Research
One image, four layers of provenance — XMSE in plain language
Why a single watermark is not enough, how four layers stack, and what gets better for users — easy to read.
One line: Provenance is not one magic stamp. It is several layers that stop different kinds of cheating.
Why we built this
As AI makes media easy to create, people ask:
- Did this file really come from our system?
- Who changed it?
- How do we find a re-photographed copy?
A watermark alone does not explain edit history.
A hash alone misses resized cousins.
A signature alone does not answer every forensic question.
So XMSE separates four layers.
How we built it
- Watermark — an invisible mark in the signal (encode runtime still unfinished — we do not claim “done”).
- Forensic cues — reading traces of suspicious transforms.
- Fingerprints — exact hash vs perceptual similarity.
- C2PA — a standard “resume” of origin and edits.
NFTs are optional registration, not a fifth magic layer. Crystal is not cash.
What gets better
Creators get a path to explain origin. Verifiers can inspect layers separately. The platform refuses to put all trust in one trick.
Public: XMSE stack · C2PA portal
Limits
Status is pilot. Neural encode, some mint/timestamp live checks, and final browser QC remain. We do not claim “all media provenance complete.”