Merge pull request #568 from mi804/nexus-gen

update nexus-gen readme

README.md

@@ -42,6 +42,12 @@ Until now, DiffSynth-Studio has supported the following models:
 * [Stable Diffusion](https://huggingface.co/runwayml/stable-diffusion-v1-5)
 
 ## News
+- **May 1, 2025** 🔥🔥🔥 We propose Nexus-Gen, a unified model that synergizes the language reasoning capabilities of LLMs with the image synthesis power of diffusion models.
+  - Paper: [Nexus-Gen: A Unified Model for Image Understanding, Generation, and Editing](https://arxiv.org/pdf/2504.21356)
+  - Github Repo: https://github.com/modelscope/Nexus-Gen
+  - Model: [ModelScope](https://www.modelscope.cn/models/DiffSynth-Studio/Nexus-Gen), [HuggingFace](https://huggingface.co/modelscope/Nexus-Gen)
+  - Online Demo: [ModelScope Nexus-Gen Studio](https://www.modelscope.cn/studios/DiffSynth-Studio/Nexus-Gen)
+
 - **March 31, 2025** We support InfiniteYou, an identity preserving method for FLUX. Please refer to [./examples/InfiniteYou/](./examples/InfiniteYou/) for more details.
 
 - **March 25, 2025** 🔥🔥🔥 Our new open-source project, [DiffSynth-Engine](https://github.com/modelscope/DiffSynth-Engine), is now open-sourced! Focused on stable model deployment. Geared towards industry. Offers better engineering support, higher computational performance, and more stable functionality.

diffsynth/pipelines/flux_image.py

@@ -202,10 +202,10 @@ class FluxImagePipeline(BasePipeline):
         return image
     
 
-    def encode_prompt(self, prompt, positive=True, t5_sequence_length=512):
-        if self.text_encoder_1 is not None and self.text_encoder_2 is not None:
+    def encode_prompt(self, prompt, positive=True, t5_sequence_length=512, image_emb=None):
+        if (self.text_encoder_1 is not None and self.text_encoder_2 is not None) or (image_emb is not None):
             prompt_emb, pooled_prompt_emb, text_ids = self.prompter.encode_prompt(
-                prompt, device=self.device, positive=positive, t5_sequence_length=t5_sequence_length
+                prompt, device=self.device, positive=positive, t5_sequence_length=t5_sequence_length, image_emb=image_emb
             )
             return {"prompt_emb": prompt_emb, "pooled_prompt_emb": pooled_prompt_emb, "text_ids": text_ids}
         else:
@@ -358,13 +358,13 @@ class FluxImagePipeline(BasePipeline):
         return eligen_kwargs_posi, eligen_kwargs_nega, fg_mask, bg_mask
 
 
-    def prepare_prompts(self, prompt, local_prompts, masks, mask_scales, t5_sequence_length, negative_prompt, cfg_scale):
+    def prepare_prompts(self, prompt, local_prompts, masks, mask_scales, t5_sequence_length, negative_prompt, cfg_scale, image_emb=None):
         # Extend prompt
         self.load_models_to_device(['text_encoder_1', 'text_encoder_2'])
         prompt, local_prompts, masks, mask_scales = self.extend_prompt(prompt, local_prompts, masks, mask_scales)
 
         # Encode prompts
-        prompt_emb_posi = self.encode_prompt(prompt, t5_sequence_length=t5_sequence_length)
+        prompt_emb_posi = self.encode_prompt(prompt, t5_sequence_length=t5_sequence_length, image_emb=image_emb)
         prompt_emb_nega = self.encode_prompt(negative_prompt, positive=False, t5_sequence_length=t5_sequence_length) if cfg_scale != 1.0 else None
         prompt_emb_locals = [self.encode_prompt(prompt_local, t5_sequence_length=t5_sequence_length) for prompt_local in local_prompts]
         return prompt_emb_posi, prompt_emb_nega, prompt_emb_locals
@@ -432,6 +432,7 @@ class FluxImagePipeline(BasePipeline):
         height=1024,
         width=1024,
         seed=None,
+        image_emb=None,
         # Steps
         num_inference_steps=30,
         # local prompts
@@ -483,7 +484,7 @@ class FluxImagePipeline(BasePipeline):
         latents, input_latents = self.prepare_latents(input_image, height, width, seed, tiled, tile_size, tile_stride)
 
         # Prompt
-        prompt_emb_posi, prompt_emb_nega, prompt_emb_locals = self.prepare_prompts(prompt, local_prompts, masks, mask_scales, t5_sequence_length, negative_prompt, cfg_scale)
+        prompt_emb_posi, prompt_emb_nega, prompt_emb_locals = self.prepare_prompts(prompt, local_prompts, masks, mask_scales, t5_sequence_length, negative_prompt, cfg_scale, image_emb)
 
         # Extra input
         extra_input = self.prepare_extra_input(latents, guidance=embedded_guidance)
@@ -679,6 +680,7 @@ def lets_dance_flux(
     step1x_mask=None,
     step1x_reference_latents=None,
     tea_cache: TeaCache = None,
+    use_gradient_checkpointing=False,
     **kwargs
 ):
     if tiled:
@@ -773,20 +775,32 @@ def lets_dance_flux(
         tea_cache_update = tea_cache.check(dit, hidden_states, conditioning)
     else:
         tea_cache_update = False
+        
+    def create_custom_forward(module):
+        def custom_forward(*inputs):
+            return module(*inputs)
+        return custom_forward
 
     if tea_cache_update:
         hidden_states = tea_cache.update(hidden_states)
     else:
         # Joint Blocks
         for block_id, block in enumerate(dit.blocks):
-            hidden_states, prompt_emb = block(
-                hidden_states,
-                prompt_emb,
-                conditioning,
-                image_rotary_emb,
-                attention_mask,
-                ipadapter_kwargs_list=ipadapter_kwargs_list.get(block_id, None)
-            )
+            if use_gradient_checkpointing:
+                hidden_states, prompt_emb = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states, prompt_emb, conditioning, image_rotary_emb, attention_mask, ipadapter_kwargs_list.get(block_id, None),
+                    use_reentrant=False,
+                )
+            else:
+                hidden_states, prompt_emb = block(
+                    hidden_states,
+                    prompt_emb,
+                    conditioning,
+                    image_rotary_emb,
+                    attention_mask,
+                    ipadapter_kwargs_list=ipadapter_kwargs_list.get(block_id, None)
+                )
             # ControlNet
             if controlnet is not None and controlnet_frames is not None:
                 hidden_states = hidden_states + controlnet_res_stack[block_id]
@@ -795,14 +809,21 @@ def lets_dance_flux(
         hidden_states = torch.cat([prompt_emb, hidden_states], dim=1)
         num_joint_blocks = len(dit.blocks)
         for block_id, block in enumerate(dit.single_blocks):
-            hidden_states, prompt_emb = block(
-                hidden_states,
-                prompt_emb,
-                conditioning,
-                image_rotary_emb,
-                attention_mask,
-                ipadapter_kwargs_list=ipadapter_kwargs_list.get(block_id + num_joint_blocks, None)
-            )
+            if use_gradient_checkpointing:
+                hidden_states, prompt_emb = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states, prompt_emb, conditioning, image_rotary_emb, attention_mask, ipadapter_kwargs_list.get(block_id + num_joint_blocks, None),
+                    use_reentrant=False,
+                )
+            else:
+                hidden_states, prompt_emb = block(
+                    hidden_states,
+                    prompt_emb,
+                    conditioning,
+                    image_rotary_emb,
+                    attention_mask,
+                    ipadapter_kwargs_list=ipadapter_kwargs_list.get(block_id + num_joint_blocks, None)
+                )
             # ControlNet
             if controlnet is not None and controlnet_frames is not None:
                 hidden_states[:, prompt_emb.shape[1]:] = hidden_states[:, prompt_emb.shape[1]:] + controlnet_single_res_stack[block_id]

diffsynth/prompters/flux_prompter.py

@@ -59,6 +59,7 @@ class FluxPrompter(BasePrompter):
         positive=True,
         device="cuda",
         t5_sequence_length=512,
+        image_emb=None,
     ):
         prompt = self.process_prompt(prompt, positive=positive)
         
@@ -66,7 +67,10 @@ class FluxPrompter(BasePrompter):
         pooled_prompt_emb = self.encode_prompt_using_clip(prompt, self.text_encoder_1, self.tokenizer_1, 77, device)
         
         # T5
-        prompt_emb = self.encode_prompt_using_t5(prompt, self.text_encoder_2, self.tokenizer_2, t5_sequence_length, device)
+        if image_emb is not None:
+            prompt_emb = image_emb
+        else:
+            prompt_emb = self.encode_prompt_using_t5(prompt, self.text_encoder_2, self.tokenizer_2, t5_sequence_length, device)
 
