by World LabsUS
Advanced world model operating natively on text, images, video and 3D. Generates video with camera control, reconstructs 3D from sparse images, simulates space-time dynamics.
Atlas released Sep 2026 in early access. Omni world model with camera control and temporal simulation.
Updated 1 Sep 2026Checked 10 Oct0 updates this week
Atlas processes text, images, video and 3D as native formats without separate encoders or post-hoc fusion, enabling efficient cross-modal reasoning.
From one or two reference images of a real scene, Atlas reconstructs full 3D geometry and texture, outperforming specialized reconstruction models.
The model predicts how scenes change over time, enabling robotics simulation, video reframing and physics prediction.
| Spec | Atlas |
|---|---|
| Maximum video generation | Up to 60 secondsR (reported) |
| Video resolution | 1440pR (reported) |
| Sparse reconstruction input | 1-2 reference imagesR (reported) |
| Native modalities | Text, images, video, 3DR (reported) |
R reported by the company
World Labs
Spatial intelligence, 3D worlds, world models, Marble, Atlas
World Labs builds spatial intelligence models. Founded by Fei-Fei Li in 2024, it released Marble (Nov 2025) for creating persistent 3D worlds from images, video or text, and Atlas (Sep 2026), an omni world model. Raised $1B in Feb 2026.
