World Labs released Atlas, trained from scratch on text, images, video and 3D in one shared spatial context.

A multimodal autoregressive diffusion transformer taking all four as native input types and generating what comes next while staying 3D-consistent with everything already seen (World Labs). Camera-controlled generation of up to one minute of 1440p video from one to six reference images, with explicit camera geometry as an input type rather than a text instruction. Spatial reconstruction from one to dozens of images outputting explicit 3D and beating specialized reconstruction models. Early access only, and it'll back future versions of Marble.