Realtime video

The question

Can a generation model keep up with the display, at 24 frames per second, and still keep its objects, lighting and motion consistent over a long clip?

Why it matters

Most video generation spends minutes of compute on seconds of footage. A model that runs in real time can be used interactively: steered by a person while it generates, used in teaching and live production, and tested in ways offline models cannot be.

Realistic generated video also raises a public problem. As it gets cheaper, it gets harder to tell what was filmed and what was made. We treat provenance as part of the research, not an afterthought.

Our approach

• A latency budget per frame. At 24 fps there are about 40 ms per frame. That budget shapes choices about tokenisation, caching and conditioning, and those choices are worth studying on their own.
• Licensed data only. Licensed and public-domain footage. The data recipe, with the licence of every source, is published with each release.
• Consistency we can measure. Three metrics for temporal consistency, each with a statement of what it misses, plus prompt adherence and latency on stated hardware. See how we evaluate.
• Provenance by default. Generated output carries provenance information, and the misuse policy is published before any release.

Status now

The architecture, data policy and evaluation suite are being defined. There is no public model yet.

What we have found

Nothing to report yet. Findings will appear here and under News when there is something measured to say.