Diffusion models localized memorized image regions without contrast maps

JaRS: Contrast-Free Localization of Memorized Regions in Diffusion Models via Jacobian Row Sums

Computer Vision and Pattern RecognitionMachine Learning

Summary

Diffusion models sometimes remember and copy parts of images they were trained on, which can cause privacy or copyright issues. The authors developed a way to find these copied regions by analyzing how the model changes pixel values, using a new technique called Jacobian row sums (JaRS). Unlike previous methods, this doesn't need extra reference images or prompts, making it easier to apply across different types of models. Their improved method, JaRS+, works better for spotting memorized content in popular diffusion models.

What this means in practice

  • For ai system auditors: Detect and localize memorized training data regions inside diffusion-generated images to assess model privacy risks and compliance.
  • For image generation developers: Improve tools for monitoring and reducing memorization in diffusion models by integrating JaRS+ localization techniques during model evaluation.

Authors

Saisuchith Mahajan, Aditey Nandan, Rohit Singh, Nishanth Shetty, Chandra Sekhar Seelamantula

Abstract

Diffusion models sometimes memorize and reproduce regions of their training images, raising privacy and copyright concerns. We aim to localize these regions. Existing contrast-based methods use the diagonal of the denoiser Jacobian, which reflects the posterior variance of each pixel. However, a small diagonal entry is ambiguous: it signals either memorization or averaging over many neighboring pixels in smooth or predictable regions. To resolve this ambiguity, these methods subtract a reference map, which requires a second evaluation with a null prompt or an earlier checkpoint. Unconditional models have no null prompt, and earlier checkpoints are unavailable in general. We show that Jacobian row sums (JaRS), unlike the diagonal, separate memorization from local averaging without a reference map. Pairing JaRS with an image-level detection score that uses the same statistic at an early denoising step yields JaRS+. Across Stable Diffusion v1.4 and v2.1, Realistic Vision v5.1, and an unconditional diffusion Transformer (DiT), JaRS+ outperforms contrast-based localization methods.