Core Concepts
DNNs trained on non-overlapping subsets converge to the same denoising function, indicating strong generalization.
Abstract
Deep neural networks (DNNs) trained for image denoising can generate high-quality samples using reverse diffusion algorithms. Recent concerns about memorization of training data by these networks have been raised. Two DNNs trained on non-overlapping subsets of a dataset learn nearly the same score function and density with a large number of training images. The inductive biases of DNNs align well with the data density, leading to distinct diffusion-generated images of high visual quality. Denoisers are biased towards geometry-adaptive harmonic bases, even for low-dimensional manifolds. The performance of networks is near-optimal when trained on regular image classes with optimal bases.
Stats
Roughly 105 images suffice for training sets to achieve strong generalization.
Denoisers operate without noise level input.
UNet architecture has 7.6m parameters.
BF-CNN architecture has 700k parameters.
Quotes
"We show that two denoisers trained on sufficiently large non-overlapping sets converge to essentially the same denoising function." - Content
"These results provide stronger and more direct evidence of generalization than standard comparisons of average performance on train and test sets." - Content
"The inductive biases of DNN denoisers encourage such bases." - Content