Original title: H3-World: Turning Language Understanding into World Control
Danze Chen, Zeqing Wang, Ziyue Lin, Xingyi Yang +1 · September 2026
Video generators create clips from text, but turning them into interactive worlds usually requires complex add-ons. Researchers wanted to see if an AI's existing language skills could directly steer what happens on screen. They built H3-World using the 33B MiniMax-H3 video generator. The system takes written instructions for both character actions and camera movements, pairing them directly with matching segments of the video. To keep actions sharp and stop commands from bleeding into the wrong moments, they used a method called temporal attention routing. This technique locks each instruction strictly to its intended time interval. The team found that this setup provides effective, precise control over characters and cameras while maintaining high video quality. Remarkably, it even generalizes well to completely unseen scenarios.
When you give instructions for a recipe, saying 'stir the pot' too early or too late can ruin the dish because timing is critical. In this paper, the challenge is that language instructions for controlling video generation must be timed precisely to match the video's timeline. Here, the system uses a method to tie each instruction to the exact moment it should affect the video, preventing commands from bleeding into wrong parts.
This lets people create interactive videos by describing actions in words, which could simplify game design and simulation. It makes precise control over digital worlds accessible without specialized tools.
The abstract doesn't specify limitations, but it notes that H3-World is trained on 8,000 gameplay samples and generalizes to unseen scenarios, though its performance on other video domains or real-time applications is not stated.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
H3-World: Turning Language Understanding into World Control — Danze Chen, Zeqing Wang, Ziyue Lin et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.