Free-form inpainting is the task of filling the image in areas defined by an arbitrary binary mask. In RePaint, this is solved purely via image condition (no text). Unfortunately, when trained, most existing approaches train on certain mask distributions, which limits their generalization capabilities to real mask types. RePaint uses a Denoising Diffusion Probabilistic Model (DDPM) under the hood, and as expected, such diffusion in the Inpainting task bypasses autoregression and gan's.
English 
Русский