H3-World adapts the large video generator MiniMax-H3 to create an interactive world model. It represents character and camera actions as structured text instructions and uses temporal attention routing to link each instruction to a specific time interval in the generated video. This method reuses the model's existing language understanding, requiring only lightweight LoRA fine-tuning on 8,000 gameplay samples.
The approach yields precise control of character behavior and camera motion while maintaining high generation quality. It generalizes to unseen action combinations and visual environments. The key finding is that large video generators already contain the semantic structure needed for control, and that language can serve as an efficient interface to access it.