World Labs, the AI startup founded by Fei-Fei Li, has released Atlas, which it calls the world's first multimodal world model that can precisely control camera movement for image and video generation while completing 3D reconstruction.
The most intuitive way to think about it is a panoramic drone — except Atlas doesn't need you to actually photograph everything around you. Give it one or a few photos, draw a camera path, and it will fill in what it never saw: what's behind walls, the backs of objects, and the surrounding space. The virtual camera then flies along your route, generating up to one minute of 1440p video.
Typical video models generate pixels based on text or images, and camera moves are usually described in words like “pan left” or “dolly in.” Atlas, by contrast, reads the camera's position and angle directly in 3D space, and it can also generate depth and a complete 3D scene. The more photos you feed it, the less it has to invent on its own.
That's why World Labs calls it a “world model.” It's not just guessing what the next frame should look like — it also needs to know where the camera is, where objects are, and what you'd see from a different angle. World Labs is also using Atlas to turn real spaces into simulated environments for robot training. Atlas is only available to select partners for now.
https://twitter.com/theworldlabs/status/2094839756329041984