Fei-Fei Li’s World Labs Unveils Atlas World Model

World Labs, the spatial-intelligence startup co-founded by Stanford's Fei-Fei Li, unveiled Atlas on Tuesday — a model the company calls an "omni world model," pretrained from scratch to operate natively on text, images, video and 3D rather than bolting a geometry module onto a video generator. In practice that means one system that will take a handful of photographs, generate a minute of 1440p video along a camera path you specify, and hand back an explicit 3D representation of the same scene as a point cloud or Gaussian splat.

Architecturally Atlas is a multimodal autoregressive diffusion transformer, which is less of a mouthful than it sounds: inputs of every type — text, reference images, camera poses, depth maps — are folded into a single shared spatial context, and outputs are generated sequentially while rectified flow denoises them. The design borrows KV-caching and cache-aware routing from large language models and diffusion distillation and classifier-free guidance from video models. The practical payoff is camera control that is geometric rather than suggested: the camera pose is a native input, not a text prompt the model interprets loosely.

On camera-path adherence, human raters preferred Atlas over every model World Labs benchmarked it against — 75% against MiniMax H3, 81% against Gemini Omni Flash, 86% against Happy Horse 1.1, 93% against FLUX 3 and 94% against Seedance 2.5. The more interesting result is on reconstruction, where a generalist model beat the specialists at their own job.

Atlas accepts anywhere from a single image to more than a hundred, and reconstructs scenes from as few as one and as many as twenty-five or more inputs, tolerating sparse coverage by inventing plausible connective geometry — doorways, hallways and transitions between otherwise unrelated photographs. Beyond stills, the company demonstrates space-time simulation: bullet-time reframing built from three to five phone cameras, and real-to-sim workflows that turn captured footage into environments a robot policy can be trained against. Output modes run to 360-degree panoramas and text-to-image generation as well.

The launch is thinner on commercial specifics than on capability. Atlas is in early access with partners World Labs declines to name, there is no published paper, and no pricing has been disclosed; access runs through a request form. The model is positioned to power future versions of Marble, the company's existing product, which already renders with the Gaussian splats Atlas emits natively. World Labs has raised roughly $1.23 billion since its 2024 founding, including a $1.2 billion round in February 2026 backed by Nvidia, AMD, Autodesk and Fidelity.

Li has argued for years that spatial understanding is not an optional module on the road to general intelligence but a prerequisite for it — that a system with no model of how space behaves cannot reason about the physical world at all. Atlas is the most direct test of that claim the company has shipped, and it arrives into a field getting crowded fast, with Odyssey working on interactive world simulation, Niantic Spatial on geospatial mapping and Yann LeCun's AMI Labs on physical planning. Whether one omni model beats four specialised ones is a question the benchmark chart answers only for World Labs' own choice of benchmarks.