What it is
Agentic video generation framework that keeps a persistent world state across shots, enabling viewpoint-consistent revisits and event recapture from screenplay input. Generate multi-shot films from plain-text screenplays while maintaining a shared world model that supports revisiting scenes and re-filming past events from new camera views.
Key results
- Supports plain-text screenplay input with optional character files
- World updates only pass after film review, registration, and state checks
- Client runs without GPU; geometry service needs an NVIDIA GPU with ~20 GB
Shows multi-shot video generation can be organized around a persistent world state instead of treating shots independently.