InSpatio-World enables real-time 4D simulation from single videos

InSpatio-World represents a breakthrough in generative world modeling by providing physically grounded, interactive 4D simulations instead of just frame-by-frame video generation.

• –Spatio-Temporal Autoregressive (STAR) architecture ensures long-horizon stability, effectively preventing the structural drift common in previous navigation models.
• –Joint Distribution Matching Distillation (JDMD) maintains 93% visual realism, bridging the fidelity gap between synthetic environments and real-world source footage.
• –Delivers 24 FPS on H-series GPUs and 10 FPS on RTX 4090, bringing high-fidelity 4D world modeling to local developer workstations for the first time.
• –Ranks #1 on the WorldScore-Dynamic leaderboard, outperforming existing real-time interactive methods in camera control and geometric consistency.
• –High utility for robotics and embodied AI, offering a scalable method to generate realistic, navigable training simulators from standard video datasets.