NVIDIA's BirdDreams Driving World Model Jointly Generates Perspective Video and BEV Layout

NVIDIA posted BirdDreams on Nov 13: an autoregressive driving world model that uses joint attention between two video diffusion transformers to generate matched first-person and bird's-eye-view video from a single camera image, no HD map or vehicle-state input required. On PhysicalAI-AV and nuScenes the model follows input ego trajectories more faithfully than prior driving world models and generates more plausible surrounding-vehicle motion than ego-action-free baselines.