Generalization in diffusion models arises from geometry-adaptive harmonic representations

Abstract

Recent diffusion models generate spectacular images, but are these truly new images or are these blends of the training set? We show that diffusion models can enter a generalization regime where the generated images are independent of the samples in the training set: networks trained on non-overlapping subsets of a dataset generate identical images when starting from the same noise sample. We also show that this generalization relies on inductive biases towards geometric regularity in images.

Florentin Guth

Postdoctoral Researcher in Deep Learning Theory

I’m a Faculty Fellow in the Center for Data Science at NYU and a Research Fellow in the Center for Computational Neuroscience at the Flatiron Institute. I’m interested in improving our scientific understanding of deep learning, e.g., understand how it escapes the curse of dimensionality, particularly in the context of image generative models.