A new paper on arXiv takes on the open question of why diffusion models memorize their training data. The authors propose loss curvature as a sampling-free proxy for detecting memorization, potentially offering a way to measure the phenomenon without generating samples.
The abstract outlines two perspectives: a model-level view and, presumably, a data-level view, though the full details are not included in the available text. The work appears to be motivated by the need to understand the mechanism behind memorization rather than just observe its effects.
Because the abstract is truncated, the article does not yet report specific results or validate the proxy. The contribution so far is the proposal itself, along with the framing of the problem from these two angles. The full paper would be needed to assess how well the proxy works in practice.