise(noise, pipe, reward_model, prompt),
+    total_eval_budget=50,
+)
+image_3 = pipe(prompt, num_inference_steps=4, initial_noise=best_noise)
+print("Score:", reward_model(image_3, prompt))
+image_3
+```
+
+![Image](https://github.com/user-attachments/assets/9a3f7598-3812-46d2-b333-cd65e49886ab)
+
+可以观察到，在同样的计算预算下，相比于随机搜索，SES 的结果在 PickScore 得分上取得了显著的提升。“素描猫猫”展现出了更精致的整体构图以及更具层次感的明暗对比。
+
+Inference-time scaling 能够以更长推理时间为代价获得更高的图像质量，那么它生成的图像数据也可以用 DPO [[4]](https://arxiv.org/abs/2311.12908)、差分训练 [[5]](https://arxiv.org/abs/2412.12888) 等方式赋予模型自身，那就是另外一个有趣的探索方向了。

docs/zh/Training/Direct_Distill.md

@@ -77,7 +77,7 @@ distill_qwen/image.jpg,"精致肖像，水下少女，蓝裙飘逸，发丝轻
 这个样例数据集可以直接下载：
 
 ```shell
-modelscope download --dataset DiffSynth-Studio/example_image_dataset --local_dir ./data/example_image_dataset
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --local_dir ./data/diffsynth_example_dataset
 ```
 
 然后开始 LoRA 蒸馏加速训练：

docs/zh/index.rst

@@ -27,6 +27,8 @@
    Model_Details/Qwen-Image
    Model_Details/FLUX2
    Model_Details/Z-Image
+   Model_Details/Anima
+   Model_Details/LTX-2
 
 .. toctree::
    :maxdepth: 2
@@ -63,6 +65,7 @@
    :caption: 学术导引
 
    Research_Tutorial/train_from_scratch
+   Research_Tutorial/inference_time_scaling
 
 .. toctree::
    :maxdepth: 2

examples/anima/model_inference/anima-preview.py

@@ -0,0 +1,19 @@
+from diffsynth.pipelines.anima_image import AnimaImagePipeline, ModelConfig
+import torch
+
+
+pipe = AnimaImagePipeline.from_pretrained(
+    torch_dtype=torch.bfloat16,
+    device="cuda",
+    model_configs=[
+        ModelConfig(model_id="circlestone-labs/Anima", origin_file_pattern="split_files/diffusion_models/anima-preview.safetensors"),
+        ModelConfig(model_id="circlestone-labs/Anima", origin_file_pattern="split_files/text_encoders/qwen_3_06b_base.safetensors"),
+        ModelConfig(model_id="circlestone-labs/Anima", origin_file_pattern="split_files/vae/qwen_image_vae.safetensors"),
+    ],
+    tokenizer_config=ModelConfig(model_id="Qwen/Qwen3-0.6B", origin_file_pattern="./"),
+    tokenizer_t5xxl_config=ModelConfig(model_id="stabilityai/stable-diffusion-3.5-large", origin_file_pattern="tokenizer_3/")
+)
+prompt = "Masterpiece, best quality, solo, long hair, wavy hair, silver hair, blue eyes, blue dress, medium breasts, dress, underwater, air bubble, floating hair, refraction, portrait."
+negative_prompt = "worst quality, low quality, monochrome, zombie, interlocked fingers, Aissist, cleavage, nsfw,"
+image = pipe(prompt, seed=0, num_inference_steps=50)
+image.save("image.jpg")