         # text_ids
         text_ids = torch.zeros(prompt_emb.shape[0], prompt_emb.shape[1], 3).to(device=device, dtype=prompt_emb.dtype)

modeling/ar/configuration_qwen2_5_vl.py

@@ -0,0 +1,258 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/qwen2_5_vl/modular_qwen2_5_vl.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_qwen2_5_vl.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# coding=utf-8
+# Copyright 2025 The Qwen Team and The HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from transformers.configuration_utils import PretrainedConfig
+from transformers.modeling_rope_utils import rope_config_validation
+
+
+class Qwen2_5_VLVisionConfig(PretrainedConfig):
+    model_type = "qwen2_5_vl"
+    base_config_key = "vision_config"
+
+    def __init__(
+        self,
+        depth=32,
+        hidden_size=3584,
+        hidden_act="silu",
+        intermediate_size=3420,
+        num_heads=16,
+        in_channels=3,
+        patch_size=14,
+        spatial_merge_size=2,
+        temporal_patch_size=2,
+        tokens_per_second=4,
+        window_size=112,
+        out_hidden_size=3584,
+        fullatt_block_indexes=[7, 15, 23, 31],
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.depth = depth
+        self.hidden_size = hidden_size
+        self.hidden_act = hidden_act
+        self.intermediate_size = intermediate_size
+        self.num_heads = num_heads
+        self.in_channels = in_channels
+        self.patch_size = patch_size
+        self.spatial_merge_size = spatial_merge_size
+        self.temporal_patch_size = temporal_patch_size
+        self.tokens_per_second = tokens_per_second
+        self.window_size = window_size
+        self.fullatt_block_indexes = fullatt_block_indexes
+        self.out_hidden_size = out_hidden_size
+
+
+class Qwen2_5_VLConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Qwen2_5_VLModel`]. It is used to instantiate a
+    Qwen2-VL model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of
+    Qwen2-VL-7B-Instruct [Qwen/Qwen2-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct).
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 152064):
+            Vocabulary size of the Qwen2_5_VL model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`Qwen2_5_VLModel`]
+        hidden_size (`int`, *optional*, defaults to 8192):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 29568):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 80):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 64):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_key_value_heads (`int`, *optional*, defaults to 8):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details checkout [this
+            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `32`.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 32768):
+            The maximum sequence length that this model might ever be used with.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-05):
+            The epsilon used by the rms normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied.
+        rope_theta (`float`, *optional*, defaults to 1000000.0):
+            The base period of the RoPE embeddings.
+        use_sliding_window (`bool`, *optional*, defaults to `False`):
+            Whether to use sliding window attention.
+        sliding_window (`int`, *optional*, defaults to 4096):
+            Sliding window attention (SWA) window size. If not specified, will default to `4096`.
+        max_window_layers (`int`, *optional*, defaults to 80):
+            The number of layers that use SWA (Sliding Window Attention). The bottom layers use SWA while the top use full attention.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        vision_config (`Dict`, *optional*):
+            The config for the visual encoder initialization.
+        rope_scaling (`Dict`, *optional*):
+            Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
+            and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
+            accordingly.
+            Expected contents:
+                `rope_type` (`str`):
+                    The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
+                    'llama3'], with 'default' being the original RoPE implementation.
+                `factor` (`float`, *optional*):
+                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
+                    most scaling types, a `factor` of x will enable the model to handle sequences of length x *
+                    original maximum pre-trained length.
+                `original_max_position_embeddings` (`int`, *optional*):
+                    Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
+                    pretraining.
+                `attention_factor` (`float`, *optional*):
+                    Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
+                    computation. If unspecified, it defaults to value recommended by the implementation, using the
+                    `factor` field to infer the suggested value.
+                `beta_fast` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 32.
+                `beta_slow` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 1.
+                `short_factor` (`List[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `long_factor` (`List[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `low_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
+                `high_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
+
+    ```python
+    >>> from transformers import Qwen2_5_VLForConditionalGeneration, Qwen2_5_VLConfig
+
+    >>> # Initializing a Qwen2_5_VL style configuration
+    >>> configuration = Qwen2_5_VLConfig()
+
+    >>> # Initializing a model from the Qwen2-VL-7B style configuration
+    >>> model = Qwen2_5_VLForConditionalGeneration(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "qwen2_5_vl"
+    sub_configs = {"vision_config": Qwen2_5_VLVisionConfig}
+    keys_to_ignore_at_inference = ["past_key_values"]
+    # Default tensor parallel plan for base model `Qwen2_5_VL`
+    base_model_tp_plan = {
+        "layers.*.self_attn.q_proj": "colwise",
+        "layers.*.self_attn.k_proj": "colwise",
+        "layers.*.self_attn.v_proj": "colwise",
+        "layers.*.self_attn.o_proj": "rowwise",
+        "layers.*.mlp.gate_proj": "colwise",
+        "layers.*.mlp.up_proj": "colwise",
+        "layers.*.mlp.down_proj": "rowwise",
+    }
+    base_model_pp_plan = {
+        "embed_tokens": (["input_ids"], ["inputs_embeds"]),
+        "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
+        "norm": (["hidden_states"], ["hidden_states"]),
+    }
+
+    def __init__(
+        self,
+        vocab_size=152064,
+        hidden_size=8192,
+        intermediate_size=29568,
+        num_hidden_layers=80,
+        num_attention_heads=64,
+        num_key_value_heads=8,
+        hidden_act="silu",
+        max_position_embeddings=32768,
+        initializer_range=0.02,
+        rms_norm_eps=1e-05,
+        use_cache=True,
+        tie_word_embeddings=False,
+        rope_theta=1000000.0,
+        use_sliding_window=False,
+        sliding_window=4096,
+        max_window_layers=80,
+        attention_dropout=0.0,
+        vision_config=None,
+        rope_scaling=None,
+        **kwargs,
+    ):
+        if isinstance(vision_config, dict):
+            self.vision_config = self.sub_configs["vision_config"](**vision_config)
+        elif vision_config is None:
+            self.vision_config = self.sub_configs["vision_config"]()
+
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.use_sliding_window = use_sliding_window
+        self.sliding_window = sliding_window
+        self.max_window_layers = max_window_layers
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.attention_dropout = attention_dropout
+        self.rope_scaling = rope_scaling
+
+        # Validate the correctness of rotary position embeddings parameters
+        # BC: if there is a 'type' field, move it to 'rope_type'.
+        # and change type from 'mrope' to 'default' because `mrope` does default RoPE calculations
+        # one can set it to "linear"/"dynamic" etc. to have scaled RoPE
+        # TODO: @raushan update config in the hub
+        if self.rope_scaling is not None and "type" in self.rope_scaling:
+            if self.rope_scaling["type"] == "mrope":
+                self.rope_scaling["type"] = "default"
+            self.rope_scaling["rope_type"] = self.rope_scaling["type"]
+        rope_config_validation(self, ignore_keys={"mrope_section"})
+
+        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)
+
+
+__all__ = ["Qwen2_5_VLConfig"]

modeling/ar/processing_qwen2_5_vl.py

@@ -0,0 +1,235 @@
+# Copyright 2025 The Qwen Team and The HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import re
+from typing import List, Union
+
+from transformers.feature_extraction_utils import BatchFeature
+from transformers.image_utils import ImageInput, VideoInput
+from transformers.processing_utils import ProcessingKwargs, ProcessorMixin, Unpack, VideosKwargs
