MultiWorld drops multi-agent, multi-view video world model

MultiWorld solves the "identity crisis" in multi-agent video generation, moving from simple scene synthesis to functional, consistent world modeling.

• –Agent Identity Embedding (AIE) uses RoPE to uniquely identify and control multiple agents simultaneously without ambiguity
• –Global State Encoder ensures 3D-aware spatial consistency across variable viewpoints via cross-attention
• –1.5x speedup from parallel view generation makes high-fidelity world modeling more computationally feasible
• –Success on high-motion datasets like It Takes Two demonstrates a new benchmark for generative video coherence