H3-World: Parameter-Efficient Language-Native Control for Video World Models

Researchers introduced H3-World, an open-source parameter-efficient framework that leverages the MiniMax-H3 video model to achieve precise temporal control over character and camera actions through text instructions. By training only 0.199% of parameters via LoRA across 8,000 gameplay clips, the model enables smooth action transitions over video latent intervals. This development proves that fine-grained, temporally aligned control in video world models can be achieved without full retraining or massive computational resources. It makes fine-tuned interactive video generation, game physics simulation, and virtual environment modeling far more accessible to open-source developers. The framework injects textual action commands directly through MiniMax-H3's pretrained text pathway, assigning distinct action prompts to each video latent interval. Trained with 10,000 LoRA steps, H3-World demonstrates strong generalizability to unseen action combinations and visual scenarios.