World Labs Unveils Atlas Spatial Foundation Model
Atlas combines generation, 3D reconstruction and simulation in one architecture aimed at filmmaking, robotics and virtual worlds.
One architecture for spatial tasks
World Labs introduced Atlas, a foundation model designed to perceive, generate, reconstruct and simulate three-dimensional environments. The company says it trained the system from scratch as a multimodal autoregressive diffusion transformer capable of processing text, images, camera poses and depth maps within a shared spatial context.
Atlas differs from conventional video generators by accepting precise camera geometry rather than relying only on written instructions such as “pan” or “crane.” From one to six reference images, it can generate controlled video lasting up to one minute at 1440p. It can also combine unrelated references placed at specified positions, filling the space between them with a geometrically consistent scene.
For reconstruction, Atlas accepts anything from a single photograph to more than 100 images. With sparse input it invents unseen regions; additional views reduce that uncertainty and produce a closer reconstruction. Outputs can include ordinary frames, depth maps, point clouds and 3D Gaussian splats, making the system relevant to game development, visual effects and robotic simulation rather than video creation alone.
World Labs reports that Atlas beat selected open-source specialist models on sparse-view 3D reconstruction and outperformed recent video systems in human judgments of camera-path adherence. Those comparisons require caution: Atlas receives camera coordinates directly, while competing video models were instructed through text, and the results have not yet been independently reproduced.
The model is entering early access with selected partners and will power future versions of World Labs’ Marble product. No public API pricing, broad release date or downloadable weights were announced.
Why it matters
Atlas treats generation and reconstruction as parts of the same spatial system. That could reduce the number of separate tools needed to move from captured footage to an editable world and then into robot simulation. Its strategic value, however, depends on whether the geometric consistency shown in curated examples survives long trajectories, moving objects and unfamiliar physical environments.