NVIDIA Lyra 2.0: Single Image to Persistent Explorable 3D World with Physics Engine Export

NVIDIA Research·medium signal

NVIDIA released Lyra 2.0 on April 15, a framework that generates persistent, explorable 3D worlds from a single image with long-horizon 3D consistency. The system solves spatial forgetting via per-frame 3D geometry routing and temporal drift via self-augmented history training. Generated scenes can be exported as 3D Gaussians and meshes to physics engines, with demo exports into NVIDIA Isaac Sim for robot navigation. Open-source on GitHub and HuggingFace. For builders working on 3D content, game prototyping, or sim-to-real robotics, this is the first open model that maintains spatial consistency across extended exploration.