World Labs, the startup founded by AI pioneer Fei‑Fei Li, has moved from research to product with Marble — a commercial “world model” that generates persistent, downloadable 3D environments from text, photos, video, panoramas or rough 3D layouts. The company says Marble is designed to give creators and developers a practical path to spatially aware AI content: editable, exportable worlds that can be used in games, visual effects, virtual reality and simulation.
What Marble is and what’s new
Marble departs from on‑the‑fly scene generators by producing persistent 3D worlds users can export and edit. The model accepts multiple input types (single images, multi‑image sets, short clips, text prompts, panoramas and coarse 3D sketches) and outputs assets users can download as Gaussian splats, meshes or video. World Labs also ships AI‑native editing tools, including an experimental hybrid 3D editor called Chisel that lets users block out spatial structure (walls, boxes, terrain) and then invoke text prompts to refine style and detail.
Key features:
- Multi‑modal input: text, images, video, panoramas and 3D layouts
- Persistent, downloadable output: Gaussian splats, meshes, and videos
- Chisel editor: block‑out spatial structure then apply AI‑driven styling
- Scene expansion and “composer” mode to stitch multiple worlds together
- Exports compatible with common 3D pipelines and viewable in VR headsets such as Vision Pro and Quest 3
- Free: four world generations from text, image or panorama
- Standard: $20/month — 12 generations, multi‑image/video input and advanced editing
- Pro: $35/month — 25 generations and commercial rights
- Max: $95/month — full feature set and 75 generations
- Game development (backgrounds, ambient spaces, rapid prototyping)
- Visual effects and film (stageable 3D assets with precise camera control)
- VR content (filling a content gap for headsets)
- Simulation and robotics (synthetic environments for training agents)
- Hobbyists and creators: The free tier offers a low‑risk way to explore Marble’s creative possibilities.
- Indie developers and VFX artists: Standard and Pro tiers provide more generations, multi‑image inputs and (for Pro) commercial rights — useful for prototyping and short‑term projects.
- Larger studios and enterprises: Marble can accelerate background and environment creation, but studios that require tight control over interactivity, provenance or custom pipelines may prefer to treat Marble outputs as one input among many and to validate rights and fidelity before adoption.
Justin Johnson, World Labs co‑founder, framed Marble as “a brand new category of model that’s generating 3D worlds,” and emphasized an editorial emphasis on "creative control," saying users should have both quick generation and deep manual control.
Pricing, availability and rights
Marble is available now in freemium and paid tiers:
The Pro tier explicitly grants commercial rights, a detail World Labs highlights for professionals who plan to monetize assets created with Marble.
How Marble compares to other world models
World Labs launched Marble after a limited beta that ran for several months and follows its own real‑time model (RTFM). Marble’s distinguishing claims are persistent, exportable worlds and a hybrid editing workflow that separates structure from style.
Other players in the space include Google’s Genie (still in limited preview), Nvidia’s Cosmos work, and demos from startups such as Decart and Odyssey. Marble aims to sit between research demos — which often render short explorable pockets with visual inconsistencies — and production tools by outputting cleaner geometry and offering practical editing/export features for downstream pipelines.
Early use cases and industry reaction
World Labs positions Marble as immediately useful for:
Fei‑Fei Li frames Marble as a first step toward “spatial intelligence,” arguing that future systems must understand three‑dimensional relationships to take robotics, simulation, science and medicine beyond current text‑centric AI capabilities.
The industry response is mixed. Some creators and studios see Marble as a fast way to generate environmental assets that can be imported into engines like Unity and Unreal and then instrumented with interactivity. World Labs’ Justin Johnson said Marble isn’t intended to replace existing pipelines but to feed them with drop‑in assets.
At the same time, skepticism among game developers and artists persists. A recent Game Developers Conference survey found an increased share of respondents worried that generative AI could harm the games industry; concerns cited include intellectual property theft, energy usage and potential quality decline. Investigations last year also flagged studios using AI to cut costs, raising questions about labor and creative standards.
Technical strengths and limitations
World Labs says Marble produces larger, stylistically diverse worlds with cleaner 3D geometry than many prior models. New inputs (multi‑image and video) help the model avoid inventing unseen details from a single photo, improving realism and fidelity for digital twins.
The product also includes tools to expand sections of a generated world when a user reaches its boundary, and a composer mode for combining multiple generated worlds — for example, stitching a stylized room into a sci‑fi environment.
But limitations remain. Beta users reported occasional morphing and rendering errors at scene edges, and TechCrunch’s beta trial noted that some prompts matched intent better in earlier previews than in the initial full launch. World Labs says those issues have been addressed to a degree, but Marble is still an early commercial product and will evolve.
Broader implications: spatial intelligence and safety questions
World models are positioned as foundational for moving AI beyond text: they let systems predict outcomes in three‑dimensional environments and plan actions. World Labs, backed by about $230 million in funding, counts investors and advisors from across tech and AI research; the company argues that spatially fluent models will be necessary for robust robotics, safer simulation and richer human‑machine collaboration.
But scaled world generation also raises policy and ethical questions: training data provenance, copyright for generated assets, energy usage, and potential misuse. Game studios have already voiced IP concerns about generative tools, and the industry debate about responsible deployment of spatially aware AI is likely to intensify as capability improves.
Who should try Marble and when
For those deciding whether to adopt Marble, two practical points matter: the Pro tier’s inclusion of commercial rights and the product’s export formats (meshes and splats) that integrate with existing 3D toolchains. Teams worried about IP should run tests with their own assets and check legal terms carefully.
Bottom line
Marble is a notable commercial step for world models: a multi‑modal generator that emphasizes persistent, editable 3D worlds and a hybrid workflow that separates structure from style. It arrives at a moment of intense industry interest — and rising scrutiny — around generative systems. For creative teams seeking faster world‑building tools, Marble offers immediate utility; for the broader AI community, it represents a tangible push toward the spatial intelligence Fei‑Fei Li has long advocated. Whether Marble becomes a standard production tool will depend on how quickly World Labs can improve fidelity, address creators’ legal and quality concerns, and prove the model’s utility inside real‑world pipelines.