Introduction
On September 1, 2026, World Labs — the spatial intelligence company Fei-Fei Li co-founded in 2024 — announced Atlas, an "omni world model" pretrained from scratch to operate natively on text, images, video and 3D. The headline capability is camera-controlled video: up to one minute at 1440p, generated from one or more reference images with the camera path supplied as a geometric input rather than described in a prompt.
Atlas is not a product launch. It is entering early access with select partners behind a request form, with no announced pricing, API or public date, and World Labs published it as a research post complete with a BibTeX citation. The company says Atlas will power future versions of Marble, its existing tool for building explorable 3D worlds.
What Atlas Does
World Labs groups the capabilities into four families, all served by one network rather than a pipeline of specialists:
- Camera-controlled generation — images and video from one or more reference images, with the camera position and angle specified exactly.
- Spatial reconstruction — rebuilding a real scene from anywhere between a single photo and over a hundred, emitting both novel-view frames and explicit 3D output as point clouds or Gaussian splats. World Labs says two or three images are typically enough for a faithful reconstruction, and frames this as progress on novel view synthesis from sparse views — a decades-old problem in computer vision.
- Space-time simulation — modeling how a scene changes over time. Two applications are shown: "bullet time" reframing from as few as three to five ordinary phone cameras on tripods, and Real-to-Sim for robotics, where a 24-frame phone video of a large space becomes a simulator that also renders what the robot's own cameras would see as it moves.
- Image generation — text-to-image, including 360 panoramas and text rendering. World Labs explicitly calls this a side capability, not the focus.
Spatial Context: the Idea Underneath
The architecture is a multimodal autoregressive diffusion transformer. Unpacked: a transformer that consumes text, images, camera poses and depth maps as one sequence, generates elements one at a time as an LLM does, and denoises each one with a rectified flow objective in the manner of a diffusion model. Videos are just sequences of images.
The idea that carries the design is what World Labs calls a spatial context. Like an LLM, Atlas encodes its inputs into a context and generates conditioned on it — but every image and depth map in that context is grounded at an explicit 3D camera pose. Position is structural, not inferred.
That distinction is what makes the demonstrations possible. Asking a video model to "orbit slowly to the left" is a text instruction it may or may not honor; handing Atlas a camera trajectory is an input it is built around. And because you control where inputs sit in space, you can place two unrelated reference images at chosen positions and have Atlas invent the hallway, doorway or transition between them.
World Labs also notes the practical dividend of the hybrid: because Atlas is autoregressive, it inherits LLM serving work such as KV caching, cache-aware routing and disaggregated serving; because it is a latent diffusion model, it can use diffusion distillation, classifier-free guidance and shifted noise schedules.
The Numbers, and How They Were Produced
Every figure below is World Labs' own, published in its announcement. None has been independently reproduced.
Camera-controlled generation. Third-party human raters were shown a single input image plus one to three cinematic camera motions and asked which model followed the intended path better. Share of voters choosing Atlas: 75% against MiniMax H3, 81% against Gemini Omni Flash, 86% against HappyHorse 1.1, 93% against FLUX 3 and 94% against Seedance 2.5.
The caveat is one World Labs states itself, and it matters more than the percentages. None of those baselines accepts a camera as a native input, so each received the camera path as a text description using standard cinematic terms. The company acknowledges that better prompt engineering might improve their camera-following. So the evaluation measures the gap between geometric camera input and text-described camera input — which is exactly Atlas's design thesis, and not the same claim as being a better video model overall. World Labs also does not publish the rater pool size or the prompt set.
3D reconstruction. Atlas is compared on mean absolute-relative pointmap error, averaged over seven benchmarks — DTU, ETH3D, KITTI, NRGBD, 7-Scenes, Tanks & Temples and ScanNet. Lower is better, and the units are AbsRel ×10⁻³:
| Model | Average error |
|---|---|
| Atlas | 25.3 |
| Pi3X (posed) | 28.7 |
| π³ | 34.7 |
| VGGT-Ω 1B | 36.4 |
| Depth Anything 3 | 39.3 |
| MapAnything | 47.7 |
World Labs says it reproduced every baseline itself under a common protocol, and states the claim precisely: Atlas beats the best specialized open-source reconstruction models. Closed systems are not in the comparison. Parameter count, frame rate and hardware requirements are not disclosed anywhere.
On scaling, the company reports that during development each new level of training compute "unlocked new model capabilities," and expects the trend to hold.
Context and Competition
World Labs is well funded for a long bet. It raised $1 billion in February 2026 — led by a $200M investment from Autodesk, with NVIDIA, AMD, Emerson Collective, Fidelity and Sea participating — bringing it to roughly $1.23B since emerging from stealth in September 2024 with $230M.
The bet is that the next capability jump comes from grounding multimodal AI in 3D space rather than from more text, an argument Li made at length in her spatial intelligence essay. She is not alone in making it: world models are now a crowded field including Odyssey-2 Max on real-time physics, Yann LeCun's JEPA work, Niantic Spatial, and domain-specific systems like the Waymo World Model.
Atlas's differentiator against video generators such as Sora and Veo is that it emits geometry alongside pixels — a scene you can re-enter from a new angle or hand to a robot simulator, rather than frames that merely look consistent. Whether that geometry holds up for production VFX or embodied AI training is what early access exists to find out.
Conclusion
Atlas stakes a real claim: that one model can generate, reconstruct and simulate 3D scenes better than three specialized ones, and that explicit camera geometry beats describing camera motion in words. The reconstruction results against open-source specialists are the more meaningful half of the evidence, because the comparison is like-for-like. The camera-control preference rates are the weaker half, since the baselines were handicapped by an input format they do not support — a limitation World Labs is upfront about.
The model is behind a partner gate with no pricing, no API and no disclosed scale, and third-party coverage so far restates the announcement rather than testing it. Treat every number here as a vendor figure until someone outside World Labs reproduces it.
Sources
- Atlas: A World Model for Spatial Intelligence — World Labs announcement, September 1, 2026
- Fei-Fei Li's World Labs debuts Atlas — SiliconANGLE, September 1, 2026
- World Labs closes $1B investment backed by Nvidia, AMD and Autodesk — SiliconANGLE, February 18, 2026
- World Labs lands $1B, with $200M from Autodesk — TechCrunch, February 18, 2026