Skip to the content
detectaicheck

AI watermark checker

Some AI image generators hide a watermark in the pixels of every picture. You cannot see it, but software can read it back. We check for the watermark of the reference Stable Diffusion code, which many apps built on it keep switched on.

Up to 10.0 MB. Kept in memory for 30 minutes at most, or deleted at once with Delete now.

What we check

  • The Stable Diffusion watermark

    The reference code of Stable Diffusion hides a short message in each picture by changing it very slightly, in a way that survives saving. We read the pixels and look for that message.

  • Content Credentials

    Newer generators often sign their pictures instead of, or as well as, watermarking them. We read and validate the record on the same file.

  • Metadata traces

    We also read the metadata, where the same tools often write their name or settings.

  • AI image classifier

    When no watermark or record is found, the classifier gives an estimate from the pixels.

How to check for a watermark

  1. Pick the largest, least edited version of the picture you have.
  2. Drop it on the box above.
  3. Read the verdict. A watermark detection is listed as a finding of its own, with the watermark it matched.

What a result means, and its limits

  • Many apps built on Stable Diffusion turn the watermark off, and most other generators use a different watermark or none at all. This check does not see those.
  • Resizing, cropping, filters, strong compression and screenshots can erase the watermark, so a missing watermark proves nothing.
  • A watermark that is found is strong evidence of AI. A clean result does not prove that a person made the picture.

Frequently asked questions

Which watermarks do you detect?

The invisible watermark of the reference Stable Diffusion code. The watermarks of other generators are usually secret and can only be read by their owners, so we do not claim to find them.

Can the watermark check raise a false alarm?

It is rare, because the check looks for one exact hidden message, but no check is perfect. A detection next to other evidence is the strongest result.

Why was no watermark found in a picture I know is AI?

The generator may not add one, the app may have turned it off, or editing may have removed it. The classifier and the metadata can still point to AI.

Can a watermark be removed?

Yes. Resizing, cropping or recompressing the picture usually weakens it until it can no longer be read.

Is a watermark the same as Content Credentials?

No. A watermark lives in the pixels and survives some copying, while Content Credentials are a signed record stored next to the pixels and lost when the file is rewritten. The two complement each other.