World Labs Unveils Atlas, an Autoregressive 3D World Model With Precise Camera Control

World Labs has released Atlas, a multimodal autoregressive diffusion transformer built to operate natively across text, images, video, and explicit 3D space. The system generates up to 1-minute videos at 1440p resolution by placing reference inputs into a shared spatial context and mathematically extrapolating the missing geometry. Atlas executes highly accurate 3D scene reconstructions using as few as 2 to 3 standard smartphone photos, bypassing the need for specialized capture hardware or dense multi-view setups. To support immediate integration into existing engineering pipelines, the model natively outputs usable spatial data formats, including explicit point clouds and 3D Gaussian splats that can render at high framerates on-device.