Sense/Net’s crowds are trained on 41 million frames from a partner in Tokyo. Its consent rules don’t cover them.
The studio’s new transparency report counts the first-person footage it licenses to train Fullfield Ambient. Who is in that footage, and whether they know, is left to the licensor.
1 min read
LOS ANGELES — Sense/Net’s 2026 transparency report, published on Oct. 14, runs to dozens of tables about the people in its recordings: how many performers gave consent, how many withdrew it, how long each withdrawal took. One table is about people who never agreed to anything.
In the year to June 30, the report says, the studio received 41,233,860 frames of “reference footage” from a single licensor. The footage is used to train and calibrate Fullfield Ambient, the tool that fills Sense/Net’s first-person recordings with generated passers-by so that no member of the public appears in a title. The report describes it as ambient first-person reference footage licensed from a Tokyo optics partner, under the partner’s consent framework. Half-year reports published in February and August give the same total in two parts, 20,297,368 and 20,936,492 frames, and say that consent records for the material are held by the licensor.
The footage falls outside the studio’s own rules. Sense/Net’s Consent Standard, which governs every recording it releases, says in its first section that it does not apply to reference material held only to develop, train or calibrate tools, which is never shown to viewers. For that material, the standard relies on the licensor’s warranty.
First-person footage is what a camera at eye level sees as a person moves through the world: streets, shops, other people’s faces. Sense/Net has written that first-person recording “has always had a bystander problem,” because such a camera sees everyone the wearer sees.
Sense/Net declined to name its partner, to describe the partner’s consent framework, or to say whether the people who appear in the footage know it is used to train Ambient. In a statement, it said reference material never appears in a title, that nobody in it appears in one, and that its licence requires the licensor to hold consent for all of it.
In March, performers who had withdrawn titles from Sense/Net told The Overseen that nobody could tell them whether footage of them had been used to train Ambient.