Diffusion Models for Image Restoration

This field studies how to recover useful images from noise, blur, low resolution, missing regions, and other degradations, including in medical and scientific imaging. Methods range from conditional diffusion and diffusion bridges to plug-and-play posterior sampling, which combines a learned image prior with consistency to the observed measurements. Recent work adapts pretrained models and uses physical, structural, or semantic guidance, while faster latent-space and one-step methods aim to make restoration practical. A central challenge is balancing fidelity to the input against realistic detail generation: strong generative priors can improve perceptual quality but may invent content, and inference cost and clinical validation remain concerns.