World Labs just made its Atlas world model explorable in real time. Frameboy, a research preview released on November 5, streams Atlas-generated scenes from cloud GPUs straight to a web browser at up to 15 frames per second while you steer the camera. Why it matters: world models have mostly worked like render farms — ask, wait, watch. Frameboy turns one into something you can walk through.
November 5, 2026
What happened
World Labs released Frameboy on November 5 as a research preview. It runs a speed-optimized variant of the company’s Atlas world model on cloud GPUs and streams the generated frames to a browser, so moving the camera produces new views continuously instead of triggering a fresh generation job.
That closes a real friction loop. World Labs notes that when a manually triggered Atlas video takes tens of seconds or several minutes, users get cautious about experimenting with scenes and camera paths. With Frameboy, you see an unwanted view and correct it immediately, the way you would in a game engine.
What you can try
The public Frameboy demo offers a curated set of scenes — Stanford Quad, Sutro Baths, the Palace of Fine Arts, plus Cyberpunk, Spaceship Interior and Frozen Fortress — and scenes can be seeded from a single image or from multiview captures of real places. World Labs has also integrated the live mode into Viewfinder, its closed-beta web app (waitlist access) where users build a scene frame by frame from their own inputs, getting color and depth for each new Atlas frame, then explore the result in a real-time Frameboy session.
Atlas itself is the omni world model World Labs introduced on September 1, 2026: a multimodal autoregressive diffusion transformer that takes text, images, video, camera poses and 3D depth as native inputs, generates up to a minute of 1440p video with camera control, and can output point clouds and 3D Gaussian splats. Frameboy builds on the lineage of RTFM, the company’s October 2025 research preview that first targeted interactive frame rates. The applications World Labs lists for Atlas include scouting viewpoints for film shoots, inspecting construction-site progress, building 3D game levels and simulating environments for robots.
The catch: cloud GPUs
World Labs is upfront that Frameboy currently needs a powerful cloud GPU server and is not cheap compared with conventional local rendering. Fifteen frames per second is also below what gamers expect from a real engine, and because each generated frame builds on earlier output, visual and geometric errors can accumulate over long camera paths. The release also lands weeks after AMD agreed to acquire World Labs in an $8.2 billion stock deal — which makes Frameboy’s speed and inference cost more than an academic question for its future owner.
Why it matters
Every frontier world model so far has been judged on clips. Frameboy is the first time one of them runs as a live, camera-controlled stream in an ordinary browser — a small frame rate with big implications for games, film pre-visualization, design and robotics simulation. Whether it becomes a product depends on the economics of cloud inference, but the direction is set: world models are moving from video generators toward interactive engines.
Frequently asked questions
What is Frameboy?
A research preview from World Labs, released November 5, 2026, that streams scenes generated by its Atlas world model from cloud GPUs to a browser at up to 15 frames per second while the user controls the camera.
How is that different from Atlas video generation?
Normally Atlas renders a clip on request, which can take tens of seconds to minutes. Frameboy generates continuously as you navigate, so you can change direction mid-stream and react to what you see.
Can anyone use it?
The Frameboy demo with curated scenes is public at frameboy.worldlabs.ai. Viewfinder, the app for building your own scenes, is in closed beta via a waitlist, and Atlas remains in early access with select partners.
What are the limits?
It requires powerful cloud GPUs, costs more than local rendering, peaks at 15 fps, and generated errors can accumulate over long camera paths.
Sources: Unite.AI, AlphaSignal, RuntimeWire