Video generators can now produce a few convincing seconds, but stretching them to minutes goes wrong: each new frame is built from earlier ones, so small errors pile up into drifting color and blur. Systems cope by keeping only a short window of recent frames, or by permanently pinning the first ones as anchors, both fixed rules rather than judgments about what is useful.
This work replaces those rules with a learned scorer, a small network that rates each past frame for how much it will help generate the next one. It is trained to distrust corrupted frames, and strong new frames gradually replace stale anchors within a fixed memory budget.
Longer clips were judged more consistent and faithful than with the fixed strategies.