World Labs Introduces Atlas, a Multimodal World Model for Spatial Intelligence

Introducing Atlas: A Next-Generation World Model

World Labs has unveiled Atlas, its latest world model designed for spatial intelligence. Atlas is an omni model, pretrained from scratch to process and generate content across text, images, video, and 3D formats. It functions as a multimodal autoregressive diffusion transformer, integrating all inputs into a unified spatial context to generate subsequent content while maintaining 3D consistency.

Core Capabilities and Features

Atlas offers a range of functionalities. It can perform camera-controlled generation, producing up to one minute of 1440p video from one or more images with precise camera control. For spatial reconstruction, Atlas reconstructs real-world scenes from multiple input images, generating novel views and explicit 3D outputs. The model also handles space-time simulation, reframing videos and enabling Real-to-Sim workflows for robotics. Additionally, Atlas generates images and 360 panoramas from text prompts, supporting complex instructions and various visual styles.

Technical Approach and Scalability

Atlas combines diverse inputs into a shared spatial context, using this context to predict and generate subsequent elements. The model maintains 3D consistency with observed data and extrapolates beyond it. World Labs states that Atlas is built for scalability, with performance improving as training compute increases, a trend expected to continue with further scaling efforts.

Applications and Future Impact

Atlas is intended to power future versions of World Labs' products, including Marble. Its ability to generate new views from reference images with specified camera positions and angles, while extrapolating scene content and geometry, makes it applicable across various scene types, visual styles, and camera motions. The model's native use of precise camera geometry as an input type allows for detailed shot framing and motion control.