World Labs Atlas bets on new view prediction

World Labs Atlas turns three iPhone photos into a bullet-time flythrough that once needed hundreds of cameras, the team told a16z.

It reframes world models as new view prediction. Not next token. Not next frame.

Atlas is the second generation model from World Labs, the two-and-a-half-year-old startup led by Fei-Fei Li with Justin Johnson and NeRF creator Ben Mildenhall. Marble, its first model, forced everything through Gaussian splats. Atlas makes the view itself the primitive.

How World Labs Atlas actually works

Every input image carries its 3D camera pose, so the model treats images as spatial context rather than loose references you argue with via text. Think of it like pinning photos to an exact map, then asking for the view from any empty point on that map.

That lets one multimodal backbone output RGB plus depth for any pose you request, doing reconstruction of what was seen and generation of what was not in the same pass.

Why this matters and what is not fixed

For creatives, a 50 to 100x cut in captures means you can restage existing photos or build a Stanford Quad flythrough from ground-level shots without rescanning. For robotics, Atlas attacks the real-to-sim bottleneck that World Labs just bought into with its Scenix acquisition.

Dynamics are still baby. Pretraining saw motion but this checkpoint was post-trained as static, so you get faint waves and moving cars not controllable 4D. Scaling, the team said, is now compute limited.

The bet is that pose control beats prompt control. If new view prediction is as general as next token prediction, control wins.