+from transformers.tokenization_utils_base import PreTokenizedInput, TextInput
+
+
+class Qwen2_5_VLVideosProcessorKwargs(VideosKwargs, total=False):
+    fps: Union[List[float], float]
+
+
+class Qwen2_5_VLProcessorKwargs(ProcessingKwargs, total=False):
+    videos_kwargs: Qwen2_5_VLVideosProcessorKwargs
+    _defaults = {
+        "text_kwargs": {
+            "padding": False,
+        },
+        "videos_kwargs": {"fps": 2.0},
+    }
+
+
+class Qwen2_5_VLProcessor(ProcessorMixin):
+    r"""
+    Constructs a Qwen2.5-VL processor which wraps a Qwen2.5-VL image processor and a Qwen2 tokenizer into a single processor.
+    [`Qwen2_5_VLProcessor`] offers all the functionalities of [`Qwen2VLImageProcessor`] and [`Qwen2TokenizerFast`]. See the
+    [`~Qwen2_5_VLProcessor.__call__`] and [`~Qwen2_5_VLProcessor.decode`] for more information.
+    Args:
+        image_processor ([`Qwen2VLImageProcessor`], *optional*):
+            The image processor is a required input.
+        tokenizer ([`Qwen2TokenizerFast`], *optional*):
+            The tokenizer is a required input.
+        chat_template (`str`, *optional*): A Jinja template which will be used to convert lists of messages
+            in a chat into a tokenizable string.
+    """
+
+    attributes = ["image_processor", "tokenizer"]
+    valid_kwargs = ["chat_template"]
+
+    image_processor_class = "AutoImageProcessor"
+    tokenizer_class = ("Qwen2Tokenizer", "Qwen2TokenizerFast")
+
+    def __init__(self, image_processor=None, tokenizer=None, chat_template=None, **kwargs):
+        self.image_token = "<|image_pad|>" if not hasattr(tokenizer, "image_token") else tokenizer.image_token
+        self.video_token = "<|video_pad|>" if not hasattr(tokenizer, "video_token") else tokenizer.video_token
+        super().__init__(image_processor, tokenizer, chat_template=chat_template)
+
+    def __call__(
+        self,
+        images: ImageInput = None,
+        text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]] = None,
+        videos: VideoInput = None,
+        **kwargs: Unpack[Qwen2_5_VLProcessorKwargs],
+    ) -> BatchFeature:
+        """
+        Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
+        and `kwargs` arguments to Qwen2TokenizerFast's [`~Qwen2TokenizerFast.__call__`] if `text` is not `None` to encode
+        the text. To prepare the vision inputs, this method forwards the `vision_infos` and `kwrags` arguments to
+        Qwen2VLImageProcessor's [`~Qwen2VLImageProcessor.__call__`] if `vision_infos` is not `None`.
+
+        Args:
+            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`):
+                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
+                tensor. Both channels-first and channels-last formats are supported.
+            text (`str`, `List[str]`, `List[List[str]]`):
+                The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
+                (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
+                `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
+            videos (`np.ndarray`, `torch.Tensor`, `List[np.ndarray]`, `List[torch.Tensor]`):
+                The image or batch of videos to be prepared. Each video can be a 4D NumPy array or PyTorch
+                tensor, or a nested list of 3D frames. Both channels-first and channels-last formats are supported.
+            return_tensors (`str` or [`~utils.TensorType`], *optional*):
+                If set, will return tensors of a particular framework. Acceptable values are:
+                - `'tf'`: Return TensorFlow `tf.constant` objects.
+                - `'pt'`: Return PyTorch `torch.Tensor` objects.
+                - `'np'`: Return NumPy `np.ndarray` objects.
+                - `'jax'`: Return JAX `jnp.ndarray` objects.
+
+        Returns:
+            [`BatchFeature`]: A [`BatchFeature`] with the following fields:
+
+            - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
+            - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
+              `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
+              `None`).
+            - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
+            - **pixel_values_videos** -- Pixel values of videos to be fed to a model. Returned when `videos` is not `None`.
+            - **image_grid_thw** -- List of image 3D grid in LLM. Returned when `images` is not `None`.
+            - **video_grid_thw** -- List of video 3D grid in LLM. Returned when `videos` is not `None`.
+            - **second_per_grid_ts** -- List of video seconds per time grid. Returned when `videos` is not `None`.
+        """
+        output_kwargs = self._merge_kwargs(
+            Qwen2_5_VLProcessorKwargs,
+            tokenizer_init_kwargs=self.tokenizer.init_kwargs,
+            **kwargs,
+        )
+        if images is not None:
+            image_inputs = self.image_processor(images=images, videos=None, **output_kwargs["images_kwargs"])
+            image_grid_thw = image_inputs["image_grid_thw"]
+        else:
+            image_inputs = {}
+            image_grid_thw = None
+
+        if videos is not None:
+            videos_inputs = self.image_processor(images=None, videos=videos, **output_kwargs["images_kwargs"])
+            video_grid_thw = videos_inputs["video_grid_thw"]
+
+            fps = output_kwargs["videos_kwargs"].pop("fps", 2.0)
+            if isinstance(fps, (int, float)):
+                second_per_grid_ts = [self.image_processor.temporal_patch_size / fps] * len(video_grid_thw)
+            elif hasattr(fps, "__len__") and len(fps) == len(video_grid_thw):
+                second_per_grid_ts = [self.image_processor.temporal_patch_size / tmp for tmp in fps]
+            else:
+                raise ValueError(
+                    f"The length of fps ({len(fps) if hasattr(fps, '__len__') else fps}) must be equal to the length of video_grid_thw ({len(video_grid_thw)}) or fps should be a single number."
+                )
+            videos_inputs.update({"second_per_grid_ts": second_per_grid_ts})
+
+        else:
+            videos_inputs = {}
+            video_grid_thw = None
+
+        if not isinstance(text, list):
+            text = [text]
+
+        if image_grid_thw is not None:
+            merge_length = self.image_processor.merge_size**2
+            index = 0
+            for i in range(len(text)):
+                while self.image_token in text[i]:
+                    text[i] = text[i].replace(
+                        self.image_token,
+                        "<|placeholder|>" * (image_grid_thw[index].prod() // merge_length),
+                        1,
+                    )
+                    index += 1
+                text[i] = text[i].replace("<|placeholder|>", self.image_token)
+
+        if video_grid_thw is not None:
+            merge_length = self.image_processor.merge_size**2
+            index = 0
+            for i in range(len(text)):
+                while self.video_token in text[i]:
+                    text[i] = text[i].replace(
+                        self.video_token,
+                        "<|placeholder|>" * (video_grid_thw[index].prod() // merge_length),
+                        1,
+                    )
+                    index += 1
+                text[i] = text[i].replace("<|placeholder|>", self.video_token)
+
+        text_inputs = self.tokenizer(text, **output_kwargs["text_kwargs"])
+
+        return BatchFeature(data={**text_inputs, **image_inputs, **videos_inputs})
+
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to Qwen2TokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
+        refer to the docstring of this method for more information.
+        """
+        return self.tokenizer.batch_decode(*args, **kwargs)
+
+    def batch_decode_all2all(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to Qwen2TokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
+        refer to the docstring of this method for more information.
+        """
+        decoded = self.tokenizer.batch_decode(*args, **kwargs)
+        pattern = r'<\|vision_start\|>.*?<\|vision_end\|>'
+        decoded_with_image_tag = [re.sub(pattern, '<image>', d, flags=re.DOTALL) for d in decoded]
+        decoded_with_image_tag = [re.sub(r'<\|im_end\|>', '', d) for d in decoded_with_image_tag]
+        return decoded_with_image_tag
+
+    def decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to Qwen2TokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
+        the docstring of this method for more information.
+        """
+        return self.tokenizer.decode(*args, **kwargs)
+
+    def post_process_image_text_to_text(
+        self, generated_outputs, skip_special_tokens=True, clean_up_tokenization_spaces=False, **kwargs
+    ):
+        """
+        Post-process the output of the model to decode the text.
+
+        Args:
+            generated_outputs (`torch.Tensor` or `np.ndarray`):
+                The output of the model `generate` function. The output is expected to be a tensor of shape `(batch_size, sequence_length)`
