Back to Project
A

ATAS

Project Oct 15, 2026

An agentic video-generation framework that keeps static scene geometry and time-indexed event observations in a persistent world state.

Visit primary link 3 sources attached
ATAS generated film frames above reconstructed world views for Winter Fire in the Cabin and Saturday Laundry, with time-indexed people shown in cyan.
Author-published generated film frames and reconstructed world views. These examples are not independently reproduced results. Source: Weijie Wang and collaborators, ATAS
ai-agentsmultimodal

What happened lately

ATAS timeline

One persistent world across shots

Any Time, Any Space, or ATAS, compiles a screenplay into bounded shot requests with explicit scene, camera, time and event references. It retrieves registered static geometry from scene memory and dynamic observations from a separate event memory. Generated clips are reconstructed and registered, but only accepted observations update the world used by subsequent requests. Film delivery and memory admission are separate decisions.

The authors report improved cross-shot consistency across four video generators. Their project page includes generated films and reconstructed world views; these are author-published examples rather than independently reproduced results.

Running the released code

The English README specifies Python 3.12 and ffmpeg for the client, plus an NVIDIA GPU geometry service with about 20 GB of free memory and Pi3X and SAM3 model files. Users configure a DashScope API key and LLM credentials. The workflow uses paid generations; the client itself needs no GPU. ATAS carries an MIT license, while third-party geometry backbones retain their own license files.