Frequency-Decoupled Guidance (FDG) for diffusion models

[Msadat97]

FDG is a new method for applying CFG in the frequency domain. It improves generation quality at low CFG scales while inherently avoiding the harmful effects of high CFG values. It could be a nice addition to the guiders part of diffusers. The implementation details for FDG are available on page 19 of the paper.

https://huggingface.co/papers/2506.19713

[yiyixuxu]

cc @a-r-r-o-w
Are we ready to work with community for guiders yet?

[a-r-r-o-w]

Nice work on the paper @Msadat97!

@yiyixuxu Yeah, I think we can open up new guiders for good-first-contributions :) The examples we have right now are simple enough to get started with and implement FDG (just need to copy CFG and do the computations as done in paper)

[yiyixuxu]

@a-r-r-o-w ok, I will mark this as community welcome, I think the contributor needs to be fairly familar with our code base, since we haven't written a good guide yet on guidere yet, so they have to be able to go around the code base to figure things out themselves

we will put out a pined issue for guiders:)

[yiyixuxu]

For these interested in contributing in this issue:

Guider is a new concept we are experimenting. We will put together more detailed guidance/info soon, but here are some pointers to get you started:

• List of current guider implementations here https://github.com/huggingface/diffusers/tree/main/src/diffusers/guiders
• How guider is used in denoising loop in Modular Diffusers

  diffusers/src/diffusers/modular_pipelines/stable_diffusion_xl/denoise.py

  Line 147 in e46e139

  class StableDiffusionXLLoopDenoiser(PipelineBlock):

• Doc on how to use guiders https://huggingface.co/docs/diffusers/main/en/modular_diffusers/modular_pipeline#customizing-guidance-techniques

[dg845]

Hi, I have opened a draft PR implementing FDG as a guider at #11976.