EndoCoT

This repository contains the official model checkpoints for EndoCoT, as presented in the paper EndoCoT: Scaling Endogenous Chain-of-Thought Reasoning in Diffusion Models.

📝TODO

• [x] Open source the training code
• [ ] Open source the training data
• [x] Open source the main task ckpt
• [ ] Open source the edit model ckpt
• [ ] Refactor the codebase for better usability and maintainability

📰News

• 🚀 [2026/3/12] We have released the EndoCoT repository and ckpts.

🌟Highlight

• EndoCoT is a reasoning paradigm for diffusion models that enables step-by-step inference. It outperforms conventional training methods on Qwen-Image-Edit-2511.

• And provide transparent, intermediate reasoning trajectories.

⚡Quick Start

Setup environment

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.

Inference

1. Download the ckpt:

   • You may find our pretrained weights at: EndoCoT

   Following the configuration of Diffthinker, we provide a customized checkpoint for Qwen-Image-Edit. This checkpoint has been merged from the original safetensors to ensure compatibility with*Diffsynth-Studio* training. Please use the checkpoint provided in this repository instead of the official version for correct loading and inference.

2. Test Single Case

   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
   

3. Eval Our Ckpt

   We follow the exact same setting as Diffthinker

   cd Maze
   bash eval/gen_and_parse.sh
   bash eval/eval_path.sh
   

Training

1. Download the datasets & metadata.csv

   • You may find our training data at: EndoCoT dataset

   Since the metadata uses relative paths, please ensure the dataset files are placed in the same directory as metadata.csv

2. Train your model

   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
   

How to change the latent reasoning steps?

Note on Customization: Since the current implementation is straightforward, you can only manually adjust the latent reasoning steps in DiffSynth-Studio/diffsynth/pipelines/qwen_image.py:

• Line 442: Modify infer_steps.
• Line 471: Modify training_steps.

We plan to optimize this in future releases.

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 =  "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.
"
            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

📖 Citation

@article{dai2026endocot,
  title={EndoCoT: Scaling Endogenous Chain-of-Thou