examples/anima/model_inference_low_vram/anima-preview.py

@@ -0,0 +1,30 @@
+from diffsynth.pipelines.anima_image import AnimaImagePipeline, ModelConfig
+import torch
+
+
+vram_config = {
+    "offload_dtype": "disk",
+    "offload_device": "disk",
+    "onload_dtype": "disk",
+    "onload_device": "disk",
+    "preparing_dtype": torch.bfloat16,
+    "preparing_device": "cuda",
+    "computation_dtype": torch.bfloat16,
+    "computation_device": "cuda",
+}
+pipe = AnimaImagePipeline.from_pretrained(
+    torch_dtype=torch.bfloat16,
+    device="cuda",
+    model_configs=[
+        ModelConfig(model_id="circlestone-labs/Anima", origin_file_pattern="split_files/diffusion_models/anima-preview.safetensors", **vram_config),
+        ModelConfig(model_id="circlestone-labs/Anima", origin_file_pattern="split_files/text_encoders/qwen_3_06b_base.safetensors", **vram_config),
+        ModelConfig(model_id="circlestone-labs/Anima", origin_file_pattern="split_files/vae/qwen_image_vae.safetensors", **vram_config),
+    ],
+    tokenizer_config=ModelConfig(model_id="Qwen/Qwen3-0.6B", origin_file_pattern="./"),
+    tokenizer_t5xxl_config=ModelConfig(model_id="stabilityai/stable-diffusion-3.5-large", origin_file_pattern="tokenizer_3/"),
+    vram_limit=torch.cuda.mem_get_info("cuda")[1] / (1024 ** 3) - 0.5,
+)
+prompt = "Masterpiece, best quality, solo, long hair, wavy hair, silver hair, blue eyes, blue dress, medium breasts, dress, underwater, air bubble, floating hair, refraction, portrait."
+negative_prompt = "worst quality, low quality, monochrome, zombie, interlocked fingers, Aissist, cleavage, nsfw,"
+image = pipe(prompt, seed=0, num_inference_steps=50)
+image.save("image.jpg")

examples/anima/model_training/full/anima-preview.sh

@@ -0,0 +1,16 @@
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --include "anima/anima-preview/*" --local_dir ./data/diffsynth_example_dataset
+
+accelerate launch examples/anima/model_training/train.py \
+  --dataset_base_path data/diffsynth_example_dataset/anima/anima-preview \
+  --dataset_metadata_path data/diffsynth_example_dataset/anima/anima-preview/metadata.csv \
+  --max_pixels 1048576 \
+  --dataset_repeat 50 \
+  --model_id_with_origin_paths "circlestone-labs/Anima:split_files/diffusion_models/anima-preview.safetensors,circlestone-labs/Anima:split_files/text_encoders/qwen_3_06b_base.safetensors,circlestone-labs/Anima:split_files/vae/qwen_image_vae.safetensors" \
+  --tokenizer_path "Qwen/Qwen3-0.6B:./" \
+  --tokenizer_t5xxl_path "stabilityai/stable-diffusion-3.5-large:tokenizer_3/" \
+  --learning_rate 1e-5 \
+  --num_epochs 2 \
+  --remove_prefix_in_ckpt "pipe.dit." \
+  --output_path "./models/train/anima-preview_full" \
+  --trainable_models "dit" \
+  --use_gradient_checkpointing

examples/anima/model_training/lora/anima-preview.sh

@@ -0,0 +1,18 @@
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --include "anima/anima-preview/*" --local_dir ./data/diffsynth_example_dataset
+
+accelerate launch examples/anima/model_training/train.py \
+  --dataset_base_path data/diffsynth_example_dataset/anima/anima-preview \
+  --dataset_metadata_path data/diffsynth_example_dataset/anima/anima-preview/metadata.csv \
+  --max_pixels 1048576 \
+  --dataset_repeat 50 \
+  --model_id_with_origin_paths "circlestone-labs/Anima:split_files/diffusion_models/anima-preview.safetensors,circlestone-labs/Anima:split_files/text_encoders/qwen_3_06b_base.safetensors,circlestone-labs/Anima:split_files/vae/qwen_image_vae.safetensors" \
+  --tokenizer_path "Qwen/Qwen3-0.6B:./" \
+  --tokenizer_t5xxl_path "stabilityai/stable-diffusion-3.5-large:tokenizer_3/" \
+  --learning_rate 1e-4 \
+  --num_epochs 5 \
+  --remove_prefix_in_ckpt "pipe.dit." \
+  --output_path "./models/train/anima-preview_lora" \
+  --lora_base_model "dit" \
+  --lora_target_modules "" \
+  --lora_rank 32 \
+  --use_gradient_checkpointing

