Fei-Fei Li's World Labs releases Atlas, first multimodal world model with pixel-perfect camera control
World Labs released Atlas on September 1, the first multimodal world model that generates camera-controlled images and video while reconstructing 3D scenes at up to 1440p resolution and one-minute length.
Spatial-intelligence startup World Labs, co-founded by AI pioneer Fei-Fei Li, released Atlas on September 1. Atlas generates camera-controlled images and video from one or several input images while reconstructing 3D scenes. The model takes precise camera geometry as a native input type rather than only as a text description of camera movement.
For generation, Atlas can produce up to one minute of continuous video at 1440p resolution from a single reference image with a specified camera trajectory. The model reconstructs 3D scenes from sparse images and outputs point clouds or 3D Gaussian splats. On seven standard 3D reconstruction benchmarks (DTU, ETH3D, KITTI, ScanNet, and others), Atlas posts an average absolute relative error (AbsRel×10⁻³) of 8.6, beating Pi3X at 11.1 and Depth Anything 3 at 11.9.
The space-time simulation capability reconstructs bullet-time multi-angle views from three to five phone or action camera videos, enabling reframing of existing footage. The Real-to-Sim workflow reconstructs simulation environments from phone video and generates the RGB and depth data a robot's sensors would observe. World Labs says Atlas performance scales with training compute. Atlas is currently in an early-access program with no public parameter count, training-data composition, or pricing disclosed.