The Neural Network That Creates Entire 3D Worlds: Testing Marble by World Labs

Registration and first login  

Marble runs in the browser and does not require installation. You can sign in with a Google account, via GitHub, or by email — on the start screen, even before you log in, you can already see a gallery of other people’s worlds with style filters: Stylized, Realism, Interior, High Quality, Fantasy, Sci-Fi.

After signing in, the service asks you to choose a username (2 to 20 characters, letters, numbers, and underscores only) and accept the terms of use.

Next comes a short survey about why you are here. The options explain the tool’s target audience better than any description: film and video production, game development, robotics, design and architecture, art, and AR/XR. You can skip this step — it does not affect functionality.

After this screen, the workspace opens with a built-in tour.

Interface  

Components  

Step 1: Input Field

The single “Imagine a world…” field at the top of the screen accepts text, images, video, and panoramas.

Step 2: Model Selection

The model selector is located next to the Create button. Here you can also see the input split into two tabs: 2D input for images, video, and panoramas, and 3D input, labeled Beta, for assembling a scene from models and primitives.

Step 3: Create and Create & edit

The arrow next to the Create button opens the creation and editing mode. Standard Create generates a world immediately from the entered data, while Create & edit launches a multi-step flow where you can tweak the intermediate result before the final generation. For your first world, the first option is enough.

Step 4: Generation Settings

The sliders icon on the right side of the input field opens additional parameters: Draft mode for quick rough iterations, a custom seed for reproducible results, and world privacy settings. Draft, however, is only available starting with the Standard plan — on the free plan, the icon is there, but the mode is not.

Step 5: 3D Input Mode

The second tab lets you assemble a scene from 3D models and primitives, and then generate a world from it. This is the most precise way to control composition and layout, but the mode is still in beta.

Step 6: Help

You can restart the tour at any time from the help menu in the bottom-left corner — that’s also where you’ll find the documentation and the bug report form.

Plans, Credits, and What’s Available for Free   

World generation on Marble 1.1 costs a fixed 1,500 credits, while Marble 1.1 Plus costs the same 1,500 plus 300 for each additional space block the model adds on its own.

The subscription is split into four tiers, and they differ not only in the number of credits.

Free ― $0. 7,000 credits, up to 4 world generations. All the core input methods are already available here — text, a single image, multiple images, and a 360 panorama — along with an advanced creation mode with editing, Chisel mode for controlling the 3D layout, and expanding existing worlds.

Standard ― $20 per month. 20,000 credits, up to 12 generations. Adds video input, export of splat files, 360 panoramas, and lightweight collider meshes, community asset uploads, Draft mode for fast rough iterations, and the ability to buy extra credits.

Pro ― $35 per month. 40,000 credits, up to 25 generations. On top of Standard — higher rendered video quality, export of high-resolution textured meshes, and commercial rights to the worlds you create.

Max ― $95 per month. 120,000 credits, up to 75 generations, the full Pro feature set with no add-ons.

You can buy extra credits starting with Standard: $5 for 5,000, $20 for 22,000, and $50 for 63,000. This is not the same as API credits — usage via the programmatic interface is billed separately on the developer platform. For larger teams, World Labs offers an Enterprise plan on request.

Two details worth keeping in mind.

1. Unused subscription credits do not roll over to the next month, and top-up packs, according to the documentation, last for a year — so it makes sense to plan generations within a month.
2. Credits are spent on more than just world generation, so “up to 4 generations” on the free plan is a ceiling, not a guarantee: four worlds at 1,500 credits each leave you with just 1,000 out of the initial 7,000 for everything else.

The practical takeaway for anyone who just wants to try it: the free plan is enough to go through the entire flow below — from prompt to walking around the world. You only have to pay when you need to take the result outside Marble.

Building a world   

Next — an end-to-end scenario on the free plan: an abandoned mountaintop observatory from a single paragraph of text. A prompt for a world model isn’t written the same way as one for an image generator: what matters here isn’t camera angle and lighting, but the layout of the space — what’s placed where, where you can walk, and what you can see in the distance. The model builds an entire room, not a single frame, so the description is structured “room by room.” 

Prompt

An abandoned mountaintop observatory at dusk. A circular main hall under a cracked dome, with a large telescope tilted toward a gap in the dome. Around the hall: a control desk covered in dust and paper charts, a spiral staircase leading up to a narrow observation balcony, and an open doorway leading to a corridor with broken windows. Through the windows and the dome gap, a ridge of snow-covered peaks and a deep blue evening sky. Cold light, volumetric haze, scattered debris on the floor, ivy growing through the cracks. Photorealistic, wide open interior space that can be walked through. 

