World Labs unveils Atlas, a single AI model that generates, reconstructs, and simulates 3D worlds from just a few photos

World Labs was founded in 2024 by Fei-Fei Li, who created ImageNet and previously led Google Cloud's AI division. The company's trajectory shows a progression from limited early systems to more capable products: a first system in late 2024 allowed only a few virtual meters of movement, while Marble followed in November 2025. A $1 billion funding round came in February 2026 from Autodesk, Andreessen Horowitz, Nvidia, and AMD, with Bloomberg previously reporting talks at a $5 billion valuation.

Atlas builds on both language model and diffusion techniques, generating output piece by piece like a language model for speedup methods such as KV caching, while using the diffusion principle to gradually filter output from noise. This hybrid approach gives it access to methods that shorten denoising or boost image quality. The company says no single benchmark captures what Atlas can do, which is why it points to separate tests for camera-controlled generation and few-view reconstruction.

The robotics angle extends World Labs' August 2026 real-to-sim-to-real engine, which came from SceniX, a startup the company acquired in July. That engine creates thousands of variants from a single real-world task and trains models entirely in simulation, with five robot platforms running for an hour each without human intervention. Atlas similarly reconstructs rooms and generates sensor data along simulated paths, aiming to produce diverse training data without capturing every situation physically. The company expects Atlas's performance to improve with more training compute as it scales, and it will power future versions of Marble and other products.