World Labs Says Its New Atlas Model Builds 3D Worlds From a Few Photos

World Labs said on Sept. 1, 2026, that it has begun early access to Atlas, an AI model the company says generates video and three-dimensional scenes from a small number of photographs. Atlas is not generally available: the announcement says the model "is entering early access with select partners," who apply for it by request.

In its post, World Labs says Atlas takes one or more reference images and generates new views from camera positions and angles the user specifies, "outputting up to 1 minute of video at 1440p." It says the model reconstructs real-world scenes "from one to dozens of input images," and can emit them as point clouds or 3D Gaussian splats, geometry formats that game engines and robotics software render directly.

For robotics, the company says it captured two large environments with cell-phone video, 24 frames each, then used Atlas to simulate robots moving through them and to generate the images their body-mounted cameras would see.

World Labs describes Atlas as "a multimodal autoregressive diffusion transformer" that it pretrained from scratch on text, images, camera poses and depth maps, a design the company says departs from the architectures used by both large language models and video generators.

The performance claims are the company's own. World Labs reports that Atlas outperforms recent video models at camera control in trials where third-party human raters compared outputs, and that it beats "the best specialized open-source reconstruction models" at rebuilding 3D scenes from sparse views in comparisons it ran itself. It calls the reconstruction work "a major step forward toward solving the problem of novel view synthesis from sparse input images." No paper accompanied the announcement, and no outside group has evaluated the model.

World Labs says Atlas will power future versions of Marble, its existing product, and other products from the company.