H3-World: Turning Language Understanding into World Control
MiniMax-H3 is being adapted into an interactive world model with language as the control surface.
H3-World represents actions as structured character and camera instructions, then aligns them to temporal video latents. The paper adds temporal attention routing so each instruction stays tied to its intended time interval, reducing spillover between actions. The authors report effective control with 8,000 gameplay samples, 10,000 LoRA steps, and 0.199% trainable parameters, while preserving generation quality and generalizing to unseen scenarios. HF Daily Papers' note
H3-World represents actions as structured character and camera instructions, then aligns them to temporal video latents. The paper adds temporal attention routing so each instruction stays tied to its intended time interval, reducing spillover between actions. The authors report effective control with 8,000 gameplay samples, 10,000 LoRA steps, and 0.199% trainable parameters, while preserving generation quality and generalizing to unseen scenarios. HF Daily Papers' note
score 6