NVIDIA Cosmos 3 for Driving: 11 Synced Camera Views and LiDAR

Loading video

Loading video

Cosmos 3 adds two driving world models. Transfer-Auto (16B, mixture-of-transformers) generates up to 11 synchronized standard and fisheye camera views plus 128-beam LiDAR from a text prompt covering weather, lighting and time of day, plus World Scenario Map controls carrying lane markings, traffic signs, lights and object cuboids. Sim-Auto is its causal distillation: chunked generation from cached history and incoming controls produced by an AV policy or a steering-wheel interface, so later frames reflect updated driving inputs. Output is 480p or 720p at 10-30 fps, with LiDAR returned as point clouds carrying per-point intensity. Both checkpoints shipped on Hugging Face and GitHub on 2026-10-20 under OpenMDW 1.1.