Abstract
It is widely believed that noise conditioning is indispensable for denoisingdiffusion models to work successfully. This work challenges this belief.Motivated by research on blind image denoising, we investigate a variety ofdenoising-based generative models in the absence of noise conditioning. To oursurprise, most models exhibit graceful degradation, and in some cases, theyeven perform better without noise conditioning. We provide a theoreticalanalysis of the error caused by removing noise conditioning and demonstratethat our analysis aligns with empirical observations. We further introduce anoise-unconditional model that achieves a competitive FID of 2.23 on CIFAR-10,significantly narrowing the gap to leading noise-conditional models. We hopeour findings will inspire the community to revisit the foundations andformulations of denoising generative models.