World Labs unveils Atlas, a 1440p video model with pixel-level camera control
- World Labs Atlas generates video up to 1440p for one minute and accepts camera pose as native conditioning, letting users specify camera trajectory, speed, and focal length at the pixel level.
- Atlas uses one multimodal autoregressive diffusion transformer with rectified flow to process text, images, video, and 3D data, producing images, video, point clouds, or 3D Gaussian splats.
- For camera-following tasks, World Labs reports users preferred Atlas over FLUX 93% of the time; it also reports a 25.3 average absolute relative error on 3D reconstruction.
- Atlas can reconstruct a scene from one to more than 100 photos, and can re-frame existing video from a new camera position for robotics Real-to-Sim workflows.
- Access is limited to selected partners through an application form, with no public API or pricing; World Labs says Atlas will power a future version of Marble.