One model where four used to be
World Labs published Atlas on 1 September, describing it as an omni world model pretrained from scratch to operate natively on text, images, video and 3D. The company said the system is a multimodal autoregressive diffusion transformer, and that every input is folded into a single shared spatial context rather than handled by a separate component.
That framing is the point of the release. Reconstructing a scene, generating video along a chosen camera path and simulating what a robot would see have generally been three different systems. World Labs is arguing they are one problem.
What it takes in and what it puts out
Atlas accepts text, images, camera poses and 3D depth maps. Video is treated as a sequence of images, each carrying explicit camera positioning. On the output side the company lists video of up to one minute at up to 1440p, point clouds and 3D Gaussian splats, and 360-degree panoramas generated from a text or image prompt.
Reconstruction is where the numbers are most concrete. World Labs said Atlas produces faithful reconstructions of a scene typically from as few as two or three images, and can also make use of more than a hundred input images when they are available.
Four capabilities in one system
The company groups the model’s work into four functions: camera-controlled generation, which it describes as pixel-perfect; spatial reconstruction from sparse inputs; space-time simulation, used for reframing video and for robotics; and image generation from text with complex prompt-following.
The robotics case is the most practical of the four. Because the model represents space and time together, it can reconstruct a real environment and then render RGB and depth observations from simulated robot viewpoints — the Real-to-Sim step that otherwise requires building a scene by hand.
The comparisons are the company’s own
World Labs said Atlas outperforms state-of-the-art models specialised for 3D reconstruction, and beats recent video models at camera-controlled generation. Neither claim comes with a published leaderboard position, a named competitor or a third-party evaluation, and the weights are not released.
Access is limited. The model is in early access with select partners and prospective users go through a request form; World Labs has not said when it will be generally available, or what it will cost.
Where this sits
World Labs was founded by Fei-Fei Li, the former director of Stanford’s AI Lab, and has raised roughly $1.2 billion from backers including Nvidia, AMD and Autodesk in a round announced in February 2026, according to SiliconANGLE. Atlas follows Marble, the company’s commercial world-model product.
What to watch
Three things will show whether the omni framing holds. Whether outside researchers can reproduce the sparse-reconstruction result on their own scenes, once anyone outside the partner list can run it. Whether robotics teams actually adopt the Real-to-Sim path in place of hand-built simulation. And whether World Labs puts a date and a price on general availability, which so far it has not.
