World Labs, the spatial-intelligence company founded by Fei-Fei Li, has unveiled Atlas, a multimodal world model that generates image and video frames with pixel-perfect camera control and reconstructs them in 3D1. The company describes Atlas as an "omni world model" aimed at advancing spatial intelligence by natively handling text, images, video, and 3D geometry within a single architecture2.
Atlas is built on an autoregressive diffusion transformer that grounds every input in a shared 3D "spatial context," a design choice that treats video not as isolated 2D pixel grids but as projections of a coherent three-dimensional scene.
The model's headline capabilities span three areas. First, what World Labs calls democratized "bullet time": by processing footage from just three to five consumer mobile phones, Atlas can reconstruct dynamic events, freeze time, and simulate fluid camera trajectories through angles that would otherwise require specialized capture rigs. Second, scalable real-to-sim transfer: Atlas simulates dynamic robot navigation and manipulation, synthesizing the exact RGB and metric depth streams that an agent's onboard sensors would capture along a given path. Third, interactive dynamics: casual real-world video can be converted into simulations that model rigid, articulated, and deformable object physics, letting researchers alter object placement, lighting, and camera paths to generate synthetic training data.
On the output side, Atlas directly produces point clouds and 3D Gaussian splats alongside novel video views. World Labs says the model outperforms dedicated 3D reconstruction baselines across standard benchmarks including DTU, ETH3D, and ScanNet.
ANALYSIS The architecture's integration of generation, reconstruction, and simulation into a single model collapses a pipeline that previously required separate systems for each task. The three-to-five-phone reconstruction path, if it holds at production quality, removes a significant hardware barrier for teams that currently depend on multi-camera rigs or LiDAR arrays. The real-to-sim capability is particularly relevant for robotics labs, where generating sensor-accurate synthetic rollouts at scale has been a persistent bottleneck.
Atlas's direct output of 3D Gaussian splats positions it at the intersection of generative AI and real-time rendering, two domains that have largely advanced on separate tracks.