Guidance in the Frequency Domain Enables High-Fidelity Sampling at Low CFG Scales

Summary

This paper talks about frequency-decoupled guidance (FDG), a method that improves image generation in diffusion models by separately controlling low-frequency and high-frequency parts of the image during the creation process.

What's the problem?

The problem is that the standard way of guiding image generation treats all frequencies the same, which can cause images to lose detail or look less diverse depending on how strong the guidance is.

What's the solution?

The researchers split the guidance into low-frequency, which controls the overall structure and alignment with the prompt, and high-frequency, which enhances fine details. By adjusting these separately, FDG improves image quality and diversity, especially at lower guidance strengths.

Why it matters?

This matters because it helps create clearer and more varied AI-generated images, making diffusion models more useful for art, design, and other creative applications.