World Models You Can Walk Through: World Labs' Frameboy Streams Atlas Live at 15 fps
On November 5 World Labs released Frameboy, a research preview that streams its Atlas world model to the browser at up to 15 fps while you steer the camera. Generation stops being 'wait for the result' and becomes 'look around while it builds.'

On November 5, World Labs — the company founded by Fei-Fei Li — released a research preview called Frameboy: it brings the company's world model, Atlas, into the browser, where you can control the camera and walk through an AI-generated 3D world in real time, with frames streamed from cloud GPUs at up to 15 frames per second. This is not another "type a prompt, wait a minute, watch a video" demo. It is a change in interaction: a world model showing up as an interactive renderer for the first time, rather than as an offline generator.
The judgment up front: for the past two months, the loudest news in world models has been about how long a video can be and how high the resolution goes. Frameboy asks a different question: once generation is fast enough to be interactive, how does the way people create change? World Labs' own answer is in its official blog post: when the cost of a bad idea is near zero, there are no bad ideas — what was formerly a costly mistake simply becomes exploration. Whether that claim holds up depends on what 15 fps costs — but the question itself points in the right direction.

1. What Frameboy actually is
First, three names that are easy to confuse. Atlas is the world model World Labs introduced on September 1, 2026 — an autoregressive model that generates frames one at a time, each conditioned on context and previously generated output, with pixel-perfect camera control. Give it a single image, a text caption, or a handful of photos taken from different viewpoints, and it can grow an explorable 3D world from that seed: up to one minute of 1440p video with camera control, reconstruction of real scenes from sparse input images, and output as point clouds and 3D Gaussian splats.
Frameboy is a real-time demo running on Atlas: it serves a variant of Atlas that has been highly optimized for the fastest, most efficient inference possible while maintaining excellent novel view synthesis, streaming frames from cloud GPUs to the browser at up to 15 frames per second. The public demo currently offers only a curated list of scenes — according to hands-on reports, real landmarks including Stanford Quad, Sutro Baths, and the Palace of Fine Arts, plus fictional settings like a cyberpunk street, a spaceship interior, and a frozen fortress — where you can move the camera freely, but cannot yet upload your own photos.
The third name is Viewfinder, a web app that also wraps Atlas: you build up your own scenes frame by frame from your own inputs, and each new frame Atlas generates comes with both color and depth, letting you generate 3D-consistent views from different camera angles as if using a virtual camera. Once a scene is built, one click triggers the Frameboy overlay and turns the static build into a live playground. Viewfinder is currently in closed beta, available through a waitlist.
Access is the friendliest part of this launch: the public Frameboy demo runs directly in the browser — according to hands-on guides, no download, no sign-up, no game engine. Open frameboy.worldlabs.ai and you can walk in. The official blog links that domain directly; after opening it, the demo first connects to a cloud GPU server, and there is a short wait before the first scene appears. That wait is itself the first lesson in understanding Frameboy: every frame you see is computed remotely by a model and streamed back, not a texture sitting on your machine. That structure also sets the experience floor: network speed and server load show up directly in frame rate, so 15 fps is a ceiling, not a guarantee.
2. Why speed matters
The official blog devotes a full section to "why speed matters," and its logic is worth restating faithfully: if Atlas takes tens of seconds or even multiple minutes to generate a video, users are forced to be cautious and methodical in building scenes and planning camera paths, which limits iteration cycles and makes it harder to home in on a desired result. In contrast, if the model can generate at a more interactive rate, users can react and adjust their input quickly and intuitively; when the cost of a bad idea is near zero, there are no bad ideas — what was formerly a costly mistake simply becomes exploration.
Behind that paragraph is a more general claim: slow generative models demand cautious, thoughtful prompting, which carries a high cognitive load, while real-time models allow quick, intuitive exploration. Once latency and cost drop low enough, users no longer even perceive a discrete act of generation — the model simply streams frames back continuously as they navigate. World Labs says that after building early prototypes around real-time variants of Atlas, they realized that interacting with the model this way "results in a completely different experience than triggering frame generation manually."
