Summary
- Fei-Fei Li's World Labs unveiled Atlas, a world model that handles generation, reconstruction, and simulation in one system
- It rebuilds 3D spaces from a few photos and lets users specify camera position directly to produce videos up to one minute long at 1440p
- It can mass-produce robot training data inside simulation, and for now it's available only through early access to select partners
Building an entire 3D world from just a handful of photos
World Labs, co-founded by AI researcher Fei-Fei Li — known for creating ImageNet — has unveiled a world model called Atlas. With just a few photos, Atlas can generate new 3D scenes, reconstruct real-world spaces, and simulate those spaces, all through a single model. Until now, generation, reconstruction, and simulation each required their own specialized model, but World Labs claims Atlas outperforms specialized models across all three tasks.
To put that in context, World Labs is the company behind Marble, a product that turns 3D spaces into video, and Atlas is essentially a bigger engine designed to power products like Marble and others. When Autodesk, NVIDIA, and other investors put $1 billion into the company in February 2026, Bloomberg reported that the deal was being negotiated at a $5 billion valuation.
World Labs describes Atlas as an "omni model" trained from the ground up on text, images, video, and 3D data. Rather than processing input as a flat sequence, it anchors every piece of information to a specific position in 3D space — an approach the company calls "spatial context." In a November 2025 essay, Fei-Fei Li pointed out that existing multimodal language models and video diffusion models struggle even with simple spatial tasks because they break data down into one- or two-dimensional sequences. Atlas is the company's answer to that problem, tackling it through 3D and 4D structure instead.
Composing scenes like you're operating a camera directly
The camera-controlled generation feature lets you feed in one or more images and pull new scenes from any camera position and angle you specify. While other video models require describing camera movement through text prompts, Atlas takes camera movement itself as a direct geometric input. The output can run up to one minute at 1440p resolution. World Labs frames this as giving users direct control over every scene, rather than pulling a slot machine lever and hoping for the right result each time.
Reviving real spaces from just two or three images
The spatial reconstruction feature rebuilds real-world scenes from anywhere between one and dozens of photos, with no special capture equipment required. The more input images you provide, the less the model has to fill in on its own — and according to World Labs, even just two or three images produced better results than specialized 3D models. In one demonstration, the team reconstructed Stanford's Main Quad, progressively increasing the input from 2 to 25 ground-level photos, and ultimately generated an aerial view looking down over the campus. By contrast, existing models like VGGT and InfiniteVGGT reportedly showed geometric distortion and blurred textures whenever the camera moved even moderately, based on comparisons using the OpenWorldLib framework.
Because Atlas processes depth information alongside RGB data, it can output not just images and video but actual 3D data as well. Supported formats include point clouds and 3D Gaussian splats — a technique that composes scenes from many small spatial data points, allowing smooth viewing from any angle. This is the same representation method used by World Labs' existing product, Marble.
Mass-producing robot training data inside simulation
In robotics, Atlas functions as a "real-to-sim" tool for translating real-world spaces into simulation. Once a room is reconstructed, it generates the images and depth data that a robot's sensors would capture along its movement path inside the simulated environment. With just a few photos, users can simulate a wide range of grasping and movement tasks by varying objects, positions, lighting, and backgrounds. The goal is to produce the diverse data robots need for training without having to physically film every real-world scenario.
World Labs previewed this approach earlier, in August, as a standalone product called the "real-to-sim-to-real" engine. That engine multiplies a single real-world task into thousands of variations, training control models entirely within simulation. According to the company, it ran unsupervised for an hour each across five different robot platforms. The underlying technology came from SceniX, a startup World Labs acquired in July.
Claims to outperform competitors in its own evaluations
World Labs says no single benchmark can fully capture what Atlas is capable of, but it did share results from two evaluations. In human-rated comparisons of camera-controlled generation, Atlas was preferred 75% of the time over MiniMax H3, 81% over Google Gemini Omni Flash, 86% over Happy Horse 1.1, and 94% over ByteDance's Seedance 2.5. The name of the model compared in the 93% figure is missing from the original material, so it couldn't be confirmed.
| Comparison | Atlas Preference Rate |
|---|---|
| MiniMax H3 | 75% |
| Gemini Omni Flash | 81% |
| Happy Horse 1.1 | 86% |
| (model name unconfirmed) | 93% |
| Seedance 2.5 | 94% |
In evaluations of 3D reconstruction from sparse photos, the company said Atlas recorded a median error of 25.3, outperforming Pi3X and VGGT-Ω 1B.
From Marble to a billion-dollar raise
World Labs was founded in 2024 by Fei-Fei Li, the researcher known for creating ImageNet and leading Google Cloud's AI division from 2017 to 2018. The company drew backing from Andreessen Horowitz, AMD, Intel, and NVIDIA from the start. Its first system, released in late 2024, was rough enough that users would hit an invisible boundary after moving just a few meters in the virtual space. Marble followed in November 2025 as its successor.
In February 2026, the company raised $1 billion from Autodesk, Andreessen Horowitz, NVIDIA, and AMD — following an earlier Bloomberg report that the deal was being negotiated at a $5 billion valuation. Atlas is set to serve as the core model powering future versions of Marble and other products, and for now it's available only through an early access program limited to select partners. There's no public entry point yet for general users to try it directly.
A debate over what counts as a "world model"
Researchers don't fully agree on what a "world model" actually is. An international research team led by Peking University proposed a unified definition through OpenWorldLib in April 2026, explicitly excluding pure text-to-video models on the grounds that they lack a feedback loop with the real world. Under that framework, 3D reconstruction and simulators like Atlas qualify as core components because they provide environments where physical laws can be verified.
Official launch video
Editor's Take
What makes Atlas interesting isn't the performance numbers — it's the approach. Until now, 3D generation, reconstruction, and robot simulation each belonged to different models made by different companies. NVIDIA has been pushing robot foundation models through Cosmos, Tencent released a workflow for generating open worlds from text prompts, and Runway put out an interface world model that draws the scene itself in real time. World Labs is essentially saying it can fold all three of these strands into a single model — and if that bet pays off, it marks the start of several separate 3D pipeline tools converging into one.
The generational comparison makes the shift tangible. World Labs' first system, from late 2024, was so limited that users hit walls after moving just a few meters in virtual space. Less than a year later, Marble produced genuinely usable 3D scenes. Now Atlas has taken on the role of the engine that powers Marble itself. What stands out is the pace at which each generation moves from "demo" to "usable tool" to "foundation that powers other tools."
Practically speaking, this is still very much a wait-and-see situation. Since it's only available through early access, general developers and studios can't get their hands on it yet, and the only benchmarks available are the company's own human evaluations and self-reported reconstruction error figures. Still, the robotics angle is worth watching. Generating large volumes of grasping and movement task variations through simulation — scenarios that are hard to capture on camera in the real world — offers a useful reference for teams working on robot foundation models who are looking to cut down on data acquisition costs.
Over the coming months, the real test for this announcement will be how much of Atlas actually makes it into Marble, and when the partner-only early access opens up to everyone else. With NVIDIA, Tencent, and Runway each taking their own approach to the 3D and world-model race, how quickly Atlas can demonstrate real-world use cases beyond its benchmark numbers will determine just how much this announcement really matters.
Comments