In a new paper, Hugging Face researchers present H3-World, an efficient framework that transforms the 33B-parameter MiniMax-H3 video generator into an interactive world model. The core insight is that as large video generators become more capable, natural language is emerging as a powerful interface for control, eliminating the need for dedicated action modules.
MiniMax-H3 already supports zero-shot control of character behavior and camera motion via natural-language instructions. H3-World builds on this by converting coarse language commands into precise, temporally grounded actions. Each action is represented as a structured combination of character and camera instructions, aligned with corresponding temporal video latents.
To achieve temporal precision, the framework introduces temporal attention routing, which restricts each instruction to its intended time interval, reducing control leakage between actions. Importantly, H3-World reuses semantic representations from large-scale video pretraining, requiring only lightweight adaptation.
The results are striking: with just 8,000 gameplay samples, 10,000 LoRA optimization steps, and only 0.199% trainable parameters, H3-World achieves effective character and camera control while preserving generation quality. It also generalizes to unseen scenarios, demonstrating that control capabilities emerging in large video generators can be efficiently turned into interactive world control.