Krea realtime video (#1011)

* krea-realtime-video * Add Krea real-time video inference and training support * Delete .gitignore * update README * update README --------- Co-authored-by: Artiprocher <wangye87v5@hotmail.com> Co-authored-by: Jintao Huang <huangjintao.hjt@alibaba-inc.com> Co-authored-by: Zhongjie Duan <35051019+Artiprocher@users.noreply.github.com>
2026-03-18 22:08:13 +00:00 · 2025-10-27 19:09:28 +08:00
parent 538017177a
commit e0eabaa426
11 changed files with 126 additions and 0 deletions
--- a/README.md
+++ b/README.md
@@ -235,6 +235,8 @@ save_video(video, "video1.mp4", fps=15, quality=5)
 |[Wan-AI/Wan2.1-VACE-1.3B](https://modelscope.cn/models/Wan-AI/Wan2.1-VACE-1.3B)|`vace_control_video`, `vace_reference_image`|[code](./examples/wanvideo/model_inference/Wan2.1-VACE-1.3B.py)|[code](./examples/wanvideo/model_training/full/Wan2.1-VACE-1.3B.sh)|[code](./examples/wanvideo/model_training/validate_full/Wan2.1-VACE-1.3B.py)|[code](./examples/wanvideo/model_training/lora/Wan2.1-VACE-1.3B.sh)|[code](./examples/wanvideo/model_training/validate_lora/Wan2.1-VACE-1.3B.py)|
 |[Wan-AI/Wan2.1-VACE-14B](https://modelscope.cn/models/Wan-AI/Wan2.1-VACE-14B)|`vace_control_video`, `vace_reference_image`|[code](./examples/wanvideo/model_inference/Wan2.1-VACE-14B.py)|[code](./examples/wanvideo/model_training/full/Wan2.1-VACE-14B.sh)|[code](./examples/wanvideo/model_training/validate_full/Wan2.1-VACE-14B.py)|[code](./examples/wanvideo/model_training/lora/Wan2.1-VACE-14B.sh)|[code](./examples/wanvideo/model_training/validate_lora/Wan2.1-VACE-14B.py)|
 |[DiffSynth-Studio/Wan2.1-1.3b-speedcontrol-v1](https://modelscope.cn/models/DiffSynth-Studio/Wan2.1-1.3b-speedcontrol-v1)|`motion_bucket_id`|[code](./examples/wanvideo/model_inference/Wan2.1-1.3b-speedcontrol-v1.py)|[code](./examples/wanvideo/model_training/full/Wan2.1-1.3b-speedcontrol-v1.sh)|[code](./examples/wanvideo/model_training/validate_full/Wan2.1-1.3b-speedcontrol-v1.py)|[code](./examples/wanvideo/model_training/lora/Wan2.1-1.3b-speedcontrol-v1.sh)|[code](./examples/wanvideo/model_training/validate_lora/Wan2.1-1.3b-speedcontrol-v1.py)|
+|[krea/krea-realtime-video](https://www.modelscope.cn/models/krea/krea-realtime-video)||[code](./examples/wanvideo/model_inference/krea-realtime-video.py)|[code](./examples/wanvideo/model_training/full/krea-realtime-video.sh)|[code](./examples/wanvideo/model_training/validate_full/krea-realtime-video.py)|[code](./examples/wanvideo/model_training/lora/krea-realtime-video.sh)|[code](./examples/wanvideo/model_training/validate_lora/krea-realtime-video.py)|
+

 </details>

@@ -385,6 +387,8 @@ https://github.com/Artiprocher/DiffSynth-Studio/assets/35051019/59fb2f7b-8de0-44

 ## Update History

+- **October 27, 2025**: We support [krea/krea-realtime-video](https://www.modelscope.cn/models/krea/krea-realtime-video) model, further expanding Wan's ecosystem.
+
 - **September 23, 2025** [DiffSynth-Studio/Qwen-Image-EliGen-Poster](https://www.modelscope.cn/models/DiffSynth-Studio/Qwen-Image-EliGen-Poster) is released! This model is jointly developed and open-sourced by us and the Taobao Design Team. The model is built upon Qwen-Image, specifically designed for e-commerce poster scenarios, and supports precise partition layout control. Please refer to [our example code](./examples/qwen_image/model_inference/Qwen-Image-EliGen-Poster.py).

