World Labs — the spatial-intelligence company founded by Fei-Fei Li with Justin Johnson, Christoph Lassner, and Ben Mildenhall — has made Marble, its multimodal world model, generally available to anyone. Marble generates editable, navigable 3D worlds from a text prompt, an image, a video clip, or a coarse 3D layout, and lets you interactively expand, edit, and combine those worlds rather than just render a fixed clip. The launch lands alongside a reported $1 billion funding round drawing from AMD, Autodesk, Emerson Collective, Fidelity, NVIDIA, and Sea — with Autodesk anchoring at $200 million — a scale of backing that signals how seriously the industry is taking 3D-native generation.
Watch: After LLMs: Spatial Intelligence and World Models — Fei-Fei Li & Justin Johnson, World Labs (YouTube)
Why “spatial” is the whole pitch
Most generative video models predict pixels forward in time — the next frame that looks plausible. Marble’s framing is different: reconstruct, generate, and simulate an actual 3D scene that persists, so a person or an agent can move through it and it stays coherent. That distinction matters for anyone who has tried to use AI video for anything requiring continuity. A world you can orbit, re-enter, and edit is closer to a 3D asset than a clip — which is exactly why a company like Autodesk, whose customers live in 3D scenes, is writing a nine-figure check rather than treating this as a novelty.
What creators actually get
For XR builders, installation designers, and 3D artists, the immediate value is prototyping speed. Describe an environment, get a navigable draft, then expand a room, swap a section, or fuse two worlds into one — the editing loop is the part that separates a toy from a tool. It sits naturally alongside the Gaussian-splatting and photogrammetry techniques this site has covered: different roads to the same destination of navigable synthetic space, with Marble coming at it from generation rather than capture. Being generally available (not a waitlisted demo) is the news here — it moves world models from conference-stage spectacle to something you can actually open and try.
The honest caveats
World models remain compute-hungry, fine-grained control is still weaker than a hand-built 3D engine (placing one specific object exactly where you want it is not yet a solved problem), and coherence can degrade across very large or very long explorations. “Generate a world” is not the same as “ship a game level,” and the gap between an evocative draft and a production-ready scene is real. But the direction — authoring 3D space by describing it, then inhabiting it by moving through it — is one of the clearest throughlines in creative technology right now, and Marble is the most accessible expression of it yet.