Early Signatures of Memorization in Diffusion Models via Basin Geometry and Cyclic Denoising
Memorization in diffusion models is encoded in their learned energy landscape before generation occurs. The research shows that training creates localized basins around specific samples, which can be detected by analyzing score divergence and basin volume. Cyclic denoising further exposes latent memorization, revealing stable states that match no training images, allowing for auditing model fidelity.