What's this about?
Atlas is a multimodal autoregressive diffusion transformer, pretrained from scratch to operate natively on text, images, video and 3D, with all inputs combined into a shared spatial context. It does camera-controlled generation (up to a minute of 1440p video with pixel-perfect camera control), spatial reconstruction from one to dozens of images with explicit 3D output, and space-time simulation. For robotics, it generates both the RGB and the depth data a simulated robot's sensors would observe as it moves.
What is a "world model", given the term is used for everything?
World Labs' own answer, from the thread, is the most useful three-way split anyone has offered: renderers output pixels or views of a world; simulators predict how the world responds to actions; planners predict what actions an agent should take. Most systems called world models do one of the three. Atlas claims all three from one pretrained model, which is the actual claim being made.
What's the community saying?
- The most enthusiastic technical comment: this looks like by far the best model yet for reconstructing 3D spaces from sparse images — you could rebuild your house from a dozen phone photos.
- The limitation spotted immediately: time appears frozen while the camera moves. An author replies that Atlas handles some scene motion — moving cars, waves — and says improving this is on the roadmap.
- A long, useful sub-thread on what "world model" even means, with the author's three-category answer as the best available definition.
- The robotics point several people flag: generating both the world and the robot's sensor view of it from the same model attacks the data-flywheel problem directly.
- A correction from the author worth noting: World Labs is not betting exclusively on explicit 3D — many results are generated frames with no underlying 3D representation, with explicit 3D optional.