 - **September 9, 2025**: Our training framework now supports multiple training modes and has been adapted for Qwen-Image. In addition to the standard SFT training mode, Direct Distill is now also supported; please refer to [our example code](./examples/qwen_image/model_training/lora/Qwen-Image-Distill-LoRA.sh). This feature is experimental, and we will continue to improve it to support comprehensive model training capabilities.
--- a/README_zh.md
+++ b/README_zh.md
@@ -235,6 +235,8 @@ save_video(video, "video1.mp4", fps=15, quality=5)
 |[Wan-AI/Wan2.1-VACE-1.3B](https://modelscope.cn/models/Wan-AI/Wan2.1-VACE-1.3B)|`vace_control_video`, `vace_reference_image`|[code](./examples/wanvideo/model_inference/Wan2.1-VACE-1.3B.py)|[code](./examples/wanvideo/model_training/full/Wan2.1-VACE-1.3B.sh)|[code](./examples/wanvideo/model_training/validate_full/Wan2.1-VACE-1.3B.py)|[code](./examples/wanvideo/model_training/lora/Wan2.1-VACE-1.3B.sh)|[code](./examples/wanvideo/model_training/validate_lora/Wan2.1-VACE-1.3B.py)|
 |[Wan-AI/Wan2.1-VACE-14B](https://modelscope.cn/models/Wan-AI/Wan2.1-VACE-14B)|`vace_control_video`, `vace_reference_image`|[code](./examples/wanvideo/model_inference/Wan2.1-VACE-14B.py)|[code](./examples/wanvideo/model_training/full/Wan2.1-VACE-14B.sh)|[code](./examples/wanvideo/model_training/validate_full/Wan2.1-VACE-14B.py)|[code](./examples/wanvideo/model_training/lora/Wan2.1-VACE-14B.sh)|[code](./examples/wanvideo/model_training/validate_lora/Wan2.1-VACE-14B.py)|
 |[DiffSynth-Studio/Wan2.1-1.3b-speedcontrol-v1](https://modelscope.cn/models/DiffSynth-Studio/Wan2.1-1.3b-speedcontrol-v1)|`motion_bucket_id`|[code](./examples/wanvideo/model_inference/Wan2.1-1.3b-speedcontrol-v1.py)|[code](./examples/wanvideo/model_training/full/Wan2.1-1.3b-speedcontrol-v1.sh)|[code](./examples/wanvideo/model_training/validate_full/Wan2.1-1.3b-speedcontrol-v1.py)|[code](./examples/wanvideo/model_training/lora/Wan2.1-1.3b-speedcontrol-v1.sh)|[code](./examples/wanvideo/model_training/validate_lora/Wan2.1-1.3b-speedcontrol-v1.py)|
+|[krea/krea-realtime-video](https://www.modelscope.cn/models/krea/krea-realtime-video)||[code](./examples/wanvideo/model_inference/krea-realtime-video.py)|[code](./examples/wanvideo/model_training/full/krea-realtime-video.sh)|[code](./examples/wanvideo/model_training/validate_full/krea-realtime-video.py)|[code](./examples/wanvideo/model_training/lora/krea-realtime-video.sh)|[code](./examples/wanvideo/model_training/validate_lora/krea-realtime-video.py)|
+

 </details>

@@ -401,6 +403,8 @@ https://github.com/Artiprocher/DiffSynth-Studio/assets/35051019/59fb2f7b-8de0-44

 ## 更新历史

+- **2025年10月27日** 支持了 [krea/krea-realtime-video](https://www.modelscope.cn/models/krea/krea-realtime-video) 模型，Wan 模型生态再添一员。
+
 - **2025年9月23日** [DiffSynth-Studio/Qwen-Image-EliGen-Poster](https://www.modelscope.cn/models/DiffSynth-Studio/Qwen-Image-EliGen-Poster) 发布！本模型由我们与淘天体验设计团队联合研发并开源。模型基于 Qwen-Image 构建，专为电商海报场景设计，支持精确的分区布局控制。 请参考[我们的示例代码](./examples/qwen_image/model_inference/Qwen-Image-EliGen-Poster.py)。

