Atlas can generate new camera views from one or a few images, reconstruct real spaces from a few dozen photos, create bullet-time videos from 3 to 5 camera feeds, and generate images from text, including 360-degree panoramas.

How does Atlas compare with other models?

In camera-control tests, Atlas was preferred over MiniMax H3 75% of the time, over Gemini Omni Flash 81%, and over Seedance 2.5 94%. For 3D reconstruction, it beat open-source models like Depth Anything 3 and MapAnything on benchmarks.

Image GenerationVideo GenerationAI Business & IndustryITmedia AI＋1h agoSep 2, 2026, 16:00 JST2 min read

World Labs unveils spatial AI model Atlas

Send on LINE

Key takeaway

World Labs unveiled Atlas, a spatial intelligence model.
It generates 3D scenes and bullet-time videos from a few images.
Atlas is not public yet, only early access for partners.

3 Key Points

What happened
World Labs released Atlas, a world model trained from scratch to handle text, images, video, and 3D natively. It will power the company's spatial generator Marble.
Why it matters
Atlas can generate new views from one or a few photos, reconstruct real spaces, create bullet-time videos, and simulate robot camera data. The company says it outperforms open-source reconstruction models on DTU, ETH3D, and ScanNet benchmarks.
What to watch
Atlas is in early access for select partners. Developers and companies can sign up via a form, and World Labs plans to integrate Atlas into future versions of its products, including Marble.

Ask the AI about this article →

Summaries like this, in your inbox every morning.

Get news like this every morning — free

Context & Analysis

World Labs, co-founded by Stanford's Fei-Fei Li, aims to build general world models that understand how spaces look, behave, and change. Atlas is the company's new step, integrating multiple data types into a single 'spatial context' where each image and depth map is tied to a 3D position. This allows the model to imagine what lies outside the visible frame, a capability useful for filmmaking, robotics, and spatial reconstruction.

The company reports that Atlas outperformed several models in user preference tests, with the gap widening for more complex camera paths. Its scaling approach also shows new abilities emerging as compute increases, suggesting further gains ahead. With Atlas not yet public and limited to early partners, World Labs is focusing on making it the standard for world models, while also hiring research and engineering staff to accelerate spatial intelligence work.

FAQ

What can Atlas do?: Atlas can generate new camera views from one or a few images, reconstruct real spaces from a few dozen photos, create bullet-time videos from 3 to 5 camera feeds, and generate images from text, including 360-degree panoramas.
How does Atlas compare with other models?: In camera-control tests, Atlas was preferred over MiniMax H3 75% of the time, over Gemini Omni Flash 81%, and over Seedance 2.5 94%. For 3D reconstruction, it beat open-source models like Depth Anything 3 and MapAnything on benchmarks.

ITmedia AI＋Read Original Article

Send on LINE

Also reported by SiliconANGLE AI