H3-World: Turning Language Understanding into World Control
H3-World adapts MiniMax-H3 into an interactive controller using language-timed character and camera instructions.
The paper says actions are encoded as structured language instructions and aligned to video latents for specific time intervals. A temporal attention routing method is used to keep each instruction from spilling into other actions. The adaptation is lightweight: 8,000 gameplay samples, 10,000 LoRA steps, and 0.199% trainable parameters. The authors report effective character and camera control, preserved generation quality, and generalization to unseen scenarios. ArXiv · AI/CL/LG's note
The paper says actions are encoded as structured language instructions and aligned to video latents for specific time intervals. A temporal attention routing method is used to keep each instruction from spilling into other actions. The adaptation is lightweight: 8,000 gameplay samples, 10,000 LoRA steps, and 0.199% trainable parameters. The authors report effective character and camera control, preserved generation quality, and generalization to unseen scenarios. ArXiv · AI/CL/LG's note
score 5