Skip to content

Invisible watermarks

An invisible watermark hides an ID in the content itself, so it stays when the metadata goes: through re-encoding, resizing and re-uploads. Kennmarq.com uses three open-source models, one for each kind of file, and reads every mark back before the file leaves.

In short

  • Images: TrustMark, by Adobe.
  • Video: Video Seal, by Meta, in every frame.
  • Audio: AudioSeal, by Meta, under the music.
  • Why: it leads back to the file's Content Credentials when platforms strip them, as the EU's Code of Practice asks.

What each one carries

The watermark in each kind of file
FileWatermarkCarriesChecked
ImagesTrustMark (variant Q), open source, by Adobe; on the C2PA's list of soft bindingsA random 61-bit ID that leads to the image's credentialAll 78 test images after JPEG at quality 50, a social-network-style resize and a resize to 25% (report)
VideoVideo Seal, open source, by MetaThe same kind of 61-bit ID, in every frameRead back from 64 frames of every clip before it leaves
AudioAudioSeal, open source, by Meta16 bits that say the sound was marked by us; the song's credential travels in its manifestRead back from the encoded file; shaped under the music, so no one hears it

What an image's watermark survives

From our robustness report, measured on 78 AI images with the code the API runs: the ID read back after JPEG at quality 50 in 100% of them, after a social network's kind of upload (long side 1080 px, JPEG 85) in 100%, after a resize to 25% in 100%, after a 2-pixel blur in 100%, and after a crop to the centre 90% in 100%. Harder edits cost more: a crop to the centre 75% kept 53%, and a 5° rotation 41%. The report has every edit, with error bars, and how often unmarked images read as marked.

Reading them

What a watermark is not

A watermark says a file was marked; it isn't an AI detector, and its absence doesn't prove a file is human-made. Very quiet recordings may not hold an audio mark: their credential then says so.