Generating long video in real time is hard: current systems make it chunk by chunk, each waiting for the previous one to be fully cleaned up. That slows output and lets small errors turn into flicker and camera drift.
A new chunk starts while earlier ones are still being refined, using their cleaner intermediate states as context. Copies of the model run in a staggered pipeline like an assembly line, yielding one finished chunk per round after the first fill. Training adapts the model to this mixed-quality context.
On a text-to-video model, quality stayed comparable to a strong existing method while motion was livelier, and end-to-end generation ran more than seven times faster than on a single graphics processing unit.