ATAS: A World-State-Centric Framework for Consistent Agentic Video Generation

Forum topic · 小凯 · 2026-10-16

Summary

ATAS (Any Time, Any Space) is a framework for long-form agentic video generation that maintains spatial and temporal consistency across nonconsecutive shots. It compiles a screenplay into shot requests with explicit location, camera, and event-time references, and maintains two forms of world-state memory: a scene memory of registered static geometry and an event memory of dynamic observations indexed by scene, event, and time. Before generation, the system retrieves relevant evidence from these memories; after generation, it reconstructs and registers the clip, updating memory only when review and state checks accept the observation, so film delivery never depends on memory admission. Evaluated across four video generators, ATAS improves all reported cross-shot consistency metrics and achieves the highest composite benchmark score for each backbone. It outperforms agentic and 3D-guided baselines on target-image-space scene revisits, leads on three event-recapture criteria, and scores highest on all six human evaluation criteria. Project page: https://lhmd.top/atas/; arXiv: 2610.20571.

Overview

Field: Computer Vision (CV)
Authors: Weijie Wang, Haomin Zhang, Ye Li, Akide Liu, Zeyu Zhang, Haoyu Zhao, Yefei He, Donny Y. Chen, Feng Chen, Bohan Zhuang
arXiv: 2610.20571
Project page: https://lhmd.top/atas/

Key points

Long-form agentic video generation must turn a screenplay into shots that remain spatially and temporally consistent across nonconsecutive requests. Passing prompts or recent frames between shots does not provide a shared spatial frame for revisiting a scene or a time-indexed record for replaying an earlier event.

ATAS addresses this by:

Compiling a screenplay into shot requests with explicit location, camera, and event-time references.
Maintaining a scene memory of registered static geometry and an event memory of dynamic observations indexed by scene, event, and time.
Retrieving relevant evidence from memory before generation.
Reconstructing and registering each generated clip after generation, updating memory only when review and state checks accept the observation — film delivery does not depend on memory admission.

Results

Improves all reported cross-shot consistency metrics across four video generators.
Achieves the highest composite benchmark score for each backbone model.
Outperforms agentic and 3D-guided baselines on target-image-space scene revisit tasks.
Leads on three event-recapture criteria and scores highest on all six human evaluation criteria.