Lyra 2.0
#17 in AI Inference Hardwarenvidia · v2.0 · since 2026-04-15 · 14× · last seen Jun 30, 2026
NVIDIA Lyra 2.0 is an AI framework from the NVIDIA Spatial Intelligence Lab for generating persistent, explorable 3D worlds from a single input image. Built on the Wan-2.1-14B diffusion transformer, it creates camera-controlled walkthrough videos that are subsequently lifted into 3D Gaussian splats and surface meshes via feed-forward reconstruction. It specifically addresses spatial forgetting and temporal drifting in long-horizon video generation, and supports export to physics engines such as NVIDIA Isaac Sim. Model weights are restricted to NVIDIA's Internal Scientific Research and Development Model License (no production use), while the source code is released under Apache 2.0.
Features
| License | Model weights: NVIDIA Internal Scientific Research and Development Model License (no production use, no commercial use, no redistribution). Source code: Apache 2.0. |
| Platform | NVIDIA GPU-accelerated systems (H100, GB200); preferred OS Linux; CUDA frameworks; no managed inference endpoint available – local operation on own hardware required |
| Price | Free (model weights on Hugging Face, source code on GitHub; no commercial inference endpoint) |
| Release Date | April 15, 2026 (paper, model weights, and inference code; 4-step DMD distillation LoRA on April 17, 2026) |
| Memory | Model checkpoint size: 97 GB (Hugging Face repository). Model parameters: 14 billion (based on Wan-2.1-14B). Training: 32 H100 nodes, ~24 billion tokens training consumption. |