Is Atlas real-time? No — it renders a camera path you hand it in advance

No. And the reason it keeps being described as if it were is worth understanding, because the same confusion decides whether a model can sit inside your loop or only at the end of it. Where World Labs has published nothing — generation time, for one — this page says so rather than filling the gap.

The difference in one line

In Atlas, the camera path is an input. In Genie 3, the camera path is a consequence.

World Labs describes Atlas as taking camera geometry as “a native input rather than something coaxed out of a text prompt” (SiliconANGLE). You hand the model a trajectory — positions and orientations over time — and it renders frames that follow it, up to 1440p and up to one minute (Superpower Daily). The path exists before the first frame does. That is what “pixel-perfect camera control” means: exact, repeatable, and decided in advance.

Google DeepMind describes Genie 3 the other way round: a world that runs in real time at 720p and 24 fps, staying consistent for several minutes, which you can alter while you are inside it by prompting for different weather or new objects (DeepMind). Nobody supplied a trajectory. The trajectory is whatever the person at the controls did, discovered one frame at a time. Runway’s GWM Worlds 2 works on the same principle, as one continuous real-time 720p stream at 24 fps with audio (Runway).

Both are legitimately world models — the category and its three meanings are laid out in what is a world model — but only one of them is something you are in while it runs.

Where the mix-up comes from

Four things push in the same direction, and none of them is anybody being careless.

The category name arrived attached to real-time. Genie 3 was, for many people, the first thing they ever saw called a world model, and it was interactive. The association stuck to the label rather than to the model, so a new world model reads as “another one of those.”

“Simulate space and time” sounds live. World Labs sums up the pitch as “model the world, move the camera, and simulate space & time” (World Labs on X). Every word is accurate about what the model does. None of them says when it does it, and “simulate” carries a strong implication of something running now.

“Camera control” has two readings. In games it means a stick under your thumb. In graphics and VFX it means authoring a camera path in advance. Atlas means the second one. A summary written for a general audience will usually be read as the first.

The robot demo looks like a simulator running. It is discussed below, and it is the most reasonable of the four misreadings.

The pattern here is worth naming: a common summary says Atlas is an interactive world model, while the primary material describes a model that renders a supplied trajectory. Nothing in the announcement contradicts the summary loudly enough to stop it spreading.

The two “minutes” are not the same measurement

This is the most consequential misreading, because it looks like a specification comparison.

• Atlas: up to one minute — the length of the video it produces.
• Genie 3: several minutes — how long a live session stays coherent while somebody is inside it.

One is output. The other is a session. Setting them beside each other in a table implies Genie 3 lasts longer, or that Atlas is a shorter version of the same thing, and neither number supports that. A minute of finished 1440p footage along an exact path and four minutes of playable 720p world are different goods; there is no exchange rate.

The comparable Atlas number would be “how long can Atlas hold a scene consistent while someone steers it,” and that number does not exist, because that is not the mode it operates in.

How long generation takes is not public

World Labs has published no paper, no price, no rate limits, no general-availability date and no latency figure (Implicator). So the honest position on “how long does a minute of 1440p take to render” is: unknown as of September 7, 2026.

That is worth stating plainly, because the gap invites filling. Any specific generation time attributed to Atlas is not coming from World Labs, and the parameter count, training data and inference cost are equally unpublished. We work through the same problem with the model’s evaluation numbers in Atlas benchmarks explained.

Why the robot demo reads as live

This is the strongest reason to think Atlas is real-time, and it is worth being precise rather than dismissive.

World Labs captured two large environments with ordinary phone video — 24 frames per environment — reconstructed them, then simulated different robots navigating those spaces and rendered the view from each robot’s onboard camera, varying lighting, backgrounds, object placement and robot behaviour to produce more training samples (Spatial Insider).

Watching the output, you see a robot moving through a room from the robot’s eyes. That is exactly what a simulator’s output looks like. But what was shown is views generated along trajectories, not a robot perceiving and deciding against a live world. The distinction is the control loop: in a simulator the robot’s next action changes what it sees next, within milliseconds. Nothing in the published material describes Atlas closing that loop. Coverage of the demo also notes that the robotics claims remain unproven outside World Labs’ own material (XenoSpectrum).

None of which makes the demo less useful. Generating large volumes of varied, labelled training views from 24 phone frames is valuable precisely because it is offline: you can run it in bulk, overnight, and keep the results. The real-to-sim guide covers the workflow.

When this bites

It decides whether the model can be in your loop. A player steering a camera, an agent acting on what it just saw, a designer walking through a space and reacting — all of that needs a model producing the next frame in response to the last input. Atlas is not built for that shape of problem, and no amount of faster hardware turns a path-conditioned renderer into an interactive one.

It also decides what you keep. The trade runs the other way too. The real-time models hand back a session; when it ends, nothing persists. Atlas produces point clouds and Gaussian splats you can open, import and reuse. If the deliverable is an asset rather than an experience, offline is not a limitation — it is the reason the assets exist. The full three-way split is in Atlas vs GWM Worlds 2 and Atlas vs Genie 3.

And it changes what “better” means. Comparisons that rank these models on one axis are usually measuring one of them on a task it was not built for.

Checking any model in one question

Ask: does my input change the next frame, or did I describe everything before it started?

If a model needs the whole camera path up front, it is offline, whatever the announcement calls it. If it produces frames in response to input arriving as it runs, it is real-time, and you should then ask the follow-up — whether anything survives the session, or whether closing the window ends it.