Fei-Fei Li's World Labs unveils Atlas, a multimodal world model that generates image/video frames with pixel-perfect camera control and reconstructs them in 3D (World Labs)
World models generate, reconstruct, and simulate any possible world. They understand how worlds appear, behave …
On September 1, 2026, World Labs unveiled Atlas, a multimodal world model with advanced capabilities in image and video generation, 3D reconstruction, and spatial intelligence. Atlas stands out as a multimodal autoregressive diffusion transformer model that can process text, images, videos, and 3D data simultaneously.
The model combines all the inputs into a shared spatial context, generating outputs that stay consistent with the entire context and imaginatively filling in missing parts based on its extensive world knowledge. World models, the core of Atlas, serve to understand how worlds appear, behave, and evolve, thus enabling the creation of imagined worlds, faithful simulations of the real world, and assisting robots in planning actions.
Atlas is particularly noted for its precise camera control, allowing users to generate new views at any camera position and angle. It can extrapolate generated views beyond the input images, projecting them into non-visible areas of the scene.
Moreover, it uses camera geometry as a native input type, which provides more precise control over framing and motion compared to text-based instructions. This capability is demonstrated in the model's ability to generate a complete scene from a single input image, extrapolating unseen parts of the scene based on the input content and world knowledge.
Atlas also excels at reconstructing real-world spaces using sparse input images. It doesn't require special capture equipment or hundreds of dense views, unlike traditional methods. With just two to three input images, Atlas provides accurate reconstructions, often outperforming state-of-the-art 3D reconstruction models.
The model's versatility extends to generating long videos with precise control, where users can design every scene and camera angle, effectively taking the director's chair. Additionally, Atlas can reconstruct real-world spaces from a variable number of input images, with faithful reconstructions emerging when more input images are provided. The more it sees, the less it imagines, striking a balance between creative generation and precise reconstruction.
Written by urgent.news from Techmeme's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
