Space should stay registered.
A return shot should know where the sofa, window and table are in relation to one another.
Any Time, Any Space
Keep the world
behind every shot.
ATAS is a world-state approach to agentic video generation. It gives a multi-shot story a persistent sense of place, time and event—so later shots can return, replay and remain grounded.
Independent English guide based on the ATAS paper, official project page and open-source code.
WORLD STATE / 01 persistent
⌂scenestatic geometry
◷eventtime-indexed
✦shotnew camera
ATASstate
scene memoryevent memoryvalidated update
The continuity problem
A long story needs more than a recent frame.
Video generators can make a convincing clip. The harder task begins when a film leaves a room, changes viewpoint, then returns to the room or replays an earlier action.
ATAS treats those earlier observations as addressable world evidence instead of a loose stack of prompts.
Events should stay retrievable.
A past action should remain tied to its scene and time, even after other shots intervene.
Generated evidence needs a gate.
A clip can belong in the film without being reliable enough to rewrite the world state.
The core idea
A shared memory for a changing film.
Before each shot, ATAS retrieves the evidence that matters. After each shot, it decides what deserves to become part of the next world state.
See the full pipelineScene memory
01Registered static geometry for revisiting a place from a new camera.
Event memory
02Dynamic observations indexed by scene, event and time.
Validation
03Only trusted observations can update persistent state.
Project film
See the world state in motion.
The official project video shows how scene geometry and event evidence support later shots. Open it with captions on when available.
Open the official videoGo deeper
A research project you can navigate.
Start with the idea, inspect the memory design, compare the results, then follow the primary sources.