Inpaint Anything

Inpaint Anything turns object-level image editing into a single click by chaining Meta's Segment Anything Model with inpainting and generation backends. Clicking an object yields a mask from SAM; the Remove Anything mode then erases the object with LaMa inpainting, Fill Anything replaces it with content described in a text prompt via Stable Diffusion, and Replace Anything keeps the foreground while regenerating the background. The same recipe extends beyond single images to video, where objects can be removed across frames, and to 3D scenes. It handles arbitrary aspect ratios and images up to 2K resolution. Built by researchers at the University of Science and Technology of China and the Eastern Institute for Advanced Study, the project is Apache 2.0 licensed, and a newer development branch updates the stack with SAM 3 and SDXL along with robotics-oriented features. It serves as a practical demonstration of composing segmentation and generative models for interactive editing.