本仓库包含EndoCoT的官方模型 checkpoint，相关研究成果已发表于论文 EndoCoT: Scaling Endogenous Chain-of-Thought Reasoning in Diffusion Models。
git clone https://github.com/InternLM/EndoCoT
cd EndoCoT
conda create -n EndoCoT python=3.10
conda activate EndoCot
# Please install the version of torch compatible with your machine.
pip install -r requirements.txt
# Please install the version of vLLM compatible with your machine.下载模型 checkpoint：
遵循 Diffthinker 的配置，我们为 Qwen-Image-Edit 提供了一个定制化的 checkpoint。此 checkpoint 已从原始
safetensors合并而来，以确保与 Diffsynth-Studio 训练的兼容性。为确保正确加载和推理，请使用本仓库提供的 checkpoint，而非官方版本。
测试单个案例
cd test
python test.py \
--task Maze \
--model_root /path/to/merged_ckpts \
--lora_path /path/to/your_lora_weight.safetensors \
--input_image ./data/sudoku_sample.png \
--output_dir ./outputs/sudoku_results评估我们的模型 checkpoint
我们采用与 Diffthinker 完全相同的设置
cd Maze
bash eval/gen_and_parse.sh
bash eval/eval_path.sh下载数据集和 metadata.csv
由于元数据使用相对路径，请确保数据集文件与
metadata.csv放置在同一目录下。
训练您的模型
cd DiffSynth-Studio
bash add/Maze/stage1.sh
python change_ckpt_prefix.py --src /path/to/the/Maze/save/dir/Maze_stage1
bash add/Maze/stage2.sh
python change_ckpt_prefix.py --src /path/to/the/Maze/save/dir/Maze_stage2自定义说明： 由于当前实现方式较为直接，您只能在
DiffSynth-Studio/diffsynth/pipelines/qwen_image.py中手动调整潜在推理步骤：
- 第 442 行： 修改
infer_steps。- 第 471 行： 修改
training_steps。我们计划在未来版本中对此进行优化。
def encode_prompt_edit(self, pipe: QwenImagePipeline, prompt, edit_image, is_final, gt_prompt=None, idx=None):
drop_idx = 64
if type(prompt[0])==str:
template = "<|im_start|>system
Describe the key features of the input image (color, shape, size, texture, objects, background), then explain how the user's text instruction should alter or modify the image. Generate a new image that meets the user's requirements while maintaining consistency with the original input where appropriate.<|im_end|>
<|im_start|>user
<|vision_start|><|image_pad|><|vision_end|>{}<|im_end|>
<|im_start|>assistant
"
txt = template.format(prompt[0])
model_inputs = pipe.processor(text=txt, images=edit_image, padding=True, return_tensors="pt").to(pipe.device)
embedding_layers = pipe.text_encoder.model.language_model.get_input_embeddings()
with torch.no_grad():
inputs_embeds = embedding_layers(model_inputs.input_ids)
self.attention_mask = model_inputs.attention_mask
self.pixel_values = model_inputs.pixel_values
self.image_grid_thw = model_inputs.image_grid_thw
else:
inputs_embeds= prompt[0]
# dxl: test use
if is_final==None or idx!=None:
print("现在在inference。或者stage2训练")
if idx!=None:
iter_times = idx-2
else:
# infer step
iter_times = 50
with torch.no_grad():
inputs_embeds = self.manual_generate_eval(
pipe,
inputs_embeds=inputs_embeds,
max_new_tokens=iter_times,
).detach()
# dxl: only update the last 2 tokens
if idx!=None:
inputs_embeds = self.manual_generate_eval(
pipe,
inputs_embeds=inputs_embeds,
max_new_tokens=2,
)
generated_embeds = inputs_embeds
... ...
# dxl：training
if is_final!=None and idx==None:
try:
generated_embeds, _ = self.manual_generate(
pipe,
inputs_embeds=inputs_embeds,
is_final=is_final,
# training steps
max_new_tokens=2,
)
except Exception as e:
print(f"Error!: {type(e).__name__} - {e}")
print(inputs_embeds.shape)
assert False
try:
return split_hidden_states, generated_embeds, eos_loss
except:
print(f"[WARNING] Prompt was not updated correctly for inference.")
return split_hidden_states@article{dai2026endocot,
title={EndoCoT: Scaling Endogenous Chain-of-Thought Reasoning in Diffusion Models},
author={Dai, Xuanlang and Zhou, Yujie and Xing, Long and Bu, Jiazi and Wei, Xilin and Liu, Yuhong and Zhang, Beichen and Chen, Kai and Zang, Yuhang},
journal={arXiv preprint arXiv:2603.12252},
year={2026}
}
