Strictly for researchers turning images into 3D scenes; the weights licence allows internal research only, with no commercial use.
Kenning's take
What it does
Lyra 2.0 takes a starting image, a chosen camera path and short text captions, generates a walkthrough video, then rebuilds it as a 3D scene that can be rendered in real time or used in simulation. NVIDIA says it keeps scenes consistent over long paths and when the camera revisits a spot. It is a research release with code on GitHub and weights on Hugging Face under a research-only licence, so it needs NVIDIA data-centre GPUs and cannot be used in production.
- Image plus camera path to explorable 3D scene
- Long camera paths that can revisit places
- Output ready for real-time rendering and robot simulation
- 4-step DMD distillation LoRA for faster inference
- GUI and training code released July 2026
Who it suits. Researchers working on 3D scene generation, simulation and robotics.
Key facts
- Category
- AI models
- Made by
- NVIDIA
- Based in
- United States
- Released
- 15 April 2026
- How to get it
- free download of code from GitHub and weights from Hugging Face; the model licence allows internal research only, not production or commercial output
- Works on
- self-hosted (NVIDIA GPUs)
- Free plan
- Free code (Apache 2.0) and weights (NVIDIA Internal Scientific Research and Development Model License)
FAQ
When was Lyra 2.0 released?
NVIDIA published the paper, weights and inference code on 15 April 2026, and added a GUI and training code on 15 July 2026.
Can Lyra 2.0 be used commercially?
No. The weights use NVIDIA's internal research licence, which bars production use and generating works for sale.
What does Lyra 2.0 need to run?
NVIDIA says it is built for its GPU systems such as H100 and GB200.
Facts from the maker's own pages and the sources listed. Checked 2026-10-11.
Similar tools
ACE-Step 1.5
Open-source music generation model that writes full songs with lyrics and runs on ordinary consumer GPUs.
AlphaGenome Atlas
Google DeepMind database of AI-predicted effects for every possible single-letter change in human DNA.
Amazon Nova 2
Amazon's second family of Nova AI models, covering reasoning, speech conversation and combined text-and-image generation.
Cisco Antares
Cisco's small open-weight models that find where known security vulnerabilities sit in a codebase, running locally.