World Labs launches Atlas AI model

AI tool generates 3D worlds from photos without special gear

World Labs launched Atlas, a single AI model, on September 2, 2026. It reconstructs real-world scenes from 1 to several dozen photos without specialized capture equipment and outputs video up to 1440p for one minute. The model achieves 75–94% human preference over competing systems in camera-controlled generation and outperforms benchmarks in few-view 3D reconstruction. Atlas produces 3D data formats like point clouds and Gaussian splats, enabling applications in virtual environments.

This could lower costs for 3D modeling in education and gaming by reducing reliance on expensive tools or manual processes. However, Atlas remains in early access for select partners and is not publicly available. Source notes performance metrics are benchmark-specific and reconstruction error is measured under controlled conditions. The model’s real-world impact depends on wider adoption beyond its current partner network.

Source: The Decoder