Atlas: A World Model for Spatial Intelligence

Atlas is a multimodal autoregressive diffusion transformer pretrained from scratch to operate on text, images, video, and 3D, generating consistent 3D content from a shared spatial context. It supports camera-controlled generation of up to one-minute 1440p videos, spatial reconstruction of scenes from multiple images with explicit 3D output, and text-to-image or panorama creation, outperforming specialized 3D reconstruction models. Scaling improves its performance, and the model will underpin future World Labs products such as Marble.