ComfyUI_GRAG_Image_Editing

GRAG-Image-Editing : Group-Relative Attention Guidance for Image Editing, on the native comfy DiT stack.

Update

• 改为comfyUI核心封装/change to comfyUI core Wrapper

Installation

---

In the ./ComfyUI/custom_nodes directory, run the following:

git clone https://github.com/smthemex/ComfyUI_GRAG_Image_Editing

Models

---

• comfyUI normal：dit, qwen-image vae and qwen_2.5_vl_7b
• lora, lightx2v 8step lora# 千问edit加速

├── ComfyUI/models/diffusion_models # or gguf
|     ├── qwen_image_edit_2511_fp8mixed.safetensors
├── ComfyUI/models/vae
|        ├─Qwen-Image.safetensors  
├── ComfyUI/models/clip
|        ├──qwen_2.5_vl_7b_fp8_scaled.safetensors
├── ComfyUI/models/loras 
|        ├──Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors

Example

GRAGPatch parameters

• cond_delta — the GRAG strength. **< 1 frees the output from the reference (stronger edit);

  1 locks onto the reference image.** Upstream recommended sweep: 0.8 ~ 1.7, step 0.01, default 1.10. Verified on 2511 (cfg 4.0, 24 steps): 1.70 → corr 0.995 with the input, 1.10 → 0.997, 0.90 → 0.928, 0.80 → 0.711.

• cond_b — bias (lambda) scale; upstream keeps it at 1.0 everywhere. Leave it alone.
• img_len — how many trailing image tokens form the reference group. comfy's TextEncodeQwenImageEdit resizes the reference to 1024×1024 → 4096 tokens, which is the default and matches upstream's hardcoded value. Use img_len = (ref_w/16) * (ref_h/16) summed over references when you deviate from that (e.g. TextEncodeQwenImageEditPlus with multiple images, or a custom non-resizing encoder). 0 = the whole image block.
• txt_delta / txt_b — same reweighting applied to the text group; upstream no-ops (1.0).
• apply_from / apply_to — sigma window (denoising progress 0→1). Defaults cover all steps, matching upstream.
• verbose — print the effective config when the graph runs.

Citation

@misc{zhang2025grouprelativeattentionguidance,
      title={Group Relative Attention Guidance for Image Editing}, 
      author={Xuanpu Zhang and Xuesong Niu and Ruidong Chen and Dan Song and Jianhao Zeng and Penghui Du and Haoxiang Cao and Kai Wu and An-an Liu},
      year={2025},
      eprint={2510.24657},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2510.24657}, 
}