World Labs Marble: Fei-Fei Li's Text-to-3D World Model, Explained (Updated August 2026)
Last updated August 7, 2026
World Labs Marble generates persistent, navigable 3D environments from text, images, panoramas, or video — Gaussian splats plus collider and visual meshes, not video clips. Built by Fei-Fei Li's ~$1.23B-funded World Labs, it ships in four variants (1.1 Plus, 1.1, 1.0, 1.0 Draft), exports to Blender/Unreal/Unity/Houdini and VR, and starts at ~$0.12 per Draft world via API. All four variants run in invideo's 3D tab.
Updated August 2026
World Labs Marble is the first commercial "world model" — an AI system that turns a text prompt, an image, or even a video clip into a persistent, navigable 3D environment you can walk through, edit, and export, rather than a flat video clip. It comes from World Labs, the spatial-intelligence startup founded by Stanford's Fei-Fei Li, and as of August 2026 it ships in four variants (Marble 1.1 Plus, 1.1, 1.0, and 1.0 Draft). The distinction that matters: the output is scene data you can re-enter, not frames you scrub — which is why filmmakers are adopting it as a previsualization tool rather than another clip generator.
What is World Labs Marble?
Marble is the flagship product of World Labs, founded by Fei-Fei Li — creator of ImageNet, often called the "godmother of AI" — with Justin Johnson, Christoph Lassner, and Ben Mildenhall, a co-author of the NeRF paper. The company's stated mission is "spatial intelligence": foundation models that, in its words, "perceive, generate, reason, and interact with the 3D world." Li laid out the thesis in her November 2025 essay From Words to Worlds.
Investors have priced that thesis aggressively. World Labs emerged from stealth in September 2024 with $230 million raised at a valuation above $1 billion, and in February 2026 it announced a further $1 billion round — roughly $1.23 billion in total — including $200 million from Autodesk, with NVIDIA, AMD, and Fidelity participating. TechCrunch reported the new valuation at around $5 billion; the company declined to confirm.
Marble itself moved fast: a first image-to-3D-scene research demo in December 2024, a limited beta in September 2025, general availability on November 12, 2025 as a "frontier multimodal world model," a developer World API in January 2026, and the Marble 1.1 generation in April 2026. In July 2026 World Labs also acquired robotics-simulation startup SceniX — a signal that generated worlds are becoming training grounds for robots, not just visuals.
What does "3D worlds from text" actually mean?
This is the part most coverage gets wrong. Marble does not generate a video that looks 3D — it generates an actual 3D scene that persists after generation. Per the official export documentation, a Marble world is delivered as:
- 3D Gaussian splats — the core representation, exportable as
.spz(World Labs' compressed native format, the default since December 2025) or.ply(the universal splat format), at roughly 2 million or 500 thousand splat tiers. - A collider mesh — a lightweight GLB triangle mesh (about 100–200k triangles) so game engines can run physics and collision against the world.
- A high-quality visual mesh — a separate GLB at roughly 600k textured triangles (or ~1M vertex-colored), for pipelines that need traditional geometry.
- 360° panoramas — 2560×1280 PNG exports.
- Video renders — flythroughs with what World Labs describes as pixel-accurate camera control, with configurable codec, frame rate, and aspect ratio.
Because the world exists as data rather than frames, you can revisit it, walk a different path through it, and get a consistent scene every time — the property World Labs calls persistence. Worlds render in the browser through Spark, World Labs' open-source Gaussian-splat renderer for THREE.js; Spark 2.0 (April 2026) added level-of-detail streaming for large worlds on the web. (World Labs' RTFM, a real-time frame-generation research preview from October 2025, is a separate project — not part of Marble.)
Marble model versions compared
Since April 2, 2026, Marble ships as four variants, per the official models documentation — the same four in invideo's model picker.
| Variant | What it is | Speed | Relative cost (API) |
|---|---|---|---|
| Marble 1.1 Plus | Most advanced; "creates the largest worlds" and auto-expands spatial coverage as needed | Standard | Variable — up to ~$2.48/world |
| Marble 1.1 | Standard quality at fixed cost; World Labs' designated future API default | Standard (~5 min full world) | ~$1.20–1.28/world |
| Marble 1.0 | The original GA model, now legacy | Standard | ~$1.20–1.28/world |
| Marble 1.0 Draft | Fast idea-exploration tier | ~20 seconds | ~10× cheaper (~$0.12–0.20/world) |
The practical workflow: iterate on composition in 1.0 Draft at pennies per world, then re-run the keeper prompt on 1.1 or 1.1 Plus. World Labs has not published quality benchmarks comparing 1.1 to 1.0 — the documented differences are world size, auto-expansion, and cost structure.
What inputs does Marble accept?
Marble is unusually multimodal for a generative model. Per the official docs, it accepts:
- Text prompts — the standard text-to-world path.
- A single image — the scene extrapolated outward from one view.
- Multiple images — photos of a space fused into one coherent world.
- 360° panoramas — an equirectangular image as the seed.
- Video — clips up to 100MB, useful for reconstructing real locations.
- Chisel layouts — coarse 3D block-outs (including imported assets) styled with a prompt, separating structure from visual style.
