World Labs, led by Fei-Fei Li, accelerates the development of world models by launching Marble, their initial commercial offering.

AI pioneer Fei-Fei Li’s startup, World Labs, is debuting its inaugural commercial world model product. Marble is now accessible through freemium and premium options, enabling users to transform text inputs, images, videos, 3D designs, or panoramas into adaptable, downloadable 3D settings.

The generative world model’s launch, initially available as a limited beta preview two months prior, occurs just over a year after World Labs emerged from stealth with $230 million in funding, positioning the startup ahead of rivals in world model development. World models are AI systems capable of generating an internal representation of an environment, utilized for predicting future outcomes and strategizing actions.

While startups like Decart and Odyssey have launched complimentary demos, and Google’s Genie is still in research preview with limitations, Marble distinguishes itself — even from World Labs’s own RTFM real-time model — through its creation of persistent, downloadable 3D environments instead of generating worlds dynamically during exploration. According to the company, this approach minimizes morphing and inconsistency, enabling users to export environments as Gaussian splats, meshes, or videos.

Marble also stands out as the first model providing AI-integrated editing tools and a hybrid 3D editor, allowing users to define spatial arrangements before the AI fills in visual aspects.  

“This represents a completely new model category for generating 3D worlds, with continuous improvements expected. We’ve already made significant enhancements,” Justin Johnson, World Labs co-founder, shared with TechCrunch. 

In December, World Labs demonstrated the ability of its early models to produce interactive 3D scenes using just one image. While notable, these somewhat cartoonish scenes had limited exploration due to movement restrictions and occasional rendering errors. 

During my beta preview trial, Marble impressed me by creating compelling worlds from mere image prompts, encompassing both game-like settings and photorealistic depictions of my living space. Although scene edges showed morphing, this issue is reportedly resolved in the current launch. Nevertheless, a world I generated in the beta with a single prompt appeared superior and more aligned with my intention than the result from the same prompt now. 

While I haven’t yet explored the editing features, Johnson indicates they enhance Marble’s practicality for immediate gaming, VFX, and virtual reality (VR) ventures. 

“A primary focus for Marble moving forward is user control,” Johnson stated. “Generating content should always be straightforward, but users should also have options for in-depth customization, preventing the machine from overshadowing their creative input.” 

Marble’s strategy for creative control starts with flexible input options. The beta was restricted to single images, requiring the model to guess unseen details for a 360-degree perspective. Now with the full version, users are able to upload several images or brief videos displaying an area from various viewpoints, enabling the model to create fairly accurate digital duplicates. 

Then there’s Chisel, an exploratory 3D editor enabling users to construct basic spatial configurations (like walls, boxes, or planes) and include text prompts to define the visual style. Marble then generates the environment, separating structure from style — much like HTML dictates a website’s layout while CSS applies the visual design. Unlike text-based editing, Chisel enables the direct manipulation of objects.  

“Users can easily select the 3D block representing a couch and move it to another location,” Johnson explained. 

The capability to broaden a world is yet another function enabling enhanced editing flexibility.  

“After generating a world, you have the option to double its size,” Johnson clarified. “If you encounter areas of the world beginning to fall apart, you can instruct the model to expand or create additional elements in that particular region, thereby enriching the detail of that space.”

Users wanting expansive environments can merge several worlds utilizing “composer mode.” Johnson demonstrated this feature, combining a cheese-themed room with grape-based seating with a futuristic meeting space situated in outer space.

The path to spatial intelligence

Marble provides four membership options: Free (limited to four generations using text, image, or panorama), Standard (at $20 monthly, including 12 generations and enabling multi-image/video input along with enhanced editing), Pro (at $35 monthly, offering 25 generations, environment expansion, and commercial licensing), and Max (at $95 monthly, providing all features and 75 generations). 

Johnson believes the primary uses for Marble will initially involve gaming, film-based visual effects, and virtual reality.  

Gaming developers hold varied perspectives on the technology. A recent survey from the Game Developers Conference revealed that a third of participants felt generative AI negatively influences the gaming sector – a 12% rise from the previous year. Major concerns expressed included intellectual property infringements, energy consumption, and a decline in quality due to AI-produced content. Furthermore, a prior Wired investigation revealed that gaming companies like Activision Blizzard utilize AI for cost reduction and to address employee turnover. 

Within gaming, Johnson anticipates that developers will utilize Marble to produce scene backgrounds and ambient settings, subsequently importing these elements into gaming platforms such as Unity or Unreal Engine for incorporating interactivity, logic, and code. 

“It’s not designed to substitute the entire gaming production line; rather, it offers elements that can be integrated,” he clarified.  

For VFX projects, Marble prevents the inconsistencies and restricted camera manipulation common with AI video creators, as Johnson notes. Its 3D assets empower artists to design scenes and manage camera motions with exact precision, he noted. 

While Johnson stated that World Labs is currently not prioritizing virtual reality (VR) applications, he acknowledged the industry’s “content scarcity” and enthusiasm surrounding the launch. Marble has immediate compatibility with the Vision Pro and Quest 3 VR headsets, enabling the viewing of any generated environment in VR.

Robotics might also discover potential applications for Marble. Johnson mentioned that robotics, unlike image and video generation, doesn’t benefit from a wide range of training data. Generators such as Marble, however, make replicating training settings more accessible.  

According to a recent statement by Fei-Fei Li, World Labs’ CEO and co-founder, Marble marks the initial phase in establishing “a truly spatially intelligent world model.” 

Li anticipates “the future generation of world models will empower machines to attain spatial intelligence at an unprecedented degree.” Just as large language models enable machines to comprehend reading and writing, Li hopes platforms such as Marble will teach them the skills of vision and construction. She suggests that grasping how objects exist and relate within three-dimensional spaces can eventually enable machines to make discoveries extending beyond gaming and robotics and into scientific and medical fields. 

“Our aspirations for genuinely intelligent machines remain incomplete without spatial intelligence,” Li concluded.

Have a confidential tip or private documents? We’re covering the inner workings of the AI sector — from the businesses shaping its future to the individuals impacted by their decisions. Contact Rebecca Bellan at [email protected] or Russell Brandom at [email protected]. For secure communication, reach them via Signal at @rebeccabellan.491 and russellbrandom.49.