ATAS framework uses world-state memory to keep long-form agentic video generation consistent
zju · hf · 2026-10-15
Researchers from Zhejiang University, Monash, UC Berkeley and Adelaide present Any Time, Any Space (ATAS), a framework for long-form agentic video generation that keeps shots spatially and temporally consistent across nonconsecutive requests.
- Screenplays are compiled into shot requests with explicit location, camera, and event-time references; a scene memory of static geometry plus an event memory indexed by scene, event and time supplies evidence before generation, and clips are reconstructed and registered afterward, updating memory only after review and state checks.
- Across four video generator backbones, ATAS improves all reported cross-shot consistency measures and tops aggregate benchmark scores, outperforming compared agents and 3D-guided methods on spatial revisit and three event-recapture criteria, with the highest human ratings on all six criteria.
- Code and videos are available on the project page.
More from Multimodal
- Stunning AI video made with Seedance 2.5 draws praise as "a masterpiece" — SimplyAnnisa · 2026-10-15
- Founder makes viral hype video in under 2 hours with Claude Opus 5.5 and Arcmira MCP — zealcaiden · 2026-10-15
- Testing how AI thinks dogs wear pants: local models fail, SOTA nails jeans — bstr3k · 2026-10-15
- 110-Page Survey From 6 Institutions Maps Memory Mechanisms in Autoregressive Video Generation — jiqizhixin · 2026-10-15
- Building a 'hypnotic piano groove' skill for AI music systems, modeled on Todd Terje — doodlestein · 2026-10-15
- Claude paints pixel art with code: a 1,000-line DSL for animated scenes — Realistic_Comfort_78 · 2026-10-15