For people building products, that logic translates into concrete scenes. Game-level concept design used to mean concept art, white-box modeling, and engine testing, with each round measured in days; if a world model can generate a walkable level draft in real time, a level designer can revise the design while walking through it. Film location scouting works the same way: instead of sending people to a site, you walk the generated scene virtually first and lock the camera positions before going on location. Robot simulation, construction progress inspection, and architectural visualization follow the same pattern — first have a world you can walk into in real time, then talk about the rest.
3. From RTFM to Atlas: what is different this time
Frameboy's name and form factor have a history. Last year World Labs released a research preview called RTFM, an efficient and lightweight frame generation model whose main skill was interpolating between a preexisting set of viewpoints. Atlas takes concepts from RTFM and scales them up: a much larger model trained on orders of magnitude more data, with the capability shifting from "interpolate between existing viewpoints" to "generate scenes from scratch" — using as little context as a single input image or text caption.
Each scene in Frameboy was seeded from a single input image: Atlas generated an additional set of viewpoints shaping the rest of the scene, and putting those viewpoints into Frameboy's spatial context makes them interactively explorable. At the same time, Atlas retains and improves on RTFM's capability for novel view synthesis from a collection of images: you can use the demo to navigate around real-world scenes such as famous landmarks — in the blog's own words, explore a famous landmark, scout out virtual viewpoints for a film shoot, inspect progress on a construction site, flesh out a concept for a 3D videogame level, or simulate an environment for a robot to navigate through.
One note on the number itself: 15 fps does not sound high, but it is a threshold figure. Below it, interaction feels stuttery and "real-time" is just a claim; at 15 fps, camera control starts to feel usable, even if it is far from smooth. That World Labs chose to ship Frameboy at this point, rather than waiting for 30 or 60 fps, is itself a statement: they consider the interaction paradigm worth validating before the frame-rate number.
4. What "research preview" actually means
The official blog labels Frameboy a "research preview," and those two words deserve to be taken seriously. They mean three things. First, there is no service commitment: a demo that runs today may slow down, change scenes, or disappear tomorrow as costs, load, or product direction shift. Second, the capability boundary is deliberately narrow — the public demo offers only curated scenes, custom input means joining the Viewfinder queue, and no timetable has been published for wider Atlas availability. Third, the feedback loop is open: World Labs says it will "continue exploring the potential of this technology," which means Frameboy is first a probe for collecting real usage feedback, and only second a product.
For anyone thinking about using it in real work, that label sets the posture directly: treat it as a source of inspiration and a reference for the interaction paradigm, but do not write it into any roadmap that carries a promise to a client or a team. The value of a research preview is in "seeing it early," not in "using it now." That is also part of why World Labs shipped at 15 fps: what they need to validate is not the frame rate, but whether people want to interact with a world model in real time at all. If the answer is yes, frame rate, cost, and access are engineering problems worth pouring resources into; if the answer is no, no frame rate will matter.
One more detail is worth noting: Frameboy's presentation — the official art shows a blue handheld-console-style device with a couple of scene cartridges — hints at how World Labs sees it: more like a "playable concept machine" than a production tool. The MOVE and LOOK sticks and the RESET and STOP buttons in the artwork are telling you this thing is "fun first" and "useful second." For vibe coders, that signal is actually welcoming: it lowers the bar for trying it out. You do not need to read a paper first or apply for API access — you can just open it and feel what interacting with a world model is like.
5. The honest side: what 15 fps costs
The official blog does not dodge Frameboy's weaknesses, and states them plainly: Frameboy currently must run on a powerful GPU server in the cloud and is "not cheap compared to local rendering methods." In other words, this real-time experience is paid for in cloud inference cost — every frame is computed live by a model in the cloud, not rasterized locally. World Labs' position is that "if history is any indication, what starts slow and expensive becomes fast and cheap as both hardware and software evolve" — which is true, but also amounts to admitting that Frameboy is a research preview right now, not a product with a price list.
