ProPainter

Removing an object from a video and having the background fill in convincingly is ProPainter's core trick, achieved by improving both halves of the classic inpainting pipeline: dual-domain propagation that combines image and feature warping to pull reliable pixels from neighboring frames, and a mask-guided sparse video Transformer that drops unnecessary tokens for efficiency. The ICCV 2023 work from NTU's S-Lab, the group behind CodeFormer, handles object removal, video completion, and video outpainting, shipping pretrained weights, inference and training code, and evaluation setups for DAVIS and YouTube-VOS. Memory-efficient inference controls, including fp16, sub-video processing, and adjustable neighbor stride, scale GPU requirements from a few gigabytes for low resolutions up to around 25 GB for 720p at fp16. Setup needs Python 3.8 or newer, PyTorch 1.7.1 or newer, and CUDA, while community WebUI and ComfyUI wrappers make it usable without touching the scripts. The S-Lab License 1.0 limits use to non-commercial purposes, with commercial terms available from the authors.