 - **2025年9月9日** 我们的训练框架支持了多种训练模式，目前已适配 Qwen-Image，除标准 SFT 训练模式外，已支持 Direct Distill，请参考[我们的示例代码](./examples/qwen_image/model_training/lora/Qwen-Image-Distill-LoRA.sh)。这项功能是实验性的，我们将会继续完善已支持更全面的模型训练功能。
--- a/diffsynth/configs/model_config.py
+++ b/diffsynth/configs/model_config.py
@@ -153,6 +153,7 @@ model_loader_configs = [
    (None, "1f5ab7703c6fc803fdded85ff040c316", ["wan_video_dit"], [WanModel], "civitai"),
    (None, "5b013604280dd715f8457c6ed6d6a626", ["wan_video_dit"], [WanModel], "civitai"),
    (None, "2267d489f0ceb9f21836532952852ee5", ["wan_video_dit"], [WanModel], "civitai"),
+    (None, "5ec04e02b42d2580483ad69f4e76346a", ["wan_video_dit"], [WanModel], "civitai"),
    (None, "47dbeab5e560db3180adf51dc0232fb1", ["wan_video_dit"], [WanModel], "civitai"),
    (None, "a61453409b67cd3246cf0c3bebad47ba", ["wan_video_dit", "wan_video_vace"], [WanModel, VaceWanModel], "civitai"),
    (None, "7a513e1f257a861512b1afd387a8ecd9", ["wan_video_dit", "wan_video_vace"], [WanModel, VaceWanModel], "civitai"),
--- a/diffsynth/models/wan_video_dit.py
+++ b/diffsynth/models/wan_video_dit.py
@@ -495,6 +495,12 @@ class WanModelStateDictConverter:
    def from_civitai(self, state_dict):
        state_dict = {name: param for name, param in state_dict.items() if not name.startswith("vace")}
        state_dict = {name: param for name, param in state_dict.items() if name.split(".")[0] not in ["pose_patch_embedding", "face_adapter", "face_encoder", "motion_encoder"]}
+        state_dict_ = {}
+        for name, param in state_dict.items():
+            if name.startswith("model."):
+                name = name[len("model."):]
+            state_dict_[name] = param
+        state_dict = state_dict_
        if hash_state_dict_keys(state_dict) == "9269f8db9040a9d860eaca435be61814":
            config = {
                "has_image_input": False,
--- a/examples/wanvideo/README.md
+++ b/examples/wanvideo/README.md
@@ -76,6 +76,8 @@ save_video(video, "video1.mp4", fps=15, quality=5)
 |[Wan-AI/Wan2.1-VACE-1.3B](https://modelscope.cn/models/Wan-AI/Wan2.1-VACE-1.3B)|`vace_control_video`, `vace_reference_image`|[code](./model_inference/Wan2.1-VACE-1.3B.py)|[code](./model_training/full/Wan2.1-VACE-1.3B.sh)|[code](./model_training/validate_full/Wan2.1-VACE-1.3B.py)|[code](./model_training/lora/Wan2.1-VACE-1.3B.sh)|[code](./model_training/validate_lora/Wan2.1-VACE-1.3B.py)|
 |[Wan-AI/Wan2.1-VACE-14B](https://modelscope.cn/models/Wan-AI/Wan2.1-VACE-14B)|`vace_control_video`, `vace_reference_image`|[code](./model_inference/Wan2.1-VACE-14B.py)|[code](./model_training/full/Wan2.1-VACE-14B.sh)|[code](./model_training/validate_full/Wan2.1-VACE-14B.py)|[code](./model_training/lora/Wan2.1-VACE-14B.sh)|[code](./model_training/validate_lora/Wan2.1-VACE-14B.py)|
 |[DiffSynth-Studio/Wan2.1-1.3b-speedcontrol-v1](https://modelscope.cn/models/DiffSynth-Studio/Wan2.1-1.3b-speedcontrol-v1)|`motion_bucket_id`|[code](./model_inference/Wan2.1-1.3b-speedcontrol-v1.py)|[code](./model_training/full/Wan2.1-1.3b-speedcontrol-v1.sh)|[code](./model_training/validate_full/Wan2.1-1.3b-speedcontrol-v1.py)|[code](./model_training/lora/Wan2.1-1.3b-speedcontrol-v1.sh)|[code](./model_training/validate_lora/Wan2.1-1.3b-speedcontrol-v1.py)|
+|[krea/krea-realtime-video](https://www.modelscope.cn/models/krea/krea-realtime-video)||[code](./model_inference/krea-realtime-video.py)|[code](./model_training/full/krea-realtime-video.sh)|[code](./model_training/validate_full/krea-realtime-video.py)|[code](./model_training/lora/krea-realtime-video.sh)|[code](./model_training/validate_lora/krea-realtime-video.py)|
+


