Atlas: one model for several spatial tasks

On 1 September 2026 World Labs introduced Atlas, a model pretrained from scratch on text, images, video and 3D that gives new views, up to a minute of 1440p video and 3D reconstruction from a few photos; access is early, for selected partners.

By the post, Atlas is a multimodal autoregressive diffusion transformer (rectified flow); each image and depth map is tied to a camera pose and a position in space, and together they make a “spatial context”. Claimed: video up to 1 minute at 1440p from one to six images along a set camera path; scene reconstruction from one to a few dozen photos, faithful “with as few as two or three”; output as images, video, point clouds and Gaussian splats, the same representation as in Marble; reframing video from three to five phone cameras; a robotics demonstration (two environments from 24 frames each). The measurements, as published. Camera-path following was judged by third-party human raters: Atlas was chosen in 75 to 94 percent of trials against five video models that were given the path in words. Reconstruction error (AbsRel ×10⁻³, lower is better) on seven benchmarks: Atlas 25.3 on average, five baselines from 28.7 to 47.7; the company says it reproduced every baseline result itself. On one benchmark, Tanks and Temples, one baseline has a lower error (40.2 against 42.4). What the record does not state. No measurement made by anyone but World Labs. A trade outlet the same day writes that the post comes with no paper, no arXiv entry, no model card and no code. Fei-Fei Li’s remark that Atlas “essentially solved” sparse reconstruction is her wording; the post says more cautiously “a major step”. That Atlas will power future versions of Marble is an intention.