World Labs debuts Atlas, an omni world model simulating space, time, and physical interaction

World Labs just announced Atlas, an omni world model aimed at advancing spatial intelligence by natively handling text, images, video, and 3D geometry within a single architecture. Rather than treating video as isolated 2D pixel grids, Atlas uses an autoregressive diffusion transformer to ground every input in a shared 3D "spatial context." The space-time simulation features demonstrate how this moves AI past basic generation into true physical simulation: Democratized "Bullet Time": By processing footage from just 3 to 5 consumer mobile phones, Atlas can reconstruct dynamic events, freeze time, and simulate fluid camera trajectories through impossible angles without specialized capture rigs. Scalable Real-to-Sim: Beyond scanning static geometry, Atlas simulates dynamic robot navigation and manipulation. As an agent moves, the model synthesizes the exact RGB and metric depth streams its onboard sensors would capture along that path. Interactive Dynamics: Casual real-world video can be turned into simulations that model rigid, articulated, and deformable object physics, allowing researchers to alter object placement, lighting, and camera paths to generate synthetic training data. Native 3D Representations: It directly outputs point clouds and 3D Gaussian splats alongside novel video views, outperforming dedicated 3D reconstruction baselines across standard benchmarks like DTU, ETH3D, and ScanNet.