World Labs, the AI startup co-founded by pioneering researcher Fei-Fei Li, has launched its first commercial world model product: Marble. Available in both free and paid tiers, Marble enables users to transform text prompts, photos, videos, 3D layouts, or panoramic images into editable and downloadable 3D environments.

The release of this generative world model follows a limited beta preview two months ago—just over a year after World Labs emerged from stealth mode with $230 million in funding, positioning it at the forefront of the race to build world models. These AI systems generate internal representations of environments to predict future outcomes and plan actions.

While startups like Decart and Odyssey have released free demos and Google’s Genie remains in restricted research preview, Marble stands apart—even from World Labs’ own real-time model RTFM—by producing persistent, downloadable 3D worlds rather than environments generated on-the-fly during exploration. According to the company, this approach minimizes distortion and inconsistencies, allowing users to export scenes as Gaussian splats, meshes, or videos.

Marble is also the first model to offer AI-native editing tools and a hybrid 3D editor, enabling users to sketch spatial structures before letting AI fill in visual details.

“This is a new category of model that generates 3D worlds and will keep improving over time. It’s something we’ve significantly refined,” said Justin Johnson, co-founder of World Labs, in an interview with TechCrunch.

Last December, World Labs demonstrated how its early model could create interactive 3D scenes from a single image. While impressive, those somewhat cartoonish environments were not fully explorable due to movement constraints within small areas and occasional rendering glitches.

In my hands-on preview of the beta, I found that Marble could produce striking worlds from image prompts alone—ranging from game-like landscapes to a realistic digital replica of my living room. Although edge distortions were noticeable in the beta, the company claims these issues have been largely addressed in the official launch. That said, worlds I generated during the beta using a single prompt appeared more aligned with my intent than those created with the same prompt in the current version.

I haven’t tested the editing features yet, though Johnson emphasized they make Marble immediately useful for gaming, visual effects (VFX), and virtual reality (VR) projects.

“One of our core themes for Marble’s future is creative control,” Johnson explained. “There should always be a fast way to generate something, but you should also have deeper control over what you’re creating. We don’t want the machine to take over the wheel and strip away your creativity.”

Marble’s approach to creative control starts with flexible input options. The beta only accepted single images, forcing the model to imagine unseen details for a full 360-degree view. With the full launch, users can now upload multiple images or short clips captured from different angles, enabling the model to generate highly realistic digital twins of physical spaces.

Enter Chisel—an experimental 3D editor that lets users draft rough spatial layouts (think walls, boxes, or planes) and then apply text prompts to define visual style. Marble separates structure from aesthetics, much like HTML defines a webpage’s layout while CSS handles styling. Unlike purely text-based editing, Chisel allows direct manipulation of 3D objects.

“I can go in and grab the 3D block representing a sofa and move it somewhere else,” Johnson said.

Another new feature enhancing user control is the ability to expand existing worlds.

“Once you’ve generated a world, you can extend it once,” Johnson noted. “When you reach a part of the world that starts to fall apart, you can instruct the model to expand there or generate more content near your current location, adding further detail to that area.”

For users aiming to build vast environments, “Composer Mode” allows them to combine multiple worlds. Johnson showed me two he’d already created—one room made entirely of cheese with grape-shaped chairs, and another depicting a futuristic conference room in outer space.

The Path to Spatial Intelligence

Marble is offered across four subscription tiers: Free (four generations per month from text, image, or panorama), Standard ($20/month for 12 generations plus multi-image/video input and advanced editing), Pro ($35/month for 25 generations with scene expansion and commercial rights), and Max ($95/month with full access and 75 generations).

Johnson sees initial use cases primarily in gaming, film VFX, and VR.

Game developers remain divided on generative AI. A recent Game Developers Conference survey found that one-third of respondents believe generative AI negatively impacts the industry—a 12% increase from the previous year. Key concerns include intellectual property theft, energy consumption, and declining quality of AI-generated content. Last year, a Wired report revealed studios like Activision Blizzard are already using AI to cut costs amid workforce reductions.

In gaming, Johnson envisions developers using Marble to rapidly generate background environments and ambient spaces, which can then be imported into engines like Unity or Unreal for adding interactivity, logic, and code.

“It’s not meant to replace existing game pipelines but to provide assets that fit into them,” he said.

For VFX workflows, Marble sidesteps the inconsistency and poor camera control that plague AI video generators, Johnson claims. Its 3D assets give artists frame-accurate control over scenes and camera movements.

Although World Labs isn’t currently prioritizing VR applications, Johnson noted the sector’s “content shortage” and expressed enthusiasm about the launch. Marble is already compatible with Vision Pro and Quest 3 headsets, and every generated world can be viewed in VR today.

Marble may also find applications in robotics. Unlike image and video generation, robotics lacks large-scale training data repositories. But with tools like Marble, creating simulated training environments becomes far more accessible.

In a recent manifesto authored by World Labs CEO and co-founder Fei-Fei Li, Marble is described as the first step toward building “truly spatially intelligent world models.”

Li argues that “the next generation of world models will enable machines to achieve spatial intelligence at an unprecedented level.” If large language models taught machines to read and write, she hopes systems like Marble will teach them to see and build. She believes that understanding how objects exist and interact in three-dimensional space could unlock breakthroughs beyond gaming and robotics—potentially transforming fields like science and medicine.

“Without spatial intelligence, our dream of truly intelligent machines will remain incomplete,” Li wrote.