Forensic watermarking for screeners.
What it traces, and what it cannot stop.

The cost of the visible mark is the picture. An email address burned across the frame is fine for a rough cut going to a producer and actively harmful for a distributor judging the grade, who is looking at shadow detail and skin tone through your overlay. That is the trade the visible mark makes, and it is why it cannot be the only control on a screener that someone is supposed to evaluate on its merits.

The cost of the forensic mark is that it does nothing until after the damage. It buys you nothing on the night of the screening. It buys you a name three weeks later, when a copy turns up somewhere it should not be.

No standards body publishes a definition of the pair that we could fetch and cite, so the definitions above are vendor documentation and are labelled that way on purpose: they come from Frame.io’s own help center. NAGRA, the detection vendor Frame.io names, uses the word imperceptible for the same idea on its own product page, which is marketing on a vendor’s site and should be read as such. Imperceptible is a design goal, not a measurement.

The mark does not lock the file. It does not block a download, does not block a screen recording, does not block a phone pointed at a monitor. Every one of those works exactly as well on a forensically marked file as on a clean one. What changes is what happens afterwards: there is something in the pixels that says which issued copy this one descended from.

That is genuinely valuable in a narrow way. Attribution only means anything when the pool of suspects is small enough to act on. A dozen festival programmers, five distributors, thirty journalists under embargo: that is the shape of the problem this was built for. A link sent to two hundred people is not a case you can close with a mark.

The second caveat is survivability. A mark is worth exactly what survives between the copy you issued and the copy you recovered, and real leaks are re-encoded, cropped, rescaled, screen-recorded and re-uploaded, often several times over. Whether an identifier still comes out at the end of that chain depends entirely on the implementation. Anyone quoting a survival rate without naming the test material and the exact chain of alterations is selling, not measuring. We have no verified source for such a number, so this page states none.

Detection is also a separate job from marking, and buyers routinely assume it is included. Frame.io documents that detection is not run by Frame.io: the customer obtains the leaked file, works with NAGRA, and then supplies the resulting Forensic ID back to Frame.io to retrieve the session details. That is a real workflow with a real third party in it. It is not a button.

Give two people two copies of the same film, each carrying its own mark. Those two people can compare. Line the files up frame by frame and the pixels that differ between them are, by definition, the ones carrying identity. Everything else is the movie. Average the two copies together and a naive per-person mark can be washed out entirely. Worse, depending on how the marks were assigned, the average can decode to a third recipient who had nothing to do with it. That second outcome is the one that matters most, because a scheme that frames an innocent recipient is worse than no scheme at all.

This is a collusion attack, and defending against it is a design problem rather than a strength problem. Making the mark harder to see or harder to remove does not help. What helps is choosing what each recipient’s mark looks like so that any combination of a small number of them still carries a readable signature of its parts. A collusion-resistant scheme is one built so that a copy assembled from several marked versions still traces back to the people who contributed to it, instead of dissolving or pointing somewhere false.

Whether you need to care depends entirely on your recipient list. One distributor is not a collusion problem. Forty jurors who all know each other and all hold the same film is. When you are evaluating a vendor, this is the question that separates the serious from the promotional: ask what happens when two of your recipients average their copies. A vendor who has not thought about it will answer that the mark is invisible, which is not an answer.

Two details in the Frame.io column are material to a buyer and easy to miss. The first is the tier: both marks sit at the top of the catalogue, session-based watermarking documented as Enterprise Prime and forensic watermarking documented as supported for Enterprise Prime accounts only. If you are on Pro or Team, the forensic mark is not a switch you can find. The static text watermark you do get on those tiers is a different feature.

The second is where the mark lands. Frame.io documents its forensic watermark as applied to proxy files rather than original uploads, working only with official Frame.io clients that support HLS. So the traceable copy is the streamed proxy, not the master. A file that leaves through the download permission is not the copy the forensic mark was written into, and for a workflow whose deliverable is the file, that decides whether the feature covers your actual risk.

Vimeo deserves a straight answer too, because it is where most of this work actually sits. Nothing in the Vimeo help pages we fetched documents watermarking of any kind. What Vimeo documents is privacy settings, and the one people reach for is Unlisted. Vimeo’s own description of it is the important sentence: videos can be accessed and shared by anyone who has the video’s unique URL. That is the vendor stating plainly that an unlisted link is transferable rather than per-recipient.

One more Vimeo detail, before you assume a link tells you who watched. Vimeo documents viewer-level analytics as a CSV download in showcases that reports the team member’s email address, and states that results will only appear for logged-in team members, not for signed-out viewers. A distributor or a journalist opening a link you emailed is not a logged-in team member. That is not a criticism. Vimeo is a portfolio and delivery platform and its documentation is clear about what it is. It is simply not a screener security product, and using it as one should be a decision rather than a default.

Two things this list deliberately does not say. FairPlay DRM is on the roadmap and has never shipped, so it is not part of the Vault today and will not be described here as though it were. And there is no published self-serve detection step: if a copy escapes, preserve it, keep the access record, and get in touch. Recovery is conditional in every implementation of forensic marking, including this one, and a page that promises otherwise is a page to distrust.

The one structural difference worth naming is that none of this costs picture quality. The player streams the exact bytes you uploaded, so a 200 Mbps master streams at 200 Mbps. Securing a film should not mean showing a director a soft proxy of their own work. The screener pillar lines every control up against the leak path it actually covers, and the security page documents the limits in full.