Can AI make a long film from a single prompt?

Every video model has a per-clip ceiling measured in seconds, and it is not a product decision waiting to be lifted. Compute cost, memory and temporal drift all rise together with length: the longer a single pass runs, the further the model wanders from what it started with, which is why a twelve-second generation often ends somewhere its first second did not predict.

So a long film is always many generations. The interesting question is not whether they are joined but what holds them together — because the joining is trivial and the holding is the entire problem.

A brief becomes a plan before it becomes footage. The planner writes the whole shot list in one pass — up to 80 shots, each up to 30 seconds, with a locked description of cast, wardrobe and location carried into every one of them.

Then it shoots sequentially. Each shot is conditioned on the closing frame of the shot before it, so a character walking out of one shot walks into the next rather than being re-imagined there. The finished shots are joined into a single file with captions written from the script rather than transcribed from the audio.

A 30-minute film is dozens of sequential provider calls, and every one is a chance to fail, drift or stall. A shot that fails does not kill the film — a run that rendered nine of its ten shots still assembled — but the length is an extrapolation from films shorter than this, and anyone attempting the maximum is finding out something new.

Money is the ceiling that actually binds. At roughly 28.7 credits per second of finished film, 30 minutes is about 51,660 credits — more than any single credit pack holds. In practice a film that long is funded by topping up partway through, which is why a film that runs out of credits pauses at its last completed shot and asks, instead of failing.