EP-CFG: Energy-Preserving Classifier-Free Guidance

TL;DR

EP-CFG introduces an energy-preserving approach to classifier-free guidance in diffusion models, effectively reducing artifacts and maintaining image quality across guidance strengths with minimal additional computation.

Contribution

The paper proposes a simple rescaling method that preserves energy distribution during guidance, improving image quality and artifact suppression in diffusion models.

Findings

EP-CFG reduces over-contrast and over-saturation artifacts.

It maintains natural image details across guidance strengths.

The method adds minimal computational overhead.

Abstract

Classifier-free guidance (CFG) is widely used in diffusion models but often introduces over-contrast and over-saturation artifacts at higher guidance strengths. We present EP-CFG (Energy-Preserving Classifier-Free Guidance), which addresses these issues by preserving the energy distribution of the conditional prediction during the guidance process. Our method simply rescales the energy of the guided output to match that of the conditional prediction at each denoising step, with an optional robust variant for improved artifact suppression. Through experiments, we show that EP-CFG maintains natural image quality and preserves details across guidance strengths while retaining CFG's semantic alignment benefits, all with minimal computational overhead.