Dynamic Beating AI News Flash: World Labs, founded by Fei-Fei Li, has released Atlas, which is officially described as the world's first multimodal world model that can precisely control the camera to generate images and videos while achieving 3D reconstruction.
The most intuitive gameplay is somewhat like a panoramic drone, but Atlas does not require actually capturing the surroundings. Give it one to several photos, then draw a camera path, and it can automatically fill in the areas not photographed, such as behind walls, behind objects, and surrounding space, allowing the virtual camera to fly along the specified path. It can generate up to a 1-minute, 1440p video.
Conventional video models mainly continue to generate pixels based on text or images, and camera motion is usually described by words like "pan left" or "zoom in." Atlas directly reads the camera's position and angle in 3D space, can also generate depth and a complete 3D scene. The more input photos it receives, the less it needs to extrapolate on its own.
This is also why World Labs calls it a "world model." It is not only guessing what the next frame will look like but also understanding where the camera is, where objects are, and what should be seen from a different angle. World Labs also uses it to turn real spaces into simulation environments for robot training. Atlas is currently only available to some partners.
