Make more film.
Keep what it established.
Videogrammer Research studies how AI can make far more film, in far more versions, without losing what a film has established. We build, measure, and publish what we find, including when it fails. Everything on this page is Videogrammer product research.
Three programmes.
One record of the film
A versioned record of people, places, props, looks and decisions, with story time, shooting order and edit order kept apart.
Planned
A check that shows its evidence
Possible breaks shown with frames and timecodes, and one of four outcomes: confirmed break, explained change, needs review, insufficient evidence.
Prototype
Which model earns which job
The same jobs done by frontier models and by open models, with a locked exam deciding which can be trusted with what.
Prototype
Every version, from one record.
Videogrammer started with a question: why is video still made one export at a time? A generator answers a prompt; it does not remember a film. Our bet runs the other way. Hold the film's approved record, and compile every format, cut and language version from it, so versions differ only where someone decided they should.
Compiled, not gambled
Re-rolling a generator until a shot looks right is a slot machine. A record of what was approved turns the next version into a build, not a bet.
One record, every screen
Vertical, square and widescreen; the 15-second cut and the full film. Each is drawn from the same decisions, so nothing drifts between them.
Many languages, one story
Language versions, starting with the Indian languages our audiences speak, keeping the same people, places and promises on screen.
The numbers, with their context.
| Test | Date | What ran | Data | Baseline | Result | Limits |
|---|---|---|---|---|---|---|
| Engine self-test | 11 Nov 2026 | Our continuity rules | The engine's own planted test scenes | None | 10 of 10 breaks caught, no false alarms | Planted cases with fixed labels; a low bar |
| First real scene | 11 Nov 2026 | Our continuity rules, then two fixes | 80 seconds of Tears of Steel (openly licensed), 17 cuts, no real breaks | Zero alarms is correct | 22 false alarms, then 8 | One scene; descriptions written by us |
| Flag triage exam v1 | 11 Nov 2026 | Two untrained local models (Qwen 2.5, 1.5B and 7B, via Ollama) | 37 written cases: 16 real breaks, 21 false alarms | Believing every flag: 16 of 37 | Both 16 of 37 | A small exam of written cases; not product validation |
Detailed reports are internal for now and available on request.
The false alarms.
What failed
On a real scene with no breaks, the rules flagged changes that were not breaks, such as a different angle, a person entering a wider shot, or a prop out of view.
Our hypothesis, and the test
We think comparing free-text descriptions is the cause. We will give the record named people, places and props with stable identities, rerun the same scene, and add a scene with a known break. We will publish both results either way.
Four rules.
Publish either way
We publish what we measure, whether it flatters us or not.
Measure, never claim
No percentages we have not measured.
Consent first
No client footage or student work trains a model without written consent.
A person decides
Review steps can share blind spots, so a person makes the final call.
Test it on your film
Studios that want to put the check through a real reel can run a scoped pilot with us.