World Labs drops Marble, its first commercial world model

If large language models taught machines how to read and write, Fei-Fei Li wants to teach them how to see and build.

World Labs, the startup launched last year by the Stanford AI pioneer with $230 million in funding, has just rolled out its first commercial product. It is called Marble, and it turns text, photos, video clips, and coarse layouts into fully interactive, downloadable 3D worlds.

Persistent 3D worlds, not dynamic video tricks

World models have become one of the hottest buzzwords in AI, with demos from startups like Decart and Odyssey, as well as research prototypes like Google’s Genie. But Marble takes a fundamentally different architectural path.

Instead of generating dynamic video frames on the fly as a user walks through an environment—which often leads to severe visual morphing whenever you turn around—Marble generates persistent 3D spaces. You can export these generated environments as Gaussian splats, textured meshes, or camera-path videos, and drop them straight into game engines like Unreal Engine and Unity.

The multimodal inputs are also significantly upgraded from the beta preview two months ago. While the early version only took single images, Marble now ingests multiple photos or short video sweeps, letting creators build digital twins of real-world rooms with much higher geometric accuracy.

Creative control via spatial blocking

The most interesting feature in this release is Chisel, an experimental 3D editor that tackles one of the biggest headaches in generative AI: steering the model without guessing text prompts.

World Labs co-founder Justin Johnson compares it to web development, where HTML handles the structure and CSS handles the styling. In Chisel, users can block out walls, boxes, and floors in 3D, and then let the generative model paint the visual details on top. If you want a couch in a different corner, you just grab the 3D proxy block and drag it over.

Users can also expand a scene once toward an edge when the geometry starts to degrade, or use a composer mode to stitch entirely separate worlds together. The tool is already compatible with Apple Vision Pro and Meta Quest 3 headsets.

Pricing is split into four tiers. A free tier offers four generations from basic prompts. The Standard tier costs $20 per month for 12 generations and multi-image inputs. Pro runs $35 per month for 25 generations, commercial rights, and scene expansion. The top-tier Max plan costs $95 per month for 75 generations.

Why it matters

We think World Labs made a genuinely smart call by anchoring Marble to exportable 3D data formats rather than chasing uneditable video generations.

AI video models have dazzled the internet, but they are notoriously frustrating for professional game designers and VFX artists who need precise camera control and editable geometry. By exporting clean meshes and Gaussian splats, Marble actually fits into existing pipelines instead of demanding that the industry throw out its workflows.

That said, it is not a silver bullet yet. Hands-on testing shows that generated spaces are still relatively small—you can only walk a few steps before hitting an invisible geometric boundary. At $95 a month for just 75 generations on the top plan, burning through quota on prompt trial-and-error will get expensive very quickly.

Culturally, the gaming sector is also pushing back. A recent Game Developers Conference survey found that a third of developers believe generative AI is harming the industry, driven by IP theft concerns and quality dilution. World Labs is wise to pitch Marble as an asset generator rather than a replacement for developers, but the model still needs to prove its generated meshes are clean enough for serious production use.