Invisible AI watermarks compared: SynthID, Meta's watermark, Adobe TrustMark and Stable Diffusion's mark
This guide compares four invisible marks: Google's SynthID, Meta's watermark, Adobe's TrustMark and the one Stable Diffusion's reference software adds. They differ in who made them and who can read them. Our checker decodes only the Stable Diffusion mark. It does not check the other three, so none of its results says anything about them.
What is an invisible AI watermark?
It is a signal written into the picture itself, not into a metadata field beside it. A person looking at the image sees no difference. A matching decoder can read the signal back.
Metadata is a different thing. The Content Authenticity Initiative's article on durable Content Credentials (April 8, 2024) says state-of-the-art watermarks can survive rebroadcasting such as screenshotting, which effectively removes secure metadata.
How do the four marks compare?
| Mark | Maker and what it says | Who can read it | Our checker |
|---|---|---|---|
| SynthID | Google says it is an invisible watermark, imperceptible to people, added to AI-generated images, audio, text or video. | Google's own checks in Gemini, Search and Chrome, and its SynthID Detector. Google says the Detector can find media only from companies that have adopted SynthID. | Not checked |
| Meta's watermark | Meta says it adds invisible watermarks, and metadata, to photorealistic images made with its Meta AI feature. | The page we read does not name a public reader. | Not checked |
| Adobe TrustMark | Adobe's repository describes open-source watermarking built for the Content Authenticity Initiative. Each mark carries 100 bits. | Anyone can run the decoding code published in the repository. Its README says the MIT license covers the code and the model files. | Not checked |
| Stable Diffusion's mark | The reference script CompVis published adds a fixed text string with a frequency-domain method called dwtDct. | Anyone, using the open-source invisible-watermark package or equivalent code. | Decoded |
What does Google say about SynthID?
Google's SynthID page says the watermark is imperceptible to humans and designed to stand up to changes such as cropping, filters and lossy compression. It says you can upload a file to Gemini and ask whether Google AI created or altered it.
The SynthID Detector page says it is not a general AI detector. It says it can detect only media from the companies that have adopted SynthID. Google's October 7, 2026 post, updated October 29, says that with the SynthID Detector anyone can check whether an image, video or audio file was made with AI "from Google or our partners", including OpenAI, NVIDIA, Kakao and Apple. The Detector page itself did not list Apple when we opened it on October 30, 2026.
What does Meta say about its watermark?
Meta's February 2024 post, updated April 1, 2025, says that when photorealistic images are created with its Meta AI feature, it adds both invisible watermarks and metadata embedded within image files.
The same post says people can strip out invisible markers. It also says it is not possible to identify all AI-generated content. The page does not say which watermarking technology Meta uses for Meta AI images.
What does Adobe's TrustMark repository say?
Adobe's TrustMark repository describes an open-source image watermarking project made for the Content Authenticity Initiative. It says the code and the model files are under the MIT license.
Each mark carries 100 bits, with error correction that can be configured. The README gives no measured robustness figures, so we give none. We have not run TrustMark on any of our files.
What is Stable Diffusion's mark?
The reference script sets the text "StableDiffusionV1" and applies it to each output with encode(img, 'dwtDct'). That method comes from the invisible-watermark package. Its README says its methods are not robust to a resize or a crop that changes the aspect ratio, and are robust to noise, color filters, brightness and JPEG compression.
Who can read each mark?
This is where the four differ most. SynthID is read through Google's tools. TrustMark and the Stable Diffusion mark can be read by anyone with the published code. For Meta's, we found no public reader.
A fifth kind exists. TikTok's November 19, 2025 post says TikTok was testing invisible watermarks and would start adding them to AI-generated content made with its own tools. It describes the mark as one "only we can read". We do not check it.
What did we measure for the Stable Diffusion mark?
Of these four marks, our checker decodes only the Stable Diffusion one. On October 27, 2026 we measured it in Chrome against 250 curated Wikimedia Commons photographs. It reports a match only at 45 of 48 bits (SDXL), 104 of 136 (SD1) or 32 of 32 (SD2). 0 of the 250 photographs, and 0 of the 31 real files in our test set that the checker evaluated, reached a threshold. The full tables are on our accuracy page.
We also embedded the mark with the real package in 7 payload and photo pairs, then saved copies. Results for the SDXL, SD1 and SD2 payloads together:
| Copy | Found |
|---|---|
| PNG | 7 of 7 |
| JPEG, quality 90, no chroma subsampling | 7 of 7 |
| JPEG, quality 90, usual 4:2:0 subsampling | 0 of 7 |
| Resize to 90% | 0 of 3 |
| Crop of 17 pixels from the top and 29 from the left | 0 of 3 |
Five unmarked control images produced no match. We did not test SynthID, Meta's mark or TrustMark, so we make no claim about how they behave. The package README's statement that the mark is robust to JPEG compression held in our test only without chroma subsampling. A not-found result on an image that has been shared online says almost nothing about this mark.
What this means when you check an image
Open the checker, choose the file, and look for the headline "Known watermark pattern matched; source unconfirmed". It means the pattern Stable Diffusion's reference software adds was found. The line under it says anyone running that software adds it, and that it is not a signed record. It does not say who made the image.
If you see "No supported provenance signals found", it means we found no signed record, no source label and no Stable Diffusion pattern. It does not mean the image is real or human-made. Every result also shows this line: "Not checked here: Google's SynthID watermark (also used by OpenAI, NVIDIA and Kakao, and by Apple according to Google), Meta's watermark, Adobe TrustMark, visible marks, and pixel-based AI detection."
For SynthID, Google's own tools can answer, within Google's limits, for example the SynthID Detector. That opens another company's site, you may need to sign in, and your image is uploaded to them, not to us. Our Privacy and method page explains what we do with your file.
Guides on this topic
- What is SynthID? is the main page on this topic.
- Can AI watermarks be removed? covers what everyday edits do to each signal.
- Watermark, metadata and fingerprint compared explains the three kinds of signal.
- Does OpenAI use SynthID? covers what OpenAI says.
- The Stable Diffusion invisible watermark covers the one mark we decode.
Frequently asked questions
- Does the checker read SynthID?
- No. It is on the not-checked line of every result. Google's tools check it, and they upload your image to Google.
- Can a mark be read by anyone?
- It depends on the mark. TrustMark and the Stable Diffusion mark have open-source code. SynthID is read through Google's tools, and TikTok's post describes its watermark as one only TikTok can read.
- Does a match on the Stable Diffusion mark mean the image was made with Stable Diffusion?
- No. It means the pattern that the reference software adds is present. Anyone running that software adds it.
- Does finding nothing mean an image has no watermark?
- No. We read one of the four marks, and in our tests our checker found 0 of 7 marks after a JPEG save with the usual 4:2:0 chroma subsampling. Finding nothing never shows an image is real or human-made.