Paste the text into the input field, keep the Marble 1.1 model selected, and click Create.

In the “Your worlds & assets” section, the world appears as a card with the status Generating. You can also see what the service did with the prompt: it generated the title itself — Abandoned Mountaintop Observatory Dusk — rewrote the scene description in its own words, and added tags for the model used and the seed. The seed is useful if you need to recreate the same world or generate a variation. 

Generation takes a few minutes. The finished card gets a preview, and it immediately shows the key point: the layout from the prompt is preserved — a circular hall under a dome, a tilted telescope in the center, and the control desk in the foreground.

One paragraph was enough to get not an image, but a room. You need to check it from the inside. 

A walk through a finished world 

The open world takes up the entire screen, and the controls feel like a 3D editor: the mouse rotates the camera, the wheel zooms, and dragging moves the viewpoint. On the right is a panel with the world’s source data: the text prompt, the panorama it was built from, a draft, and the world itself, with an Expand button below to push the boundaries outward. At the bottom is a toolbar for editing, switching styles, composition, export, screenshots, and a sharing link. In the bottom-right corner there’s a separate Open VR button — the same world opens in a headset.

From the prompt, the model nailed almost everything: the round hall, windows with a mountain ridge beyond the glass, a spiral staircase up to the balcony, a desk with papers, ivy on the walls, and debris on the floor. It improvised the walls, though — adding ornamentation that wasn’t in the description. Spatial consistency works as advertised: you can walk to the far wall, turn around, and come back — the furniture and staircase stay in the same places.

But when you move in closer, the picture changes. The telescope, the scene’s main object, looks convincing at a distance, but up close it falls apart into a smeared dark blot with no clear geometry. The same goes for thin elements: the balcony railing and the telescope stand warp and drift. That’s a direct consequence of how the output is built: the world is assembled from Gaussian splats reconstructed from an image, not modeled as discrete objects, so the detail holds only up to the distance from which the scene was “captured.”

The practical takeaway from the walkthrough is simple: you should view a world like this from the distances it was designed for — wide shots, walk-throughs, and backgrounds. For previs, level blocking, or a backdrop, it’s usable material, and that’s exactly the form it gets exported in.

What to do with a finished world next 

Export is what makes Marble more than a theme-park ride, and it’s the first feature that runs into the pricing tier. Standard unlocks splat files, 360 panoramas, and lightweight collider meshes, while Pro adds high-resolution textured meshes, better video quality, and commercial rights to the output.

From there, the scene lives on like a normal asset. You can refine it manually — for example, in a Blender + Claude workflow — or send it into simulation: NVIDIA describes importing Marble worlds into Isaac Sim to train robots. The second route is to stay inside Marble: expand the boundaries with the Expand button, assemble a shared scene from multiple worlds in the composer, or switch styles without touching the geometry. Expansion also works on the free plan, so you can grow a world piece by piece without paying anything.

The choice between the two routes depends on the task: for previs, the built-in tools are enough, while for production, the world gets exported and refined externally. 

Limitations 

The tool works, but you can see its limits from the very first world. Some are inherent to the technology, and some come down to plan restrictions.

1. Detail holds up only at medium and long distances. Up close, the geometry falls apart, and small objects and text get distorted. For a final asset, the output needs manual cleanup.
2. On the free plan, you can create, edit, and expand a world, but you can’t take it out: exporting splat files and collider meshes starts with Standard, while high-resolution textured meshes and commercial rights to the output start with Pro.
3. Unused subscription credits don’t roll over to the next month, and a single Marble 1.1 generation costs a fixed 1,500 credits, regardless of whether you like the result or not.
4. The 3D input mode, which gives the most precise control over layout, is still labeled as beta.
5. The prompt sets the structure, but doesn’t guarantee it. The model fills in details at its own discretion — and the only way to reproduce a specific result is with a saved seed.

None of this changes the main point: one paragraph of text really is enough to get a navigable 3D space. The question is which stage of the workflow it’s good for.

Who Marble Is Useful for Right Now 

The list of roles from the registration survey — film, game dev, robotics, architecture, AR/XR — captures the audience accurately. These use cases all have one thing in common: what you need is not a final asset, but a fast space to validate an idea. Storyboard a scene, rough out a level layout, show a client the volume of a room, assemble an environment for training a robot — in every case, a world generated from a single paragraph saves days of manual work, and its shortcomings at this stage do not get in the way.

Marble is not suitable for final imagery yet, and World Labs is not promising that either. But it shows what working with 3D space looks like when you do not model it, but describe it in words.