NVIDIA's Lyra 2.0 turns a single image into explorable, consistent 3D worlds
NVIDIAAI · x · 2026-11-07
NVIDIA's Spatial Intelligence Lab unveiled Lyra 2.0 (accepted to SIGGRAPH Asia 2026): it generates camera-controlled walkthrough videos from a single image and lifts them to 3D via feed-forward reconstruction. Per-frame geometry routes information to combat spatial forgetting, and self-augmented training teaches the model to correct temporal drift, keeping worlds consistent when revisiting areas. Ships with an interactive GUI for trajectory planning, exports to 3DGS/meshes, and Isaac Sim demos for robot navigation. Paper, weights, inference, GUI and training code are public.
Related event: NVIDIA Unveils Lyra 2.0: Explorable 3D Worlds from a Single Image(4 posts)→
More from Embodied
- DaxAI's rideable Qiji robot horse goes on sale for $43,000 — luisdans · 2026-11-07
- ROS weekly: ROSCon Italy 2026 a hit, ManyForge and Lopsy projects debut — kscottz · 2026-11-07
- Dream-1 update adds SO-101 and Franka support; user runs it zero-shot on a Roomba — scott_e_reed · 2026-11-07
- Eka Robotics heads to CoRL and opens hiring for its technical team — pulkitology · 2026-11-07
- XM4D: certifiable geometric optimization for 4D reconstruction from monocular video — kwangmoo_yi · 2026-11-07
- Cartwheel's Swing Choreo model makes robots perform any pose from a text prompt — andrew_n_carr · 2026-11-07