World Labs just released Atlas, a multimodal AI model that generates, reconstructs, and simulates 3D worlds from text, images, video, and 3D inputs combined in a shared spatial context. Unlike specialized models built for single tasks, Atlas handles camera-controlled generation, spatial reconstruction, space-time simulation, and image generation all through the same architecture. The key technical move is treating each input image as grounded at a specific 3D position in space rather than as isolated 2D data—this spatial context lets you do things like place two unrelated images in 3D space and have the model generate a coherent world that smoothly connects them. It can generate up to 1 minute of video at 1440p with pixel-perfect camera control, meaning you specify exact camera positions and angles rather than relying on text descriptions. The reconstruction capabilities are where Atlas shows real muscle. It can faithfully rebuild real-world scenes from as few as two or three input images, outperforming models specialized solely for 3D reconstruction. The model handles the sparse input problem that's plagued computer vision for decades—when parts of a scene aren't visible in your reference images, it uses its world knowledge to fill gaps plausibly. You can feed it a hundred images if you want exact fidelity, or just a couple if you're okay with educated guesses about hidden areas. Atlas outputs both 2D frames and explicit 3D data (point clouds, depth maps, 3D Gaussian splats), which matters for robotics, gaming, and VFX workflows that need actual geometry, not just pretty pictures. On the simulation side, Atlas understands how worlds evolve over time, not just their spatial structure. You can feed it video from multiple cameras and have it reframe shots with "bullet time" effects—freezing moments and viewing them from new angles. This bridges the gap between capturing video in the real world and simulating it for robotics applications. The architecture scales with compute, meaning performance improves as they throw more training resources at it, and they're planning to integrate Atlas into future versions of their Marble product and other World Labs tools.