NVIDIA’s Lyra builds explorable 3D worlds from a single photo

What it does Project Lyra is NVIDIA Spatial Intelligence Lab’s open research codebase for generative 3D world models. Lyra 1.0 produces static 3D and dynamic 4D scenes from a single image or video via video diffusion model self-distillation. Lyra 2.0 pushes further into explorable, long-horizon worlds that stay 3D-consistent as you move through them.

The interesting bit Rather than classic multi-view reconstruction, the project teaches a video diffusion model to distill itself into spatial representations—essentially making the model hallucinate coherent geometry from minimal input. The goal is not a static mesh, but a persistent world you can actually explore.

Key highlights

• Two versions in one repo: Lyra 1.0 for feed-forward 3D/4D generation, Lyra 2.0 for explorable long-horizon environments.
• Pre-trained weights live on HuggingFace; the code is Apache 2.0, though model licenses vary by version.
• Lyra 2.0’s academic paper is currently “temporarily anonymized,” so the README literally tells you to Google for the citation.
• Backed by NVIDIA’s Spatial Intelligence Lab with an associated ICLR paper.

Caveats

• The top-level README is a sparse index; real technical details, benchmarks, and system requirements are tucked into the Lyra-1/ and Lyra-2/ subdirectories.
• Model licenses are separate from the open-source code, so check each folder before shipping anything commercial.
• Lyra 2.0 has no arXiv or paper link yet.

Verdict A strong candidate if you research generative 3D, world models, or neural rendering. Look elsewhere if you want a polished game engine or plug-and-play consumer tool—this is research code with NVIDIA-scale ambitions.