Every number has a method behind it.
Where a number comes from an eval, the protocol is published. Where it comes from a product decision, we say so. Nothing here is a marketing round-up.
140 fields. 17 tiers.
One annotation per shot.
Every frame in the training set carries a structured annotation — shot size, lens, movement, cut rate, hook archetype, product placement, CTA position, voice-over language, music bed, colour temperature, and the beat where the ask arrives. Supervision at that resolution is why Origin can explain a decision instead of only producing pixels. Pick a tier.
The choices a director makes: blocking, eyeline, reveal timing, hold length.
From one sentence to a signed file.
Audio is generated inside the same pass as picture, not stitched afterwards — which is why lip timing survives a re-cut and why a re-voiced market doesn’t need a new render.
What we measure, and against whom.
Four internal evals plus blind human preference. The protocol, the prompt set and the raters’ brief are published with the report — including the two evals where a filmmaker-first model beats us.
Predicted vs observed three-second retention on 4,200 held-out ads across 12 markets.
Human raters judging whether ten ads from one memory read as one brand.
Word error rate on burned-in captions, measured back off the rendered frame.
Shot-by-shot art direction on a single hero clip. Filmmaker-first models win this one.
Where Origin still gets it wrong.
Published because you will find them anyway, and because a model you can’t predict is a model you can’t ship on.
If you need a single hero clip with total shot-by-shot control for a film festival, use a filmmaker-first model — they are excellent at that and we are not trying to be. Origin is for teams who need a hundred good ads live this quarter, and need to know which ones worked.