ProPainter
Video inpainting model for object removal, video completion, and video outpainting, released with code and pretrained weights.
At a Glance
Code and models are available for non-commercial use, reproduction and distribution under the NTU S-Lab License 1.0 (not an OSI-approved license). Commercial use requires permission from the authors.
Engagement
Available On
Alternatives
Listed Nov 2026
About ProPainter
ProPainter is a video inpainting model from Shangchen Zhou, Chongyi Li, Kelvin C.K. Chan and Chen Change Loy at S-Lab, Nanyang Technological University, presented at ICCV 2023. It takes a video and frame-wise masks and fills the masked regions, supporting object removal, video completion, and video outpainting. Code and pretrained models are published on GitHub under a non-commercial S-Lab license.
What It Is
ProPainter is a research model and inference codebase for video inpainting. Users supply an mp4 video (or split frames) plus frame-wise masks, run a Python inference script, and receive the inpainted result in a results folder. The project page lists three tasks: removing objects from a video, completing a masked video, and expanding the view of a video.
How the Method Works
The paper describes three components: a recurrent flow completion network that repairs corrupted optical flow, dual-domain propagation that combines image and feature warping to exploit global correspondences, and a mask-guided sparse video Transformer that discards redundant tokens to save memory. The authors report a 1.46 dB PSNR improvement over prior approaches while keeping efficiency.
Running It
Setup uses a Conda environment with Python 3.8, CUDA, PyTorch and the packages in requirements.txt. Pretrained weights can be downloaded from the v0.1.0 release or are fetched automatically on first inference. Memory-efficient options include fp16 inference, resizing, smaller width and height, fewer local neighbors, a larger reference stride, and shorter sub-videos. The repository also includes training configs and evaluation scripts for YouTube-VOS and DAVIS, and an interactive Gradio demo for selecting objects to remove, available on Hugging Face or locally.
Current Status and Licensing
The README notes that its Hugging Face demo now supports SAM 2.1 for interactive video segmentation. The authors removed watermark-removal demos in 2023 to discourage misuse. The code and models are licensed for non-commercial use only; commercial use requires contacting the authors.
Community Discussions
Be the first to start a conversation about ProPainter
Share your experience with ProPainter, ask questions, or help others learn from your insights.
Pricing
Non-Commercial Use
Code and models are available for non-commercial use, reproduction and distribution under the NTU S-Lab License 1.0 (not an OSI-approved license). Commercial use requires permission from the authors.
- Non-commercial use only
- Licensed under NTU S-Lab License 1.0 (source-available, not OSI)
- Commercial use requires contacting the authors
- Free online demos on Hugging Face and OpenXLab
Capabilities
Key Features
- Object removal from videos
- Video completion of masked regions
- Video outpainting
- Dual-domain (image and feature) propagation
- Mask-guided sparse video Transformer
- Recurrent flow completion network
- Memory-efficient inference options including fp16 and sub-video chunking
- Interactive Gradio demo for object selection
- Training and evaluation scripts for YouTube-VOS and DAVIS
Integrations
Demo Video