+                or `(sequence_length,)`.
+            skip_special_tokens (`bool`, *optional*, defaults to `True`):
+                Whether or not to remove special tokens in the output. Argument passed to the tokenizer's `batch_decode` method.
+            Clean_up_tokenization_spaces (`bool`, *optional*, defaults to `False`):
+                Whether or not to clean up the tokenization spaces. Argument passed to the tokenizer's `batch_decode` method.
+            **kwargs:
+                Additional arguments to be passed to the tokenizer's `batch_decode method`.
+
+        Returns:
+            `List[str]`: The decoded text.
+        """
+        return self.tokenizer.batch_decode(
+            generated_outputs,
+            skip_special_tokens=skip_special_tokens,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+            **kwargs,
+        )
+
+    @property
+    def model_input_names(self):
+        tokenizer_input_names = self.tokenizer.model_input_names
+        image_processor_input_names = self.image_processor.model_input_names
+        names_from_processor = list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))
+        return names_from_processor + ["second_per_grid_ts"]
+
+
+__all__ = ["Qwen2_5_VLProcessor"]

modeling/decoder/flux_decoder.py

@@ -0,0 +1,64 @@
+import torch
+from diffsynth import ModelManager
+from .flux_image_pipeline import FluxImagePipelineAll2All
+
+class FluxDecoder:
+
+    def __init__(self, flux_all2all_modelpath, flux_path, device='cuda', torch_dtype=torch.bfloat16):
+        self.device = device
+        self.torch_dtype = torch_dtype
+        self.pipe, self.adapter = self.get_pipe(flux_all2all_modelpath, flux_path, device, torch_dtype)
+
+    def get_pipe(self, flux_all2all_modelpath, flux_path, device="cuda", torch_dtype=torch.bfloat16):
+        model_manager = ModelManager(torch_dtype=torch_dtype, device=device)
+        model_manager.load_models([
+            f"{flux_path}/FLUX/FLUX.1-dev/text_encoder/model.safetensors",
+            f"{flux_path}/FLUX/FLUX.1-dev/text_encoder_2",
+            f"{flux_path}/FLUX/FLUX.1-dev/ae.safetensors",
+            f"{flux_path}/FLUX/FLUX.1-dev/flux1-dev.safetensors"
+        ])
+
+        state_dict = torch.load(flux_all2all_modelpath, weights_only=True, map_location='cpu')
+        adapter_states = ['0.weight', '0.bias', '1.weight', '1.bias', '3.weight', '3.bias', '4.weight', '4.bias']
+        adapter_state_dict = {}
+        for key in adapter_states:
+            adapter_state_dict[key] = state_dict.pop(key)
+
+        in_channel = 3584
+        out_channel = 4096
+        expand_ratio = 1
+        adapter = torch.nn.Sequential(torch.nn.Linear(in_channel, out_channel * expand_ratio),
+                                      torch.nn.LayerNorm(out_channel * expand_ratio), torch.nn.ReLU(),
+                                      torch.nn.Linear(out_channel * expand_ratio, out_channel),
+                                      torch.nn.LayerNorm(out_channel))
+        adapter.load_state_dict(adapter_state_dict)
+        adapter.to(device, dtype=torch_dtype)
+
+        pipe = FluxImagePipelineAll2All.from_model_manager(model_manager)
+        pipe.dit.load_state_dict(state_dict)
+
+        return pipe, adapter
+
+    @torch.no_grad()
+    def decode_image_embeds(self,
+                            output_image_embeddings,
+                            height=512,
+                            width=512,
+                            num_inference_steps=50,
+                            seed=42,
+                            negative_prompt="",
+                            cfg_scale=1.0,
+                            **pipe_kwargs):
+        output_image_embeddings = output_image_embeddings.to(device=self.device, dtype=self.torch_dtype)
+        image_embed = self.adapter(output_image_embeddings)
+        image = self.pipe(prompt="",
+                          image_embed=image_embed,
+                          num_inference_steps=num_inference_steps,
+                          embedded_guidance=3.5,
+                          negative_prompt=negative_prompt,
+                          cfg_scale=cfg_scale,
+                          height=height,
+                          width=width,
+                          seed=seed,
+                          **pipe_kwargs)
+        return image

modeling/decoder/flux_image_pipeline.py

@@ -0,0 +1,192 @@
+from typing import List
+from tqdm import tqdm
+import torch
+from diffsynth.models import ModelManager
+from diffsynth.controlnets import ControlNetConfigUnit
+from diffsynth.prompters.flux_prompter import FluxPrompter
+from diffsynth.pipelines.flux_image import FluxImagePipeline, lets_dance_flux, TeaCache
+
+
+class FluxPrompterAll2All(FluxPrompter):
+    def encode_prompt(
+        self,
+        prompt,
+        positive=True,
+        device="cuda",
+        t5_sequence_length=512,
+        clip_only=False
+    ):
+        prompt = self.process_prompt(prompt, positive=positive)
+        # CLIP
+        pooled_prompt_emb = self.encode_prompt_using_clip(prompt, self.text_encoder_1, self.tokenizer_1, 77, device)
+        if clip_only:
+            return None, pooled_prompt_emb, None
+        # T5
+        prompt_emb = self.encode_prompt_using_t5(prompt, self.text_encoder_2, self.tokenizer_2, t5_sequence_length, device)
+        # text_ids
+        text_ids = torch.zeros(prompt_emb.shape[0], prompt_emb.shape[1], 3).to(device=device, dtype=prompt_emb.dtype)
+        return prompt_emb, pooled_prompt_emb, text_ids
+
+
+class FluxImagePipelineAll2All(FluxImagePipeline):
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        self.prompter = FluxPrompterAll2All()
+
+    def encode_prompt(self, prompt, positive=True, t5_sequence_length=512, clip_only=False):
+        prompt_emb, pooled_prompt_emb, text_ids = self.prompter.encode_prompt(
+            prompt, device=self.device, positive=positive, t5_sequence_length=t5_sequence_length, clip_only=clip_only
+        )
+        return {"prompt_emb": prompt_emb, "pooled_prompt_emb": pooled_prompt_emb, "text_ids": text_ids}
+
+
+    @staticmethod
+    def from_model_manager(model_manager: ModelManager, controlnet_config_units: List[ControlNetConfigUnit]=[], prompt_refiner_classes=[], prompt_extender_classes=[], device=None, torch_dtype=None):
+        pipe = FluxImagePipelineAll2All(
+            device=model_manager.device if device is None else device,
+            torch_dtype=model_manager.torch_dtype if torch_dtype is None else torch_dtype,
+        )
+        pipe.fetch_models(model_manager, controlnet_config_units, prompt_refiner_classes, prompt_extender_classes)
+        return pipe
+
+
+    def prepare_prompts(self, prompt, image_embed, local_prompts, masks, mask_scales, t5_sequence_length, negative_prompt, cfg_scale):
+        # Extend prompt
+        self.load_models_to_device(['text_encoder_1', 'text_encoder_2'])
+        prompt, local_prompts, masks, mask_scales = self.extend_prompt(prompt, local_prompts, masks, mask_scales)
+
+        # Encode prompts
+        if image_embed is not None:
+            image_embed = image_embed.to(self.torch_dtype)
+            prompt_emb_posi = self.encode_prompt("", positive=True, clip_only=True)
+            if len(image_embed.size()) == 2:
+                image_embed = image_embed.unsqueeze(0)
+            prompt_emb_posi['prompt_emb'] = image_embed
+            prompt_emb_posi['text_ids'] = torch.zeros(image_embed.shape[0], image_embed.shape[1], 3).to(device=self.device, dtype=self.torch_dtype)
+        else:
+            prompt_emb_posi = self.encode_prompt(prompt, t5_sequence_length=t5_sequence_length)
+        prompt_emb_nega = self.encode_prompt(negative_prompt, positive=False, t5_sequence_length=t5_sequence_length) if cfg_scale != 1.0 else None
+        prompt_emb_locals = [self.encode_prompt(prompt_local, t5_sequence_length=t5_sequence_length) for prompt_local in local_prompts]
+        return prompt_emb_posi, prompt_emb_nega, prompt_emb_locals
+
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        # Prompt
+        prompt,
+        negative_prompt="",
+        cfg_scale=1.0,
+        embedded_guidance=3.5,
+        t5_sequence_length=512,
+        # Image
+        input_image=None,
+        denoising_strength=1.0,
+        height=1024,
+        width=1024,
+        seed=None,
+        # image_embed
+        image_embed=None,
+        # Steps
+        num_inference_steps=30,
+        # local prompts
+        local_prompts=(),
+        masks=(),
+        mask_scales=(),
+        # ControlNet
+        controlnet_image=None,
+        controlnet_inpaint_mask=None,
+        enable_controlnet_on_negative=False,
+        # IP-Adapter
+        ipadapter_images=None,
+        ipadapter_scale=1.0,
+        # EliGen
+        eligen_entity_prompts=None,
+        eligen_entity_masks=None,
+        enable_eligen_on_negative=False,
+        enable_eligen_inpaint=False,
+        # TeaCache
+        tea_cache_l1_thresh=None,
+        # Tile
+        tiled=False,
+        tile_size=128,
