World Labs Unveils Atlas: Real-Time 3D Flythrough of NVIDIA HQ from 32 Photos

On September 18, World Labs (co-founded by Fei-Fei Li) unveiled and demonstrated Atlas, its next-generation world model: given only 32 photos of NVIDIA's Voyager headquarters as 3D spatial context, it generates novel views in real time and lets users fly through the scene with pixel-precise camera control. NVIDIA researcher Jim Fan and Fei-Fei Li reshared the demo.

Confirmed

• Atlas is a natively pretrained omni spatial intelligence world model handling text, images, video, and 3D within a unified spatial context (per World Labs' official posts).
• Its architecture is a multimodal autoregressive diffusion Transformer that merges all inputs.
• The model was trained on NVIDIA Blackwell GPUs.
• Demonstrated capabilities include scene reconstruction from 32 photos, novel view generation, world simulation, and real-time free-flight with pixel-precise camera control.
• jonstephens85, who has visited the campus, tested the reconstruction and said it captured distinctive buildings, occlusions, and lighting differences from just 32 training images — "Atlas NAILED IT."

Why it matters

• Atlas shows the potential of the "few photos + world model" approach for 3D scene reconstruction and real-time exploration, pushing generative AI from 2D media toward interactive spatial intelligence.
• Its unified omni spatial-context architecture offers a new foundation-model blueprint for robotics, simulation, and other spatially-aware applications.

2026-09-18 ~ 2026-09-19 · 6 related posts

• [source] World Labs' Atlas turns 32 photos into a real-time flythrough of NVIDIA HQ, trained on Blackwell GPUs — theworldlabs · 2026-09-18
• [source] World Labs launches Atlas, an omni world model that rebuilds scenes in 3D from 32 images — theworldlabs · 2026-09-18
• [source] NVIDIA Campus World Model Rebuilt from Just 32 Photos, Eyewitness Impressed — jonstephens85 · 2026-09-18

3 near-duplicate retellings: drfeifei · DrJimFan · TheTuringPost