NVIDIA Research released Lyra 2.0, a framework that generates persistent, explorable 3D worlds from single images at scale. The system maintains per-frame 3D geometry to prevent temporal inconsistencies and uses self-augmented training to correct drift, allowing users to walk through generated environments, look back, and drop robots into them for real-time simulation. This solves the longstanding problem of AI models 'forgetting' spatial layouts over time, enabling reliable 3D environments for robotics simulation, VR/AR applications, and immersive content creation.