+        tile_stride=64,
+        # Progress bar
+        progress_bar_cmd=tqdm,
+        progress_bar_st=None,
+    ):
+        height, width = self.check_resize_height_width(height, width)
+
+        # Tiler parameters
+        tiler_kwargs = {"tiled": tiled, "tile_size": tile_size, "tile_stride": tile_stride}
+
+        # Prepare scheduler
+        self.scheduler.set_timesteps(num_inference_steps, denoising_strength)
+
+        # Prepare latent tensors
+        latents, input_latents = self.prepare_latents(input_image, height, width, seed, tiled, tile_size, tile_stride)
+
+        # Prompt
+        prompt_emb_posi, prompt_emb_nega, prompt_emb_locals = self.prepare_prompts(prompt, image_embed, local_prompts, masks, mask_scales, t5_sequence_length, negative_prompt, cfg_scale)
+
+        # Extra input
+        extra_input = self.prepare_extra_input(latents, guidance=embedded_guidance)
+
+        # Entity control
+        eligen_kwargs_posi, eligen_kwargs_nega, fg_mask, bg_mask = self.prepare_eligen(prompt_emb_nega, eligen_entity_prompts, eligen_entity_masks, width, height, t5_sequence_length, enable_eligen_inpaint, enable_eligen_on_negative, cfg_scale)
+
+        # IP-Adapter
+        ipadapter_kwargs_list_posi, ipadapter_kwargs_list_nega = self.prepare_ipadapter(ipadapter_images, ipadapter_scale)
+
+        # ControlNets
+        controlnet_kwargs_posi, controlnet_kwargs_nega, local_controlnet_kwargs = self.prepare_controlnet(controlnet_image, masks, controlnet_inpaint_mask, tiler_kwargs, enable_controlnet_on_negative)
+
+        # TeaCache
+        tea_cache_kwargs = {"tea_cache": TeaCache(num_inference_steps, rel_l1_thresh=tea_cache_l1_thresh) if tea_cache_l1_thresh is not None else None}
+
+        # Denoise
+        self.load_models_to_device(['dit', 'controlnet'])
+        for progress_id, timestep in enumerate(progress_bar_cmd(self.scheduler.timesteps)):
+            timestep = timestep.unsqueeze(0).to(self.device)
+
+            # Positive side
+            inference_callback = lambda prompt_emb_posi, controlnet_kwargs: lets_dance_flux(
+                dit=self.dit, controlnet=self.controlnet,
+                hidden_states=latents, timestep=timestep,
+                **prompt_emb_posi, **tiler_kwargs, **extra_input, **controlnet_kwargs, **ipadapter_kwargs_list_posi, **eligen_kwargs_posi, **tea_cache_kwargs,
+            )
+            noise_pred_posi = self.control_noise_via_local_prompts(
+                prompt_emb_posi, prompt_emb_locals, masks, mask_scales, inference_callback,
+                special_kwargs=controlnet_kwargs_posi, special_local_kwargs_list=local_controlnet_kwargs
+            )
+
+            # Inpaint
+            if enable_eligen_inpaint:
+                noise_pred_posi = self.inpaint_fusion(latents, input_latents, noise_pred_posi, fg_mask, bg_mask, progress_id)
+            
+            # Classifier-free guidance
+            if cfg_scale != 1.0:
+                # Negative side
+                noise_pred_nega = lets_dance_flux(
+                    dit=self.dit, controlnet=self.controlnet,
+                    hidden_states=latents, timestep=timestep,
+                    **prompt_emb_nega, **tiler_kwargs, **extra_input, **controlnet_kwargs_nega, **ipadapter_kwargs_list_nega, **eligen_kwargs_nega,
+                )
+                noise_pred = noise_pred_nega + cfg_scale * (noise_pred_posi - noise_pred_nega)
+            else:
+                noise_pred = noise_pred_posi
+
+            # Iterate
+            latents = self.scheduler.step(noise_pred, self.scheduler.timesteps[progress_id], latents)
+
+            # UI
+            if progress_bar_st is not None:
+                progress_bar_st.progress(progress_id / len(self.scheduler.timesteps))
+        
+        # Decode image
+        self.load_models_to_device(['vae_decoder'])
+        image = self.decode_image(latents, **tiler_kwargs)
+
+        # Offload all models
+        self.load_models_to_device([])
+        return image

requirements.txt

@@ -1,7 +1,7 @@
 torch>=2.0.0
 torchvision
 cupy-cuda12x
-transformers==4.46.2
+transformers==4.49.0
 controlnet-aux==0.0.7
 imageio
 imageio[ffmpeg]
@@ -11,3 +11,4 @@ sentencepiece
 protobuf
 modelscope
 ftfy
+qwen_vl_utils

run_single.sh

@@ -0,0 +1,4 @@
+accelerate launch \
+    train.py \
+    --output_path models/nexus_v3 \
+    --steps_per_epoch 4000

test.py

@@ -0,0 +1,312 @@
+from transformers import AutoConfig, AutoTokenizer
+import torch, json, os, torchvision
+from modeling.ar.modeling_qwen2_5_vl import Qwen2_5_VLForConditionalGeneration
+from modeling.ar.processing_qwen2_5_vl import Qwen2_5_VLProcessor
+from diffsynth import ModelManager, FluxImagePipeline, load_state_dict, hash_state_dict_keys
+from qwen_vl_utils import smart_resize
+from PIL import Image
+import numpy as np
+from torchvision.transforms import v2
+
+
+
+class SingleTaskDataset(torch.utils.data.Dataset):
+    def __init__(
+        self,
+        base_path,
+        keys=(("image_1", "image_2", "editing_instruction"), ("image_2", "image_1", "reverse_editing_instruction"), (None, "image_1", "prompt")),
+        height=1024, width=1024, random=True, steps_per_epoch=1000, metadata_path=None
+    ):
+        self.base_path = base_path
+        self.keys = keys
+        self.metadata = []
+        self.bad_data = []
+        self.height = height
+        self.width = width
+        self.random = random
+        self.steps_per_epoch = steps_per_epoch
+        self.image_process = v2.Compose([
+            v2.CenterCrop(size=(height, width)),
+            v2.ToImage(),
+            v2.ToDtype(torch.float32, scale=True),
+            v2.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
+        ])
+        if metadata_path is None:
+            self.search_for_data("", report_data_log=True)
+            self.report_data_log()
+        else:
+            with open(metadata_path, "r", encoding="utf-8-sig") as f:
+                self.metadata = json.load(f)
+
+
+    def report_data_log(self):
+        print(f"{len(self.metadata)} valid data, {len(self.bad_data)} invalid data.")
+
+
+    def dump_metadata(self, path):
+        with open(path, "w", encoding="utf-8") as f:
+            json.dump(self.metadata, f, ensure_ascii=False, indent=4)
+        
+        
+    def parse_json_file(self, absolute_path, relative_path):
+        data_list = []
+        with open(absolute_path, "r") as f:
+            metadata = json.load(f)
+            for image_1, image_2, instruction in self.keys:
+                image_1 = os.path.join(relative_path, metadata[image_1]) if image_1 is not None else None
+                image_2 = os.path.join(relative_path, metadata[image_2])
+                instruction = metadata[instruction]
+                data_list.append((image_1, image_2, instruction))
+        return data_list
+    
+        
+    def search_for_data(self, path, report_data_log=False):
+        now_path = os.path.join(self.base_path, path)
+        if os.path.isfile(now_path) and path.endswith(".json"):
+            try:
+                data_list = self.parse_json_file(now_path, os.path.dirname(path))
+                self.metadata.extend(data_list)
+            except:
+                self.bad_data.append(now_path)
+        elif os.path.isdir(now_path):
+            for sub_path in os.listdir(now_path):
+                self.search_for_data(os.path.join(path, sub_path))
+                if report_data_log and os.path.isdir(os.path.join(self.base_path, path, sub_path)):
+                    self.report_data_log()
+                
+                
+    def load_image(self, image_path, skip_process=False):
+        image_path = os.path.join(self.base_path, image_path)
+        image = Image.open(image_path).convert("RGB")
+        width, height = image.size
+        scale = max(self.width / width, self.height / height)
+        image = torchvision.transforms.functional.resize(
+            image,
+            (round(height*scale), round(width*scale)),
+            interpolation=torchvision.transforms.InterpolationMode.BILINEAR
+        )
+        if skip_process:
+            return image
+        image = self.image_process(image)
+        return image
+                
+                
+    def load_data(self, data_id):
+        image_1, image_2, instruction = self.metadata[data_id]
+        image_1 = self.load_image(image_1, skip_process=True) if image_1 is not None else None
+        image_2 = self.load_image(image_2)
+        return {"image_1": image_1, "image_2": image_2, "instruction": instruction}
+        
+        
+    def __getitem__(self, data_id):
+        if self.random:
+            data_id = (torch.randint(0, len(self.metadata), (1,))[0] + data_id) % len(self.metadata)
+            data = self.load_data(data_id)
+            return data
+        else:
+            return self.load_data(data_id)
+
+
+    def __len__(self):
+        return self.steps_per_epoch if self.random else len(self.metadata)
+    
+    
+    
+class MultiTaskDataset(torch.utils.data.Dataset):
+    def __init__(self, dataset_list, dataset_weight, steps_per_epoch=1000):
