Is Noise Conditioning Necessary for Denoising Generative Models?
This work addresses a foundational assumption in generative modeling for AI researchers, potentially simplifying model design, though it is incremental as it builds on existing denoising frameworks.
This paper challenges the necessity of noise conditioning in denoising diffusion models, showing that many models degrade gracefully or improve without it, and introduces a noise-unconditional model achieving a competitive FID of 2.23 on CIFAR-10.
It is widely believed that noise conditioning is indispensable for denoising diffusion models to work successfully. This work challenges this belief. Motivated by research on blind image denoising, we investigate a variety of denoising-based generative models in the absence of noise conditioning. To our surprise, most models exhibit graceful degradation, and in some cases, they even perform better without noise conditioning. We provide a theoretical analysis of the error caused by removing noise conditioning and demonstrate that our analysis aligns with empirical observations. We further introduce a noise-unconditional model that achieves a competitive FID of 2.23 on CIFAR-10, significantly narrowing the gap to leading noise-conditional models. We hope our findings will inspire the community to revisit the foundations and formulations of denoising generative models.