Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

GradCheck: Analyzing classifier guidance gradients for conditional diffusion sampling

Philipp Vaeth, Alexander M. Fruehwald, Benjamin Paassen, Magda Gregorova

TL;DR

A gradient analysis comparing robust and non-robust classifiers, as well as multiple gradient stabilization techniques, demonstrates that these techniques significantly improve the quality of class-conditional samples for non-robust classifiers by providing more stable and informative classifier guidance gradients.

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.

GradCheck: Analyzing classifier guidance gradients for conditional diffusion sampling

TL;DR

A gradient analysis comparing robust and non-robust classifiers, as well as multiple gradient stabilization techniques, demonstrates that these techniques significantly improve the quality of class-conditional samples for non-robust classifiers by providing more stable and informative classifier guidance gradients.

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.
Paper Structure (18 sections, 5 equations, 8 figures, 1 table)

This paper contains 18 sections, 5 equations, 8 figures, 1 table.

Table of Contents

  1. Introduction
  2. Classifier guidance
  3. $\hat{{\bm{x}}}_0^{({\bm{x}}_t)}\text{-prediction}$
  4. Adaptive moment estimation (ADAM)
  5. Experimental setup
  6. Data set
  7. DDPM
  8. Classifiers
  9. Conditional sampling
  10. Gradient analysis
  11. Metrics
  12. Experimental results & discussion
  13. Conclusion
  14. Data set
  15. DDPM
  16. ...and 3 more sections

Figures (8)

  • Figure 1: Classifier guidance cosine similarity between time steps for the robust classifier (red) and the non-robust classifier (blue). The values are presented as the mean over a batch of 64 generations, with the lighter colors indicating the standard deviation.
  • Figure 2: Classifier guidance without modifications (blue), with only $\hat{{\bm{x}}}_0^{({\bm{x}}_t)}\text{-prediction}$ (orange), with only ADAM stabilization (purple), and with both modifications (green) cosine similarity between time steps for the non-robust classifier. The values are presented as the mean over a batch of 64 generations, with the lighter colors indicating the standard deviation.
  • Figure 3: Robust classifier guidance cosine similarity between time steps with ADAM gradient stabilization and $\hat{{\bm{x}}}_0^{({\bm{x}}_t)}\text{-prediction}$ (purple) and without stabilization (red). The values are presented as the mean and standard deviation over a batch of 64 generations.
  • Figure 4: SportBalls images
  • Figure 5: Trade-off between the probability w.r.t. the target class image distribution (FID) and the classifier guidance scale ($s$)
  • ...and 3 more figures