2025-06-28 15:46:52 +09:00
2025-06-28 15:46:52 +09:00
2025-06-28 03:08:34 +09:00
2025-06-28 03:38:10 +09:00

ComfyUI-FDG

Implementation of Guidance in the Frequency Domain Enables High-Fidelity Sampling at Low CFG Scales for ComfyUI.

Thanks to the FDG researchers for making their code public.

FDG improves image quality at low guidance scales and avoids the drawbacks of high CFG scales by design.

Paper: https://arxiv.org/abs/2506.19713

Usage

To use FDG, just place the FDGNode (located in the advanced/model category) in front of the KSampler node in your workflow.

Basically disables cfg values ​​within KSampler.

Testing on SDXL only.

Inputs

  • guidance_scale_high: Guidance scale for high-frequency details.
  • guidance_scale_low: Guidance scale for low-frequency structures.
  • levels: Number of pyramid levels for frequency decomposition. For levels higher than 2, it operates using linear interpolation for each frequency.
  • fdg_steps: Number of initial steps to apply FDG before switching to CFG. After the limit, the cfg value of the KSampler node is used. If the cfg is set 1, the value of guidance_scale_high is applied.

Citation

If you use this implementation in your research, please cite the original paper:

@misc{sadat2025guidance, title={Guidance in the Frequency Domain Enables High-Fidelity Sampling at Low CFG Scales}, author={Seyedmorteza Sadat and Tobias Vontobel and Farnood Salehi and Romann M. Weber}, year={2025}, eprint={2506.19713}, archivePrefix={arXiv}, primaryClass={cs.LG} }

S
Description
No description provided
Readme
34 KiB
Languages
Python 100%