Image Inpainting

Image inpainting studies how to fill missing or damaged regions while keeping the result coherent with the visible parts of an image. Methods range from convolutional and transformer networks and GANs to diffusion and flow models, which may use masks, text, sketches, reference images, or medical and scientific priors to guide completion. Researchers evaluate both visual similarity and perceptual or task-specific outcomes on natural-image benchmarks and specialized settings such as medical scans, satellite imagery, and 3D scenes. A central challenge is balancing plausible, detailed synthesis against faithful preservation of observed content, with consistency across large gaps, video frames, or multiple views and efficient, reliable performance beyond training data still difficult.