Diffusion Models for Image Restoration and Super Resolution: Methods and Evaluation

Abstract

Image restoration asks for the clean picture behind a blurred, noisy, downsampled or partly missing one. That question has no single answer, and the honest output is a distribution rather than a point estimate. Diffusion models produce exactly that, which explains why they moved into restoration so quickly. This review covers thirty three works and separates the field into two routes. One trains a conditional model on paired data. The other reuses an unconditional prior and enforces agreement with the measurement at every sampling step, which needs no retraining but does need a forward model. We describe how each route handles blur, noise, downsampling and missing pixels, and we set out the sampling cost that both share. On evaluation we take a firm position: distortion metrics and perceptual metrics disagree by construction, so a paper reporting only one of them has not reported a result. We end with four open problems, of which blind degradation and hallucinated detail are the ones that matter most for scientific and medical use.