EP-CFG: Energy-Preserving Classifier-Free Guidance

TL;DR

EP-CFG (Energy-Preserving Classifier-Free Guidance), which addresses issues of over-contrast and over-saturation artifacts at higher guidance strengths by preserving the energy distribution of the conditional prediction during the guidance process, is presented.

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.

Figures (12)

• Figure 1: Comparison of generation results with different CFG strengths (increasing from top to bottom). Without EP-CFG (left column), higher CFG strengths lead to over-saturation and contrast artifacts, especially in the background. EP-CFG (right column) maintains natural image quality across all CFG strengths while preserving the benefits of stronger guidance.
• Figure 2: Comparison of generation results with different CFG strengths. Without EP-CFG (left column), higher CFG strengths cause severe over-contrast and light bloom effects in the forest scene. EP-CFG (right column) preserves natural lighting and details across all guidance scales while maintaining the motorcycle's structure and forest atmosphere.
• Figure 3: Comparison of generation results with different CFG strengths (increasing from top to bottom). Without EP-CFG (left column), higher CFG strengths lead to over-contrast and color saturation in the dragon's features and lantern decorations. EP-CFG (right column) maintains natural image quality and preserves intricate details of the dragon's ornate design across all CFG strengths while retaining the festive Lunar New Year ambiance.
• Figure 4: Comparison of generation results with different CFG strengths (increasing from top to bottom). Without EP-CFG (left column), higher CFG strengths lead to over-contrast and color saturation in the flower petals and suburban background. EP-CFG (right column) maintains natural image quality and preserves delicate details of the flower and foliage across all CFG strengths while retaining the serene windowsill setting.
• Figure 5: Comparison of generation results with different CFG strengths (increasing from top to bottom). Without EP-CFG (left column), higher CFG strengths lead to over-contrast and color distortion in the horse's form and Van Gogh's signature swirling sky. EP-CFG (right column) maintains natural image quality and preserves the characteristic brushstroke details across all CFG strengths while retaining the distinctive Starry Night artistic style.