World Labs has unveiled Atlas, a model that generates new camera views, reconstructs 3D scenes and simulates how those scenes change over time. It can start from one or several images, anchor inputs to positions in 3D space and let users specify camera movement geometrically rather than through a text prompt.
Outputs include video up to one minute at 1440p, point clouds and 3D Gaussian splats, a representation that renders a scene from many spatial data points. For robotics, Atlas can rebuild a room and generate the image and depth readings a simulated robot would encounter, allowing developers to vary objects, lighting and positions without recording every case in the physical world.
World Labs says external evaluators preferred Atlas over several specialized generation systems and that adding more input photos reduces how much the model must invent. Those are company-reported comparisons, and the broad set of tasks makes a single benchmark difficult. Atlas is currently available only through an early-access program for selected partners and is expected to power future versions of World Labs’ Marble product.