NVIDIA Lyra 2.0 Generates Explorable 3D Worlds from a Single Image

NVIDIA researchers released Lyra 2.0, a framework that turns a single image into an explorable 3D world by generating camera-controlled walkthrough videos and lifting them to 3D via feed-forward reconstruction. The method maintains per-frame 3D geometry for information routing to counter spatial forgetting and uses self-augmented training to correct temporal drift, enabling longer, 3D-consistent trajectories. Paper, model weights, and inference code are publicly available, with a GUI and training code released on July 15, 2026.
Event version 1.
Key insights
Single-image to explorable 3D world
Lyra 2.0 turns a single input image into a 3D world that users can explore along a chosen camera path. It combines video generation with feed-forward 3D reconstruction to produce environments from the generated walkthrough.
Spatial forgetting addressed via geometry routing
The framework maintains per-frame 3D geometry and uses it solely for information routing, retrieving relevant past frames and establishing dense correspondences with target viewpoints. Appearance synthesis relies on the generative prior.
Temporal drifting corrected by self-augmented training
Lyra 2.0 trains with self-augmented histories that expose the model to its own degraded outputs, teaching it to correct drift rather than propagate it. This supports substantially longer and 3D-consistent video trajectories.
Outputs exportable to simulation engines
Generated video can be lifted into 3D Gaussian splats and meshes, which can be exported to physics engines. Examples include exporting scenes into Isaac Sim for physically grounded robot navigation and interaction.
Public release of code and weights
The paper, model weights, and inference code are publicly available. A 4-step DMD distillation LoRA was released on April 17, 2026, and GUI and training code were released on July 15, 2026.