GradCheck: Analyzing classifier guidance gradients for conditional diffusion sampling

TL;DR

This paper analyzes the stability of classifier guidance gradients in diffusion models and introduces techniques to improve sample quality by stabilizing these gradients, especially for non-robust classifiers.

Contribution

It provides a detailed gradient analysis comparing robust and non-robust classifiers and proposes stabilization methods to enhance class-conditional sampling quality.

Findings

Gradient stabilization techniques improve sample quality.

Non-robust classifiers benefit significantly from stabilization.

Stable gradients lead to more informative guidance during sampling.

Abstract

To sample from an unconditionally trained Denoising Diffusion Probabilistic Model (DDPM), classifier guidance adds conditional information during sampling, but the gradients from classifiers, especially those not trained on noisy images, are often unstable. This study conducts a gradient analysis comparing robust and non-robust classifiers, as well as multiple gradient stabilization techniques. Experimental results demonstrate that these techniques significantly improve the quality of class-conditional samples for non-robust classifiers by providing more stable and informative classifier guidance gradients. The findings highlight the importance of gradient stability in enhancing the performance of classifier guidance, especially on non-robust classifiers.