World Labs launches Atlas, an omni world model that rebuilds scenes in 3D from 32 images

theworldlabs · x · 2026-09-18

World Labs (Fei-Fei Li) has introduced Atlas, a next-generation omni world model for spatial intelligence, pretrained from scratch to natively operate on text, images, video, and 3D. It's a multimodal autoregressive diffusion transformer: all inputs combine into a shared spatial context, from which it generates what comes next — staying 3D-consistent with everything seen and imagining beyond it. Performance reportedly improves with training compute.

Four capability pillars:

Camera-controlled generation: images/video from one or more images with pixel-perfect camera control, up to 1 minute of 1440p video
Spatial reconstruction: rebuilds real scenes from 1–dozens of images, outputting novel-view frames plus explicit 3D, claimed to outperform specialized 3D reconstruction SOTA
Space-time simulation: models space and time from video, enabling dramatic reframing and Real-to-Sim workflows for robotics
Image generation: text-to-image and 360 panoramas with complex prompt following and text rendering

In the demo, 32 input images were enough to train real-time flight through NVIDIA's Voyager headquarters on Blackwell GPUs. Atlas will power future versions of Marble and other products.

Original post →

World Labs launches Atlas, an omni world model that rebuilds scenes in 3D from 32 images

More from Multimodal