Skip to content
← Models
Open weightsMMLab, The Chinese University of Hong Kong (Shangchen Zhou et al.)/10 trustVerified

ProPainter

Video inpainting via improved flow propagation and sparse transformers — object removal, watermark cleanup.

Visit official page →

About

ProPainter inpaints masked regions in video by first completing optical flow (recurrent flow completion), then propagating pixels and features along those flows in dual domains, and finally refining with a mask-guided sparse video transformer that only attends where masks actually are. This design reports 34.47 dB PSNR on the DAVIS benchmark — well ahead of its predecessor E2FGVI at 33.01 dB — while staying compute-competitive.

With ~7,000 GitHub stars it is the default open baseline for video object removal, and it has been ported to Apple Silicon (MLX) and wrapped by several community tools. Inference is self-hosted PyTorch; there is no official API. It is superb for background restoration and object removal but is not a semantic editor — it cannot invent new content a diffusion inpainting model could.

Specs

Context window

Not applicable — processes sliding windows of ~20 frames at inference

Pricing

Free / open weights

Free quota

Unlimited (self-hosted)

API access

No official API; self-host from GitHub weights

Weights

Open weights

Released

Not disclosed

Benchmarks

  • Provider-reported (ICCV 2023 paper, arXiv 2309.03897): 34.47 dB PSNR on DAVIS (stationary masks) vs 33.01 dB for E2FGVI, with better SSIM/VFID/flow-warping error
  • Third-party community benchmarks on DAVIS50 (jh5117-debug) report ~33.02 dB PSNR / 0.9627 SSIM with released weights

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

video-inpaintingobject-removalvideo-editingwatermark-removal

Trust score — 8/10

/10 trustVerified

Scored 8/10 (TIER verified, metrics-v1). Evidence: ICCV 2023 paper (arXiv 2309.03897); ~7,000 GitHub stars; independent community benchmark reproductions on DAVIS50; MLX port exists. Strengths: huge adoption, reproducible evals, de-facto open standard for video inpainting. Docked for: (1) custom research-oriented license (no standard SPDX); (2) not a generative editor — limited to reconstructive inpainting. License: custom research-oriented license.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
9.0/10
Maintenance · recency, releases, issue load
8.0/10
Openness · license + docs
7.0/10
Manual review · sanity read + red-flag check
8.0/10
  • · ICCV 2023 paper (arXiv 2309.03897); ~7,000 GitHub stars (verified 2026-10-16)
  • · Independent community DAVIS50 benchmark reproductions (~33.02 dB PSNR)
  • · Community ports exist (MLX, wrappers); de-facto open video-inpainting baseline

Sources