World Labs releases Atlas, a multimodal world model with pixel-perfect camera control and 3D reconstruction

FinanceYF5 · x · 2026-09-02

World Labs has introduced Atlas, its next-generation omni world model pretrained from scratch. It is a multimodal autoregressive diffusion transformer that natively handles text, images, video and 3D, combining all inputs into a shared spatial context and generating what comes next while staying 3D-consistent.

Key capabilities:

Camera-controlled generation: generates images/videos from one or more images with pixel-perfect camera control, up to 1 minute at 1440p; a demo uses just 7 reference images to design a camera trajectory.
Spatial reconstruction: reconstructs real scenes from 1 to dozens of images, outputting novel-view frames and explicit 3D, outperforming state-of-the-art open 3D reconstruction models.
Space-time simulation: models space and time from video, enables bullet-time effects from ordinary cameras, and supports Real-to-Sim robotics workflows — reconstructing spaces from a few photos and rendering realistic RGB and depth along arbitrary robot trajectories.
Image generation: follows complex prompts, renders text, and produces 360 panoramas.

The company says Atlas scales with training compute and will power future versions of Marble and other products.

Original post →

World Labs releases Atlas, a multimodal world model with pixel-perfect camera control and 3D reconstruction

More from Embodied