Google Imagen and ImageFX Image Detector: SynthID and What to Check Beyond It

How to detect Google Imagen, ImageFX and Gemini images: check for the SynthID watermark, what it survives, and what to run when the check comes back empty.

To detect an image made with Google Imagen, ImageFX or Gemini, check it for SynthID, the invisible watermark Google embeds in the pixels of every image its models generate. Upload the image to the Gemini app and ask whether it was made or altered by Google AI, or use Google's SynthID Detector portal if you have access. If the answer is no watermark, that rules out unedited Google output and nothing else, so the next step is a pixel-level AI image detector.

Google's approach is the opposite of Adobe's, and knowing the difference is most of the skill here.

Key Takeaways

• Every Google image model embeds SynthID. Imagen, ImageFX and Gemini's image output carry it, and Google says over 10 billion pieces of content have been watermarked.
• It lives in the pixels, not the metadata. According to Google DeepMind, SynthID is imperceptible and designed to stand up to cropping, filters and lossy compression, which is what kills Content Credentials.
• You can check it yourself in Gemini. Upload the image and ask whether it was created or altered by Google AI. The SynthID Detector portal, announced May 20, 2025, is on a waitlist for journalists and researchers.
• A negative result is narrow. No watermark means either not Google, real, or heavily edited. It does not mean real.
• Google also attaches metadata, which a screenshot strips. The watermark is the layer that survives.

What SynthID actually is

SynthID is a watermarking system from Google DeepMind. For images, it modifies pixel values in a pattern that is invisible to people and detectable by Google's own model. According to Google DeepMind's SynthID page, the watermark does not change image quality and is designed to survive modifications like cropping, adding filters and lossy compression. That resilience is the whole point. Metadata is a label stuck on the outside of the file. SynthID is dye in the water.

Google applies it across its generative products: Imagen (the model), ImageFX (the consumer tool built on it), Gemini's image generation, and Veo for video. There is no user setting to turn it off, and it is applied whether the image was generated from a prompt or edited from an uploaded photo, which is why the Gemini check asks whether Google AI created or altered the image rather than only whether it created it. Our SynthID explainer covers the mechanism in more depth.

How to detect an Imagen or ImageFX image, step by step

1. Ask Gemini. Open the Gemini app, upload the image, and ask whether it was created or altered by Google AI. Google states this works for images, video and audio made with its tools. It is the only public route to a SynthID check today.

2. Use the SynthID Detector portal if you have access. Google announced the portal on May 20, 2025 for images, audio, video and text. It highlights the regions of an image most likely to be watermarked. Access is rolling out through a waitlist aimed at journalists, media professionals and researchers.

3. Read the answer narrowly. Watermark found: a Google model made or edited it. No watermark: the image is real, made by a non-Google model, or edited so heavily that the watermark no longer reads. Google does not publish a false-negative rate for the public check, so treat a clean result as one signal.

4. Check the metadata too. An unedited Imagen file also carries provenance metadata identifying it as AI-generated. Drop the original into a Content Credentials or metadata viewer. If it is a screenshot, expect nothing; the metadata is gone and the watermark is what you have left.

5. When both are empty, read the pixels. Upload to the Google Imagen image detector at GPTOne for a confidence score and a region heatmap showing which parts of the image look generated. It does not depend on any vendor's watermark, which is what you need for the two-thirds of the internet that was not made by Google. The scan also reads C2PA credentials in the same pass. A new free account includes one image scan on its starting credits, no card required.

What SynthID survives, and what it does not

Google's design goal was survival through ordinary handling: crop, resize, compress, filter, re-upload. Independent testers report the watermark reading reliably on original files and less reliably after heavy social media compression, which is consistent with Google's own wording of "designed to stand up to" rather than "guaranteed". Heavy edits, generative inpainting in another tool, or re-generation through a different model can remove it.

There is a second limit that matters for anyone outside Google. Only Google can read SynthID. The detection model is not public, there is no open-source verifier, and third-party sites that advertise a SynthID checker are either wrapping the Gemini app or guessing. If a tool outside Google tells you an image has a SynthID watermark, ask how it knows.

The practical rule: a Google image that was downloaded and re-posted probably still carries SynthID. A Google image that was screenshotted, cropped and run through three apps may not. In both cases the metadata is long gone, which is why the watermark, not the credential, is Google's primary signal.

Imagen versus Firefly versus OpenAI: three provenance designs

The trade-off is visible in the table. Metadata is open and anyone can verify it, but it is fragile. A watermark is durable, but only its owner can read it, so a Google image can only be confirmed as Google by Google. Adobe's case is in our Firefly image detector guide, and the standard behind the metadata column in how to verify C2PA Content Credentials.

Visual tells for Imagen and ImageFX output

When the watermark check is unavailable and the metadata is gone, look at the image. Recent Imagen models are among the most photorealistic available, so the tells are subtle:

• Text and typography. Imagen renders text better than most, which makes the failures stranger: correct words in a font that no printer produced.
• Lighting that is too consistent. Every surface lit from a plausible direction with no stray reflections or mixed colour temperatures.
• Hands, teeth and jewellery. Improved, but still where extra digits and merged shapes appear.
• Backgrounds that resolve into nothing. Crowds, shelves and skylines that look right at a glance and dissolve when you zoom.
• Symmetry without imperfection. Faces, buildings and objects more regular than photography produces.

Score them with the heatmap rather than by eye, because the tells shift with every model release and the checklist you learned on last year's outputs is already partly out of date. Our 12 signs an image is AI-generated is the longer checklist, and the detector points at the region so you know where to look.

FAQ

How do I check if an image was made with Google AI?

Upload it to the Gemini app and ask whether it was created or altered by Google AI. That runs a SynthID check. Journalists and researchers can also apply for Google's SynthID Detector portal.

Does ImageFX add a watermark?

Yes. ImageFX runs on Imagen and every output carries the invisible SynthID watermark, with no option to disable it.

Can SynthID be removed?

Heavy editing, re-generation through another model, or extreme compression can degrade it to the point it no longer reads. Ordinary cropping, resizing and filters are what it was designed to survive.

If Gemini says no watermark, is the image real?

Not necessarily. It means Google's models did not make it, or it was edited enough to break the watermark. Run a pixel-level AI image detector for the general case.

Can GPTOne detect SynthID?

No tool outside Google can read SynthID. GPTOne detects generated images from their pixel patterns regardless of generator, reads C2PA credentials where present, and shows a region heatmap.

Check for the watermark first, the metadata second, and the pixels third. The GPTOne AI image detector covers the third step for any generator, with a heatmap and C2PA reading on the free tier.