World Labs Introduces Atlas World Model for 3D Reconstruction

World Labs, the AI startup cofounded by Fei-Fei Li, introduced Atlas, a multimodal world model detailed in a company blog post. The model was pretrained from scratch to operate natively on text, images, video and 3D, combining all inputs into a shared spatial context and generating what comes next while staying consistent with everything it has already seen.

The claim with the most weight behind it concerns reconstruction from very few images. World Labs says Atlas typically produces faithful reconstructions of a real scene from as few as two or three photographs, and that this outperforms results from models trained specifically for 3D reconstruction. It presents that as a step toward solving novel view synthesis from sparse inputs, a problem it describes as decades old in 3D computer vision. Where a scene is not fully visible, the model fills gaps from its own world knowledge, and the company says the more images it receives, the less it imagines. One example assembles Stanford's Main Quad from between two and twenty five ground level photographs, then generates camera paths flying above the campus.

Atlas generates up to one minute of video at 1440p, using camera geometry as a native input rather than text instructions about camera movement. It can output explicit 3D as well as flat frames, producing point clouds and 3D Gaussian splats, the same representation used in the company's Marble product. From a single image it generates new views and estimates their geometry together; from video of a real space it predicts the depth of every frame and combines them.

The company positions the model as a simulator as well as a generator. It says Atlas can turn as few as three ordinary cameras into a bullet time capture setup that freezes and reframes an event, and it points to real to sim workflows for robotics, where footage of a physical space becomes an environment for training and testing.

World Labs says Atlas will power future versions of Marble and its other products. Early access is available to selected enterprises, with no general release date announced.