What is the Atlas world model? World Labs’ omni world model explained

Atlas is the world model that World Labs announced on September 1, 2026. The company describes it as “the world’s first multimodal world model that generates image and video frames with pixel-perfect camera control and reconstructs them in 3D,” and sums up the pitch as “model the world, move the camera, and simulate space and time” (World Labs on X).

This page is an independent explainer. It is not affiliated with World Labs. Every specification below links to its source and carries a date, because the details of a brand-new model change quickly and some of them have not been independently verified.

What Atlas is (and what it is not)

The word “Atlas” is crowded. Boston Dynamics builds a humanoid robot called Atlas, OpenAI ships a browser called ChatGPT Atlas, and dozens of software products carry the name. The Atlas discussed here is none of those. It is a world model: a neural network trained to predict how a scene looks from any viewpoint and how it evolves over time, so that it can generate new views, reconstruct real places, and serve as a simulator.

World Labs calls Atlas an “omni world model for spatial intelligence” (AI Weekly). The “omni” part matters. Earlier generations of tools split the problem into specialised models: one for text-to-video, one for photogrammetry, one for novel-view synthesis, one for 3D reconstruction. Atlas is presented as a single model that takes text, images, video and 3D data as inputs, and produces both video frames and explicit 3D geometry as outputs.

It is also not a product you can sign up for today. As of September 3, 2026, Atlas is in early access for select partners, and World Labs has not published a paper, a price, or a general-availability date. The product you can use right now is Marble, which we compare in Atlas vs Marble.

Architecture: a multimodal autoregressive diffusion transformer

World Labs describes Atlas as “a multimodal autoregressive diffusion transformer” that was “pretrained from scratch” (The Decoder). That phrase packs three design decisions into one line.

Transformer. The backbone is a transformer, the same family of architecture behind large language models and most modern video generators. Transformers scale well with data and compute and handle mixed sequences of tokens naturally, which is exactly what a multimodal model needs.

Diffusion. Frames are produced by a diffusion process, in which the model starts from noise and iteratively denoises toward an image conditioned on the inputs. Diffusion is the dominant approach for high-fidelity image and video synthesis because it produces sharp, detailed pixels rather than blurry averages.

Autoregressive. Instead of generating a whole clip in one shot, Atlas generates sequentially, conditioning each new step on what came before. Autoregression is what lets a model keep a scene consistent as the camera moves through it for up to a minute, and it is what makes the same model usable for reconstruction: given real frames as context, it predicts the missing views.

The word “multimodal” refers to the inputs. According to World Labs, Atlas natively handles text, images, video and 3D data such as depth maps and point clouds (The Decoder). Crucially for the 3D use cases, camera geometry is “treated as a native input rather than something coaxed out of a text prompt” (SiliconANGLE). In other words, you hand the model a camera path, not a sentence that describes one.

World Labs has not released a technical report, so parameter count, training data, token layout and inference cost are unknown as of September 3, 2026. Treat any figure you see for those as speculation.

What Atlas generates

World Labs highlights four capabilities. The first two are generation tasks, the third is reconstruction, and the fourth is an application built on the other three.

Video generation with pixel-perfect camera control

Atlas generates image and video frames at resolutions up to 1440p and lengths up to one minute, with the camera path supplied as an explicit input (Superpower Daily). The distinction from prompt-driven video models is the control surface: with Sora-style models you write “the camera slowly orbits the table,” and hope; with Atlas you provide the orbit as geometry and the model renders frames that follow it exactly. That is what World Labs means by “pixel-perfect.”

For creators this turns a world model into something closer to a virtual camera. Because the same world can be rendered from any path, a single generated scene can yield multiple shots that agree with each other, which is a long-standing weakness of clip-by-clip video generators.

Images from a single reference

Atlas “excels in creating expansive and highly-detailed simulated 3D environments from a single image input,” viewable from any angle (Dealroom). Give it one photo and it extrapolates the rest of the space. This is the same task Marble users know from image-to-world generation, and the image to 3D world tool on this site is built around it.

3D reconstruction from a few photos

The reconstruction side is where Atlas departs most from video models. It accepts anywhere from one to more than 100 reference images and, according to World Labs, “with just two or three images, Atlas delivers faithful results and outperforms specialized 3D models” (The Decoder). Classical photogrammetry and Gaussian-splatting pipelines typically need dozens to hundreds of overlapping photos. A model that reconstructs from a handful is a different category of tool.

