Any Time, Any Space brings persistent 3D memory to AI video

Any Time, Any Space (ATAS) is a new open-source framework designed to fix the worst problem in long-form AI video: scene drift.

Today, when an AI character walks out of a room and comes back three shots later, the furniture, lighting, and layout almost always morph into something else. ATAS gives models a persistent spatial memory so the room stays identical no matter where the camera goes in between.

How it builds 3D memory

Most video models only pass prompt text or a few recent frames between generation passes. That fails because flat frames cannot provide a stable 3D coordinate frame for revisiting an earlier location.

ATAS compiles screenplays into shot requests tagged with explicit locations, camera coordinates, and event timestamps. Before generating a shot, it pulls 3D geometry from a registered scene memory, alongside time-indexed event records from dynamic observations.

After each clip is generated, ATAS reconstructs the new footage in 3D and registers it back into the world state. Crucially, updates only enter memory if they pass state and review checks, ensuring temporary visual hallucinations do not corrupt the master scene.

Compatible models and open code

ATAS acts as a modular controller rather than a standalone diffusion model, meaning it plugs directly into existing video engines.

In benchmark tests across four video generators, including Alibaba Wan and MiniMax H3, ATAS improved every reported cross-shot consistency metric and delivered the highest aggregate benchmark scores across all tested backbones.

It beat competing agent workflows and 3D-guided baselines on target-image spatial revisit tasks, led three event-recapture categories, and earned the highest human evaluation scores across all six testing criteria.

Both the code and sample video outputs are publicly accessible, with repository documentation explaining how to download the framework and run it on local hardware.

Why it matters

Storytellers wanting to produce coherent AI films have hit a wall with spatial continuity, usually resorting to heavy post-production or hundreds of re-rolls when shots cut back and forth. ATAS proves you do not need a brand new foundation model to achieve spatial coherence; you just need structured 3D world memory wrapped around the model.

For indie animators and developers exploring narrative video generation, this is an immediate project to pull and test locally. The extra 3D reconstruction loop will add compute overhead to your pipeline, but trading raw generation speed for guaranteed room continuity is an easy trade-off.