Understand the watermark
What SynthID actually is, and how removal works
What is SynthID and why it doesn't wash out
SynthID is Google DeepMind's invisible watermarking system. Instead of drawing a logo in the corner, it nudges image pixels by amounts too small to see, encoding a signature across the frequency domain of the picture. Every image from Gemini, Nano Banana, Veo and other participating generators carries one — and unlike a visible watermark, it is deliberately built to survive crops, resizes, filters, color grading and re-saving. That robustness is the point: Google wants the mark to persist through ordinary editing so AI-generated images stay identifiable.
Because the signal lives in the pixels rather than beside them, the usual tricks do nothing. Re-saving a JPEG, stripping EXIF data, or taking a screenshot all leave the frequency-domain signature intact. Removing it requires actually transforming the pixel data itself — which is a different job from metadata cleanup, and why "metadata eraser" tools do not affect it at all.
How SynthID detection works
Detection runs a classifier over the image's frequency characteristics and reports a probability that a SynthID watermark is present. Google keeps the official detector closed — it is offered through limited access to verified partners, not as a public API. That is worth knowing when you evaluate any SynthID detector, including this one: a claim to "detect SynthID with 100% certainty" from a third-party site is not something Google enables technically.
Our scan takes the honest route. C2PA manifests and EXIF/XMP AI tags are deterministic — we parse the bytes and tell you exactly what is there. For SynthID itself we infer likelihood from what the file's metadata and origin say about which generator produced it, and we label that result as inferred. The removal pass, however, runs on every file regardless: rewriting the pixels is cheap insurance even when inference is uncertain.
Why metadata tools cannot remove it
An image can carry two unrelated layers of provenance. The metadata layer — C2PA manifests, EXIF blocks, XMP packets — is a set of side-cards attached to the file. To remove C2PA metadata you simply delete those cards; it is easy and deterministic, and plenty of free tools do it. Platforms that rely on metadata alone will stop labeling your file once the cards are gone.
The pixels are the hard half. SynthID's frequency-domain signal survives metadata deletion by design, so a file with pristine metadata can still be identified as AI-generated when a platform or lab inspects the pixels. A real remover has to do both jobs in one pass: strip every metadata card, then re-encode the pixel data through a transform that breaks the embedded signature without visibly degrading the image. That two-layer pass is exactly what runs in your browser on this site.
How to use a remover without losing quality
Quality loss in watermark removal comes from three places, and all three are avoidable. First, resolution: some pipelines shrink your image or re-encode at a lower quality setting — insist on tools that keep your original dimensions. Second, compression artifacts: every JPEG re-save loses a little; exporting to PNG for archival or using quality 90+ for JPEG keeps the cost invisible. Third, the removal transform itself: aggressive pixel rewriting can leave smearing or noise in flat areas, which is why the rewrite here is tuned to stay below the threshold of perception.
A practical workflow for any SynthID watermark remover: keep your originals, clean a copy, and compare the two side by side at 100% zoom before you publish. A proper cleanup changes nothing you can see — same colors, same sharpness, same crop. If a tool visibly alters your image to remove an invisible watermark, you are paying too much.
Limitations you should know about
Being straight about what this can and cannot do. SynthID is an evolving system, and Google publishes new versions tuned to survive new attacks; a rewrite that breaks today's signal may be less effective against a future encoder. Video is handled frame by frame in your browser with no length cap, at roughly the speed of the clip itself. And no tool, ours included, can prove a negative: we can clear every known signal and still not guarantee how a closed, unpublished detector run by a platform will score your file tomorrow.
Use this on files you generated or hold rights to. Clearing a watermark from your own work is asserting control over your output; using it to pass off someone else's AI work as your own, or to evade disclosure rules you are legally bound by, is on you. Our terms spell this out, and they exist so the tool stays available for the people it is built for.