The outputs are explicit 3D: point clouds and Gaussian splats. World Labs says Atlas “fills gaps to turn point clouds into complete splat scenes that render on device” (Dealroom). Gaussian splats are the format Marble already exports and the format you can open in the free splat viewer on this site.

Real-to-sim for robotics

The fourth capability is real-to-sim: capture a real space, reconstruct it, and use the reconstruction as a simulator. World Labs showed two large environments captured with ordinary phone video, using 24 frames per environment, then simulated different robots navigating them and rendered what the robots’ onboard cameras would see (Spatial Insider). They also varied lighting, backgrounds, object placement and robot behaviour to generate extra training samples. Coverage of the demo notes that the robotics claims “remain unproven” outside World Labs’ own material (XenoSpectrum). We cover the workflow in the robotics real-to-sim guide.

Camera control explained

Camera control is the feature World Labs leads with, so it is worth being precise about it.

A camera in 3D graphics is a position, an orientation and a set of lens parameters. A camera path is a sequence of those over time. Traditional video generators do not have a camera; they have a text encoder, and any motion you get is the model’s interpretation of words like “pan” or “dolly.” The result drifts, and two generations of the same prompt produce different geometry.

Atlas takes the camera path as an input alongside the images or text. Because the model was trained with camera geometry as a native modality, it can render frames that correspond to those exact poses. Three things follow:

1. Shot consistency. Multiple renders of the same world from different paths show the same geometry, so you can cut between them.
2. Reconstruction and generation share one mechanism. Reconstruction is just generation where most of the frames are given and the poses are known. That is why one model does both.
3. Simulation. A robot’s planned trajectory is a camera path. Feed it in and you get the frames the robot would see, which is the basis of the real-to-sim demo.

The limit is that camera control is only as good as the poses you supply. For generated worlds that is easy: you author the path. For reconstruction from your own photos the model must estimate poses, and World Labs has not published how accurately it does that.

Benchmarks: what World Labs reported

World Labs evaluated Atlas with human preference studies on two tasks, camera-controlled generation and few-view reconstruction. External raters preferred Atlas over the comparison systems in 75% to 94% of head-to-head comparisons, depending on the task (The Decoder, AlphaSignal).

Those numbers are self-reported by World Labs and have not been independently reproduced. No paper, dataset or full list of baselines has been published as of September 3, 2026. Human preference is also a soft metric: it measures which output raters liked, not geometric accuracy. We unpack what the figures do and do not tell you in Atlas benchmarks explained.

Access, pricing and roadmap

As of September 3, 2026:

• Access. Early access only, through a request form, for “select partners” (Implicator). World Labs has not named the partners.
• Pricing. Not disclosed. Marble’s World API, by contrast, sells credits at $1 for 1,250 credits, with a Marble 1.1 world costing 1,500 credits, roughly $1.20 (Radiance Fields). See the Atlas and Marble pricing page.
• API. None public. The World API launched on January 21, 2026 for Marble (World Labs), and Atlas “builds on the Marble model” (Dealroom), so a future Atlas endpoint on the same platform is plausible but unannounced. Track it on the Atlas World Model API page.
• Roadmap. No general-availability date. World Labs has raised $1.23 billion, including a $1 billion round in February 2026 with Autodesk, AMD, NVIDIA and Fidelity participating (Dealroom), which is enough to fund a long private beta.

If you want to be told when any of this changes, the waitlist at the bottom of this page sends one email per material update. Our step-by-step notes on the request form are in how to get Atlas early access.

Atlas vs Marble in brief

World Labs now has two world models, and the distinction confuses people.

Marble launched on November 12, 2025 as the company’s first commercial product (TechCrunch). It generates navigable 3D worlds from text, images, panoramas and video, has a public API, an editor, and exports Gaussian splats, meshes and video. Marble 1.1 and Marble 1.1 Plus shipped on April 2, 2026 (Radiance Fields).

Atlas is the newer, broader model. It adds camera-controlled video up to 1440p, few-view reconstruction with explicit 3D output, and simulation, but is not publicly available.

The practical rule: use Marble for anything you need to ship this quarter, and follow Atlas for what comes next. The full breakdown, dimension by dimension, is in Atlas vs Marble. For how Atlas stacks up against other labs’ models, see the comparison hub, starting with Atlas vs Genie 3.

Where Atlas sits among world models

“World model” has become a label for several quite different products, so it helps to place Atlas on the map.

