NVIDIA Just Released a Framework That Turns One Image Into a Persistent 3D World You Can Explore

NVIDIA Research has unveiled Lyra 2.0, a framework that transforms single images into persistent, navigable 3D environments. The system allows users to walk through generated worlds, look back at previous areas, and deploy robots for real-time simulation, addressing a critical limitation in current AI models. Traditional AI systems struggle with large-scale 3D environments because they lose geometric consistency over time. Objects shift, blur, or become inconsistent as users move through spaces. Lyra 2.0 tackles this problem by maintaining per-frame 3D geometry to retrieve past frames and establish spatial correspondences, whilst using self-augmented training to correct temporal drifting. The release marks significant progress from the original Lyra, published at ICLR 2026, which used video diffusion model self-distillation for 3D scene reconstruction. Version 2.0 scales this approach to full explorable environments, making it particularly valuable for robotics and autonomous driving applications where spatial consistency is crucial. This announcement coincides with Tencent's release of HY-World 2.0, which tackles similar challenges but focuses on engine-ready mesh output. The convergence of these developments from major labs suggests 3D world generation has become a competitive priority, with different companies pursuing complementary approaches to solving spatial navigation and simulation challenges.