How Each Watermark Is Actually Removed
How the Visible Gemini Logo Comes Off — Reverse Alpha Blending
The sparkle is undone with the inverse of the formula that applied it, not blurred or painted over.
The sparkle isn't painted over or blurred — it's mathematically undone. Gemini stamps the logo with alpha blending: watermarked = α × logo + (1 − α) × original, where α is the logo's per-pixel transparency. Because that operation is deterministic, it inverts exactly: original = (watermarked − α × logo) / (1 − α). The engine loads the calibrated alpha map for your image size (48×48 px under 1024 px, 96×96 px above) and reconstructs the pixels that were underneath.
That's why every pixel outside the small corner region is byte-for-byte identical between the two images above — there is no smear, patch, or AI re-paint. You can run the same removal on the Gemini image watermark remover on the homepage.
Regeneration Fidelity: When Nothing Changing Is the Point
For the hidden layer, a convincing before-and-after is one where you cannot tell which image is which.
The examples above show a mark disappearing, which is the easy kind of proof. Clearing SynthID works the other way round: it targets the carrier inside the picture rather than anything you can see, and on stubborn images the pipeline regenerates every pixel outright — so the thing to judge is not whether something vanished but whether anything else moved.
That sets a demanding standard, and it is the right one. Hand your picture to a general-purpose image model and you get a lookalike: the same subject, a visibly different image — different hands, different texture, a face that is nearly right. Useless as a replacement for your file. A pass worth paying for holds the output close enough that placed side by side with the original, you cannot reliably say which is which. Faces, text, fine texture and edges are where the difference shows first, so those are what to check.
Judge it on your own work rather than on a picture we chose. Run an image through, open both at 100%, and look at the details you care about. If you can spot the regenerated one, it did not clear the bar.
The Invisible Layers: What Proof Actually Looks Like
A screenshot cannot show an invisible watermark, so the evidence is detector output — and you can run it yourself.
Gemini embeds two invisible markers alongside the logo: Google's SynthID signal in the pixels and a C2PA content-credentials manifest in the file header. Neither shows up in a screenshot. A "before and after" of an invisible watermark would be two identical images, so we do not stage one — a tool showing you a dramatic "SynthID removed" picture is showing you nothing.
The real evidence is what a detector says before and after, and you do not have to take our word for any of it. Check the credentials with the C2PA detector, confirm SynthID with the SynthID detector, then clear both with the SynthID remover and re-run the same two checks on the result. Same file, same detectors, two readings.
One limit we will not bury: clearing the provenance check platforms run at upload is not the same as being invisible to every forensic classifier, and we do not claim universal undetectability. What the detectors show you is exactly what we claim, and nothing beyond it.
Video: Rebuilding the Logo Frame by Frame
Re-encoded once at source resolution and framerate, with the audio copied across untouched.
Video can't be undone with a single alpha map — the logo is burned into thousands of frames. The video example above is rebuilt frame by frame and re-encoded once at the source resolution and framerate, with the audio copied across untouched. The hidden content credentials are dropped in the same pass.
The full tool is the Gemini & Veo video watermark remover. One honest limit shown here: the in-pixel SynthID signal on video is not removed yet — only the visible logo and the hidden credentials are, and credentials are what platform upload checks actually read.
Why Any of This Matters for Your Reach
These markers are what platforms read at upload to decide whether your post gets an AI label.
The visible sparkle is a presentation problem — it makes a client deliverable look unfinished. The invisible layers are the ones with consequences. Instagram, TikTok and X read C2PA and SynthID on upload and attach an AI label automatically, and on most platforms you cannot remove that label once the post is live.
That is why the fidelity standard above is not a technicality. A cleaned file is only useful if it is still the picture you made. If you are here because something you posted got labelled, start with how AI labels work — it explains which of the two layers caused it, and which fix you actually need.
How These Examples Were Made
Generated in Gemini, run through this site's tool, nothing else done to them.
Every image and clip on this page was generated with Google Gemini, then run through this site's tool with no other editing — no retouching, no recompositing, no cherry-picked crops to hide artifacts. The originals and the cleaned results are the same files, identical everywhere except the watermark. Where the tool can't do something cleanly (video SynthID), we say so rather than hide it.