Back to timeline

World Labs Introduces Atlas for Spatial World Modeling

World Labs introduces Atlas, a multimodal world model with camera-conditioned generation and explicit 3D outputs.

Architecture

What Happened

World Labs introduced Atlas on September 1, 2026, and opened requests for early access. The announcement described a model intended to power future products; it did not announce general availability or downloadable weights.

Why It Matters

Atlas brings scene generation, reconstruction, and simulation into one spatially conditioned model, connecting generative media with the explicit geometry needed by robotics and 3D workflows.

Technical Details

The architecture combines autoregressive sequencing with a rectified-flow diffusion transformer. Text, images, camera poses, and depth maps share a spatial context. Outputs include new views, videos, point clouds, and Gaussian splats. Demonstrations include camera-controlled video up to one minute at 1440p. Unobserved regions are inferred rather than measured; reconstruction and benchmark comparisons remain claims reported by World Labs.