<ul><li>Diffusion probabilistic models are essential in modern generative AI, but their generalization mechanisms are not well understood.</li><li>In highly overparameterized diffusion models, generalization in natural data domains is achieved during training before memorization occurs.</li><li>Results show that the time taken for memorization is proportional to the dataset size, highlighting a competition between generalization and memorization time scales.</li><li>A principled early-stopping criterion scaling with dataset size can optimize generalization and prevent memorization, with implications for hyperparameter transfer and privacy-sensitive applications.</li></ul>

Bigger Isn't Always Memorizing: Early Stopping Overparameterized Diffusion Models

Discover more