Interactive world models such as Google DeepMind’s Genie 3 and Odyssey-2 render a playable environment in real time. Genie 3 runs at 720p and 24 frames per second and keeps a scene consistent for several minutes while you move with a controller (DeepMind); Odyssey-2 Pro streams 720p at 22 frames per second with roughly 50 ms latency (Odyssey). Their strength is interactivity; their limitation is that they output video frames, not geometry you can export.

Video generation models such as OpenAI’s Sora 2 and MiniMax H3 optimise for cinematic clips. Sora 2 produced 15 to 25 second clips at up to 1080p (OpenAI), and MiniMax H3 generates up to 15 seconds at native 2K with stereo audio (MarkTechPost). They have audio and strong aesthetics, but camera motion is prompt-driven and there is no 3D output.

3D world generators such as Marble produce a static, explorable scene as Gaussian splats that you can export to an engine. They are the best fit today for assets, but they do not generate long video or simulate dynamics.

Atlas is World Labs’ attempt to cover the middle ground: offline rather than real-time, longer and higher-resolution than clip generators, and with explicit 3D output like a world generator. Each of those trade-offs is explored in the comparison hub, including Atlas vs Sora 2 and Atlas vs MiniMax H3.

Limitations and open questions

A launch post is a marketing document, so it is worth listing what is not yet known.

• No paper. Architecture details, training data, model size and compute are undisclosed. The phrase “multimodal autoregressive diffusion transformer” is the whole public description.
• No independent evaluation. The 75 to 94 percent preference figures were produced by World Labs. Nobody outside the company has run Atlas on a public benchmark.
• No latency figures. A one-minute 1440p clip is not real-time. World Labs has not said how long generation takes or what hardware it needs.
• Pose estimation is unspecified. Reconstruction from two or three photos requires estimating where those photos were taken. How robust that is on casual phone captures is unknown.
• No audio. Unlike Sora 2 or MiniMax H3, nothing in the announcement mentions sound.
• Dynamic scenes. The demos show mostly static environments with the camera moving. Whether Atlas handles moving people, vehicles or deformable objects as well is unclear.
• Robotics claims are demonstrations, not results. No policy trained in an Atlas simulation has been shown transferring to a real robot in public material.

None of these are unusual for a model at the early-access stage. They are listed here so that the specification table above is read with the right level of confidence.

Glossary

World model. A model that learns an internal representation of an environment well enough to predict future observations, render new viewpoints, or simulate the consequences of actions.

Spatial intelligence. World Labs’ term for the ability to perceive, reason about and generate the 3D world, as opposed to the language-centred intelligence of LLMs.

Autoregressive. Generating a sequence one step at a time, with each step conditioned on the previous ones.

Diffusion. A generative method that learns to reverse a noising process, producing images by iterative denoising.

Gaussian splat. A 3D scene representation made of millions of small, oriented, semi-transparent ellipsoids with colour and opacity. Splats render fast on consumer GPUs and are the export format for Marble and the output format of Atlas. Common file types are .ply and the compressed .spz.

Point cloud. A set of 3D points, usually with colour, that describes surface geometry without connectivity. Atlas produces point clouds and then fills them into splat scenes.

Novel view synthesis. Rendering an image of a scene from a viewpoint that was not captured. Reconstruction and generation are both forms of it.

Real-to-sim. Turning a capture of a real place into a simulation environment, typically to train or test robots before deploying them.

Camera path. The sequence of camera positions and orientations over time. In Atlas it is an explicit input; in prompt-driven video models it is implied by text.

How to prepare today

You do not need Atlas access to start building the pipeline it will slot into.

1. Learn the World API. Marble’s API accepts the same kinds of inputs Atlas does, and returns Gaussian splats, which is the output format Atlas produces. The Marble World API guide walks through generation and export, and the TypeScript quickstart gets you to a first world in an afternoon.
2. Get comfortable with splats. Gaussian splats are becoming the interchange format for generated 3D. Drop any .ply or .spz file into the online splat viewer to see how they behave, and read the game development guide for importing them into Unity and Unreal.
3. Capture with reconstruction in mind. Whether it is Atlas or Marble, results depend on well-lit, overlapping views. The capture guidelines in the robotics guide apply to any space you want to rebuild.
4. Budget in credits. At today’s Marble pricing a world is about $1.20. Atlas pricing is unknown, but planning around per-world costs rather than per-second video costs is the right mental model.

Atlas is the most ambitious world model any lab has shown so far: one network for generation, reconstruction and simulation, with the camera as a first-class input. Whether it lives up to the self-reported numbers will only be clear once independent users get access. Until then, this site tracks every change, and the tools here run on the model you can use today.