The most awkward moment in long-form AI video: every shot looks fine on its own, but stitch them into a film and things break — the character's outfit changes, the room rearranges itself, yesterday's event replays as a different version. With single-shot quality already pushing 0.99, the bottleneck is no longer "can it render well" but "can it remember."
Previous multi-shot pipelines pass prompts or recent frames between shots. The ATAS team — Zhejiang University with Monash, UC Berkeley BAIR and Adelaide — argues this is not enough: passing prompts provides no shared spatial frame for revisiting a scene, and passing frames gives no time-indexed record for replaying an earlier event. Their answer, ATAS (Any Time, Any Space), landed on arXiv (2610.20571) on October 14 and reached the top of Hugging Face Daily Papers the next day, now at 271 upvotes.
The framework: compile the screenplay, keep the world
The central object is a persistent world state. A world-state compiler recursively splits the screenplay into bounded shot requests — each carrying an explicit scene, spatial anchor, camera, time interval and optional event; only the description becomes the prompt for the video model. Before generation, each request retrieves evidence from two memories: scene memory holds registered static geometry as point clouds in one shared coordinate frame, while event memory holds dynamic observations indexed by scene, event and time. After generation, the new clip must pass reconstruction, registration and review before it is written back into the world.
One design choice deserves attention: film delivery is decoupled from memory admission. A clip can stay in the film even if its observation fails validation and never becomes evidence for later shots. Dirty data stays out of memory — a discipline database engineers know well.
Results: first place on all four backbones
The team compared ATAS against I2V Extension, VGoT, Mora, MovieAgent and ReCA on four backbones: MiniMax-H3, Wan 2.7, Wan 3.0 and HappyHorse 1.0. ATAS takes the best average on every backbone: 0.734 vs ReCA's 0.648 on Wan 3.0, 0.702 vs 0.659 on MiniMax-H3, 0.694 vs 0.656 on HappyHorse 1.0, and 0.656 vs 0.612 on Wan 2.7. The widest gap is on cross-shot consistency: ST-Bench 0.719 vs 0.602 and ViStory-Cross 0.584 vs 0.481 on Wan 3.0.
Two targeted protocols make the point sharper. In the spatial revisit test (return to an earlier location, compare against a target image), ATAS reaches 15.968 PSNRC, ahead of ViewCrafter at 15.255 and every agent baseline. In event recapture, ATAS leads event identity, temporal order and viewpoint realization, while ReCA stays slightly ahead on action and state (0.700 vs 0.688). In human evaluation, ATAS has the highest mean on all six criteria — script faithfulness 4.40, character consistency 4.22 — against ReCA's range of 2.62 to 3.67.
The cost: fewer retries, more compute
Memory is not free. Ablations show that moving from no memory to the full world state raises per-story wall-clock time from 51.9 to 64.9 minutes and drops throughput from 7.28 to 6.21 shots per hour, with 98 seconds of memory overhead per shot; in exchange, retries fall from 1.00 to 0.54 per shot and the film-review rejection rate drops from 52.6% to 36.8%. Another detail: the state admission check has a sweet spot — the default setting admits 67.4% of candidates and scores best on all three measures, while permissive (90.0%) and strict (36.0%) both lose quality. Admitting too much dirty data and rejecting too much good data cost symmetrically.
Code and results are open-sourced under MIT (arXiv:2610.20571, GitHub:ziplab/ATAS).
So what? As single-shot generation saturates, competition shifts to state management — whoever maintains a trustworthy shared world turns AI video from a clip generator into a narrative tool. ATAS brings database-style admission discipline into the generation pipeline, and that idea is more worth copying than the scores themselves.