World Labs / Atlas

Atlas is a world model from World Labs, a company founded in 2024 around spatial intelligence whose first product, Marble, generates 3D worlds from text, images, video or panoramas. Introduced on 1 September 2026, Atlas is a multimodal autoregressive diffusion transformer pretrained from scratch on text, images, video and 3D; every input is placed at a 3D position in a shared spatial context, and the model generates what comes next consistently with everything it has seen.

Its tasks are generation, reconstruction and simulation. It generates images and video with explicit camera control, up to one minute at 1440p from a few reference images. It reconstructs real scenes from one to over a hundred photographs and outputs novel views, point clouds and 3D Gaussian splats; the page's examples include Stanford's Main Quad rebuilt from two to twenty-five ground-level photos and flown over from the air. From three to five ordinary phone cameras it can freeze and reframe a video; from a phone video of a space it builds a simulation in which robots can navigate or manipulate objects, with the robot's camera views generated by the same model. Its relevance to models of places is the reconstruction and simulation side; it is not a geospatial or earth-observation model.

Atlas is in early access with selected partners; access is by request on the page. The splat representation it outputs is the one Marble uses, so it integrates with World Labs' existing product.

Status: early access. On 28 September 2026 World Labs announced a definitive agreement to join AMD, expected to close by the end of 2026 subject to regulatory approval, with Fei-Fei Li joining AMD as Executive Vice President and Chief Scientist.