World Labs' Atlas turns a few photos into minute-long, camera-controlled 3D video

World Labs' Atlas takes one or a handful of reference photos and generates minute-long, 1440p video you can shoot from any camera angle you specify. The company calls this a "world model," an AI system that builds an internal representation of how a 3D scene looks and changes over time so it can render new views and simulate what comes next rather than guess pixels frame by frame.

The distinguishing move is treating 3D geometry as a first-class input instead of a post-processing step. The model takes precise camera position and angle as a native signal, which is why a still photo becomes a navigable scene rather than a stitched flythrough. In the launch post, World Labs describes Atlas as a multimodal autoregressive diffusion transformer that combines text, images, video, and 3D into a shared spatial context. The architecture matters for understanding what the demos can do, but it is not the lead. The lead is the input: hand the model a photo, hand it a camera path, get a minute of video that stays consistent as the viewpoint moves.

The launch bundles four task families into one model. Camera-Controlled Generation extrapolates new views beyond the reference frames. Spatial Reconstruction rebuilds real-world scenes from up to dozens of photos and outputs both novel-view frames and explicit 3D geometry. Space-Time Simulation models how a scene changes over time and supports a Real-to-Sim workflow for robotics, in which a reconstructed environment becomes a training ground where a simulated robot moves through generated RGB and depth observations. Image Generation rounds it out with text-to-image and 360 panoramas, including text rendering and varied visual styles.

Two caveats matter before reading the demos as proof. World Labs' claim that Atlas is "outperforming state-of-the-art models specialized for 3D reconstruction" is a self-report; the captured launch materials do not include third-party benchmarks, so the comparison is the vendor's framing rather than an independent result. A Hacker News discussion raises a temporal-consistency objection: in some of the posted demos, time appears effectively frozen while the camera moves, with the model snapping back to a ground-truth view before advancing. That makes the Space-Time Simulation label read more like a roadmap than a delivered capability, and it complicates the headline-grabbing claim.

There is also the product-status caveat. The launch post says Atlas "will power future versions of Marble," the company's existing 3D tool, which means this is a research release rather than something a paying user can plug in. World Labs has not announced a public API, pricing, or a Marble integration date in the captured material. The "spatial intelligence" framing is the company's own category claim, not a settled scientific term, and the launch reads as a positioning move for a foundation model the team is still scaling.

What is genuinely new here is the input. Earlier 3D-from-image systems produced geometry that downstream tools then had to render or animate. Atlas folds the geometry into the generation loop, so the camera path is part of the prompt and the output stays 3D-consistent as a consequence of how the model was trained rather than as a cleanup pass. World Labs says performance scales with training compute and expects the trend to hold as the team continues to scale. The mechanism, treating 3D as a native input rather than a downstream filter, is the test future "world model" launches will be measured against: does the model actually let you move a camera through a generated scene, or is "3D" a sticker on pixel-only video that looks right only while the camera is on the rail?