World Labs' Frameboy lets you explore AI-generated worlds live in your browser

What's going on

World Labs, the AI startup founded by Stanford researcher Fei-Fei Li, released a demo called Frameboy on November 5, 2026. It looks like a handheld game console in your browser: you pick a "game cartridge," meaning one of several prebuilt worlds, and then move around freely inside it. The twist is that the frames don't come from a traditional 3D engine. They're generated continuously by AI, specifically a heavily optimized version of the company's Atlas world model. A world model is an AI that generates or predicts an environment, such as what a scene looks like from a different angle.

The frames are generated on powerful cloud GPUs and streamed to your browser at up to 15 frames per second. World Labs explicitly calls Frameboy a research preview. Fei-Fei Li kept her comment on X short: "Real time Atlas! Try it now!"

What Frameboy feels like

Chinese tech site ITHome tested the demo. After you click a cartridge, it takes a little over ten seconds to load, and then you can move. You control it by touchscreen or keyboard: the D-pad moves the camera through the space, and the stick turns your view. ITHome says the controls feel a bit choppy. That's no surprise at a maximum of 15 frames per second, when games typically run at 30 to 60.

According to Unite.AI, the curated scenes include Stanford Quad, the Sutro Baths and the Palace of Fine Arts in San Francisco, plus fictional worlds like "Cyberpunk," "Spaceship Interior" and "Frozen Fortress." World Labs says each of these worlds was generated from a single input image, with Atlas filling in the missing viewpoints itself. Real places photographed from multiple angles can be made explorable the same way.

Why real time matters

Until now, Atlas worked on an "order and wait" basis: you set a camera path, start the job, and get a video back after anywhere from tens of seconds to several minutes. World Labs argues that this wait makes users cautious. They plan every move carefully and experiment less. With a continuous stream, you can react instantly to an unwanted angle and change direction. The blog post puts it this way:

"Once latency and cost are low enough, users no longer perceive a discrete act of 'generation' at all – the model simply streams frames continuously as they move through the scene."

The company lists use cases including virtually touring landmarks, scouting camera positions for film shoots, checking construction progress, designing video game levels and simulating environments where robots learn to navigate. Remio magazine cautions that faster feedback doesn't automatically mean better results. What matters is whether the world stays consistent as you move around. If a door changes shape between two viewpoints, exploring could actually lead you astray. And the frame rate is the company's own figure, not an independent measurement.

Under the hood: Atlas and its predecessor RTFM

World Labs unveiled Atlas on September 1, 2026, as an "omni world model" trained from the ground up on text, images, video and 3D. Technically, it's a multimodal autoregressive diffusion transformer. The model first anchors all inputs in a spatial context and then generates frame by frame, each one conditioned on the previous ones. Every frame and every depth map (a kind of grayscale image that encodes each point's distance from the camera) is tied to a precisely defined camera position. According to the company, Atlas can produce up to a minute of 1440p video with camera control, reconstruct real scenes from just a few photos, and generate point clouds and 3D Gaussian splats, a way of representing scenes as lots of small, colored blobs.

Frameboy builds on RTFM, a research preview from October 16, 2025. RTFM ran interactively on a single Nvidia H100 GPU and worked as a "learned renderer": it generated new views without building an explicit 3D model, using previously generated frames as spatial memory. But RTFM was mostly limited to interpolating between existing viewpoints. Atlas is much larger, was trained on orders of magnitude more data, and can create scenes from a single image or a text description.

Viewfinder: building your own worlds

If you want more than the prebuilt scenes, you need Viewfinder, World Labs' web app for building scenes. There, you assemble a scene frame by frame from your own inputs, generate views from different camera angles, and get color and depth for every Atlas frame. You can then explore a scene built this way live in Frameboy mode. However, Viewfinder is only available in a closed beta via a waitlist, and Atlas itself was only available to select partners at launch.

The catch: compute and cost

Every single frame is a new model call on a cloud server, not something your device's graphics card draws. World Labs itself says Frameboy is "not cheap compared to local rendering." The company hasn't disclosed pricing, GPU type, resolution, maximum session length or how many users a server can handle. It's confident, though: what starts out slow and expensive, it says, quickly becomes fast and cheap with better hardware and software.

World Labs had already put a number on the scale of the problem in its RTFM post: an interactive 4K stream at 60 frames per second would require more than 100,000 tokens per second, which it called neither feasible nor economical with today's infrastructure. FourWeekMBA does the math: at up to 900 frames per minute, Frameboy delivers a quarter of that frame rate, at an unknown resolution.

That ties into the hardware angle. In September, AMD agreed to acquire World Labs in an all-stock deal worth about $8.2 billion. The collaboration is set to start with training and inference optimization on AMD GPUs. The deal is expected to close by the end of 2026, pending regulatory approval. World Labs hasn't said which chips the Frameboy demo runs on.

The bigger picture

Frameboy is mainly a showcase: you get to steer an AI world model in real time yourself, with no sign-up and no hardware of your own. The main competition comes from Google DeepMind, which unveiled Genie 4 on October 27 as the first real-time world model with natively generated audio, though without disclosing frame rate or pricing. World Labs at least gives a concrete frame rate and openly admits that running it is expensive.

There are still real caveats: 15 frames per second feels choppy, access to custom worlds is restricted, and it remains to be seen whether scenes stay consistent as you roam freely. The vision of everyone soon creating their own virtual worlds and wandering through them in a browser depends on the cost per generated frame dropping dramatically. Still, for film, game development, architecture and robotics, the shift from "render and wait" to "explore live" is a meaningful signal.