Diffusion Classifier Guidance for Non-robust Classifiers
This work advances conditional sampling in generative models by enabling a broader range of classifiers to be used, though it is incremental as it builds on existing guidance techniques.
The paper tackled the problem of classifier guidance in diffusion models being limited to robust classifiers trained on noise, by extending it to work with non-robust classifiers trained without noise. The result showed that their method improved stability while maintaining sample diversity and visual quality, as demonstrated on datasets like CelebA, SportBalls, and CelebA-HQ.
Classifier guidance is intended to steer a diffusion process such that a given classifier reliably recognizes the generated data point as a certain class. However, most classifier guidance approaches are restricted to robust classifiers, which were specifically trained on the noise of the diffusion forward process. We extend classifier guidance to work with general, non-robust, classifiers that were trained without noise. We analyze the sensitivity of both non-robust and robust classifiers to noise of the diffusion process on the standard CelebA data set, the specialized SportBalls data set and the high-dimensional real-world CelebA-HQ data set. Our findings reveal that non-robust classifiers exhibit significant accuracy degradation under noisy conditions, leading to unstable guidance gradients. To mitigate these issues, we propose a method that utilizes one-step denoised image predictions and implements stabilization techniques inspired by stochastic optimization methods, such as exponential moving averages. Experimental results demonstrate that our approach improves the stability of classifier guidance while maintaining sample diversity and visual quality. This work contributes to advancing conditional sampling techniques in generative models, enabling a broader range of classifiers to be used as guidance classifiers.