World Labs unveils Atlas, a model that generates minute-long videos and reconstructs 3D scenes from photos
Prepared by Sintelligent with the help of SI. Publication time restored from the build log. . Updated: .
SI is learning to understand and build the three-dimensional world, not just flat images, a direct path to cheap robot training in virtual copies of reality.
Atlas is a single multimodal "world model" (an autoregressive diffusion transformer): it generates videos up to a minute long in 1440p with precise camera control, reconstructs real scenes from one to several dozen photos into point clouds and 3D Gaussians, reframes finished videos and creates simulations for training robots. According to the company, the model outperforms specialized systems in camera-controlled generation and in 3D reconstruction. For now, access is limited to select partners.
Sources 1