Step 0: understand what is actually in the file
Before anything else, it helps to know that a generated video can carry more than one kind of marker, and they behave completely differently. Google documents two: SynthID, an invisible watermark embedded directly in the content, and Content Credentials, a signed provenance record based on the C2PA standard.
A generated video may also carry a visible overlay — a corner logo or badge. That is the only one of the three this method addresses, and the distinction is not academic:
- Visible overlay — a composition problem. Solvable arithmetically if the original file survives.
- SynthID — a signal-processing problem. Google states it is designed to stand up to cropping, filters, frame-rate changes and lossy compression. Not addressed here.
- Content Credentials — a signed metadata chain from the C2PA project. Not addressed here, and tampering with it invalidates the credentials.
If you need to know what is in a file, ask Google. Gemini can check for SynthID and report which parts of a video carry it, and will read Content Credentials if present. That is the right tool for the question, rather than a remover.
Step 1: get the original export
This is the step that decides whether the rest works. A visible overlay is applied with ordinary alpha blending, which means the original pixels are still present underneath it, just mixed. Recover them means undoing the mix, which requires knowing the overlay’s exact geometry and opacity profile. That information is only fully intact in the file as exported.
Each of these destroys some of it:
- A screenshot — resamples the frame to screen resolution and resharpens it. The overlay edge and the underlying texture are now interleaved at pixel level.
- A re-save from a player or editor — recompresses with a different codec and quality target. Compression noise concentrates exactly where the overlay boundary is, because that is the highest-contrast region in the frame.
- A file that has been cropped or scaled — moves the overlay off its calibrated anchor, so the geometry no longer matches what the method expects.
- Social re-encoding — platforms transcode aggressively. If the clip has been through one, assume the overlay is visible but no longer recoverable.
Step 2: check whether it is worth doing
A visible mark is not proof that the pixels underneath survive. If detection comes back with low confidence, that usually means one of the above already happened. Running the pipeline on a file that has been transcoded will produce a plausible-looking result that is subtly wrong, which is worse than an honest failure.
Step 3: run the cleanup
- Load the original file into Video Studio.
- Confirm the detected region matches the mark you can see.
- Run the frame-by-frame cleanup.
- Inspect several frames before exporting — the first, the last, and a few in between. Watermarks sometimes change position or opacity across a clip, and a method calibrated to one placement will not handle another.
- Export and confirm the audio still plays. Some export paths carry audio through untouched; some do not.
Everything runs locally in the browser, so the file is never transmitted. There is no upload step, which also means there is no server-side queue and no account.
Step 4: what “lossless” actually claims
It claims one specific thing: outside the detected watermark region, pixels are passed through unmodified. Inside the region, the blend is inverted arithmetically to recover the underlying values. Nothing is synthesised, and no texture is invented.
This is the whole difference from a generative approach. A diffusion-based inpainter asks “what texture plausibly belongs here?” and answers with something that looks right and is fabricated — smeared eyelashes, melted type, softened detail. Inverting a known equation does not have that failure mode, because there is nothing being guessed. The failure modes that remain are different: wrong geometry estimate, and codec damage the original already carried.
Common questions
Does turning off the watermark setting clean old videos?
No. That setting governs whether newly generated output is marked. It does not modify files that have already been exported.
Why is my result slightly soft around the mark?
Almost always an input problem rather than a method problem. A file that has already been through one re-encode has compression artefacts concentrated at the overlay boundary, because that boundary is the sharpest edge in the frame. Recovering the blend cannot recover detail that the earlier encode destroyed.
Can you remove SynthID from a video?
No, and we do not attempt it. Google states the watermark usually persists through re-scaling, re-colouring and compression, and notes there is still a chance it will not be detected after many alterations — which is a statement about degradation, not an invitation. The full reasoning is on the SynthID page.
What are the limits?
Practical limits come from your device rather than a server: available memory, and whether your browser can decode the codec. Long clips and 4K sources scale roughly with resolution times duration.
Before you use any of this
This page describes a technical method, not an endorsement of every way it might be used. Removing a visible overlay from media you do not own is very likely unlawful, and removing Google’s own branding may breach Google’s Terms of Service even when you hold the rights to the underlying content. The sourced legal analysis covers the statutes and the specific policy clauses, and is worth reading before relying on this for anything commercial.