World Labs Unveils Atlas: Fei-Fei Li’s 3D World Model
Fei-Fei Li's World Labs introduces Atlas, a multimodal 3D world model generating 1440p video and spatial assets with precise camera control from single images.
👁 112 reads
Update September 18, 2026: For products and models that have changed names or been discontinued since publication, I have added their current status and links to the updated guide in parentheses. The rest of the text remains as published.
World Labs, the startup founded by Fei-Fei Li, unveiled Atlas on September 1, 2026 - a model described as the first multimodal world model capable of generating images and video with pixel-accurate camera control alongside full 3D scene reconstruction. You can find a comparison of Atlas with Genie 3, Odyssey-3, and Cosmos in our guide to world models.
Atlas can take a single 2D photo and generate up to one minute of video at 1440p resolution while maintaining rigid geometric consistency, allowing you to view the scene from any angle. Beyond video, the model outputs native 3D assets - point clouds and 3D Gaussian splats - fusing video, text, camera poses, and depth maps into a unified spatial context.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →How Atlas differs from standard AI video generators
Traditional video generators like OpenAI’s Sora (discontinued by OpenAI on April 26, 2026; what’s next) rely on imprecise text prompts to direct camera angles and movement. Atlas works differently: it accepts camera trajectories and geometric data as native inputs, delivering vastly superior perspective control. World Labs describes it as built on a multimodal autoregressive diffusion transformer architecture that is significantly more sophisticated than early video generation systems.
Early user experiments shared on X look promising - developer Ian Curtis demonstrated a scene reconstructed from a single input image using Atlas alongside spark.js and three.js.
The debut of Atlas delivers on the core premise Li established when launching the company in February 2024. She argued that achieving AGI - artificial intelligence that matches or exceeds human cognitive abilities - is impossible without spatial intelligence. That requires native reasoning about 3D objects, physical environments, and the spatial consequences of interactions between them, ensuring AI truly understands physical reality rather than simply processing text descriptions.
Li previously directed the Stanford AI Lab and co-founded the Stanford Institute for Human-Centered AI. World Labs has secured $1.2 billion in funding from prominent backers including Nvidia, AMD, and Autodesk - underscoring the industry’s serious commitment to physically grounded world models.
Whether Atlas truly bridges the gap between simulated environments and real-world physical reasoning remains to be seen as wider real-world testing unfolds.
