Generalization in diffusion models arises from geometry-adaptive harmonic representations | OpenReview

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 function, and thus the same density, when the number of training images is large enough. In this regime of strong generalization, diffusion-generated images are distinct from the training set, and are of high visual quality, suggesting that the inductive biases of the DNNs are well-aligned with the data density. We analyze the learned denoising functions and show that the inductive biases give rise to a shrinkage operation in a basis adapted to the underlying image. Examination of these bases reveals o

Verifying your browser | OpenReview Verifying your browser Complete the check below to continue to OpenReview Please complete the verification above. Have an OpenReview account? Sign in to skip this check.

related reading

• What are Diffusion Models? | Lil'Loglilianweng.github.io
• ⭐️ Diffusion Modelsandrewkchan.dev
• Yang Songyang-song.net
• Diffusion models from scratchchenyang.co
• Bare-bones Diffusion Modelsmadebyoll.in
• Diffusion model - Wikipediaen.wikipedia.org
• [2006.11239] Denoising Diffusion Probabilistic Modelsarxiv.org
• Diffusion is spectral autoregression – Sander Dielemansander.ai
• 2409.02908arxiv.org
• Large Language Diffusion Modelsarxiv.org
• Generative modelling in latent space – Sander Dielemansander.ai
• [2010.02502] Denoising Diffusion Implicit Modelsarxiv.org