Generalization in diffusion models arises from geometry-adaptive harmonic representations

arXiv:2310.02557v2 Announce Type: replace-cross
Abstract: Deep neural networks (DNNs) trained for image denoising are able to generate high-quality samples with score-based reverse diffusion algorithms. These impressive capabilities seem to imply an escape from the curse of dimensionality, but recent reports of memorization of the training set raise the question of whether these networks are learning the "true" continuous density of the data. Here, we show that two DNNs trained on non-overlapping subsets of a dataset learn nearly the same score …

abstract algorithms arxiv capabilities cs.cv cs.lg denoising diffusion diffusion models dimensionality generate geometry image imply networks neural networks quality question raise reports samples set the curse of dimensionality training type