The limits do not stop at cost. The public demo's scenes are curated by World Labs — you cannot upload your own photos, and building worlds from your own inputs means joining the Viewfinder waitlist while it remains in closed beta. Atlas itself was only opened to select partners for early access when it launched in September, with no published timetable for wider availability. So what an ordinary user can do today is open Frameboy in a browser, walk around the handful of scenes World Labs provides, and get a feel for the interaction itself.
There is also context worth knowing: in late September 2026, AMD announced an all-stock deal to acquire World Labs for about $8.2 billion, expected to close by the end of 2026, with Fei-Fei Li becoming AMD's executive vice president and chief scientist. That acquisition is the biggest off-stage variable around the Frameboy launch — it means Atlas's future compute and commercialization path will likely be tied to AMD's hardware roadmap. For people vibe coding, the direct significance is limited, but it explains why World Labs felt confident putting a real-time demo in public at this point: the acquisition carries expectations of compute and funding behind it.
6. For vibe coders, world models are becoming generative renderers
Why is VibeFix writing about a world model? Because it bears on what vibe coding can do next. Building a 3D game or interactive experience with AI coding today still follows the traditional pipeline: generate assets, import them into an engine, write logic, test. World models point toward a different pipeline: let the model generate a world first, then walk through it and revise as you go — levels, scenes, and camera positions settled inside real-time interaction with the model. The phrase World Labs uses in its blog is "generative renderer": the model stops being merely a content producer and becomes a renderer you can interact with to build and visualize rich, complex 3D worlds.
But the boundaries need stating plainly. Frameboy today is not a tool you can integrate into a product: no API, no programmable interface, only curated scenes in the public demo, and a waitlist for Viewfinder. It is more accurately understood as a validation of the interaction paradigm for world models than as a new tool launch. For vibe coders working on games, 3D experiences, or virtual spaces, the most worthwhile thing to do right now is not to wait for Atlas to open up — it is to try the public Frameboy demo and feel what "camera control plus real-time generation" is like as an interaction. Because once this paradigm matures, it will change how you organize 3D content production: from "model first, render, then test" to "grow a world, walk into it, revise while walking."
There is also a more practical judgment to make. A 15 fps real-time world model still has several gaps before it belongs in a product: frame rate has to climb, cost has to fall, scenes have to accept the user's own inputs, and generated content has to export reliably into assets an engine can use. World Labs' roadmap points at exactly that "user's own inputs" step through Viewfinder — build a scene frame by frame there, then roam it live with Frameboy. Once that path works, world models move from "demo" to "tool." Until then, Frameboy's value is in showing you what the interaction paradigm looks like ahead of time, not in letting you ship with it today.
Three things you can do today: first, open frameboy.worldlabs.ai and walk through each of the provided scenes, paying attention to two things — whether the scene's 3D consistency breaks when you turn the camera, and whether the picture keeps up when you move fast. Second, join the Viewfinder waitlist — it is currently the only way in to building scenes from your own inputs, joining costs nothing, and you can decide how seriously to try it when your turn comes. Third, according to reports, Fei-Fei Li is scheduled to present Atlas at the Conference on Robot Learning (CoRL) on November 11 and share the stage with AMD CEO Lisa Su — what she says there about Atlas's next steps is worth revisiting after the talk, because it will decide whether Frameboy is a small demo from World Labs or a formal piece of Atlas's commercialization path.
In one sentence: Frameboy's significance is not the number 15 fps itself — it is the proof that a world model can run in a browser, in real time, interactively, so that generation stops being "wait for the result" and becomes "look around while it builds." Once that step is proven, the question left for vibe coding is: when a world can be generated and explored on the spot, does your product still need to be modeled in advance?