Fei-Fei Li's Worldlabs Unveils Atlas: Single-Image 3D Reconstruction and 1440p Video Generation

智东西 · wechat · 2026-09-02

Worldlabs, founded by Fei-Fei Li, has released Atlas, a new world model utilizing a multimodal autoregressive diffusion Transformer architecture focused on 3D spatial understanding and generation.

Core Capabilities

Camera Control Generation: Supports pixel-level precision to generate up to 1 minute of 1440p video from a single image, capable of extrapolating scenes beyond the original view.
Spatial Reconstruction: Reconstructs 3D scenes from as few as one image to dozens, outputting explicit 3D data like Gaussian Splats, outperforming specialized models in sparse input scenarios.
Spatiotemporal Simulation: Jointly models space and time, enabling "bullet time" multi-angle filming and robot navigation simulation.
Image Generation: Generates images and 360° panoramas from text, following complex instructions and rendering text accurately.

Technical Architecture

Atlas blends the strengths of LLMs and diffusion models, natively supporting text, images, camera poses, and 3D depth maps. By anchoring input images within a spatial context, it achieves high adherence to camera instructions and high-fidelity reconstruction, providing a unified world model foundation for robotics, gaming, and VFX.

Original post →

Fei-Fei Li's Worldlabs Unveils Atlas: Single-Image 3D Reconstruction and 1440p Video Generation

More from Multimodal