examples/anima/model_training/train.py

@@ -0,0 +1,145 @@
+import torch, os, argparse, accelerate
+from diffsynth.core import UnifiedDataset
+from diffsynth.pipelines.anima_image import AnimaImagePipeline, ModelConfig
+from diffsynth.diffusion import *
+os.environ["TOKENIZERS_PARALLELISM"] = "false"
+
+
+class AnimaTrainingModule(DiffusionTrainingModule):
+    def __init__(
+        self,
+        model_paths=None, model_id_with_origin_paths=None,
+        tokenizer_path=None, tokenizer_t5xxl_path=None,
+        trainable_models=None,
+        lora_base_model=None, lora_target_modules="", lora_rank=32, lora_checkpoint=None,
+        preset_lora_path=None, preset_lora_model=None,
+        use_gradient_checkpointing=True,
+        use_gradient_checkpointing_offload=False,
+        extra_inputs=None,
+        fp8_models=None,
+        offload_models=None,
+        device="cpu",
+        task="sft",
+    ):
+        super().__init__()
+        # Load models
+        model_configs = self.parse_model_configs(model_paths, model_id_with_origin_paths, fp8_models=fp8_models, offload_models=offload_models, device=device)
+        tokenizer_config = self.parse_path_or_model_id(tokenizer_path, ModelConfig(model_id="Qwen/Qwen3-0.6B", origin_file_pattern="./"))
+        tokenizer_t5xxl_config = self.parse_path_or_model_id(tokenizer_t5xxl_path, ModelConfig(model_id="stabilityai/stable-diffusion-3.5-large", origin_file_pattern="tokenizer_3/"))
+        self.pipe = AnimaImagePipeline.from_pretrained(torch_dtype=torch.bfloat16, device=device, model_configs=model_configs, tokenizer_config=tokenizer_config, tokenizer_t5xxl_config=tokenizer_t5xxl_config)
+        self.pipe = self.split_pipeline_units(task, self.pipe, trainable_models, lora_base_model)
+
+        # Training mode
+        self.switch_pipe_to_training_mode(
+            self.pipe, trainable_models,
+            lora_base_model, lora_target_modules, lora_rank, lora_checkpoint,
+            preset_lora_path, preset_lora_model,
+            task=task,
+        )
+        
+        # Other configs
+        self.use_gradient_checkpointing = use_gradient_checkpointing
+        self.use_gradient_checkpointing_offload = use_gradient_checkpointing_offload
+        self.extra_inputs = extra_inputs.split(",") if extra_inputs is not None else []
+        self.fp8_models = fp8_models
+        self.task = task
+        self.task_to_loss = {
+            "sft:data_process": lambda pipe, *args: args,
+            "direct_distill:data_process": lambda pipe, *args: args,
+            "sft": lambda pipe, inputs_shared, inputs_posi, inputs_nega: FlowMatchSFTLoss(pipe, **inputs_shared, **inputs_posi),
+            "sft:train": lambda pipe, inputs_shared, inputs_posi, inputs_nega: FlowMatchSFTLoss(pipe, **inputs_shared, **inputs_posi),
+            "direct_distill": lambda pipe, inputs_shared, inputs_posi, inputs_nega: DirectDistillLoss(pipe, **inputs_shared, **inputs_posi),
+            "direct_distill:train": lambda pipe, inputs_shared, inputs_posi, inputs_nega: DirectDistillLoss(pipe, **inputs_shared, **inputs_posi),
+        }
+        
+    def get_pipeline_inputs(self, data):
+        inputs_posi = {"prompt": data["prompt"]}
+        inputs_nega = {"negative_prompt": ""}
+        inputs_shared = {
+            # Assume you are using this pipeline for inference,
+            # please fill in the input parameters.
+            "input_image": data["image"],
+            "height": data["image"].size[1],
+            "width": data["image"].size[0],
+            # Please do not modify the following parameters
+            # unless you clearly know what this will cause.
+            "cfg_scale": 1,
+            "rand_device": self.pipe.device,
+            "use_gradient_checkpointing": self.use_gradient_checkpointing,
+            "use_gradient_checkpointing_offload": self.use_gradient_checkpointing_offload,
+        }
+        inputs_shared = self.parse_extra_inputs(data, self.extra_inputs, inputs_shared)
+        return inputs_shared, inputs_posi, inputs_nega
+    
+    def forward(self, data, inputs=None):
+        if inputs is None: inputs = self.get_pipeline_inputs(data)
+        inputs = self.transfer_data_to_device(inputs, self.pipe.device, self.pipe.torch_dtype)
+        for unit in self.pipe.units:
+            inputs = self.pipe.unit_runner(unit, self.pipe, *inputs)
+        loss = self.task_to_loss[self.task](self.pipe, *inputs)
+        return loss
+
+
+def anima_parser():
+    parser = argparse.ArgumentParser(description="Training script for Anima models.")
+    parser = add_general_config(parser)
+    parser = add_image_size_config(parser)
+    parser.add_argument("--tokenizer_path", type=str, default=None, help="Path to tokenizer.")
+    parser.add_argument("--tokenizer_t5xxl_path", type=str, default=None, help="Path to tokenizer_t5xxl.")
+    return parser
+
+
+if __name__ == "__main__":
+    parser = anima_parser()
+    args = parser.parse_args()
+    accelerator = accelerate.Accelerator(
+        gradient_accumulation_steps=args.gradient_accumulation_steps,
+        kwargs_handlers=[accelerate.DistributedDataParallelKwargs(find_unused_parameters=args.find_unused_parameters)],
+    )
+    dataset = UnifiedDataset(
+        base_path=args.dataset_base_path,
+        metadata_path=args.dataset_metadata_path,
+        repeat=args.dataset_repeat,
+        data_file_keys=args.data_file_keys.split(","),
+        main_data_operator=UnifiedDataset.default_image_operator(
+            base_path=args.dataset_base_path,
+            max_pixels=args.max_pixels,
+            height=args.height,
+            width=args.width,
+            height_division_factor=16,
+            width_division_factor=16,
+        )
+    )
+    model = AnimaTrainingModule(
+        model_paths=args.model_paths,
+        model_id_with_origin_paths=args.model_id_with_origin_paths,
+        tokenizer_path=args.tokenizer_path,
+        tokenizer_t5xxl_path=args.tokenizer_t5xxl_path,
+        trainable_models=args.trainable_models,
+        lora_base_model=args.lora_base_model,
+        lora_target_modules=args.lora_target_modules,
+        lora_rank=args.lora_rank,
+        lora_checkpoint=args.lora_checkpoint,
+        preset_lora_path=args.preset_lora_path,
+        preset_lora_model=args.preset_lora_model,
+        use_gradient_checkpointing=args.use_gradient_checkpointing,
+        use_gradient_checkpointing_offload=args.use_gradient_checkpointing_offload,
+        extra_inputs=args.extra_inputs,
+        fp8_models=args.fp8_models,
+        offload_models=args.offload_models,
+        task=args.task,
+        device=accelerator.device,
+    )
+    model_logger = ModelLogger(
+        args.output_path,
+        remove_prefix_in_ckpt=args.remove_prefix_in_ckpt,
+    )
+    launcher_map = {
+        "sft:data_process": launch_data_process_task,
+        "direct_distill:data_process": launch_data_process_task,
+        "sft": launch_training_task,
+        "sft:train": launch_training_task,
+        "direct_distill": launch_training_task,
+        "direct_distill:train": launch_training_task,
+    }
+    launcher_map[args.task](accelerator, dataset, model, model_logger, args=args)

