Contrastive CFG: Improving CFG in Diffusion Models by Contrasting Positive and Negative Concepts

TL;DR

This paper introduces Contrastive CFG, a novel method that enhances negative guidance in diffusion models using contrastive loss, effectively removing unwanted features while preserving sample quality.

Contribution

It presents a new contrastive guidance technique that improves negative CFG by aligning or repelling denoising directions, overcoming limitations of existing methods.

Findings

Effectively removes undesirable concepts from samples

Maintains high sample quality across diverse scenarios

Aligns guidance directions with traditional CFG for positive guidance

Abstract

As Classifier-Free Guidance (CFG) has proven effective in conditional diffusion model sampling for improved condition alignment, many applications use a negated CFG term to filter out unwanted features from samples. However, simply negating CFG guidance creates an inverted probability distribution, often distorting samples away from the marginal distribution. Inspired by recent advances in conditional diffusion models for inverse problems, here we present a novel method to enhance negative CFG guidance using contrastive loss. Specifically, our guidance term aligns or repels the denoising direction based on the given condition through contrastive loss, achieving a nearly identical guiding direction to traditional CFG for positive guidance while overcoming the limitations of existing negative guidance methods. Experimental results demonstrate that our approach effectively removes undesirable concepts while maintaining sample quality across diverse scenarios, from simple class conditions to complex and overlapping text prompts.