Process video locally in your browser: remove fixed-position watermarks with automatic temporal reconstruction or MI-GAN and LaMa AI inpainting, an optional HQ mode for large marks, exported as H.264 MP4 with the original audio. Files never leave your device.
Videos are decoded, rebuilt and encoded locally in your browser and never uploaded; AI inpainting models (about 28MB / 198MB) are downloaded on demand.
These tools decode and re-encode video with WebCodecs in your browser and re-add audio with a WebAssembly build of ffmpeg, so watermark removal never uploads your file. AI inpainting models load on demand: lightweight MI-GAN for small marks, and the LaMa HQ mode for large watermarks or marks over faces and building edges.
The compute comes from your device: processing time scales with video length, resolution and hardware, so test a clip under one minute before running a full video.
Trade-offs between the three reconstruction modes: speed, model size and best-fit scenarios.
| Mode | Model size | Speed | Best for |
|---|---|---|---|
| Temporal rebuild | No model | Fastest (about the clip duration) | Moving backgrounds or camera movement |
| MI-GAN inpainting (default) | About 28MB | About 0.4s per frame with WebGPU | Small marks on static backgrounds |
| LaMa HQ inpainting | About 198MB | About 1/7 of MI-GAN speed | Large marks over faces or structures |
Only fixed-position watermarks are supported; marks that move or change shape cannot be covered by a single selection, and rolling credits are better handled by cropping.
Marks over faces, text, or continuously changing texture are rebuilt by estimation and may not match the real content — inspect the output frame by frame.
Export re-encodes to H.264, adding slight second-generation compression; use original project files when maximum fidelity matters.