Two studies examine how diffusion models memorize and generalize

Researchers sharing work in Nature Communications say diffusion models behave like dense associative memory and identify three computational phases: memorization, generalization and spurious pattern generation. They say they extended their analysis to Stable Diffusion. Separately, researchers behind an information-theoretic paper say measuring how much a model learns from data can predict when memorization begins and how generation quality, measured by FID, scales with training.