World Labs Debuts Atlas to Unify 3D and Video Generation

World Labs, the spatial intelligence startup founded by Fei-Fei Li, has introduced Atlas. This multimodal autoregressive diffusion transformer is trained from scratch to natively process text, images, video, and 3D data within a shared spatial context. Unlike traditional video generators that rely on imprecise text prompts for camera movement, Atlas accepts exact camera geometry as an input. This allows users to direct camera trajectories with pixel-perfect accuracy, generating up to one minute of 1440p video from as few as seven reference images.

The model spans four task families: camera-controlled generation, spatial reconstruction, space-time simulation, and image generation. In human evaluation trials for camera-controlled generation, raters preferred Atlas over MiniMax H3 in 75% of cases, Gemini Omni Flash in 81%, Happy Horse 1.1 in 86%, FLUX 3 in 93%, and Seedance 2.5 in 94%. On sparse-view 3D reconstruction benchmarks, Atlas achieved a mean absolute-relative pointmap error of 25.3, outperforming the next-best open-source specialist model's score of 28.7. It showed notable margins on the DTU benchmark (8.6 versus 11.1) and the T&T benchmark (12.4 versus 15.7) against baselines like VGGT, Depth Anything 3, MapAnything, and Pi3.

For practitioners, Atlas collapses the boundary between 3D reconstruction and simulation. In robotics, developers can walk through a physical space with a smartphone to generate accurate RGB and depth data for real-to-sim workflows, bypassing expensive LiDAR setups. In visual effects, filmmakers can create complex bullet-time sequences using footage from just three phone cameras. The model outputs explicit 3D geometry, such as point clouds and Gaussian splats, and will soon power future versions of World Labs' Marble platform. Atlas is entering early access in the coming weeks, joining a competitive landscape that includes Google DeepMind's Genie 3 and Tencent's HunyuanWorld.