ProPainter

ProPainter is an AI video inpainting tool for people who need to remove unwanted objects or fill missing areas in footage. It processes videos with masks that identify the regions to replace.

• Object removal: Remove selected objects from a video, with an interactive demo for choosing what to erase.
• Video completion: Fill masked regions in MP4 videos or sequences of individual frames.
• Memory controls: Reduce GPU memory use through half-precision processing, lower processing resolution, and shorter video segments.
• Training and evaluation: Train the video inpainting and flow completion models, then evaluate results using YouTube-VOS and DAVIS datasets.

The Python tool runs locally with PyTorch and CUDA, and the interactive demo can also run on your own machine. Hugging Face and OpenXLab provide hosted demos. Local processing uses downloaded pretrained models and saves results locally.

GPU memory needs vary substantially with resolution and segment length. The supplied estimates put half-precision processing of 720p footage at roughly 19 to 25 GB of GPU memory; smaller resolutions use less. Outside the interactive object selection demo, processing requires masks for the input frames.

The code and models use the NTU S-Lab License 1.0, which permits non-commercial use, reproduction, and distribution. Commercial use requires permission from the licensor.