Gemini Omni Video Watermark Remover

An MP4 is a container holding separate video and audio streams. Removing a watermark from the picture without disturbing the sound means taking that container apart, working on one stream, and rebuilding it — which is what Erasio does entirely inside your browser.

First the file is demuxed: the video and audio streams are separated, and the audio is set aside untouched. This is why audio quality is never a question with Erasio — the audio is never decoded, never re-encoded, and is copied into the output exactly as it arrived.

Next, every video frame is decoded through WebCodecs and patched individually. The sparkle occupies the same region on each frame, so each one gets the same inverse alpha transform applied to that region: the arithmetic that recovers the original pixels from the blended ones. A 10-second clip at 24fps means 240 of these operations, all local.

Finally the patched frames are re-encoded and re-muxed with the original audio into a new MP4. Re-encoding the picture is unavoidable — every frame genuinely changed — but the encode targets a bitrate derived from your source file rather than a fixed default, so a high-quality input produces a high-quality output. Resolution and frame rate pass through unchanged; there is no downscaling anywhere in the pipeline.

On a still image, watermark removal is one operation on one array of pixels, and the output is lossless PNG. Nothing is approximated and nothing is thrown away.

Video cannot work that way. Codecs compress across frames, so changing any frame requires re-encoding, and re-encoding is by definition lossy. The honest framing is not "no quality loss" in an absolute sense — it is that the loss is bounded by the encode settings, which is why matching the output bitrate to the source matters so much. A fixed low bitrate is exactly how naive tools turn a crisp clip into a soft one.

The second difficulty is consistency. A tool that detects the watermark per frame can drift, leaving the mark flickering in and out across the clip. Erasio locks the pattern once and applies it uniformly to every frame, so the result is stable rather than shimmering — which is the failure mode most visible to viewers even when a single frame looks fine.

The third is audio. Many tools re-encode the whole file, degrading sound that never needed touching. Separating the streams and copying the audio across is straightforward engineering, but it has to be done deliberately.

A cleaned clip has no visible sparkle, keeps its original resolution and frame rate, and carries the same audio it always had. For the common cases — an ad creative a platform would reject for a visible watermark, a client previz reel, a portfolio piece — that is the whole of what was needed.

What has not changed is provenance. SynthID, the invisible signal Google distributes across generated frames, survives re-encoding and is untouched by Erasio. Your clip still identifies as AI-generated to anything that reads for it, and platforms that check provenance rather than looking for a logo will still see what it is.

That boundary is intentional. Erasio exists so that video you generated and own is usable where you meant to publish it, not so that AI-generated footage can pass as something else.