World Labs' Atlas steers video from a single photo

World Labs released Atlas on 1 September, and it is the clearest statement yet of the company's argument that video generation is a spatial problem. Atlas is a single model the company calls an omni world model, pretrained from scratch on text, images, video and 3D rather than built as a video generator with a geometry module attached.

That difference shows up in the input. Ask most video models for a camera move and you describe it in words and hope. Atlas takes camera geometry directly, grounds every input image at a position in space, and generates along the path you specify. With one reference image, or as many as six, it produces up to one minute of video at 1440p from any viewpoint in the scene.

The reconstruction side is the one that reads as a change of category. Atlas rebuilds a 3D scene, with point clouds and Gaussian splats, from as few as two or three images, and can hold anywhere from a handful to more than a hundred posed images in a single consistent world. Where coverage is sparse it invents connective geometry: a hallway between two photographs of rooms that were never next to each other.

What the numbers are, and what they are not

World Labs reports human raters preferring Atlas on camera-path adherence in head-to-head comparisons against every model it benchmarked, at 75% against MiniMax H3, 81% against Gemini Omni Flash, 86% against Happy Horse 1.1, 93% against FLUX 3 and 94% against Seedance 2.5. On reconstruction it reports pointmap errors lower than the reproducible open-source specialists on three established benchmarks.

All of it is the company's own testing. There is no independent leaderboard placement, no published paper and no third-party replication with the announcement, so the margins are a starting claim rather than a settled one.

What you can actually use today

Nothing yet, in the ordinary sense. Atlas runs through an early access request form for partners World Labs has not named, there are no public weights, and the public API documentation lists Marble models with no Atlas entry. Pricing has not been disclosed. The company says Atlas will power future versions of Marble, its browser-based world generation product, which is the route most creators would eventually reach it by.

Atlas also does the work World Labs was founded for. It turns captured footage into simulated environments that robot policies can be trained against, generating the images and depth a robot's sensors would see along a movement path. That work has almost nothing to do with the videos SLOP TV's readers publish, and it is the reason a company backed by roughly $230 million is spending on spatial intelligence.

For a creator, the practical question is whether direct camera control and cheap reconstruction arrive in the tools you already pay for. If Atlas's architecture is right, they will, because the physics of the scene stops being something a model has to improvise one frame at a time.

To ask for early access, use World Labs' request form; there is no sign-up link, no waitlist date and no price published.