NVIDIA showcased Lyra 2.0, developed by its Spatial Intelligence Lab. This framework can start with a single photograph to generate new, actively explorable viewpoints, then reconstruct the visuals into a persistent 3D environment that can be rendered in real time from different perspectives.
Users can control the camera trajectory; even when moving beyond the boundaries of the original photo, the system continues to generate subsequent viewpoints while striving to maintain consistency within the same virtual environment. The generated video sequences can be further reconstructed into explicit 3D Gaussian representations and polygonal meshes, which can then be imported into applications like NVIDIA Isaac Sim for robot training and simulation.
Lyra 2.0 focuses on addressing spatial forgetting and temporal drift in generative systems. Spatial forgetting causes the system to generate different structures when returning to previously visited locations, while temporal drift leads to the continuous accumulation of visual and geometric errors. The research team introduced a spatial memory mechanism guided by 3D geometric information to retrieve relevant historical frames and trained the model on its own imperfect outputs to learn how to correct errors.
In terms of performance, the distilled model compresses the denoising process for video generation from 35 steps to just 4, reducing the required time to approximately one-thirteenth of the original. According to the paper, generating an 80-frame clip takes about 15 seconds using a single NVIDIA GB200 GPU, whereas the full model requires around 194 seconds for the same task.
Current performance is not yet sufficient to generate entire worlds in real time, as the computational load remains heavy and the system is primarily suited for static environments. Dynamic scene generation and occasional visual inconsistencies remain current limitations, meaning its significance for game development lies mainly in demonstrating the potential to assist with environment creation, rather than directly generating a complete, playable game from a single image.
This research was initially published in April 2026, with NVIDIA highlighting it again ahead of SIGGRAPH Asia 2026. The implementation code is now publicly available on GitHub, and model weights have been released on Hugging Face, subject to NVIDIA's research-use license.