andreas128/RePaint

Leverages a resampling-based noise schedule that iteratively reintroduces noise during the reverse diffusion process to harmonize generated content with known image regions, enabling flexible inpainting across diverse mask types (thick, thin, super-resolution, expansion). Built on OpenAI's guided-diffusion models, it operates as a pure inference scheme without requiring model retraining, with configurable schedules to trade-off quality for speed through adjustable denoising steps and resampling frequency. Supports multiple datasets (ImageNet, CelebA-HQ, Places2) via YAML-based configuration and extends to custom images through template configs.

2,247 stars. No commits in the last 6 months.