Localizing Memorized Regions in Diffusion Models via Coordinate-Wise Curvature Differences

Researchers propose a geometric method to localize memorized regions in diffusion model outputs.

It compares coordinate-wise curvature against an underfitted baseline to identify where memorization appears in images.

On Stable Diffusion, the approach outperforms prior attention-based localization methods using ground-truth masks.

The work could improve auditing for privacy, copyright, and training-data overfitting in generative models.