examples/anima/model_training/validate_full/anima-preview.py

@@ -0,0 +1,21 @@
+from diffsynth.pipelines.anima_image import AnimaImagePipeline, ModelConfig
+from diffsynth.core import load_state_dict
+import torch
+
+
+pipe = AnimaImagePipeline.from_pretrained(
+    torch_dtype=torch.bfloat16,
+    device="cuda",
+    model_configs=[
+        ModelConfig(model_id="circlestone-labs/Anima", origin_file_pattern="split_files/diffusion_models/anima-preview.safetensors"),
+        ModelConfig(model_id="circlestone-labs/Anima", origin_file_pattern="split_files/text_encoders/qwen_3_06b_base.safetensors"),
+        ModelConfig(model_id="circlestone-labs/Anima", origin_file_pattern="split_files/vae/qwen_image_vae.safetensors"),
+    ],
+    tokenizer_config=ModelConfig(model_id="Qwen/Qwen3-0.6B", origin_file_pattern="./"),
+    tokenizer_t5xxl_config=ModelConfig(model_id="stabilityai/stable-diffusion-3.5-large", origin_file_pattern="tokenizer_3/")
+)
+state_dict = load_state_dict("./models/train/anima-preview_full/epoch-1.safetensors", torch_dtype=torch.bfloat16)
+pipe.dit.load_state_dict(state_dict)
+prompt = "a dog"
+image = pipe(prompt=prompt, seed=0)
+image.save("image.jpg")

