Denoising Diffusion Models Without Diffusions

Explore an innovative approach to denoising diffusion models that eliminates the need for continuous-time diffusion processes in this conference talk by Francis Bach from INRIA Paris. Learn about a novel framework that reframes denoising-based sampling through an alternative perspective, focusing on the two-step process of sampling noisy data and denoising it using the score function and optimal denoiser mapping. Discover how this approach naturally extends to discrete settings including binary data, offering both fresh conceptual insights and practical applications. The presentation covers joint research work that challenges traditional diffusion model paradigms while maintaining the remarkable generative modeling capabilities that have advanced the field across various domains.