Fei-Fei Li's World Labs unveils Atlas, an omni world model turning 1-6 photos into navigable 3D worlds
APPSO · wechat · 2026-09-02
World Labs, founded by Fei-Fei Li, has released Atlas, an omni world model for spatial intelligence. Pretrained from scratch, it natively handles text, images, video, camera poses and 3D depth, unifying world generation, spatial reconstruction and spatiotemporal simulation in one model.
Key capabilities
- Camera-controlled generation: camera pose is a native input; 1–6 ordinary photos yield up to 1-minute, 1440p video with precise camera moves and consistent geometry. Even two unrelated photos can be stitched into a continuous, navigable 3D scene.
- Spatial reconstruction: the Stanford Main Quad case needs only 2–25 tourist-level photos. Sparse-view reconstruction error (AbsRel×10⁻³) hits 25.3, beating Pi3X (28.7), DepthAnything3 (39.3) and MapAnything (47.7); outputs connect to point clouds, 3DGS and World Labs' Marble platform.
- Reframing video: a few phones or action cams capture a scene, then viewpoints can be freely switched in virtual space — bullet-time without camera arrays.
Architecture: a multimodal autoregressive diffusion Transformer combining autoregressive spatial-state prediction, Rectified Flow denoising, and Transformer scaling, enabling KV cache, distributed inference and diffusion distillation reuse. In blind tests of camera control it beats MiniMax H3 75%, Gemini Omni Flash 81%, FLUX3 93% and Seedance2.5 94% of the time.
For robotics: Atlas targets embodied AI's sim-data bottleneck — traditional data collection is estimated to cost on the order of $100 trillion at scale. Two 24-frame phone videos can now yield a high-fidelity 3D physical space with simulated RGB, depth and force feedback. World Labs reports emergent abilities with scaling. Atlas is in Early Access for select partners and will underpin Marble and other products.
Related event: World Labs Launches Atlas, an Omni World Model for Spatial Intelligence(61 posts)→
More from Embodied
- AlphaChip creators' Ricursive claims 1000x faster chip design, recursive self-improvement underway — annadgoldie · 2026-09-03
- Uber launches London's first robotaxi service using Wayve's self-driving tech — Polymarket · 2026-09-03
- Etienne joins Hugging Face as ML Advocate Engineer for Robotics, starting with Asimov — JiliJeanlouis · 2026-09-03
- Leaked photos reveal Waymo's Cybercab rival 'Firefly': two seats, ~100-mile range — reed · 2026-09-03
- Unitree G1 taught to drive an office chair with its feet: seated humanoid locomotion — ChongZzZhang · 2026-09-03
- Splat2Mesh converts 3D Gaussian Splatting to printable meshes, demoed with Mimaki 3D print — janusch_patas · 2026-09-03