ziplab/ATAS is a video generation framework that splits a single script into multiple shots and maintains one shared world state that those shots share. The repository description reads "Any Time, Any Space: A World-State-Centric Framework for Agentic Video Generation." It was made public on October 15, 2026 with its first commit, b86aab0, which added 84,958 lines in that single commit.
The structure is divided into four axes. The World-State Compiler turns the script into shot requests that explicitly specify scenes, cameras, moments, and events. Scene Memory holds the static geometry of all locations in a single shared coordinate system, so that any location can be shown again at any point in time. Event Memory indexes observations that change over time by scene, event, and moment, making it possible to re-shoot past scenes from new angles. Validated Updates prevent the world state from being changed until a generated shot has been reconstructed, registered, and verified.
The runtime targets Python 3.12. The client runs without a GPU, whereas the geometry service requires an NVIDIA GPU and roughly 20GB of free memory, and it must separately download the Pi3X and SAM3 weights. The client side needs ffmpeg and keys such as DASHSCOPE_API_KEY, and re-running the same command continues from where it left off without being charged again for results that were already generated at a cost. You can also run with `--memory none`, with no world state, and the example scripts bookstore, clockshop, and record are included.
That said, because this is the first commit, there is no release history or version label. All the results in the README are images: a human evaluation on a 1–5 scale (versus a multi-shot baseline), spatial revisits and event re-shoots, and runtime charts by memory configuration are presented, but there are no textual figures. It still requires a 20GB GPU environment, an API key, and separate weights all at once. Judging only by the scope released under the MIT license, it is close to a starting point for the experiment of producing video that continues across multiple shots, but it is too early to judge quality by the numbers today.