Any Time, Any Space: Coherent Video Generation (Weijie Wang et al.)
from Marginalia
Any Time, Any Space tackles the problem of keeping generated video shots consistent across changes of place, viewpoint, and time. The framework turns a screenplay into shot requests with explicit scene, camera, and event-time references. Scene memory stores registered static geometry, while event memory retains dynamic observations indexed by scene, event, and time. Before generating a shot, the system retrieves relevant evidence; afterward, review and state checks decide whether new observations enter memory. The authors report improvements in cross-shot consistency, spatial revisiting, and event replay across four video generators. Series: Weekly AI Papers · Recent research Authors: Weijie Wang, Haomin Zhang, Ye Li, Akide Liu, Zeyu Zhang, Haoyu Zhao, Yefei He, Donny Y. Chen, Feng Chen, and Bohan Zhuang Source: Any Time, Any Space: A World-State-Centric Framework for Agentic Video Generation · Weijie Wang, Haomin Zhang, Ye Li et al. Explore: AI & Computer Science · AI Papers Related episodes: V-JEPA: Learning Visual Representations from Video (Adrien Bardes, Yann LeCun et al.) · VisionHOPE: Visual Backbones as Self-Modifying Learning Systems Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode!
Embed this episode
No transcript yet.