examples/anima/model_training/validate_lora/anima-preview.py

@@ -0,0 +1,19 @@
+from diffsynth.pipelines.anima_image import AnimaImagePipeline, ModelConfig
+import torch
+
+
+pipe = AnimaImagePipeline.from_pretrained(
+    torch_dtype=torch.bfloat16,
+    device="cuda",
+    model_configs=[
+        ModelConfig(model_id="circlestone-labs/Anima", origin_file_pattern="split_files/diffusion_models/anima-preview.safetensors"),
+        ModelConfig(model_id="circlestone-labs/Anima", origin_file_pattern="split_files/text_encoders/qwen_3_06b_base.safetensors"),
+        ModelConfig(model_id="circlestone-labs/Anima", origin_file_pattern="split_files/vae/qwen_image_vae.safetensors"),
+    ],
+    tokenizer_config=ModelConfig(model_id="Qwen/Qwen3-0.6B", origin_file_pattern="./"),
+    tokenizer_t5xxl_config=ModelConfig(model_id="stabilityai/stable-diffusion-3.5-large", origin_file_pattern="tokenizer_3/")
+)
+pipe.load_lora(pipe.dit, "./models/train/anima-preview_lora/epoch-4.safetensors")
+prompt = "a dog"
+image = pipe(prompt=prompt, seed=0)
+image.save("image.jpg")

examples/flux/model_training/full/FLEX.2-preview.sh

@@ -1,6 +1,8 @@
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --include "flux/FLEX.2-preview/*" --local_dir ./data/diffsynth_example_dataset
+
 accelerate launch --config_file examples/flux/model_training/full/accelerate_config.yaml examples/flux/model_training/train.py \
-  --dataset_base_path data/example_image_dataset \
-  --dataset_metadata_path data/example_image_dataset/metadata.csv \
+  --dataset_base_path data/diffsynth_example_dataset/flux/FLEX.2-preview \
+  --dataset_metadata_path data/diffsynth_example_dataset/flux/FLEX.2-preview/metadata.csv \
   --max_pixels 1048576 \
   --dataset_repeat 200 \
   --model_id_with_origin_paths "ostris/Flex.2-preview:Flex.2-preview.safetensors,black-forest-labs/FLUX.1-dev:text_encoder/model.safetensors,black-forest-labs/FLUX.1-dev:text_encoder_2/*.safetensors,black-forest-labs/FLUX.1-dev:ae.safetensors" \

examples/flux/model_training/full/FLUX.1-Kontext-dev.sh

@@ -1,6 +1,8 @@
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --include "flux/FLUX.1-Kontext-dev/*" --local_dir ./data/diffsynth_example_dataset
+
 accelerate launch --config_file examples/flux/model_training/full/accelerate_config.yaml examples/flux/model_training/train.py \
-  --dataset_base_path data/example_image_dataset \
-  --dataset_metadata_path data/example_image_dataset/metadata_kontext.csv \
+  --dataset_base_path data/diffsynth_example_dataset/flux/FLUX.1-Kontext-dev \
+  --dataset_metadata_path data/diffsynth_example_dataset/flux/FLUX.1-Kontext-dev/metadata.csv \
   --data_file_keys "image,kontext_images" \
   --max_pixels 1048576 \
   --dataset_repeat 400 \

examples/flux/model_training/full/FLUX.1-Krea-dev.sh

@@ -1,6 +1,8 @@
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --include "flux/FLUX.1-Krea-dev/*" --local_dir ./data/diffsynth_example_dataset
+
 accelerate launch --config_file examples/flux/model_training/full/accelerate_config.yaml examples/flux/model_training/train.py \
-  --dataset_base_path data/example_image_dataset \
-  --dataset_metadata_path data/example_image_dataset/metadata.csv \
+  --dataset_base_path data/diffsynth_example_dataset/flux/FLUX.1-Krea-dev \
+  --dataset_metadata_path data/diffsynth_example_dataset/flux/FLUX.1-Krea-dev/metadata.csv \
   --max_pixels 1048576 \
   --dataset_repeat 400 \
   --model_id_with_origin_paths "black-forest-labs/FLUX.1-Krea-dev:flux1-krea-dev.safetensors,black-forest-labs/FLUX.1-dev:text_encoder/model.safetensors,black-forest-labs/FLUX.1-dev:text_encoder_2/*.safetensors,black-forest-labs/FLUX.1-dev:ae.safetensors" \

examples/flux/model_training/full/FLUX.1-dev-AttriCtrl.sh

@@ -1,6 +1,8 @@
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --include "flux/FLUX.1-dev-AttriCtrl/*" --local_dir ./data/diffsynth_example_dataset
+
 accelerate launch examples/flux/model_training/train.py \
-  --dataset_base_path data/example_image_dataset \
-  --dataset_metadata_path data/example_image_dataset/metadata_attrictrl.csv \
+  --dataset_base_path data/diffsynth_example_dataset/flux/FLUX.1-dev-AttriCtrl \
+  --dataset_metadata_path data/diffsynth_example_dataset/flux/FLUX.1-dev-AttriCtrl/metadata.csv \
   --data_file_keys "image" \
   --max_pixels 1048576 \
   --dataset_repeat 100 \

