Single-image super-resolution (or zoom) is a crucial problem in image restoration. The goal is to recover the high frequency information that has been lost through im- age downsampling and compression. Deep learning meth- ods are now producing very impressive solutions to this problem. Numerous super-resolution methods have been proposed in the computer vision community. Much of the early work on super-resolution is regression based and trained with an MSE loss. As such, they effectively esti- mate the posterior mean, yielding blurry images when the posterior is multi-modal. In this project, we will investigate a family of models called ”denoising diffusion probabilistic models” (DDPM) which are nowadays of great interest for image generation. The goal of this project is to get familiar with this type of method and to understand how they are applied to super- resolution in this paper (SR3).