+        self.dataset_list = dataset_list
+        self.dataset_weight = torch.tensor(dataset_weight, dtype=torch.float)
+        self.steps_per_epoch = steps_per_epoch
+
+        
+    def __getitem__(self, data_id):
+        dataset_id = torch.multinomial(self.dataset_weight, 1).tolist()[0]
+        data_id = torch.randint(0, len(self.dataset_list[dataset_id]), (1,))[0]
+        data = self.dataset_list[dataset_id][data_id]
+        return data
+
+
+    def __len__(self):
+        return self.steps_per_epoch
+
+
+
+class NexusGenQwenVLEncoder(torch.nn.Module):
+    def __init__(self, model_path, torch_dtype="auto", device="cpu"):
+        super().__init__()
+        model_config = AutoConfig.from_pretrained(model_path)
+        self.tokenizer = AutoTokenizer.from_pretrained(model_path)
+        self.model = Qwen2_5_VLForConditionalGeneration.from_pretrained(model_path, config=model_config, trust_remote_code=True, torch_dtype=torch_dtype, device_map=device)
+        self.processor = Qwen2_5_VLProcessor.from_pretrained(model_path)
+        self.t2i_template = "Here is an image based on the description: <|vision_start|><|image_pad|><|vision_end|>"
+        self.i2i_template = "Here is the image: <|vision_start|><|image_pad|><|vision_end|>"
+    
+    @staticmethod
+    def from_pretrained(model_path, torch_dtype="auto", device="cpu"):
+        return NexusGenQwenVLEncoder(model_path, torch_dtype=torch_dtype, device=device).eval()
+    
+    def process_images(self, images=None):
+        if images is None:
+            return None
+        # resize input to max_pixels to avoid oom
+        for j in range(len(images)):
+            input_image = images[j]
+            input_w, input_h = input_image.size
+            resized_height, resized_width = smart_resize(
+                input_h,
+                input_w,
+                max_pixels=262640,
+            )
+            images[j] = input_image.resize((resized_width, resized_height))
+        return images
+
+    def forward(self, prompt, images=None, num_img_tokens=81):
+        messages = [
+            {
+                "role": "user",
+                "content": [{
+                    "type": "text",
+                    "text": prompt
+                },],
+            },
+            {
+                "role": "assistant",
+                "content": [{
+                    "type": "text",
+                    "text": self.t2i_template if images is None else self.i2i_template
+                },],
+            }
+        ]
+        images = self.process_images(images)
+        target_image = Image.fromarray(np.zeros((252, 252, 3), dtype=np.uint8))
+        if images is None:
+            images = [target_image]
+        else:
+            images = images + [target_image]
+        text = self.processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=False)
+        inputs = self.processor(
+            text=[text],
+            images=images,
+            padding=True,
+            return_tensors="pt",
+        )
+        inputs = inputs.to(self.model.device)
+
+        input_embeds = self.model.model.embed_tokens(inputs['input_ids'])
+        image_embeds = self.model.visual(inputs['pixel_values'], grid_thw=inputs['image_grid_thw'])
+        ground_truth_image_embeds = image_embeds[-num_img_tokens:]
+        input_image_embeds = image_embeds[:-num_img_tokens]
+
+        image_mask = inputs['input_ids'] == self.model.config.image_token_id
+        indices = image_mask.cumsum(dim=1)
+        input_image_mask = torch.logical_and(indices <= (image_embeds.shape[0] - ground_truth_image_embeds.shape[0]), image_mask)
+        gt_image_mask = torch.logical_and(image_mask, ~input_image_mask)
+        input_image_mask = input_image_mask.unsqueeze(-1).expand_as(input_embeds)
+        input_embeds = input_embeds.masked_scatter(input_image_mask, input_image_embeds)
+
+        position_ids, _ = self.model.get_rope_index(inputs['input_ids'],
+                                                    inputs['image_grid_thw'],
+                                                    attention_mask=inputs['attention_mask'])
+        position_ids = position_ids.contiguous()
+        outputs = self.model(inputs_embeds=input_embeds, position_ids=position_ids, attention_mask=inputs['attention_mask'], return_dict=True)
+        output_image_embeddings = outputs.image_embeddings[:, :-1, :] # shift right
+        output_image_embeddings = output_image_embeddings[gt_image_mask[:, 1:]]
+        output_image_embeddings = output_image_embeddings.unsqueeze(0)
+        return output_image_embeddings
+
+
+
+model_manager = ModelManager(torch_dtype=torch.bfloat16, device="cuda")
+model_manager.load_models([
+    "models/FLUX/FLUX.1-dev/text_encoder/model.safetensors",
+    "models/FLUX/FLUX.1-dev/ae.safetensors",
+    "models/FLUX/FLUX.1-dev/flux1-dev.safetensors"
+])
+pipe = FluxImagePipeline.from_model_manager(model_manager)
+
+# state_dict = load_state_dict("models/DiffSynth-Studio/Nexus-Gen/decoder_81_512.bin", torch_dtype=torch.bfloat16)
+# pipe.dit.load_state_dict(state_dict, strict=False)
+
+adapter = torch.nn.Sequential(torch.nn.Linear(3584, 4096), torch.nn.LayerNorm(4096), torch.nn.ReLU(), torch.nn.Linear(4096, 4096), torch.nn.LayerNorm(4096)).to(dtype=torch.bfloat16, device="cuda")
+# adapter.load_state_dict(state_dict, strict=False)
+
+qwenvl = NexusGenQwenVLEncoder.from_pretrained('models/DiffSynth-Studio/Nexus-Gen').to("cuda")
+
+sd = {}
+for i in range(1, 6):
+    print(i)
+    sd.update(load_state_dict(f"models/nexus_v3/epoch-19/model-0000{i}-of-00005.safetensors", torch_dtype=torch.bfloat16))
+pipe.dit.load_state_dict({i.replace("pipe.dit.", ""): sd[i] for i in sd if i.startswith("pipe.dit.")})
+qwenvl.load_state_dict({i.replace("qwenvl.", ""): sd[i] for i in sd if i.startswith("qwenvl.")})
+adapter.load_state_dict({i.replace("adapter.", ""): sd[i] for i in sd if i.startswith("adapter.")})
+
+
+dataset = MultiTaskDataset(
+    dataset_list=[
+        SingleTaskDataset(
+            "/shark/zhongjie/data/image_pulse_datasets/task1/data/dataset_change_add_remove",
+            keys=(("image_1", "image_2", "editing_instruction"), ("image_2", "image_1", "reverse_editing_instruction"), (None, "image_1", "prompt")),
+            height=1024, width=1024,
+            metadata_path="/shark/zhongjie/data/image_pulse_datasets/task1/data/metadata/20250507_dataset_change_add_remove.json",
+        ),
+        SingleTaskDataset(
+            "/shark/zhongjie/data/image_pulse_datasets/task1/data/dataset_style_transfer",
+            keys=(("image_1", "image_4", "editing_instruction"), ("image_4", "image_1", "reverse_editing_instruction"), (None, "image_1", "prompt")),
+            height=1024, width=1024,
+            metadata_path="/shark/zhongjie/data/image_pulse_datasets/task1/data/metadata/20250507_dataset_style_transfer.json",
+        ),
+        SingleTaskDataset(
+            "/shark/zhongjie/data/image_pulse_datasets/task1/data/dataset_faceid",
+            keys=(("image_1", "image_2", "editing_instruction"), ("image_2", "image_1", "reverse_editing_instruction")),
+            height=1024, width=1024,
+            metadata_path="/shark/zhongjie/data/image_pulse_datasets/task1/data/metadata/20250507_dataset_faceid.json",
+        ),
+    ],
+    dataset_weight=(4, 2, 1,),
+    steps_per_epoch=100000
+)
+
+
+torch.manual_seed(0)
+for data_id, data in enumerate(dataset):
+    image_1 = data["image_1"]
+    image_2 = data["image_2"].cpu().float().permute(1, 2, 0).numpy()
+    image_2 = Image.fromarray(((image_2 / 2 + 0.5).clip(0, 1) * 255).astype("uint8"))
+    instruction = data["instruction"]
+
+    print(instruction)
+    if image_1 is None:
+        with torch.no_grad():
+            instruction = f"Generate an image according to the following description: {instruction}"
+            emb = qwenvl(instruction, images=None)
+            emb = adapter(emb)
+            image_3 = pipe("", image_emb=emb)
+    else:
+        with torch.no_grad():
+            instruction = f"<|vision_start|><|image_pad|><|vision_end|> {instruction}"
+            emb = qwenvl(instruction, images=[image_1])
+            emb = adapter(emb)
+            image_3 = pipe("", image_emb=emb)
+    
+    if image_1 is not None:
+        image_1.save(f"data/output/{data_id}_1.jpg")
+    image_2.save(f"data/output/{data_id}_2.jpg")
+    image_3.save(f"data/output/{data_id}_3.jpg")
+    if data_id >= 100:
+        break
+
+
+
+# with torch.no_grad():
+#     instruction = "Generate an image according to the following description: hyper-realistic and detailed 2010s movie still portrait of Josip Broz Tito, by Paolo Sorrentino, Leica SL2 50mm, clear color, high quality, high textured, dramatic light, cinematic"
+#     emb = qwenvl(instruction, images=None)
+#     emb = adapter(emb)
+#     image = pipe("", image_emb=emb)
+#     image.save("image_1.jpg")
+    
+# with torch.no_grad():
+#     instruction = "<|vision_start|><|image_pad|><|vision_end|> transform the image into a cartoon style with vibrant colors and a confident expression."
