World Labs unveils Atlas, a multimodal world model that generates and reconstructs 3D scenes

World Labs, the spatial-intelligence startup, has introduced Atlas, an ‘omni’ world model pretrained from scratch to work natively across text, images, video, and 3D. Architecturally it’s a multimodal autoregressive diffusion transformer: every input is folded into a shared spatial context in which each image is anchored to a 3D position, and the model generates what comes next while staying geometrically consistent with everything it has already seen. World Labs frames it as a scaling play, claiming performance keeps improving with more training compute.

Atlas targets three broad capabilities. For generation, it offers pixel-perfect camera control—treating explicit camera geometry as a first-class input rather than vague text prompts—and can produce up to a minute of 1440p video from a handful of reference images, along with text-to-image and 360° panorama output. For reconstruction, it synthesizes novel views and explicit 3D geometry (point clouds and Gaussian splats) from as few as two or three photos, filling unseen regions from learned world knowledge; the company claims it beats models purpose-built for 3D reconstruction while scaling up to a hundred-plus input images for faithful recreation. For simulation, it models space and time together, enabling ‘bullet time’ multiview reframing from as few as three cameras and real-to-sim pipelines for robotics.

The practical hook is control: by letting users stage scenes, place unrelated reference images in 3D space, and hand-design camera paths, World Labs pitches Atlas as putting creators ‘in the director’s chair’ rather than gambling on a generative slot machine. The Gaussian-splat output matches the representation already used in the company’s Marble product, and Atlas is slated to power future versions of Marble and other World Labs offerings—signaling a bet that general-purpose world models, not just image or video generators, are the substrate for creative tools, VFX, gaming, and robotics.