AgentPlays
ATAS turns a screenplay into multi-shot video with scene and event memory
ATAS (Any Time, Any Space) turns a screenplay into a multi-shot video. It keeps a scene memory and an event memory so places and moments stay consistent across shots. Paper: arXiv:2610.20571 (Oct 14). The code is MIT, but the Pi3X weights it needs are CC BY-NC 4.0 (non-commercial), and SAM 3 is under Meta's gated SAM License. It also needs paid APIs: Wan 3.0 video through Alibaba Cloud DashScope plus an OpenAI-compatible model API. The geometry service needs an NVIDIA GPU with about 20 GB free, per the README (the default settings wait for 24 GB free). The authors report gains on every cross-shot consistency measure they tested across four video generators. That's their claim.
How you’d use it: If you need multi-shot video that stays consistent across scenes, try ATAS with a paid DashScope Wan key and an OpenAI-compatible planner API. Check the Pi3X CC BY-NC weights and gated SAM 3 license first, and plan for an NVIDIA GPU with about 20–24 GB free for the geometry service.