World Labs has released a new model called Atlas, which it positions as an "omni-modal" model for spatial intelligence: text, images, video, and 3D data all run through the same architecture, and the output can be an image, a video, a point cloud, or a 3D Gaussian splat. Until now, the company's public moves have centered on fundraising and its Marble platform, with the underlying model itself never fully shown.
Technically, Atlas is a multimodal autoregressive diffusion transformer, with the diffusion component using rectified flow. Its toughest piece of engineering is treating camera pose as a native conditioning input. When generating a video, where the camera moves, how fast, and at what focal length can all be specified at the pixel level — no need to type a prompt like "slow push-in" and hope the model gets it.
One model, four jobs
World Labs lists four groups of capabilities. The first is camera-controllable generation: starting from text or a single image, the model generates video along a camera trajectory the user specifies. The second is sparse reconstruction: feed in anywhere from one to over a hundred photos, and the model outputs a 3D representation of the scene. The third is spatio-temporal simulation — re-framing and re-shooting existing video from a different camera position, which the company specifically calls out for robotics' Real-to-Sim use case. The fourth is more conventional: text-to-image and 360-degree panoramas.
The video output caps out at 1440p and one minute. That's not an aggressive number by today's video-model standards — products dedicated to video generation have already pushed well past it on duration. But duration isn't the metric Atlas is pitching itself on; the comparison it offers is on camera-following tasks, where users preferred Atlas's output 93% of the time over FLUX. On 3D reconstruction, the company says Atlas beats dedicated SOTA reconstruction models, with an average absolute relative error of 25.3.
A generalist model holding its own against specialists in two separate tracks — that's the story Atlas wants to tell. Whether it holds up will depend on third-party results from early testers.
Why camera control is the dividing line
For the past two years, camera movement in video generation models has basically been a matter of semantic guesswork. Write "orbit shot" and the model produces something that roughly rotates, but how many degrees, where the center is, and whether it drifts are all up in the air. People doing previsualization for film typically deal with this by generating a dozen-plus clips and picking one; people doing robotics simulation don't use it at all, because simulation needs reproducible camera extrinsics — "looks like it's moving" has no value there.
Turning pose into a conditioning input effectively downgrades this from a generation problem to a control problem. The model is still inventing the frame's content, but the camera dimension is no longer random. The most direct downstream effect is that the generated output can be aligned with 3D reconstruction: the same scene can yield both a video and a point cloud or Gaussian splat, sharing one coordinate system. World Labs calls this path spatial intelligence — the same term Fei-Fei Li used when the company raised $1 billion in funding for it.
Early testing, and what's next for Marble
Atlas is currently open only to a limited set of partners through an application form — there's no public API and no pricing. The company has stated explicitly that it will power future versions of Marble, World Labs' already-live spatial generation platform, which currently still runs on the previous-generation stack.
The pacing here is deliberate. World Labs has moved slowly on product since it was founded — Marble was the first thing anyone could actually try, and Atlas is skipping a public launch altogether. For a company carrying a high valuation and a large funding round, handing the model to a small number of robotics and film-industry partners first pays off more than opening a demo everyone can try: the former can tell you whether the simulation accuracy is good enough, while the latter would most likely just produce a batch of short clips for social media.
Do a rough estimate of Atlas's real barrier to entry: a one-minute, 1440p video plus a 3D output of the same scene means the compute cost of a single inference run is hard to cover with a consumer-grade subscription — which is probably part of why the pricing field is still blank. The model's form has converged for now; how to sell it is still an open question.
Sources: World Labs official blog, CocoLoop; model architecture, video duration and resolution, camera-following preference rate, and reconstruction error figures have all been checked against officially published figures.