What it does

Video-subtitle-remover (VSR) detects hard-coded subtitles and text-like watermarks in videos or images, then fills the erased regions using AI inpainting. It outputs at the original resolution—no downscaling—and can optionally extract the subtitle text it removed. There’s a GUI, a CLI, and pre-built packages for Windows plus Docker images for NVIDIA, AMD/Intel, and CPU-only runs.

The interesting bit

The project doesn’t just slap a blur over text. It offers three distinct inpainting algorithms—STTN for live-action speed, LaMA for stills and animation, and ProPainter for heavy motion—each with tunable parameters for quality-versus-VRAM tradeoffs. You can also skip detection entirely and brute-force STTN across the whole frame if you’re in a hurry and willing to risk collateral damage.

Key highlights

• Supports targeted removal via coordinate boxes, or full-frame auto-detection of all text
• Cross-platform: Windows, Linux, macOS (Apple Silicon), with CUDA 11.8/12.6/12.8, DirectML, and CPU backends
• Ships with pre-built 7z archives and Docker tags matched to GPU generations (10–30 series, 40 series, 50 series)
• Can extract subtitle text for reuse with the author’s companion project, video-subtitle-extractor
• Retrainable: includes training code and instructions in the design folder for rolling your own models

Caveats

• macOS Apple Silicon users get a warning that the PP-OCRv4-Server detection model “seems not ideal”; other models are recommended
• ProPainter mode is explicitly noted as slow and VRAM-hungry
• The README’s install guide is thorough to the point of exhaustion—expect to wrangle CUDA, cuDNN, PaddlePaddle, and Torch versions manually if not using Docker or pre-builts

Verdict

Worth a look for video archivists, fansubbers working in reverse, or anyone cleaning up screen recordings. Skip it if you need real-time processing or a one-click SaaS experience.