examples/flux/model_training/full/FLUX.1-dev-Controlnet-Inpainting-Beta.sh

@@ -1,6 +1,8 @@
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --include "flux/FLUX.1-dev-Controlnet-Inpainting-Beta/*" --local_dir ./data/diffsynth_example_dataset
+
 accelerate launch --config_file examples/flux/model_training/full/accelerate_config.yaml examples/flux/model_training/train.py \
-  --dataset_base_path data/example_image_dataset \
-  --dataset_metadata_path data/example_image_dataset/metadata_controlnet_inpaint.csv \
+  --dataset_base_path data/diffsynth_example_dataset/flux/FLUX.1-dev-Controlnet-Inpainting-Beta \
+  --dataset_metadata_path data/diffsynth_example_dataset/flux/FLUX.1-dev-Controlnet-Inpainting-Beta/metadata.csv \
   --data_file_keys "image,controlnet_image,controlnet_inpaint_mask" \
   --max_pixels 1048576 \
   --dataset_repeat 400 \

examples/flux/model_training/full/FLUX.1-dev-Controlnet-Union-alpha.sh

@@ -1,6 +1,8 @@
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --include "flux/FLUX.1-dev-Controlnet-Union-alpha/*" --local_dir ./data/diffsynth_example_dataset
+
 accelerate launch --config_file examples/flux/model_training/full/accelerate_config.yaml examples/flux/model_training/train.py \
-  --dataset_base_path data/example_image_dataset \
-  --dataset_metadata_path data/example_image_dataset/metadata_controlnet_canny.csv \
+  --dataset_base_path data/diffsynth_example_dataset/flux/FLUX.1-dev-Controlnet-Union-alpha \
+  --dataset_metadata_path data/diffsynth_example_dataset/flux/FLUX.1-dev-Controlnet-Union-alpha/metadata.csv \
   --data_file_keys "image,controlnet_image" \
   --max_pixels 1048576 \
   --dataset_repeat 400 \

examples/flux/model_training/full/FLUX.1-dev-Controlnet-Upscaler.sh

@@ -1,6 +1,8 @@
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --include "flux/FLUX.1-dev-Controlnet-Upscaler/*" --local_dir ./data/diffsynth_example_dataset
+
 accelerate launch --config_file examples/flux/model_training/full/accelerate_config.yaml examples/flux/model_training/train.py \
-  --dataset_base_path data/example_image_dataset \
-  --dataset_metadata_path data/example_image_dataset/metadata_controlnet_upscale.csv \
+  --dataset_base_path data/diffsynth_example_dataset/flux/FLUX.1-dev-Controlnet-Upscaler \
+  --dataset_metadata_path data/diffsynth_example_dataset/flux/FLUX.1-dev-Controlnet-Upscaler/metadata.csv \
   --data_file_keys "image,controlnet_image" \
   --max_pixels 1048576 \
   --dataset_repeat 400 \

examples/flux/model_training/full/FLUX.1-dev-IP-Adapter.sh

@@ -1,6 +1,8 @@
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --include "flux/FLUX.1-dev-IP-Adapter/*" --local_dir ./data/diffsynth_example_dataset
+
 accelerate launch examples/flux/model_training/train.py \
-  --dataset_base_path data/example_image_dataset \
-  --dataset_metadata_path data/example_image_dataset/metadata_ipadapter.csv \
+  --dataset_base_path data/diffsynth_example_dataset/flux/FLUX.1-dev-IP-Adapter \
+  --dataset_metadata_path data/diffsynth_example_dataset/flux/FLUX.1-dev-IP-Adapter/metadata.csv \
   --data_file_keys "image,ipadapter_images" \
   --max_pixels 1048576 \
   --dataset_repeat 100 \

