ziplab team releases open ATAS framework for script-based video...

ziplab team releases open ATAS framework for script-based video generation

The international ziplab team released ATAS (Any Time, Any Space), an open-source MIT-licensed framework for agentic script-based video generation: it compiles scripts into shot prompts with explicit location, camera, and event time, and maintains consistency through scene memory and event memory, outperforming baselines on four video generators.

Compiled by Sergey KostenchukPublished 2026-10-15Updated 2026-10-15

2026-10-15 Coding NVIDIA

🛠 ATAS: agentic script-based video with world memory

The ziplab team (Zhejiang University, Monash, BAIR/UC Berkeley, Adelaide) released ATAS (Any Time, Any Space) — a script-based video generation framework: the code is open under MIT, and the arXiv paper is 2610.20571 from October 14. The script is compiled into shot prompts with location, camera, and time, while the world state is stored in scene memory and event memory.

🌍 The framework targets the main pain point of long agentic video generation — consistency of non-adjacent frames. On four generators (MiniMax-H3, Wan 2.7, HappyHorse 1.0, Wan 3.0), it outperforms baselines in consistency and receives the highest human ratings on all six criteria.

👤 Demo at lhmd.top/atas, code on GitHub. The client is installed via pip with a DASHSCOPE_API_KEY; the geometry service (Pi3, SAM 3) requires an NVIDIA GPU with ~20 GB of memory, while the client itself works without a GPU.

Source 1: https://arxiv.org/abs/2610.20571

Source 2: https://github.com/ziplab/ATAS

Sources