+#     emb = qwenvl(instruction, images=[Image.open("image_1.jpg")])
+#     emb = adapter(emb)
+#     image = pipe("", image_emb=emb)
+#     image.save("image_2.jpg")

train.py

@@ -0,0 +1,421 @@
+from diffsynth import ModelManager, FluxImagePipeline, load_state_dict
+from diffsynth.trainers.text_to_image import LightningModelForT2ILoRA, add_general_parsers, launch_training_task
+from diffsynth.models.lora import FluxLoRAConverter
+import torch, os, argparse
+from diffsynth.pipelines.flux_image import lets_dance_flux
+from accelerate import Accelerator
+from tqdm import tqdm
+import torch, os, json, torchvision
+from PIL import Image
+from torchvision.transforms import v2
+from transformers import AutoConfig, AutoTokenizer
+import torch
+from modeling.ar.modeling_qwen2_5_vl import Qwen2_5_VLForConditionalGeneration
+from modeling.ar.processing_qwen2_5_vl import Qwen2_5_VLProcessor
+from diffsynth import ModelManager, FluxImagePipeline, load_state_dict, hash_state_dict_keys
+from qwen_vl_utils import smart_resize
+from PIL import Image
+import numpy as np
+import lightning as pl
+os.environ["TOKENIZERS_PARALLELISM"] = "True"
+
+
+
+class SingleTaskDataset(torch.utils.data.Dataset):
+    def __init__(
+        self,
+        base_path,
+        keys=(("image_1", "image_2", "editing_instruction"), ("image_2", "image_1", "reverse_editing_instruction"), (None, "image_1", "prompt")),
+        height=1024, width=1024, random=True, steps_per_epoch=1000, metadata_path=None
+    ):
+        self.base_path = base_path
+        self.keys = keys
+        self.metadata = []
+        self.bad_data = []
+        self.height = height
+        self.width = width
+        self.random = random
+        self.steps_per_epoch = steps_per_epoch
+        self.image_process = v2.Compose([
+            v2.CenterCrop(size=(height, width)),
+            v2.ToImage(),
+            v2.ToDtype(torch.float32, scale=True),
+            v2.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
+        ])
+        if metadata_path is None:
+            self.search_for_data("", report_data_log=True)
+            self.report_data_log()
+        else:
+            with open(metadata_path, "r", encoding="utf-8-sig") as f:
+                self.metadata = json.load(f)
+
+
+    def report_data_log(self):
+        print(f"{len(self.metadata)} valid data, {len(self.bad_data)} invalid data.")
+
+
+    def dump_metadata(self, path):
+        with open(path, "w", encoding="utf-8") as f:
+            json.dump(self.metadata, f, ensure_ascii=False, indent=4)
+        
+        
+    def parse_json_file(self, absolute_path, relative_path):
+        data_list = []
+        with open(absolute_path, "r") as f:
+            metadata = json.load(f)
+            for image_1, image_2, instruction in self.keys:
+                image_1 = os.path.join(relative_path, metadata[image_1]) if image_1 is not None else None
+                image_2 = os.path.join(relative_path, metadata[image_2])
+                instruction = metadata[instruction]
+                data_list.append((image_1, image_2, instruction))
+        return data_list
+    
+        
+    def search_for_data(self, path, report_data_log=False):
+        now_path = os.path.join(self.base_path, path)
+        if os.path.isfile(now_path) and path.endswith(".json"):
+            try:
+                data_list = self.parse_json_file(now_path, os.path.dirname(path))
+                self.metadata.extend(data_list)
+            except:
+                self.bad_data.append(now_path)
+        elif os.path.isdir(now_path):
+            for sub_path in os.listdir(now_path):
+                self.search_for_data(os.path.join(path, sub_path))
+                if report_data_log and os.path.isdir(os.path.join(self.base_path, path, sub_path)):
+                    self.report_data_log()
+                
+                
+    def load_image(self, image_path, skip_process=False):
+        image_path = os.path.join(self.base_path, image_path)
+        image = Image.open(image_path).convert("RGB")
+        width, height = image.size
+        scale = max(self.width / width, self.height / height)
+        image = torchvision.transforms.functional.resize(
+            image,
+            (round(height*scale), round(width*scale)),
+            interpolation=torchvision.transforms.InterpolationMode.BILINEAR
+        )
+        if skip_process:
+            return image
+        image = self.image_process(image)
+        return image
+                
+                
+    def load_data(self, data_id):
+        image_1, image_2, instruction = self.metadata[data_id]
+        image_1 = self.load_image(image_1, skip_process=True) if image_1 is not None else None
+        image_2 = self.load_image(image_2)
+        return {"image_1": image_1, "image_2": image_2, "instruction": instruction}
+        
+        
+    def __getitem__(self, data_id):
+        if self.random:
+            while True:
+                try:
+                    data_id = (torch.randint(0, len(self.metadata), (1,))[0] + data_id) % len(self.metadata)
+                    data = self.load_data(data_id)
+                    return data
+                except:
+                    continue
+        else:
+            return self.load_data(data_id)
+
+
+    def __len__(self):
+        return self.steps_per_epoch if self.random else len(self.metadata)
+    
+    
+    
+class MultiTaskDataset(torch.utils.data.Dataset):
+    def __init__(self, dataset_list, dataset_weight, steps_per_epoch=1000):
+        self.dataset_list = dataset_list
+        self.dataset_weight = torch.tensor(dataset_weight, dtype=torch.float)
+        self.steps_per_epoch = steps_per_epoch
+
+        
+    def __getitem__(self, data_id):
+        dataset_id = torch.multinomial(self.dataset_weight, 1).tolist()[0]
+        data_id = torch.randint(0, len(self.dataset_list[dataset_id]), (1,))[0]
+        data = self.dataset_list[dataset_id][data_id]
+        return data
+
+
+    def __len__(self):
+        return self.steps_per_epoch
+    
+    
+    
+class NexusGenQwenVLEncoder(torch.nn.Module):
+    def __init__(self, model_path, torch_dtype="auto", device="cpu"):
+        super().__init__()
+        model_config = AutoConfig.from_pretrained(model_path)
+        self.tokenizer = AutoTokenizer.from_pretrained(model_path)
+        self.model = Qwen2_5_VLForConditionalGeneration.from_pretrained(model_path, config=model_config, trust_remote_code=True, torch_dtype=torch_dtype, device_map=device)
+        self.processor = Qwen2_5_VLProcessor.from_pretrained(model_path)
+        self.t2i_template = "Here is an image based on the description: <|vision_start|><|image_pad|><|vision_end|>"
+        self.i2i_template = "Here is the image: <|vision_start|><|image_pad|><|vision_end|>"
+    
+    @staticmethod
+    def from_pretrained(model_path, torch_dtype="auto", device="cpu"):
+        return NexusGenQwenVLEncoder(model_path, torch_dtype=torch_dtype, device=device).eval()
+    
+    def process_images(self, images=None):
+        if images is None:
+            return None
+        # resize input to max_pixels to avoid oom
+        for j in range(len(images)):
+            input_image = images[j]
+            input_w, input_h = input_image.size
+            resized_height, resized_width = smart_resize(
+                input_h,
+                input_w,
+                max_pixels=262640,
+            )
+            images[j] = input_image.resize((resized_width, resized_height))
+        return images
+
+    def forward(self, prompt, images=None, num_img_tokens=81):
+        messages = [
+            {
+                "role": "user",
+                "content": [{
+                    "type": "text",
+                    "text": prompt
+                },],
+            },
+            {
+                "role": "assistant",
+                "content": [{
+                    "type": "text",
+                    "text": self.t2i_template if images is None else self.i2i_template
+                },],
+            }
+        ]
+        images = self.process_images(images)
+        target_image = Image.fromarray(np.zeros((252, 252, 3), dtype=np.uint8))
+        if images is None:
+            images = [target_image]
+        else:
+            images = images + [target_image]
+        text = self.processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=False)
+        inputs = self.processor(
+            text=[text],
+            images=images,
+            padding=True,
+            return_tensors="pt",
+        )
+        inputs = inputs.to(self.model.device)
+
+        input_embeds = self.model.model.embed_tokens(inputs['input_ids'])
+        image_embeds = self.model.visual(inputs['pixel_values'], grid_thw=inputs['image_grid_thw'])
+        ground_truth_image_embeds = image_embeds[-num_img_tokens:]
+        input_image_embeds = image_embeds[:-num_img_tokens]
+
+        image_mask = inputs['input_ids'] == self.model.config.image_token_id
+        indices = image_mask.cumsum(dim=1)
