Zhejiang-led team open-sources ATAS for cross-shot consistent AI filmmaking
lmoroney · x · 2026-10-16
- Teams from Zhejiang University, Monash, UC Berkeley, and Adelaide open-sourced ATAS (Any Time, Any Space), which turns a plain-text screenplay into a multi-shot film while maintaining a running world state to fix cross-shot inconsistency.
- A scene memory stores each location's geometry in a shared 3D frame; an event memory indexes earlier moments for re-filming from new cameras; generated clips are reconstructed and verified before updating memory.
- Authors report gains across four video generators and top human scores on all six criteria, at a cost of runtime rising from 52 to 65 minutes. MIT-licensed, uses Wan 3.0 by default, needs 20GB GPU for the geometry service.
More from Multimodal
- Shanghai AI Lab unveils Intern Lumina U2, a diffusion LLM unifying image, video and 3D tasks — Shanghai-AI-Laboratory · 2026-10-16
- NAVER AI Lab proposes Speculative Forcing to speed up autoregressive video generation — naver-ai · 2026-10-16
- MFVSR: A Multi-Style Benchmark with 3,000 Videos and 16 Languages for Video Subtitle Removal — Xiyan Huang · 2026-10-16
- OmniCompass: PolyU launches a holistic benchmark for omni-modal video grounding and QA — VCLab-HKPU · 2026-10-16
- Researcher admits his prediction that text-to-image would incubate AI control research was falsified — jd_pressman · 2026-10-16
- AI turns Kuroko's Basket rivals into muscle-bound shoppers in viral clip — EternalSnow05 · 2026-10-16