ComfyUI InvSR

This project is an unofficial ComfyUI implementation of InvSR (Arbitrary-steps Image Super-resolution via Diffusion Inversion)

Installation

Navigate to the ComfyUI /custom_nodes directory

git clone https://github.com/yuvraj108c/ComfyUI_InvSR
cd ComfyUI_InvSR

# requires diffusers>=0.28
pip install -r requirements.txt

Usage

• Load example workflow
• Diffusers model (stabilityai/sd-turbo) will download automatically to ComfyUI/models/diffusers
• InvSR model (noise_predictor_sd_turbo_v5.pth) will download automatically to ComfyUI/models/invsr
• To deal with large images, e.g, 1k---->4k, set chopping_size 256
• If your GPU memory is limited, please set chopping_batch_size to 1

Parameters

• num_steps: number of inference steps
• cfg: classifier-free guidance scale
• batch_size: Controls how many complete images are processed simultaneously
• chopping_batch_size: Controls how many patches from the same image are processed simultaneously
• chopping_size: Controls the size of patches when splitting large images
• color_fix: Method to fix color shift in processed images

Updates

03 February 2025

• Add cfg parameter
• Make image divisible by 16
• Use mm to set torch device

31 January 2025

• Merged yuvraj108c#5 by wfjsw
  • Compatibility with diffusers>=0.28
  • Massive code refactoring & cleanup

Citation

@article{yue2024InvSR,
  title={Arbitrary-steps Image Super-resolution via Diffusion Inversion},
  author={Yue, Zongsheng and Kang, Liao and Loy, Chen Change},
  journal = {arXiv preprint arXiv:2412.09013},
  year={2024},
}

License

This project is licensed under NTU S-Lab License 1.0

Star History