ChainThink News, April 16: According to Beating's monitoring, NVIDIA has released the open-source Lyra 2.0 framework, which enables the generation of exploratory 3D worlds from a single image. After users upload a photo, Lyra 2.0 first generates a walkthrough video controlled by camera trajectory, then reconstructs the video into a 3D Gaussian Splatting (Gaussian Splats) and mesh model, directly importable into game engines and simulators for real-time rendering.

The model weights and code are released under the Apache 2.0 license on Hugging Face and GitHub, permitting commercial use. Its core technical breakthrough lies in addressing two degradation issues in long-distance walkthroughs: first, "spatial forgetting," where Lyra 2.0 maintains 3D geometric information per frame, resolving inconsistencies in scene depth when the camera retraces its path; second, "temporal drift," mitigated through self-enhanced training that enables the model to correct errors, preventing cumulative frame-by-frame inaccuracies that distort the scene. The framework is built upon Wan 2.1-14B diffusion Transformer as its underlying architecture, with an output resolution of 832×480.

One of Lyra 2.0’s primary application scenarios is robotic simulation. In NVIDIA’s demonstration, the generated 3D scenes were imported into its proprietary physics simulator, Isaac Sim, enabling robots to perform navigation and interaction tasks. Previously, a major bottleneck in embodied intelligence training was the high cost and limited variety of 3D environment creation—Lyra 2.0 provides a scalable pipeline for batch-generating training environments from photographs. Compared to Lyra 1.0, released in September last year, Lyra 2.0 extends the generation scope to continuous long-range exploration. While Google’s earlier Genie 3 demonstrated similar capabilities, it remained non-open-source. Lyra 2.0 is currently the most comprehensive open-source solution in this domain.