 ## Model Inference
--- a/examples/wanvideo/README_zh.md
+++ b/examples/wanvideo/README_zh.md
@@ -76,6 +76,8 @@ save_video(video, "video1.mp4", fps=15, quality=5)
 |[Wan-AI/Wan2.1-VACE-1.3B](https://modelscope.cn/models/Wan-AI/Wan2.1-VACE-1.3B)|`vace_control_video`, `vace_reference_image`|[code](./model_inference/Wan2.1-VACE-1.3B.py)|[code](./model_training/full/Wan2.1-VACE-1.3B.sh)|[code](./model_training/validate_full/Wan2.1-VACE-1.3B.py)|[code](./model_training/lora/Wan2.1-VACE-1.3B.sh)|[code](./model_training/validate_lora/Wan2.1-VACE-1.3B.py)|
 |[Wan-AI/Wan2.1-VACE-14B](https://modelscope.cn/models/Wan-AI/Wan2.1-VACE-14B)|`vace_control_video`, `vace_reference_image`|[code](./model_inference/Wan2.1-VACE-14B.py)|[code](./model_training/full/Wan2.1-VACE-14B.sh)|[code](./model_training/validate_full/Wan2.1-VACE-14B.py)|[code](./model_training/lora/Wan2.1-VACE-14B.sh)|[code](./model_training/validate_lora/Wan2.1-VACE-14B.py)|
 |[DiffSynth-Studio/Wan2.1-1.3b-speedcontrol-v1](https://modelscope.cn/models/DiffSynth-Studio/Wan2.1-1.3b-speedcontrol-v1)|`motion_bucket_id`|[code](./model_inference/Wan2.1-1.3b-speedcontrol-v1.py)|[code](./model_training/full/Wan2.1-1.3b-speedcontrol-v1.sh)|[code](./model_training/validate_full/Wan2.1-1.3b-speedcontrol-v1.py)|[code](./model_training/lora/Wan2.1-1.3b-speedcontrol-v1.sh)|[code](./model_training/validate_lora/Wan2.1-1.3b-speedcontrol-v1.py)|
+|[krea/krea-realtime-video](https://www.modelscope.cn/models/krea/krea-realtime-video)||[code](./model_inference/krea-realtime-video.py)|[code](./model_training/full/krea-realtime-video.sh)|[code](./model_training/validate_full/krea-realtime-video.py)|[code](./model_training/lora/krea-realtime-video.sh)|[code](./model_training/validate_lora/krea-realtime-video.py)|
+

