NVIDIA Lyra

NVIDIA’s Spatial Intelligence Lab advances the state of the art in 3D perception, reconstruction, and generative world modeling. The team publishes open research assets, datasets, and codebases that prioritize reproducibility and measurable progress. Their work bridges graphics and vision to enable practical, evaluation‑ready tools for scientists and developers.

Project Lyra is an open research codebase from NVIDIA for generative 3D world models. Lyra 1.0 introduced feed‑forward 3D and 4D scene reconstruction from a single image or video via video‑diffusion self‑distillation. Lyra 2.0 extends this to explorable worlds with long‑horizon, 3D‑consistent generation, enabling camera motion and scene traversal without per‑scene optimization. The repository ships official implementations for both versions, a GUI for sampling and inspection, and training code to reproduce reported results or adapt models to custom datasets. Source code is released under Apache‑2.0, with model licenses defined per version. Practitioners use Lyra to generate navigable scenes for previsualization, research baselines, and dataset bootstrapping. The focus is scientific clarity: reproducible scripts, documented configuration, and assets to evaluate 3D consistency and temporal stability across trajectories. It provides a consistent baseline for multi‑view coherence studies and controlled comparisons across datasets, training regimes, and trajectory lengths.