The editing stack extends beyond generation: world editing (remove, swap, or restyle elements), Chisel for structure-first authoring, Expand for growing a region of an existing world, and Composer for stitching multiple worlds together, with mesh import supported since January 2026.
What can you do with a Marble world?
Export is the point. World Labs maintains official export guides for Blender, Unreal Engine, Unity, and Houdini, plus Spark for web delivery — a generated world can become a set in a real DCC pipeline rather than staying trapped in an app. Every world also gets a shareable VR link from the viewer; TechCrunch reported headset support naming Quest 3 and Apple Vision Pro.
Exports carry production-minded details: real-world unit scaling, OpenGL or OpenCV coordinate conventions (added January 2026), and free PLY splat export via the API. On the research side, World Labs' real-to-sim-to-real work (July 2026) uses Marble-style worlds to train robot policies, with integrations spanning NVIDIA Isaac Sim, MuJoCo, and RoboSuite.
How long does a world take to generate?
Documented times, as of August 2026: a navigable preview appears in about 30 seconds from text, image, or panorama inputs (about 2 minutes for multi-image or video). A Draft world completes in roughly 20 seconds; a full-quality world takes around 5 minutes, as does expanding a region. The outlier is the high-quality mesh export — up to an hour, capped at 4 per hour.
What does Marble cost?
Official numbers only, from World Labs' billing and API pricing pages, as of August 2026:
| Plan | Price | What you get |
|---|---|---|
| Free | $0 | 4 generations, basic inputs |
| Standard | $20/mo | 20,000 credits (~12 worlds); multi-image and video input, editing, export |
| Pro | $35/mo | ~25 worlds; adds Expand, HQ mesh export, commercial usage rights |
| Max | $95/mo | ~75 worlds (tier reported by press; confirm on the live pricing page) |
The World API is pay-as-you-go: $1 buys 1,250 credits ($5 minimum). That works out to roughly $0.12–0.20 per Draft world, ~$1.20–1.28 for Marble 1.0 or 1.1, and up to ~$2.48 for a maximum-size 1.1 Plus world. A high-quality mesh export runs about $2.80; PLY splat export is free. App and API credits are separate pools; subscription credits don't roll over, and top-ups last one year.
What are Marble's honest limitations?
- Worlds are environments, not animations. Marble generates static spaces — no characters or simulated dynamics inside a generated world. Life gets added downstream, in a game engine or renderer, after export; World Labs' own materials frame Marble as environment generation.
- Expansion can show seams. Regions grown with Expand can show artifacts relative to the originally generated core.
- Physics is your job. The collider mesh enables physics in an engine; the world itself simulates nothing.
- Commercial rights are gated. App-generated worlds require the Pro tier or above for commercial use; API builders should review World Labs' current output-licensing terms before shipping.
Why should filmmakers care about Marble?
Because a walkable set changes previsualization. A location scout's photo — or a written premise — becomes an environment a director can move through, frame shots inside, and revise before anyone builds or books anything. Two Marble outputs slot directly into video workflows: the camera-path video render, which turns a world into establishing shots and b-roll with exact, repeatable camera moves, and the mesh export, which drops the set into Unreal or Blender for virtual production. For AI-first filmmakers, that solves the consistency problem video models struggle with: generate the location once, then shoot it from as many angles as the edit needs. invideo's storyboard generator pairs naturally here — plan the sequence, then pull establishing footage from a world that never changes between shots.
Frequently asked questions
Is Marble a video generator? No. Marble generates persistent 3D environments (Gaussian splats plus meshes). It can render videos from a world via camera paths, but the underlying output is a scene you can re-enter — a video model produces frames and forgets the space.
Can I export Marble worlds to Blender or Unreal? Yes. Worlds export as Gaussian splats (.spz/.ply) and GLB meshes, and World Labs publishes official guides for Blender, Unreal, Unity, and Houdini, as of August 2026.
Can I use Marble worlds commercially? On the app side, commercial rights come with the Pro plan ($35/mo) and above. For API usage, check World Labs' current licensing terms — the fine print matters if outputs ship in paid work.
How long does Marble take to generate a world? About 20 seconds for a Draft world, ~30 seconds for a first navigable preview, ~5 minutes for a full-quality world, and up to an hour for a high-quality mesh export, per official docs.
Can I view Marble worlds in VR? Yes — every world has a shareable VR link, with press coverage naming Quest 3 and Apple Vision Pro support.
Which Marble version should I use? Draft for cheap iteration, 1.1 for standard fixed-cost worlds, 1.1 Plus for the largest environments with auto-expansion. 1.0 remains as the legacy model.
Where can you use Marble today?
All four Marble variants — 1.1 Plus, 1.1, 1.0, and 1.0 Draft — are available inside invideo, the AI video platform that gives serious creatives every major model in one place. They live in the agent's 3D tab — as of August 2026, the only tab of its kind in a mainstream creation product — with text-to-3D-worlds sitting alongside the video and image models you'd use to shoot inside them. Block out a location in Marble, then cut it into a sequence with the AI video generator or extend the style into motion with the AI animation generator, without leaving one workspace.
Version history: Marble GA (1.0) Nov 12, 2025 · World API Jan 21, 2026 · Marble 1.1 & 1.1 Plus Apr 2, 2026; prior models renamed 1.0 / 1.0 Draft.