No SynthID watermark remover can do what its headline says. SynthID is written into the pixels while the image is being generated, so any tool that reaches the mark has to change the image, and a promise of removal with no quality loss is describing a header strip instead. The useful question is not which of them works — it is which of them says what it actually does. Four sites were read on 2026-10-16 and their own wording is quoted below; two of the four are honest about the boundary, and the clearest evidence that the boundary is knowable is that those two state it correctly.
Google DeepMind says SynthID "embeds digital watermarks directly into AI-generated images, audio, text or video" and that the marks are "imperceptible to humans" but "can be detected by SynthID's technology". On images it says the watermark "doesn't change the image or video quality", is added "the moment content is created", and is "designed to stand up to modifications like cropping, adding filters, changing frame rates, or lossy compression". OpenAI says it embeds Google's SynthID into the pixels of images generated through ChatGPT, Codex and the OpenAI API, alongside a separate Content Credentials record that is metadata and does come off. That is the whole technical picture, and it is what makes the claims below checkable.
Each quote below is from the site's own page, read on 2026-10-16. Nothing was uploaded to any of them and no removal was attempted, so nothing here reports how any of them performed.
synthidremove.com — headline: "SynthID Remover | #1 Google SynthID Watermark Removal Tool Online"
"Our free AI watermark remover tool erases hidden identifiers and AI fingerprints without losing quality."
"Our platform offers the most advanced SynthID watermark removal engine, specifically designed to identify and neutralize invisible content markers in generative media."
"Google SynthID is embedded as a specific 'noise' signal in the frequency domain of an image bitstream. Our SynthID remover tool applies an inverse spectral filter that targets only these specific frequencies, effectively neutralizing the watermark without altering the visual appearance of the image pixels."
What is verifiable in that last claim: the site's own security claim is precise and checkable — processing runs locally over WebAssembly, no uploads, no cookies. "Discrete Cosine Transform to identify high-frequency delta noise patterns characteristic of SynthID" is a description of an operation that must change every coefficient it touches. A filter that neutralises a signal spread across an image's frequency domain cannot leave the image bit-identical, and "without altering the visual appearance" is a much weaker claim than "without quality loss", which appears elsewhere on the same page.
removesynthid.io — headline: "Remove SynthID: Delete AI Watermarks Without Quality Loss"
"Our Gemini Image Watermark Remover is designed to disrupt SynthID watermark patterns through small, structured adjustments to image pixels. Instead of trying to remove metadata, the tool targets the kinds of embedded patterns used by robust invisible watermarking. In simple terms, we apply extremely ..."
What is verifiable in that: this is the most candid of the four, because it says "disrupt" and "structured adjustments to image pixels" rather than "remove". It also explicitly distinguishes itself from metadata removal. That is the honest framing, and it is the same conclusion our measurement reaches by a different route: the only operation that disturbs a pixel-level mark is one that changes pixels.
notebooklmremover.org — a Gemini sparkle remover, quoting itself
"This tool targets the supported visible overlay only. It is not designed to remove Google's imperceptible SynthID watermark or Content Credentials."
"A different mark and a different method entirely. The sparkle is alpha-blended, so the pixels underneath can be solved for rather than repainted: original = (watermarked − α × 255) / (1 − α)."
Why this one matters: a tool in the same keyword space draws the line correctly, in one sentence, and then explains the method it does use. A visible alpha-blended overlay is mathematically invertible; a signal written into the generation process is not. The two sentences side by side are the clearest possible statement of why the other three promises cannot hold.
geminieraser.com — a Gemini eraser, quoting itself
"This tool targets the supported visible overlay only. It is not designed to remove Google's imperceptible SynthID watermark or Content Credentials."
"What Is the Gemini Watermark? Some Google Gemini image outputs include a visible sparkle icon near the bottom-right corner. This tool recognizes calibra..."
What is verifiable in that: the same honest boundary, drawn by a second independent site, in almost the same sentence. When two tools in one keyword space both say "we do not touch SynthID", that is the boundary speaking, not modesty.
Every one of those promises is about a mark inside the pixels, so the way to test them is to price the operation rather than argue about it. The remove-synthid page runs five routes over one fixture and hashes the decoded samples each time:
The two halves cannot be the same operation. The left column is a header operation that never decodes the picture and so cannot reach a mark inside it; the right column reaches the mark and changes every decoded sample. A site that promises the right column at the left column's price is not describing an operation, and that is a claim you can check from the outside, without an account and without a marked file.
The page with the measurement is remove SynthID: what actually comes off, and the ChatGPT-specific two-layer split is on the ChatGPT image page. If your file has a visible corner mark rather than an invisible one, the logo page prices that route, and the metadata page prices the one that is genuinely lossless.
None of them can do what their headline says. SynthID is embedded in the pixels while the image is generated, so a tool that reaches the mark must change the pixels, and a tool that claims removal with no quality loss is describing a header strip instead. Three of the four sites quoted here really do strip EXIF, XMP and C2PA metadata, and then attribute that work to the pixel mark. The measurement on our remove-synthid page prices each route: the only one that leaves the decoded pixels identical never decodes the image.
Because the honest part of their pipeline genuinely is lossless, and they are describing it while implying it covers the mark. Metadata lives in header segments before the picture data, so removing it changes nothing a decoder draws. A mark inside the pixels is the opposite case. One of the sites quoted below makes the distinction correctly in its own words, which is the clearest evidence that the distinction is available to the tools that blur it.
Two of the four here are. The NotebookLM remover states plainly that its Gemini tool "targets the supported visible overlay only" and is "not designed to remove Google's imperceptible SynthID watermark or Content Credentials", and a second comparable site states in the same sentence that the mark is not removed and why it is not. A page that names what it cannot do is the version worth reading; the ones that promise pixel removal with zero loss are not describing an operation that exists.
For the metadata half, it removes real bytes: our measurement on the remove-synthid page took 4,415 bytes out of a header and left the decoded pixel hash identical, at PSNR infinity. For the pixel half it either re-encodes, crops, or perturbs every pixel, and all three changed the decoded hash at 45.9 to 49.4 dB. Whether a given site does the honest part, the dishonest part, or both is visible in its own wording, which is why this page quotes them rather than summarising them.
That is a question about the site, not about watermarks, and the sites here are split on it. One advertises that processing is entirely local and that the page can be checked in a network panel, which is a claim you can verify in about thirty seconds. The tool on this site runs on your device with no upload at all, because there is no backend to upload to — which is a property of the architecture rather than a promise you are asked to believe.
Yes, with one of the official detectors rather than a remover site. Google runs the SynthID Detector at synthid.com for files made with Google or its partners including OpenAI, NVIDIA and Kakao, and OpenAI runs openai.com/verify for content generated or exported by OpenAI tools. Both look for a signal written at generation. Neither is a general AI detector, and a negative result does not mean a person made the file.
Sources, all read 2026-10-16: synthidremove.com · removesynthid.io · notebooklmremover.org · geminieraser.com · Google DeepMind, Google SynthID · OpenAI Help Center, Provenance signals in OpenAI-generated content. The measurement quoted above is reproduced by dev/synthid_edits.py in this repository.