World Labs Atlas

Atlas is World Labs' next-generation world model, introduced on September 1, 2026. According to World Labs, Atlas is an omni model that was pretrained from scratch to natively operate on text, images, video and 3D inputs. It is described as a multimodal autoregressive diffusion transformer: inputs are encoded into a context, and outputs are generated conditioned on that context. What distinguishes Atlas from video-only generative models is spatial grounding. Each image is grounded at a 3D position in space, forming what World Labs calls a spatial context, so the model stays consistent in 3D with everything it has seen and can imagine what lies beyond the current view. This supports camera-controlled generation and spatial reconstruction from sparse input images. Atlas natively operates on both 2D image frames and 3D representations, producing outputs such as depth maps and complete splat scenes that render on-device at high resolution and frame rates. This is the same representation used in Marble, which allows Atlas to integrate with the rest of the World Labs product line. The company states that Atlas will power future versions of Marble and other World Labs products. Relevance to embodied AI: World Labs states that world models can help robots plan actions, and that Atlas supports Real-to-Sim workflows for robotics. World Labs reports that Atlas performance improves with increased training compute. Its claim that Atlas outperforms state-of-the-art models specialized for 3D reconstruction is the company's own statement and has not been independently verified.