Sora video watermark detector

Sora video has carried marking at two levels that are structurally unrelated to each other. At the pixel level there is a visible mark composited into the frames, its position varying across the clip. At the container level there is a C2PA manifest: a cryptographically signed provenance record stating what produced the file.

This route reads both and reports them separately, because collapsing them into one answer destroys the only interesting information in the result. The clip is uploaded to our media workflow, since container metadata cannot be inspected without the file.

Durability follows location. Anything encoded in the picture inherits the durability of the picture, and anything stored beside the picture inherits the fragility of metadata.

Pixel layer

Survives transcoding, scaling and compression, because those operations preserve the image content they are applied to. Removed only by altering the affected pixels.

Container layer

Discarded by most re-muxing and re-encoding paths, frequently without notice, because carrying unknown metadata forward is not the default behaviour of most pipelines.

Consequence

The two layers decay at wildly different rates, so the pair of results is more informative than either result alone.

The provenance of the copy under test bounds what any check can establish.

Step 1

Select the earliest generation of the file available to you. Each re-encode in the chain is an opportunity for the container layer to have been dropped.

Step 2

Submit it above. Our media workflow parses the container for a provenance manifest and examines frames for the visible mark.

Step 3

Interpret the two findings independently, and record the file's handling history alongside them, because that history is what makes a negative reading meaningful or meaningless.

A present manifest and an absent manifest are not two sides of one measurement. A present manifest is signed, which makes forgery expensive and makes the reading strong evidence about the file. An absent manifest is consistent with the file never having had one and equally consistent with routine stripping, and those two explanations are not distinguishable from the file alone.

The practical rule that follows: treat a positive reading as informative and a negative reading as a statement about the file's history rather than its origin. Any page that presents the absence of provenance as evidence of human authorship has inverted the logic.

The intuition that a visible mark indicates a heavily marked file and a clean picture indicates a clean file does not hold. A download issued without the visible logo can carry a complete signed manifest, while a clip that obviously shows the mark can have lost its metadata to a single platform upload.

The visible mark therefore correlates with the delivery path, not with how much provenance data the file holds. It is a poor proxy, and a detector reporting only pixels or only metadata will mislead in opposite directions on those two files.

A mark whose position varies across frames turns detection into a sampling problem. A thumbnail, a screenshot or a first-frame check samples the clip once, and a mark that is small or against a high-detail background at that instant is easy to miss.

The same sampling error distorts evaluation of a removal pass. Inspecting a handful of frames and concluding the mark is gone overstates the result, since the frames where removal is hardest are precisely the frames nobody checks.

It reads two signals. It does not perform statistical analysis of frame content, does not fingerprint motion, does not model temporal artefacts and does not output a probability that a clip is synthetic. Those are different techniques with different error profiles, and none of them is running here.

It also makes no claim about what any third-party detector or platform will conclude about the same file. A signed manifest naming a generator is strong evidence about the file; everything beyond that is inference, and this page does not present inference as measurement.

The useful applications share one trait: they need a defensible record of what the file declared, not a headline verdict.

Verification work

Establishing and documenting what a circulating file declares about itself, including the common case of declaring nothing.

Rights and delivery review

Checking supplied footage against delivery requirements before it is cut into a published piece.

Pipeline auditing

Determining at which stage of an internal chain provenance metadata is being discarded.