RePaint uses denoising diffusion probabilistic models to reconstruct missing regions of images. The method iteratively denoises from pure noise while preserving known image regions via a resampling strategy. It provides pretrained models for ImageNet, CelebA-HQ, and Places2 datasets.