Project Lyra is NVIDIA’s open research project on generative 3D world models, exploring how diffusion models can synthesize explorable three-dimensional scenes from video or a single image. The GitHub repository from NVIDIA’s Toronto AI lab has collected around 2,780 stars and is actively updated in 2026. Lyra turns casual captures into 3D Gaussian representations that can be rendered and navigated in real time.
Lyra addresses a core problem in world modeling: video diffusion models generate beautiful footage but no persistent 3D structure. The project distills knowledge from video generators into feed-forward models that predict 3D Gaussian splats directly, producing self-consistent scenes you can move a camera through. The research page showcases results, and the repository provides code and model releases so other researchers can reproduce and extend the work. It sits at the intersection of generative video, neural rendering, and spatial computing.
Lyra’s releases include the Lyra 1.0 work on generating 3D scenes from a single image or video and follow-up research extending quality and controllability, with project pages that let you compare generated worlds interactively. The work comes out of NVIDIA’s Spatial Intelligence Lab in Toronto, a group known for influential generative-graphics research. Outputs in standard 3D Gaussian formats can be dropped into common viewers and engines, which helps the results plug into existing 3D pipelines.
Lyra is a free, open research release from NVIDIA under the Apache-2.0 license. There is no commercial product or paid tier; you supply your own GPU compute.
Researchers in generative 3D, neural rendering, and world models, plus graphics engineers curious about how diffusion models can produce navigable virtual environments.