Fei-Fei Li's World Labs unveils Atlas, billed as the world's first multimodal world model

World Labs, the company founded by AI pioneer Fei-Fei Li, has released Atlas, which it calls the world's first multimodal world model. According to a report by Chinese outlet QbitAI, the model goes beyond generating interactive video clips — its stated mission is to "model the world, move the camera, simulate space and time," producing images and video frames with pixel-level camera control while performing 3D reconstruction.

From a single image, Atlas can generate images and video with pixel-level camera control, at up to one minute of 1440p footage. Given anywhere from one to dozens of input images, it can reconstruct real-world scenes, producing both novel-view image frames and explicit 3D representations that reportedly outperform state-of-the-art models trained specifically for 3D reconstruction. It can also generate images and 360-degree panoramas from text, following complex prompts and rendering text accurately.

For robotics, the model's spatiotemporal simulation is the standout capability: Atlas can model both space and time from an input video, change the viewing angle of existing footage, and support real-to-sim workflows. Feeding it a few photos yields realistic RGB and depth data, allowing robots to train and be tested in many more simulated environments.

Technically, Atlas is an "omni model" built on a newly designed multimodal autoregressive diffusion Transformer architecture. Text, images, video and 3D data are anchored in three-dimensional space to form a spatial context, from which the model generates multimodal output. As a Rectified Flow diffusion model, it can balance generation speed and quality by adjusting the number of denoising steps.

In quantitative evaluations on camera-controlled generation and 3D reconstruction, Atlas beat state-of-the-art video models on camera control — with the advantage growing as camera trajectories become more complex — and outperformed the best open-source 3D reconstruction models on sparse-view reconstruction. The team says Atlas was designed for scaling from the start, with early evidence that its capabilities keep improving as it grows.

Atlas ties together Fei-Fei Li's world-model agenda this year: in June she classified world models into renderers, simulators and planners, arguing the simulator matters most; in July World Labs acquired robot-simulation startup SceniX and published its Real-to-Sim-to-Real system, identifying the lack of cheap, controllable training experience as robotics' biggest bottleneck. Atlas now completes the path from real photos and video to 3D space to robot sensor views. NVIDIA's head of robotics, Jim Fan, publicly called it a big step for real-to-sim in robotics.

Atlas is currently in early access for select partners, with applications open on the World Labs website. The company says Atlas will become the foundation model for future versions of Marble and other World Labs products, hinting at a general world-model platform that can simulate space and time.