toolgarden.xyz
中文

Video Watermark Remover

Remove fixed watermarks from videos in your browser: temporal reconstruction on moving backgrounds or MI-GAN AI inpainting, with audio kept.

Video watermark settings

Only process videos you own or are authorized to edit. Decoding, reconstruction and encoding all happen locally; nothing is uploaded and data is cleared when the page closes.

Method
12px

Source preview

Waiting for a video

After uploading, drag a box over the watermark in the preview; the box can be moved and resized.

Output video

The rendered clean video will appear here

About this tool

The video watermark remover targets watermarks that stay in a fixed position (corner marks, logos, platform badges). The tool first samples pixels inside the selection across time: when the background moves, a temporal median rebuilds the clean background directly — no flicker and no AI model needed; when the background is static, the MI-GAN model inpaints one representative frame and the repaired patch is applied to the whole clip. For large watermarks or marks over faces and building edges, the optional LaMa HQ mode gives better structural reconstruction at the cost of a bigger download and slower inference.

Everything runs locally in your browser; the video is never uploaded. Output is an H.264 MP4, and the original audio track is added back with a local ffmpeg pass. Browsers without WebCodecs encoding cannot process video.

Use this tool only on videos you own, are explicitly authorized to edit, or are legally permitted to modify. Removing an attribution mark does not grant distribution or commercial rights.

How to use it

  1. Confirm editing rights

    Upload only a video you may modify and retain any required attribution and license records.

  2. Select the watermark area

    Drag a box over the complete mark and its edge; a tight selection gives steadier results.

  3. Choose a method and render

    Auto decides for you: temporal reconstruction on moving backgrounds, AI inpainting on static ones.

Supported range and limits

How it works
Temporal median rebuild on moving backgrounds, single-frame MI-GAN inpainting by default, or opt-in LaMa HQ inpainting
Output
H.264 MP4, with the original audio track re-added as AAC when possible
Works well on
Fixed static watermarks: corner badges, logos, platform marks
Works poorly on
Moving watermarks, rolling credits, large translucent overlays on subjects
Privacy
Everything runs locally in the browser; the video is never uploaded
Rights reminder
Removing a watermark does not change who owns the video; confirm you have the right to use it

When you would use it

  • Cleaning an owned asset

    Remove an outdated logo, internal mark, or test watermark from your own exports.

  • Repairing an authorized clip

    With permission from the rights holder, clear a specified mark over a simple background before publishing.

  • Unifying branding

    Material from different sources gets its old badges removed before a shared watermark is applied.

What to know before you start

  • Only fixed-position watermarks are supported; watermarks that move or change shape cannot be covered by a single selection.
  • Marks over faces, text, or continuously changing texture are rebuilt by estimation and may not match the real content.
  • Export re-encodes to H.264, adding slight second-generation compression; use original project files when maximum fidelity matters.

Related concepts

Temporal median
A per-pixel median across sampled times; on moving backgrounds it recovers what a static watermark hid.
Video inpainting
Reconstructing missing or occluded regions from neighbouring frames; ProPainter and similar models do this.

Frequently asked questions

视频去水印会上传文件吗？
No. The video is decoded, reconstructed and encoded locally, and the output is generated locally too; nothing is uploaded. AI mode downloads the about-28MB MI-GAN model on first use, which the browser then usually caches.
什么水印去不掉？
Watermarks that move or change shape cannot be covered by one selection, and rolling credits are better handled by cropping. Marks over faces or continuously changing texture are rebuilt by estimation and may not match the real content.
导出是什么格式？画质有损失吗？
Output is an H.264 MP4 with the original audio re-added when possible. The picture is re-encoded, so there is slight second-generation compression depending on the source bitrate and content complexity. Default AI inpainting uses lightweight MI-GAN (about 28MB); the LaMa HQ mode (about 198MB) suits large watermarks but is much slower.
Why does Auto sometimes skip the AI model?
The tool samples how pixels inside the selection change over time. When the background moves, the same pixel shows what is behind the watermark at different moments, so a temporal median restores the clean background directly — faster and steadier than per-frame AI. AI is only needed when the background is static and the median is still the watermark itself.
How long does processing take?
Temporal mode takes roughly the playback duration of the clip; AI mode adds one model load and a single-frame inference. Longer and higher-resolution videos take longer to encode, so validate on a clip under one minute first.
Does the output keep sound?
Normally yes: after the picture is re-encoded, a local ffmpeg pass adds the original audio track back into the MP4. If the container or audio codec cannot be muxed, a silent video is exported and the result details show whether audio was kept.
Can I remove someone else's watermark?
Technically yes, but removing a watermark does not change who owns the video. Confirm you have the rights before using someone else's work; what a tool can do is not the same as what you are permitted to do.

Related tools