examples/flux/model_training/full/FLUX.1-dev-InfiniteYou.sh

@@ -1,6 +1,8 @@
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --include "flux/FLUX.1-dev-InfiniteYou/*" --local_dir ./data/diffsynth_example_dataset
+
 accelerate launch --config_file examples/flux/model_training/full/accelerate_config.yaml examples/flux/model_training/train.py \
-  --dataset_base_path data/example_image_dataset \
-  --dataset_metadata_path data/example_image_dataset/metadata_infiniteyou.csv \
+  --dataset_base_path data/diffsynth_example_dataset/flux/FLUX.1-dev-InfiniteYou \
+  --dataset_metadata_path data/diffsynth_example_dataset/flux/FLUX.1-dev-InfiniteYou/metadata.csv \
   --data_file_keys "image,controlnet_image,infinityou_id_image" \
   --max_pixels 1048576 \
   --dataset_repeat 400 \

examples/flux/model_training/full/FLUX.1-dev-LoRA-Encoder.sh

@@ -1,6 +1,8 @@
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --include "flux/FLUX.1-dev-LoRA-Encoder/*" --local_dir ./data/diffsynth_example_dataset
+
 accelerate launch examples/flux/model_training/train.py \
-  --dataset_base_path data/example_image_dataset \
-  --dataset_metadata_path data/example_image_dataset/metadata_lora_encoder.csv \
+  --dataset_base_path data/diffsynth_example_dataset/flux/FLUX.1-dev-LoRA-Encoder \
+  --dataset_metadata_path data/diffsynth_example_dataset/flux/FLUX.1-dev-LoRA-Encoder/metadata.csv \
   --data_file_keys "image" \
   --max_pixels 1048576 \
   --dataset_repeat 100 \

examples/flux/model_training/full/FLUX.1-dev.sh

@@ -1,6 +1,8 @@
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --include "flux/FLUX.1-dev/*" --local_dir ./data/diffsynth_example_dataset
+
 accelerate launch --config_file examples/flux/model_training/full/accelerate_config.yaml examples/flux/model_training/train.py \
-  --dataset_base_path data/example_image_dataset \
-  --dataset_metadata_path data/example_image_dataset/metadata.csv \
+  --dataset_base_path data/diffsynth_example_dataset/flux/FLUX.1-dev \
+  --dataset_metadata_path data/diffsynth_example_dataset/flux/FLUX.1-dev/metadata.csv \
   --max_pixels 1048576 \
   --dataset_repeat 400 \
   --model_id_with_origin_paths "black-forest-labs/FLUX.1-dev:flux1-dev.safetensors,black-forest-labs/FLUX.1-dev:text_encoder/model.safetensors,black-forest-labs/FLUX.1-dev:text_encoder_2/*.safetensors,black-forest-labs/FLUX.1-dev:ae.safetensors" \

examples/flux/model_training/full/Nexus-Gen.sh

@@ -1,6 +1,8 @@
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --include "flux/Nexus-Gen/*" --local_dir ./data/diffsynth_example_dataset
+
 accelerate launch --config_file examples/flux/model_training/full/accelerate_config_zero2offload.yaml examples/flux/model_training/train.py \
-  --dataset_base_path data/example_image_dataset \
-  --dataset_metadata_path data/example_image_dataset/metadata_nexusgen_edit.csv \
+  --dataset_base_path data/diffsynth_example_dataset/flux/Nexus-Gen \
+  --dataset_metadata_path data/diffsynth_example_dataset/flux/Nexus-Gen/metadata.csv \
   --data_file_keys "image,nexus_gen_reference_image" \
   --max_pixels 262144 \
   --dataset_repeat 400 \

examples/flux/model_training/full/Step1X-Edit.sh

@@ -1,6 +1,8 @@
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --include "flux/Step1X-Edit/*" --local_dir ./data/diffsynth_example_dataset
+
 accelerate launch --config_file examples/flux/model_training/full/accelerate_config.yaml examples/flux/model_training/train.py \
-  --dataset_base_path data/example_image_dataset \
-  --dataset_metadata_path data/example_image_dataset/metadata_step1x.csv \
+  --dataset_base_path data/diffsynth_example_dataset/flux/Step1X-Edit \
+  --dataset_metadata_path data/diffsynth_example_dataset/flux/Step1X-Edit/metadata.csv \
   --data_file_keys "image,step1x_reference_image" \
   --max_pixels 1048576 \
   --dataset_repeat 400 \

