Visible overlay vs invisible signal
These are two different technical problems, and conflating them is the source of most confusion in this area.
A visible Gemini logo is a semi-transparent raster overlay composited on top of your image using ordinary alpha blending. It is a composition problem: you have the composite, you can measure the overlay, and you can solve backwards for what was underneath. That is what the image and video studios on this site do, and it is why they can work without guessing.
SynthID is not an overlay. It is an imperceptible signal embedded in the content itself. Google’s own description states that SynthID “embeds digital watermarks directly into AI-generated images, audio, text or video” and that the watermark “doesn’t change the image or video quality”. There is no layer to subtract, no bounding box to crop, and no region whose absence would remove the signal.
What Google says it resists
The official SynthID documentation states that for image and video, the watermark is added at the moment of creation and is “designed to stand up to modifications like cropping, adding filters, changing frame rates, or lossy compression”.
Read that list as the specification a removal tool would have to defeat. It is a deliberately demanding target, because those are exactly the operations a normal user performs on an image: crop it to post it, filter it to look right, convert it to fit a platform, re-encode it to save space. A signal that survived none of those would not survive contact with real use.
For audio, Google states the watermark “can’t be altered by common modifications like adding noise, MP3 compression, or changing the speed of the track”.
How the text watermark works
Text is the most interesting case, because there is no visible mark at all. Google explains that large language models generate one token at a time, and each token carries a probability score reflecting how likely it was to come next. SynthID adjusts those probability scores in a way that encodes the watermark. The signal is therefore carried in the statistical texture of the text.
The practical implication is that the text humanizer on this site does not and cannot affect SynthID. It swaps stock filler phrases drawn from its own pattern list and strips invisible padding characters. That is a lexical pass over surface strings. SynthID lives in token probabilities chosen upstream, and a lexical rewrite does not reach it.
How to check whether a file carries it
If you need to know, use Google’s own detection paths rather than guessing:
- Ask Gemini. Google states you can upload an image, video, or audio clip to Gemini and ask whether it was created or altered by Google AI, and it will check for a SynthID watermark. See Google’s documentation on verifying AI-generated content.
- SynthID Detector. Google operates a dedicated verification portal for checking whether a piece of content is watermarked.
Note the limit of what a detector can tell you. Absence of a detected SynthID watermark is not proof that content is human-made, because absence of signal is not the same as evidence of absence.
Why we do not attempt it
This is a boundary, not a limitation we failed to engineer around. Four reasons, in the order they matter:
- It would be a contractual breach. Google’s Generative AI Prohibited Use Policy prohibits circumventing abuse protections and safety filters, and the Terms of Service prohibit misleading people into thinking generative content was created by a human. A tool built to defeat a provenance signal is aimed squarely at both.
- The statutory framework points the same way. The U.S. Copyright Office describes 17 U.S.C. § 1202 as making it unlawful to remove or alter copyright management information in certain circumstances. Content Credentials are metadata by construction, which makes them a stronger fit for that description than a decorative overlay is. See is watermark removal legal for the sourced analysis.
- It breaks the reason this tool exists. The entire claim of this site is that it recovers original pixels by solving a known equation rather than inventing replacement texture. That claim depends on the overlay being a genuine composition. A tool that attacked provenance signals would be doing something categorically different, and the honest name for it would not be “lossless”.
- Defeating it is not the same as removing it. Even a successful attack degrades a signal designed to survive modification. A watermark that sometimes fails is worse than one that reliably works, because it produces false confidence rather than a correct answer.
What this site does instead
The studios on this site target visible transparent overlays only. Specifically:
- Does touch: a composited visible logo or corner badge applied over the image, where the underlying pixels are still intact beneath the blend.
- Does not touch: SynthID in any modality, C2PA Content Credentials, EXIF or IPTC metadata, or any other invisible or cryptographic signal.
- Does not verify: anything about your rights to the file. Processing is local, so there is nothing to check against.
If you need a provenance signal removed from your own work, the responsible path is to go to whoever embedded it. For Google-generated content that means Google Gemini support.