Fei-Fei Li's World Labs Demos Frameboy, Streaming AI-Generated 3D Worlds Into a Browser at Up to 15 fps

World Labs, the spatial-intelligence company founded by AI researcher Fei-Fei Li, has unveiled Frameboy, a research preview that turns its world model into a real-time system: instead of generating a fixed video clip and making you wait, it streams an explorable 3D environment into a browser at up to 15 frames per second. Li shared the demo on X with a short pitch — "real-time Atlas! Try it now!"

The project is built on Atlas, the autoregressive world model World Labs released in September, with the company's engineering work focused on squeezing the inference loop down to interactive latency. Atlas supports pixel-level camera control and can generate new viewpoints from a single image or from multiple reference photos of the same real location, using spatial context to keep the scene three-dimensionally consistent. Frameboy takes that capability and makes the camera live: users move freely through the scene, and the model generates each new view based on the camera's position and orientation, the way a game engine renders a level — except here the level is being invented frame by frame.

That is a meaningful break from how AI-generated video works today. Conventional tools require the camera path to be decided up front; the output is a movie, not a place. Frameboy's output is closer to a place, and the company sees the difference as the point — the technology moves Atlas from a model you wait on to one that responds continuously to user input.

World Labs has also wired the feature into Viewfinder, its web application for building scenes frame by frame with Atlas. Once a scene is constructed, users can switch into Frameboy's real-time rendering mode and walk around inside what they just made. Viewfinder remains in closed beta with a waitlist, so most people will encounter Frameboy through the standalone demo first.

The company is candid about the limits. Frameboy is a research preview, not a product, and the 15 fps figure is a peak: demanding scenes will not hold that rate. The heavy inference runs on cloud GPU servers, which World Labs acknowledges costs significantly more than local rendering — a structural constraint on how widely this can be deployed while the economics of real-time generative inference remain what they are. Performance claims beyond the demo have not been independently verified.

World Labs positions the technology for film production, game-scene design, architectural visualization and robot simulation — domains where an explorable generated environment is more useful than a rendered clip. The company has been one of the most visible champions of "spatial intelligence," the argument that the next wave of AI needs to model and act in 3D space rather than only predict pixels or tokens.

The demo also arrives with a new corporate backdrop. In September, AMD agreed to acquire World Labs in an all-stock deal valued at $8.2 billion, a wager on spatial intelligence as a product line. Frameboy is the first notable public artifact since that announcement, and it doubles as evidence for the acquirer's thesis: world models that respond in real time are moving from research demos toward something that feels like an application.