+        input_image_mask = torch.logical_and(indices <= (image_embeds.shape[0] - ground_truth_image_embeds.shape[0]), image_mask)
+        gt_image_mask = torch.logical_and(image_mask, ~input_image_mask)
+        input_image_mask = input_image_mask.unsqueeze(-1).expand_as(input_embeds)
+        input_embeds = input_embeds.masked_scatter(input_image_mask, input_image_embeds)
+
+        position_ids, _ = self.model.get_rope_index(inputs['input_ids'],
+                                                    inputs['image_grid_thw'],
+                                                    attention_mask=inputs['attention_mask'])
+        position_ids = position_ids.contiguous()
+        outputs = self.model(inputs_embeds=input_embeds, position_ids=position_ids, attention_mask=inputs['attention_mask'], return_dict=True)
+        output_image_embeddings = outputs.image_embeddings[:, :-1, :] # shift right
+        output_image_embeddings = output_image_embeddings[gt_image_mask[:, 1:]]
+        output_image_embeddings = output_image_embeddings.unsqueeze(0)
+        return output_image_embeddings
+
+
+
+class UnifiedModel(pl.LightningModule):
+    def __init__(self, flux_text_encoder_path, flux_vae_path, flux_dit_path, flux_decoder_path, qwenvl_path):
+        super().__init__()
+        # Load models
+        model_manager = ModelManager(torch_dtype=torch.bfloat16, device="cpu")
+        model_manager.load_models([
+            flux_text_encoder_path,
+            flux_vae_path,
+            flux_dit_path
+        ])
+        self.pipe = FluxImagePipeline.from_model_manager(model_manager)
+
+        state_dict = load_state_dict(flux_decoder_path, torch_dtype=torch.bfloat16)
+        self.pipe.dit.load_state_dict(state_dict, strict=False)
+
+        self.adapter = torch.nn.Sequential(torch.nn.Linear(3584, 4096), torch.nn.LayerNorm(4096), torch.nn.ReLU(), torch.nn.Linear(4096, 4096), torch.nn.LayerNorm(4096)).to(dtype=torch.bfloat16)
+        self.adapter.load_state_dict(state_dict, strict=False)
+
+        self.qwenvl = NexusGenQwenVLEncoder.from_pretrained(qwenvl_path)
+        
+        self.pipe.vae_decoder.requires_grad_(False)
+        self.pipe.vae_encoder.requires_grad_(False)
+        self.pipe.text_encoder_1.requires_grad_(False)
+        self.qwenvl.requires_grad_(False)
+        self.qwenvl.model.visual.requires_grad_(False)
+        self.pipe.train()
+        self.adapter.train()
+        self.qwenvl.train()
+        self.qwenvl.model.visual.eval()
+        # self.qwenvl.model.model.gradient_checkpointing = True
+        
+        self.pipe.scheduler.set_timesteps(1000, training=True)
+        
+        
+    def training_step(self, batch, batch_idx):
+        # Data
+        text, image = batch["instruction"], batch["image_2"]
+        image_ref = batch["image_1"]
+        image = image.unsqueeze(0)
+
+        # Prepare input parameters
+        self.pipe.device = self.device
+        latents = self.pipe.vae_encoder(image.to(dtype=self.pipe.torch_dtype, device=self.device))
+        noise = torch.randn_like(latents)
+        timestep_id = torch.randint(0, self.pipe.scheduler.num_train_timesteps, (1,))
+        timestep = self.pipe.scheduler.timesteps[timestep_id].to(self.device)
+        extra_input = self.pipe.prepare_extra_input(latents)
+        noisy_latents = self.pipe.scheduler.add_noise(latents, noise, timestep)
+        training_target = self.pipe.scheduler.training_target(latents, noise, timestep)
+
+        # Compute loss
+        if image_ref is None:
+            instruction = f"Generate an image according to the following description: {text}"
+            images_ref = None
+        else:
+            instruction = f"<|vision_start|><|image_pad|><|vision_end|> {text}"
+            images_ref = [image_ref]
+        emb = self.qwenvl(instruction, images=images_ref)
+        emb = self.adapter(emb)
+        prompt_emb = self.pipe.encode_prompt("", positive=True, image_emb=emb)
+        
+        noise_pred = lets_dance_flux(
+            self.pipe.denoising_model(),
+            hidden_states=noisy_latents, timestep=timestep, **prompt_emb, **extra_input,
+            image_emb=emb,
+            use_gradient_checkpointing=False
+        )
+        loss = torch.nn.functional.mse_loss(noise_pred.float(), training_target.float())
+        loss = loss * self.pipe.scheduler.training_weight(timestep)
+        return loss
+
+
+    def forward(self, batch):
+        return self.training_step(batch, 0)
+
+
+
+def search_for_last_checkpoint(path):
+    if not os.path.exists(path):
+        return None, 0
+    checkpoint_list = os.listdir(path)
+    checkpoint_list = [int(checkpoint.split("-")[1]) for checkpoint in checkpoint_list if checkpoint.startswith("epoch")]
+    if len(checkpoint_list) == 0:
+        return None, 0
+    else:
+        max_epoch_id = max(checkpoint_list)
+        return f"{path}/epoch-{max_epoch_id}/model.safetensors", max_epoch_id + 1
+    
+    
+
+def parse_args():
+    parser = argparse.ArgumentParser(description="Simple example of a training script.")
+    parser.add_argument(
+        "--gradient_accumulation_steps",
+        type=int,
+        default=1,
+        help="gradient_accumulation_steps",
+    )
+    parser.add_argument(
+        "--steps_per_epoch",
+        type=int,
+        default=1000,
+        help="steps_per_epoch",
+    )
+    parser.add_argument(
+        "--output_path",
+        type=str,
+        default="./models",
+        help="output_path",
+    )
+    parser.add_argument(
+        "--learning_rate",
+        type=float,
+        default=1e-5,
+        help="learning_rate",
+    )
+    args = parser.parse_args()
+    return args
+
+
+
+if __name__ == '__main__':
+    args = parse_args()
+    model = UnifiedModel(
+        "models/FLUX/FLUX.1-dev/text_encoder/model.safetensors",
+        "models/FLUX/FLUX.1-dev/ae.safetensors",
+        "models/FLUX/FLUX.1-dev/flux1-dev.safetensors",
+        "models/DiffSynth-Studio/Nexus-Gen/decoder_81_512.bin",
+        "models/DiffSynth-Studio/Nexus-Gen",
+    )
+    # dataset and data loader
+    accelerator = Accelerator(gradient_accumulation_steps=args.gradient_accumulation_steps)
+    dataset = MultiTaskDataset(
+        dataset_list=[
+            SingleTaskDataset(
+                "/shark/zhongjie/data/image_pulse_datasets/task1/data/dataset_change_add_remove",
+                keys=(("image_1", "image_2", "editing_instruction"), ("image_2", "image_1", "reverse_editing_instruction"), (None, "image_1", "prompt")),
+                height=1024, width=1024,
+                metadata_path="/shark/zhongjie/data/image_pulse_datasets/task1/data/metadata/20250507_dataset_change_add_remove.json",
+            ),
+            SingleTaskDataset(
+                "/shark/zhongjie/data/image_pulse_datasets/task1/data/dataset_style_transfer",
+                keys=(("image_1", "image_4", "editing_instruction"), ("image_4", "image_1", "reverse_editing_instruction"), (None, "image_1", "prompt")),
+                height=1024, width=1024,
+                metadata_path="/shark/zhongjie/data/image_pulse_datasets/task1/data/metadata/20250507_dataset_style_transfer.json",
+            ),
+            SingleTaskDataset(
+                "/shark/zhongjie/data/image_pulse_datasets/task1/data/dataset_faceid",
+                keys=(("image_1", "image_2", "editing_instruction"), ("image_2", "image_1", "reverse_editing_instruction")),
+                height=1024, width=1024,
+                metadata_path="/shark/zhongjie/data/image_pulse_datasets/task1/data/metadata/20250507_dataset_faceid.json",
+            ),
+        ],
+        dataset_weight=(4, 2, 1,),
+        steps_per_epoch=args.steps_per_epoch * accelerator.num_processes,
+    )
+    train_loader = torch.utils.data.DataLoader(
+        dataset,
+        shuffle=True,
+        batch_size=1,
+        num_workers=1,
+        collate_fn=lambda x: x[0]
+    )
+    # train
+    pretrained_path, start_epoch_id = search_for_last_checkpoint(args.output_path)
+    if pretrained_path is not None:
+        print(f"pretrained_path: {pretrained_path}")
+        model.load_state_dict(load_state_dict(pretrained_path, torch_dtype=torch.bfloat16), assign=True, strict=False)
+
+    model.to(torch.bfloat16)
+    model.to(accelerator.device)
+    
+    trainable_modules = filter(lambda p: p.requires_grad, model.parameters())
+    optimizer = torch.optim.AdamW(trainable_modules, lr=args.learning_rate)
+    
+    model, optimizer, train_loader = accelerator.prepare(model, optimizer, train_loader)
+
+    for epoch in range(start_epoch_id, 100000):
+        for batch in tqdm(train_loader, desc=f"epoch-{epoch}", disable=not accelerator.is_local_main_process):
+            with accelerator.accumulate(model):
+                optimizer.zero_grad()
+                loss = model(batch)
+                accelerator.backward(loss)
+                optimizer.step()
+        path = args.output_path
+        os.makedirs(path, exist_ok=True)
+        accelerator.wait_for_everyone()
+        accelerator.save_model(model, f"{path}/epoch-{epoch}", max_shard_size="10GB", safe_serialization=True)