World Labs introduces Atlas, a world model that takes camera pose and depth as native inputs

Sources: primary, discussion
Summary: The World Labs post describes Atlas as a world model that accepts camera pose and 3D depth maps as architecture-level input types rather than as text descriptions of camera motion, and it states that outputs include point clouds and 3D Gaussian splats. Every performance comparison in the post is World Labs' own, and the post names its baselines and a third-party human-rater method without publishing figures in the page text. World Labs states the model is in early access with selected partners rather than generally available.
Why it matters: Camera geometry and depth enter the model as native input types rather than as prose, which changes what a world model can be driven with, and no independent evaluation of the claims exists yet.
Follow-up: Track whether Atlas reaches general availability and whether any independent evaluation publishes figures.