Fei‑Fei Li Unveils Atlas, Ushering a New Era for World Models

Introduction

World Labs, co‑founded in 2024 by AI pioneer Fei‑Fei Li, Justin Johnson, Christoph Lassner and Ben Mildenhall, announced on September 1 the release of Atlas, a next‑generation world model trained from scratch to process four modalities—text, image, video and 3D—natively.

Capabilities

Camera‑control generation : Users can specify camera position and angle like a director; given 1‑6 reference images, Atlas generates up to one‑minute 1440p video that matches the input view and fills unseen regions.

Spatial reconstruction : The model accepts anywhere from a single photo to over one hundred images; more inputs yield reconstructions closer to the real scene, while with only 2‑3 photos Atlas leverages its world knowledge to “hallucinate” missing parts, useful for rapid concept design.

Spatiotemporal simulation : By understanding both geometry and time, Atlas can recreate “bullet‑time” effects that traditionally require dozens of synchronized professional cameras; only 3‑5 ordinary smartphones shot from different angles are needed to rebuild a full 3D space that can be replayed from any viewpoint and frozen at any moment.

Image and panorama synthesis : Text prompts can drive high‑quality image generation and 360° panorama creation, supporting a range of visual styles from photorealism to illustration.

Core Design and Architecture

Atlas is built on a “multimodal autoregressive diffusion transformer”. The autoregressive component generates tokens sequentially, similar to large language models, while the diffusion component decodes those tokens into pixels. All inputs—text, image, video or 3D—are embedded into a shared 3D spatial context, anchoring each image and depth map to a concrete location in space. Generation proceeds from this unified context, ensuring consistency with previously observed content and allowing the model to infer unseen viewpoints.

Benchmark Results

In human‑preference tests, Atlas outperformed mainstream models: 93 % of participants preferred Atlas over Flux 3, 94 % over Seedance 2.5, and 81 % over Gemini Omni Flash. For 3D reconstruction, Atlas achieved an absolute relative error of 8.6 × 10⁻³, lower than Pi3X (11.1 × 10⁻³) and VGGT‑Ω 1B (9.7 × 10⁻³). Independent blind evaluations reported win rates above 75 % on camera‑control generation tasks, peaking at 94 %, and an average reconstruction error of 8.6 × 10⁻³ across several standard benchmarks, surpassing specialized open‑source 3D‑reconstruction models.

Positioning and Outlook

Co‑founder Justin Johnson categorises world‑model systems into three tiers—renderer, simulator and planner. He places Atlas between renderer and simulator: it can produce high‑quality 3D views but does not yet support physical interaction planning. This candid positioning contrasts with more sensational AGI claims from other companies. The first commercial product from World Labs, Marble, is slated for release in November 2025, while Atlas remains in early‑access mode for a limited set of partners.