World Labs' Atlas mixes observed and invented geometry in one measurable 3D output

World Labs, the company co-founded by Fei-Fei Li, calls the goal spatial intelligence: getting a model to represent how objects and environments relate in three dimensions instead of as a flat stream of words or pixels ^[2]. The input is where that starts. Atlas takes text, images, camera positions, depth information, and video handled as sequences of images, and places them in a common spatial context that records where a camera sat when each image was captured. The company says that shared context unlocks a very different kind of control ^[6].

One piece of that control is the camera itself. Where a typical video generator makes you ask for a shot that will "crane upward and orbit around the subject" and hope it reads the filmmaking language, Atlas takes the route directly: a user designs a path through the generated scene, and Atlas produces video along that trajectory ^[7].

The amount of invention is tied to the input. Give Atlas more images and it has to infer less; give it only a few, and some of the resulting world is the model's best guess about what should be there ^[5].

The output type decides where Atlas is safe to use. A point cloud or a splat stores explicit positions, and a downstream tool reads those positions as measurements ^[8]. Some of the positions come from photographs, and some come from inference ^[13]. By the article's account, the finished scene can make the two look equally real ^[12].

For content work that is not a problem. A filmmaker does not care whether the alley behind a building ever existed; the shot only has to work, and a game artist can let the model invent a room joining two photographed spaces as long as it looks coherent ^[10].

A robot moving through the same building is in a different position. It needs the wall to sit where the geometry says it sits ^[11]. In an Atlas scene a fabricated wall, doorway, distance, or surface is part of an output that looks measurable and reads the same as the photographed parts ^[13].

The article calls Atlas's benchmarks impressive but includes no figures ^[14]. In my view content pipelines can adopt the inferred geometry as it stands, while a perception or simulation stack cannot until the output separates measured space from guessed space.