NVIDIA brings Lyra 2.0 research to SIGGRAPH Asia 2026

NVIDIA says its SIGGRAPH Asia 2026 research program will include Lyra 2.0.[1] The company says the broader program will cover AI, neural rendering, world models and simulation, and points attendees to information about technical sessions and hands-on labs.[2][1] The event is scheduled for Dec. 1–4 at KLCC in Kuala Lumpur, Malaysia.[2]

NVIDIA describes Lyra 2.0 as a system that can turn a single image into an explorable 3D world by combining video generation with 3D reconstruction.[3] The system generates environments along a camera path chosen by the user.[3] Its project description says it retrieves relevant past frames and establishes dense correspondences with target viewpoints, while self-augmented training teaches the model to correct drift.[4] The paper says these methods enable substantially longer, 3D-consistent video trajectories.[5] The project description says generated videos are reconstructed into 3D Gaussians or meshes.[4]

Users can plan camera trajectories to revisit parts of an environment already explored or to venture into areas that have not been observed.[4] NVIDIA’s project page also describes exporting the resulting scene into Isaac Sim for physically grounded robot navigation and interaction.[4] That makes the work relevant to teams assessing generated environments for simulation workflows, although the cited description does not establish how well those environments perform in a particular deployment. The Decoder reported that Lyra 2.0 supports static scenes only.[6] That reported limitation is important for teams whose intended use depends on moving objects or changing environments.

NVIDIA’s project page says the Lyra 2.0 paper, model weights and inference code became publicly available on April 15, 2026, and that the paper was accepted to SIGGRAPH Asia 2026.[4] The Decoder reported that scenes can span roughly 90 meters and that NVIDIA said Lyra 2.0 outperformed six other methods across nearly all measured criteria, including image quality, style consistency and camera control.[6] It also reported a faster variant that generates videos roughly 13 times faster at comparable quality.[6] These are reported research comparisons, not evidence in this material of performance in a specific professional workflow.