 ## 模型推理

--- a/examples/wanvideo/model_inference/krea-realtime-video.py
+++ b/examples/wanvideo/model_inference/krea-realtime-video.py
@@ -0,0 +1,25 @@
+import torch
+from diffsynth import save_video
+from diffsynth.pipelines.wan_video_new import WanVideoPipeline, ModelConfig
+
+
+pipe = WanVideoPipeline.from_pretrained(
+    torch_dtype=torch.bfloat16,
+    device="cuda",
+    model_configs=[
+        ModelConfig(model_id="krea/krea-realtime-video", origin_file_pattern="krea-realtime-video-14b.safetensors", offload_device="cpu"),
+        ModelConfig(model_id="Wan-AI/Wan2.1-T2V-14B", origin_file_pattern="models_t5_umt5-xxl-enc-bf16.pth", offload_device="cpu"),
+        ModelConfig(model_id="Wan-AI/Wan2.1-T2V-14B", origin_file_pattern="Wan2.1_VAE.pth", offload_device="cpu"),
+    ],
+)
+pipe.enable_vram_management()
+
+# Text-to-video
+video = pipe(
+    prompt="a cat sitting on a boat",
+    num_inference_steps=6, num_frames=81,
+    seed=0, tiled=True,
+    cfg_scale=1,
+    sigma_shift=20,
+)
+save_video(video, "video1.mp4", fps=15, quality=5)
--- a/examples/wanvideo/model_training/full/krea-realtime-video.sh
+++ b/examples/wanvideo/model_training/full/krea-realtime-video.sh
@@ -0,0 +1,12 @@
+accelerate launch --config_file examples/wanvideo/model_training/full/accelerate_config_14B.yaml examples/wanvideo/model_training/train.py \
+  --dataset_base_path data/example_video_dataset \
+  --dataset_metadata_path data/example_video_dataset/metadata.csv \
+  --height 480 \
+  --width 832 \
+  --dataset_repeat 100 \
+  --model_id_with_origin_paths "krea/krea-realtime-video:krea-realtime-video-14b.safetensors,Wan-AI/Wan2.1-T2V-14B:models_t5_umt5-xxl-enc-bf16.pth,Wan-AI/Wan2.1-T2V-14B:Wan2.1_VAE.pth" \
+  --learning_rate 1e-5 \
+  --num_epochs 2 \
+  --remove_prefix_in_ckpt "pipe.dit." \
+  --output_path "./models/train/krea-realtime-video_full" \
+  --trainable_models "dit"
--- a/examples/wanvideo/model_training/lora/krea-realtime-video.sh
+++ b/examples/wanvideo/model_training/lora/krea-realtime-video.sh
@@ -0,0 +1,14 @@
+accelerate launch examples/wanvideo/model_training/train.py \
+  --dataset_base_path data/example_video_dataset \
+  --dataset_metadata_path data/example_video_dataset/metadata.csv \
+  --height 480 \
+  --width 832 \
+  --dataset_repeat 100 \
+  --model_id_with_origin_paths "krea/krea-realtime-video:krea-realtime-video-14b.safetensors,Wan-AI/Wan2.1-T2V-14B:models_t5_umt5-xxl-enc-bf16.pth,Wan-AI/Wan2.1-T2V-14B:Wan2.1_VAE.pth" \
+  --learning_rate 1e-4 \
+  --num_epochs 5 \
+  --remove_prefix_in_ckpt "pipe.dit." \
+  --output_path "./models/train/krea-realtime-video_lora" \
+  --lora_base_model "dit" \
+  --lora_target_modules "q,k,v,o,ffn.0,ffn.2" \
+  --lora_rank 32
--- a/examples/wanvideo/model_training/validate_full/krea-realtime-video.py
+++ b/examples/wanvideo/model_training/validate_full/krea-realtime-video.py
@@ -0,0 +1,28 @@
+import torch
+from PIL import Image
+from diffsynth import save_video, VideoData, load_state_dict
+from diffsynth.pipelines.wan_video_new import WanVideoPipeline, ModelConfig
+
+
+pipe = WanVideoPipeline.from_pretrained(
+    torch_dtype=torch.bfloat16,
+    device="cuda",
+    model_configs=[
+        ModelConfig(model_id="krea/krea-realtime-video", origin_file_pattern="krea-realtime-video-14b.safetensors", offload_device="cpu"),
+        ModelConfig(model_id="Wan-AI/Wan2.1-T2V-14B", origin_file_pattern="models_t5_umt5-xxl-enc-bf16.pth", offload_device="cpu"),
+        ModelConfig(model_id="Wan-AI/Wan2.1-T2V-14B", origin_file_pattern="Wan2.1_VAE.pth", offload_device="cpu"),
+    ],
+)
+state_dict = load_state_dict("models/train/krea-realtime-video_full/epoch-1.safetensors")
+pipe.dit.load_state_dict(state_dict)
+pipe.enable_vram_management()
+
+# Text-to-video
+video = pipe(
+    prompt="a cat sitting on a boat",
+    num_inference_steps=6, num_frames=81,
+    seed=0, tiled=True,
+    cfg_scale=1,
+    sigma_shift=20,
+)
+save_video(video, "output.mp4", fps=15, quality=5)
--- a/examples/wanvideo/model_training/validate_lora/krea-realtime-video.py
+++ b/examples/wanvideo/model_training/validate_lora/krea-realtime-video.py
@@ -0,0 +1,28 @@
+import torch
+from PIL import Image
+from diffsynth import save_video, VideoData, load_state_dict
+from diffsynth.pipelines.wan_video_new import WanVideoPipeline, ModelConfig
+
+
+pipe = WanVideoPipeline.from_pretrained(
+    torch_dtype=torch.bfloat16,
+    device="cuda",
+    model_configs=[
+        ModelConfig(model_id="krea/krea-realtime-video", origin_file_pattern="krea-realtime-video-14b.safetensors", offload_device="cpu"),
+        ModelConfig(model_id="Wan-AI/Wan2.1-T2V-14B", origin_file_pattern="models_t5_umt5-xxl-enc-bf16.pth", offload_device="cpu"),
+        ModelConfig(model_id="Wan-AI/Wan2.1-T2V-14B", origin_file_pattern="Wan2.1_VAE.pth", offload_device="cpu"),
+    ],
+)
+
+pipe.load_lora(pipe.dit, "models/train/krea-realtime-video_lora/epoch-4.safetensors", alpha=1)
+pipe.enable_vram_management()
+
+# Text-to-video
+video = pipe(
+    prompt="a cat sitting on a boat",
+    num_inference_steps=6, num_frames=81,
+    seed=0, tiled=True,
+    cfg_scale=1,
+    sigma_shift=20,
+)
+save_video(video, "output.mp4", fps=15, quality=5)