World Labs' Atlas fuses text, images, video, and 3D into one spatial model
World Labs is shipping Atlas, an omni model pretrained from scratch to handle text, images, video, and 3D as native inputs in one shared spatial context. It generates up to a minute of 1440p video with pixel-perfect camera control, reconstructs scenes into explicit 3D, and runs real-to-sim workflows for robotics. It will power future versions of Marble. The bet: a single world model that scales with compute, replacing task-specific pipelines.