# Lyra bundles two NVIDIA 3D world models in one repository, and the root README stops at links > Lyra 1.0 turns a single image or video into a 3D or 4D scene, and Lyra 2.0 adds explorable worlds. The code, model cards and licenses live in two subdirectories, not in the front page. **nv-tlabs/lyra** — Project Lyra: Open Generative 3D World Models - Repository: https://github.com/nv-tlabs/lyra - Website: https://research.nvidia.com/labs/sil/projects/lyra2/ - Stars: 2,828 · Forks: 289 - Language: Python - License: Apache-2.0 - Published: 2026-10-12 · Updated: 2026-10-30 · Language: en - Canonical page: https://hysenlabs.com/projects/nv-tlabs-lyra ## One repository holds two separate generations of the model The repository is the official implementation of both Lyra 1.0 and Lyra 2.0, developed by the NVIDIA Spatial Intelligence Lab. Each generation has its own directory, Lyra-1/ and Lyra-2/, its own paper, its own project page and its own checkpoint on Hugging Face. The root README is a table of contents: a table links the arXiv paper, project page, Hugging Face model and code folder for each version. Lyra 1.0 maps to arXiv 2509.19296 and the Hugging Face repository nvidia/Lyra. Lyra 2.0 maps to arXiv 2604.13036 and nvidia/Lyra-2.0. The top level also holds .gitmodules, ATTRIBUTIONS.md, CONTRIBUTING.md and LICENSE. The presence of .gitmodules means at least one dependency is pulled in as a git submodule, so a plain download of the repository as an archive will not be complete for every directory. ## Lyra 1.0 generates 3D and 4D scenes from a single image or video through self-distillation The 2025-09-23 news entry describes Lyra 1.0 as feed-forward 3D and 4D scene generation from a single image or video, done by video diffusion model self-distillation. The paper title in the citation block is Lyra: Generative 3D Scene Reconstruction via Video Diffusion Model Self-Distillation, accepted at the International Conference on Learning Representations in 2026. Feed-forward here means the scene comes out of one pass of the network, with no per-scene optimization loop described. The root README gives nothing beyond that summary: no input resolution, no output format, no runtime and no GPU memory figure. Those answers sit in the paper and in the Lyra-1/ directory. Anyone comparing it with per-scene reconstruction methods has to read the paper to learn what 3D representation the model emits. ## Lyra 2.0 targets long-horizon, 3D-consistent worlds you can explore The 2026-04-15 entry introduces Lyra-2.0 as explorable generative 3D worlds with long-horizon, 3D-consistent generation. The paper is Lyra 2.0: Explorable Generative 3D Worlds, listed in the citation as arXiv preprint arXiv:2604.13036 for 2026. Fifteen authors are named there, led by Tianchang Shen and Sherwin Bahmani, with Sanja Fidler, Huan Ling, Jun Gao and Xuanchi Ren among them. The shift from 1.0 is in the framing: 1.0 produces a scene from one input, while 2.0 aims at worlds a user moves through while the geometry stays consistent. The root README does not say how 2.0 is built, what it takes as input, or whether it reuses 1.0 weights. Treat the two as separate models with separate checkpoints, and check the Lyra-2 folder before assuming code is shared. ## The 2026-07-20 release added a GUI and training code for Lyra 2.0 The newest news entry says the Lyra-2.0 GUI and training code were released on 2026-07-20. It points to Lyra-2/gui/README.md for GUI instructions and Lyra-2/TRAINING.md for training instructions. The last push to the repository carries the same date, 20 July 2026. Two consequences follow for a reader. Inference with released weights and retraining are now both possible for 2.0, and the instructions are split into two documents that the root README only links. Training anything of this kind implies a large GPU budget, but the root README gives no hardware requirement, so read TRAINING.md before planning compute. The repository has no GitHub releases, so the news entries and the Hugging Face model pages are the only record of what changed between versions. ## Apache 2.0 covers the source code, and each model carries its own license The README states that Lyra source code is released under the Apache 2.0 License and tells readers to refer to Lyra-1 and Lyra-2 for their respective model licenses. That sentence is the main trap for anyone planning commercial use. The code and the weights are licensed separately, and the weights may carry different terms from the Apache grant. The repository metadata reports Apache-2.0 for the code, and an ATTRIBUTIONS.md file lists third-party material. The root README does not reproduce the model license text, so the only way to know what you may do with a checkpoint is to open the license in the version's folder and the model card on Hugging Face. A pipeline that ships the weights inside a product should be reviewed against those documents, not against the front-page badge. ## What the front page does not tell you: install steps, hardware and a first run There is no install command in the root README, so this article shows no code block for installation. There is no dependency list, no Python version, no CUDA requirement and no example command that generates a scene. The only code in the file is two BibTeX entries, which are citations and not instructions. The primary language is Python, and the Homepage field points to the Lyra 2.0 project page on research.nvidia.com. A reader who clones the repository and looks for a quick start will find none at the top level. The per-version README files in the subfolders are the next place to look, and the GUI README for 2.0 is named in the news. Expect to read two papers and two folders before running anything. ## Choosing between Lyra 1.0 and 2.0 comes down to the kind of output you need The two versions answer different questions. Lyra 1.0 takes a single image or video and produces a 3D or 4D scene, and its pitch is distilling a video diffusion model into a feed-forward generator. Lyra 2.0 is framed around worlds that can be explored over a long horizon with consistent 3D structure, and it is the version with the GUI and the training instructions. If you only need a scene from a photograph, start with the Lyra 1.0 paper and model card. If you want to move through a generated world or train your own variant, start in Lyra-2. Either way, the project page for the version is the place to see the generated results the video at the top of the README only previews. ## Conclusion Read Lyra if you work on 3D scene generation and want the official code behind the ICLR 2026 paper and the Lyra 2.0 report, with the 2.0 GUI and training code now public. Do not expect a quick start from the front page, because installation, hardware needs and model terms are all in the subfolders. Before anything else, open Lyra-1/ and Lyra-2/ and read the model license of the version you plan to use, since the Apache 2.0 grant covers source code only. ## FAQ ### What is the difference between Lyra 1.0 and Lyra 2.0? Lyra 1.0 does feed-forward 3D and 4D scene generation from a single image or video via video diffusion model self-distillation. Lyra 2.0 produces explorable generative 3D worlds with long-horizon, 3D-consistent generation. ### Is the Lyra code free to use? The source code is released under the Apache 2.0 License. The model weights have their own licenses, which the README says to look up in the Lyra-1 and Lyra-2 directories. ### Can I train Lyra 2.0 myself? Training code for Lyra 2.0 was released on 2026-07-20 along with a GUI. The instructions are in Lyra-2/TRAINING.md and Lyra-2/gui/README.md. ## Sources - [Issues](https://github.com/nv-tlabs/lyra/issues) - [License: Apache-2.0](https://github.com/nv-tlabs/lyra/blob/main/LICENSE) - [nv-tlabs/lyra on GitHub](https://github.com/nv-tlabs/lyra) - [Project website](https://research.nvidia.com/labs/sil/projects/lyra2/) - [README](https://github.com/nv-tlabs/lyra/blob/main/README.md) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/nv-tlabs-lyra