Lyra 2.0 is a 3D world generation framework from Nvidia: give it one image and it builds a persistent environment, explorable along a camera path and exportable as Gaussian splats or meshes. Everything is a free download, code on GitHub and weights on Hugging Face, with no hosted service. Released in April 2026 by Nvidia's Spatial Intelligence Lab, it followed Lyra 1.0 from September 2025 and quickly caught on among robotics and simulation researchers.
Lyra 2.0 works in two stages: it first generates a walkthrough video along a camera trajectory you define, then lifts that footage into 3D in a single feed-forward pass, with no per-scene optimization.
In practice, you load a photo of an alley into the GUI, draw a path toward the back of the scene, and the model extends the environment chunk by chunk as the camera moves forward. Video synthesis relies on a diffusion model, Wan 2.1-14B, before the sequence gets converted into 3D Gaussian splats and a surface mesh.
Video generators have short memories: once the camera looks away and comes back, they hallucinate fresh content, and tiny errors pile up until the whole scene warps. Lyra 2.0 tackles both problems by keeping per-frame 3D geometry to retrieve past views, and by training on its own degraded outputs so it learns to correct drift rather than spread it.
According to the research paper, the model led GEN3C, VMem and HY-WorldPlay on the DL3DV and Tanks and Temples benchmarks at publication time. Against Google's Genie 3 or World Labs' Marble, Lyra 2.0 bets on explicit 3D, the kind that feeds AI simulation worlds inside real physics engines.
Lyra 2.0 is a free download: Apache 2.0 source code, weights, GUI and training scripts included. There is no online service, so everything runs on your own machine, ideally under Linux, with Nvidia recommending H100 or GB200 cards for its 14-billion-parameter video backbone.
Good news for the impatient: the full model denoises in 35 steps and takes around 194 seconds per 80-frame segment on a GB200, while the distilled variant drops to 4 steps and roughly 15 seconds (your coffee stays warm).
| Version | Release | Main contribution | Weights license |
|---|---|---|---|
| Lyra 1.0 | September 2025 | 3D and 4D scenes in one pass, from an image or a video | NVIDIA Open Model License, commercial use permitted |
| Lyra 2.0 | April 2026 | persistent worlds over long trajectories, 3DGS and mesh export | research license, production excluded |
| Lyra 2.0 distilled | April 2026 | 4 denoising steps instead of 35, about 13x faster | same as Lyra 2.0 |
Yes, entirely. Code, weights, GUI and training scripts all download at no cost from GitHub and Hugging Face. The real expense sits in hardware: without a high-end Nvidia GPU, generation crawls, and no hosted version lets you try the model without installing anything on your own machine.
Not with the Lyra 2.0 weights: the Hugging Face card places them under a research and development license that rules out production and commercial output. The source code itself is Apache 2.0, and the older Lyra 1.0 weights sit under the NVIDIA Open Model License, which does permit commercial use.
Both generate explorable worlds, but in different shapes. Genie 3 streams an interactive video simulation, available only to Google AI Ultra subscribers in the US. Lyra 2.0 outputs open 3D files, Gaussian splats and meshes, ready for game engines and robotics simulators like Isaac Sim.
Verdict: Persistent scenes from one photograph instead of weeks of manual capture: robotics labs, embodied-AI researchers and well-equipped 3D tinkerers get an open playground here, with commercial-friendly weights already sitting on the Lyra 1.0 side.