examples/flux/model_training/lora/FLEX.2-preview.sh

@@ -1,6 +1,8 @@
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --include "flux/FLEX.2-preview/*" --local_dir ./data/diffsynth_example_dataset
+
 accelerate launch examples/flux/model_training/train.py \
-  --dataset_base_path data/example_image_dataset \
-  --dataset_metadata_path data/example_image_dataset/metadata.csv \
+  --dataset_base_path data/diffsynth_example_dataset/flux/FLEX.2-preview \
+  --dataset_metadata_path data/diffsynth_example_dataset/flux/FLEX.2-preview/metadata.csv \
   --max_pixels 1048576 \
   --dataset_repeat 50 \
   --model_id_with_origin_paths "ostris/Flex.2-preview:Flex.2-preview.safetensors,black-forest-labs/FLUX.1-dev:text_encoder/model.safetensors,black-forest-labs/FLUX.1-dev:text_encoder_2/*.safetensors,black-forest-labs/FLUX.1-dev:ae.safetensors" \

examples/flux/model_training/lora/FLUX.1-Kontext-dev.sh

@@ -1,6 +1,8 @@
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --include "flux/FLUX.1-Kontext-dev/*" --local_dir ./data/diffsynth_example_dataset
+
 accelerate launch examples/flux/model_training/train.py \
-  --dataset_base_path data/example_image_dataset \
-  --dataset_metadata_path data/example_image_dataset/metadata_kontext.csv \
+  --dataset_base_path data/diffsynth_example_dataset/flux/FLUX.1-Kontext-dev \
+  --dataset_metadata_path data/diffsynth_example_dataset/flux/FLUX.1-Kontext-dev/metadata.csv \
   --data_file_keys "image,kontext_images" \
   --max_pixels 1048576 \
   --dataset_repeat 400 \

examples/flux/model_training/lora/FLUX.1-Krea-dev.sh

@@ -1,6 +1,8 @@
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --include "flux/FLUX.1-Krea-dev/*" --local_dir ./data/diffsynth_example_dataset
+
 accelerate launch examples/flux/model_training/train.py \
-  --dataset_base_path data/example_image_dataset \
-  --dataset_metadata_path data/example_image_dataset/metadata.csv \
+  --dataset_base_path data/diffsynth_example_dataset/flux/FLUX.1-Krea-dev \
+  --dataset_metadata_path data/diffsynth_example_dataset/flux/FLUX.1-Krea-dev/metadata.csv \
   --max_pixels 1048576 \
   --dataset_repeat 50 \
   --model_id_with_origin_paths "black-forest-labs/FLUX.1-Krea-dev:flux1-krea-dev.safetensors,black-forest-labs/FLUX.1-dev:text_encoder/model.safetensors,black-forest-labs/FLUX.1-dev:text_encoder_2/*.safetensors,black-forest-labs/FLUX.1-dev:ae.safetensors" \

examples/flux/model_training/lora/FLUX.1-dev-AttriCtrl.sh

@@ -1,6 +1,8 @@
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --include "flux/FLUX.1-dev-AttriCtrl/*" --local_dir ./data/diffsynth_example_dataset
+
 accelerate launch examples/flux/model_training/train.py \
-  --dataset_base_path data/example_image_dataset \
-  --dataset_metadata_path data/example_image_dataset/metadata_attrictrl.csv \
+  --dataset_base_path data/diffsynth_example_dataset/flux/FLUX.1-dev-AttriCtrl \
+  --dataset_metadata_path data/diffsynth_example_dataset/flux/FLUX.1-dev-AttriCtrl/metadata.csv \
   --data_file_keys "image" \
   --max_pixels 1048576 \
   --dataset_repeat 100 \

examples/flux/model_training/lora/FLUX.1-dev-Controlnet-Inpainting-Beta.sh

@@ -1,6 +1,8 @@
+modelscope download --dataset DiffSynth-Studio/diffsynth_example_dataset --include "flux/FLUX.1-dev-Controlnet-Inpainting-Beta/*" --local_dir ./data/diffsynth_example_dataset
+
 accelerate launch examples/flux/model_training/train.py \
-  --dataset_base_path data/example_image_dataset \
-  --dataset_metadata_path data/example_image_dataset/metadata_controlnet_inpaint.csv \
+  --dataset_base_path data/diffsynth_example_dataset/flux/FLUX.1-dev-Controlnet-Inpainting-Beta \
+  --dataset_metadata_path data/diffsynth_example_dataset/flux/FLUX.1-dev-Controlnet-Inpainting-Beta/metadata.csv \
   --data_file_keys "image,controlnet_image,controlnet_inpaint_mask" \
   --max_pixels 1048576 \
   --dataset_repeat 100 \