The News That Matters
World Labs, the spatial intelligence company founded by Fei-Fei Li, has announced Atlas — a next-generation world model designed to generate, reconstruct, and simulate 3D environments. The announcement, made on September 1, 2026, represents a significant step toward what the company calls "spatial intelligence": AI that can reason about and interact with the three-dimensional world.
What Atlas Can Do
Atlas is a multimodal autoregressive diffusion transformer that was pretrained from scratch to natively handle text, images, video, and 3D data. Unlike previous models that required separate systems for different modalities, Atlas processes everything through a shared spatial context, grounding each input at a specific 3D position.
The model's capabilities span several domains:
- Camera-Controlled Generation: Atlas generates images and videos from reference images with pixel-perfect camera control, outputting up to one minute of video at 1440p resolution. It can extrapolate beyond input images to imagine parts of a scene that were never photographed.
- Spatial Reconstruction: Atlas reconstructs real-world scenes from as few as one to three input images, outperforming specialist 3D reconstruction models on benchmarks including DTU, ETH3D, KITTI, and ScanNet.
- Space-Time Simulation: Atlas models both spatial structure and temporal evolution, enabling video reframing effects and Real-to-Sim workflows for robotics.
- Image Generation: Atlas generates images and 360-degree panoramas from text prompts, following complex instructions and rendering text.
Why It Matters
The implications extend well beyond creative tools. Atlas's ability to reconstruct 3D environments from sparse input images has direct applications in robotics, autonomous vehicles, and augmented reality. The model's Real-to-Sim capabilities could dramatically reduce the cost of training robots by generating realistic simulation environments from casual phone recordings.
For the creative industry, Atlas puts directors in control of camera movements rather than leaving generation to chance. The spatial context system means multiple unrelated reference images can be placed in 3D space, with Atlas generating coherent worlds that interpolate between them.
The Bigger Picture
Atlas represents a convergence of two major AI research directions: large language models and video generation. By combining autoregressive sequence modeling with diffusion-based generation, World Labs has created a foundation that can benefit from advances in both fields. The company reports strong scaling evidence, with performance improving predictably as compute increases.
The model is entering early access with select partners and will power future versions of Marble, World Labs' existing product for 3D world creation. The company is also hiring across research and engineering, signaling that this is just the beginning of its spatial intelligence roadmap.
