Firefly video watermark detector
What the Adobe Firefly video watermark detector does
Firefly's signal on video is Content Credentials: a signed C2PA manifest carried inside the video container, declaring which tools contributed to the file and whether a model produced the content. This detector reads that record.
Drop a video into the upload control and it goes into our media workflow, where the container is parsed, any manifest is located and checked, and what it declares is reported back. The file is genuinely uploaded, because a manifest lives inside the file and reading one means reading the file.
How it works
How to detect a Firefly video watermark in three steps
One pass through the upload control on this page.
- Step 1
- Use the file itself, not a screen recording of it. A screen recording is a new video and carries none of the original container's metadata.
- Step 2
- Drop it into the box above. The container is read and any C2PA manifest is parsed and checked.
- Step 3
- Read the verdict against the three outcomes below, because found and not found are not the only two answers.
Mechanics
Where the C2PA record sits inside a video file
An MP4 is a stack of boxes. Some hold video, some hold audio, some hold timing, and one box type exists to carry data the format itself does not define: the uuid box, identified by a 16-byte identifier saying what the contents are. C2PA claims one of those identifiers, D8FEC3D6-1B0B-4872-BD80-E04F827E80D9, and the manifest goes inside it.
That is why the record is separable from the picture. It sits beside the encoded frames rather than inside them. It is also why a screen recording never carries credentials: a recording captures what was played back, and boxes the player never rendered are not part of what it captured.
Reading the result
The three answers a Content Credentials check can give
A detector reporting yes or no cannot explain its own output. There are three states, and they are worth very different amounts.
- A manifest present and valid
- The signature verifies and the content still matches what was signed. Strong evidence about what the file declares, because signatures are hard to forge and cheap to check.
- A manifest present but invalid
- A record exists but does not verify, or the file has diverged from what was signed. That points at modification after signing rather than at no provenance.
- No manifest at all
- The weakest of the three. Equally consistent with a video that never carried credentials and a Firefly video that lost them somewhere along the way.
The important part
Why a blank result on video means even less
Credentials are metadata, and metadata does not survive the ordinary life of a video file. Three separate things strip it and none of them is an attempt to remove a watermark: the generating tool may never have embedded it, a screen recording or re-save discards it, and platform re-encoding writes a new container without it.
Video takes that harder than stills. Every platform transcodes on upload, share sheets re-wrap files, and screen recording is a far more common way to capture video than screenshotting is to capture an image. By the time a clip has been through a feed, the odds that any original manifest survived are poor. Treat a blank result as no information rather than an all clear.
Two different checks
Finding a marker is not the same as verifying it
Some detectors scan a file for the C2PA identifier and report a hit. That answers a narrow question: does a C2PA structure exist in this container. It does not answer whether the signature is good, whether the certificate behind it is trusted, or whether the video still matches what was signed.
Those are the checks that turn a hit into evidence. A present-but-invalid result is genuinely informative, because it says a record was written and then the file changed, which is a different situation from a file that never had one. A page that collapses both into found is telling you less than it appears to.
Stated scope
What this detector does not examine
This route reads the provenance record. It does not inspect frames, does not score how AI-generated a clip looks, and does not infer a generator from resolution, duration or bitrate patterns the way screening tools do. Those are different techniques with different failure modes, and merging them behind one verdict makes the result impossible to read.
It also will not tell you which model made a video in the absence of credentials. Detectors that attempt that say so themselves in their own limitations sections, and nothing here does better.
Who it's for
Who checks a video for a Firefly watermark and when
Editorial and broadcast teams doing provenance review on supplied footage, where a declared generation record changes how a clip is labelled or whether it runs. Agencies auditing delivered video against a brief that restricted generated material.
Anyone about to reuse a clip who wants to know what history is still attached to the file, and anyone who ran a removal and wants to confirm the manifest is actually gone before publishing.