Cold Diffusion

Pavol Harar, 16. Nov 2022

Standard diffusion models are built around two main components. The image degradation (forward) operator that contaminates images with Gaussian noise and a restoration (backward) operator that is trained to perform denoising. In this presentation we are going to take a look on a preprint that explores generalized diffusion that is able to use completely deterministic degradations (e.g., blur, masking, and more) instead of just the Gaussian noise.