Reproduce “Detecting and Mitigating Memorization in Diffusion Models through Anisotropy of the Log-Probability” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Rohan Asthana, Vasileios Belagiannis

Researchers from Friedrich-Alexander-Universität Erlangen-Nürnberg developed a denoising-free metric for detecting memorization in diffusion models, specifically addressing the anisotropic nature of the log-probability distribution in the low-noise regime. This method outperforms previous baselines in both speed and accuracy, and effectively mitigates memorization during inference while preserving image quality and text-image alignment.

Key insights

  • The log-probability distribution in diffusion models transitions from an isotropic to an anisotropic state as noise levels decrease, indicating varying curvature across directions during fine detail synthesis.
  • Memorized samples in the anisotropic low-noise regime exhibit a high angular alignment between the conditional guidance term and the unconditional score, where guidance primarily reinforces existing memorized directions.
  • Effective and efficient memorization detection requires considering both the overall curvature (isotropic regime) and the directional alignment (anisotropic regime) of the log-probability density.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.