3D Data Licensing for AI,
Robotics & Spatial Computing

Most 3D training data is inconsistent by design

Every major stock 3D marketplace is an aggregator - thousands of independent artists, each with their own scale conventions, topology habits, UV workflows and material setups. For a single render, that inconsistency is a minor annoyance. For a machine learning pipeline, it's a structural problem: a model trained on data with no consistent standard learns the noise along with the signal.

If you've ever had to filter, re-scale, re-topologize or clean a "licensed" 3D dataset before it was usable for training, you already know the cost.

The inconsistency is not random, and it is not a quality problem on the authors' side. Each of those artists is producing a correct model for their own purpose. The four things that break in aggregate are always the same:

• Scale drifts between authors. One models a door handle at real-world millimetres, another at a convenient unit size. Train a size or grasp estimator on both and the label distribution is meaningless.
• Pivots sit wherever the render needed them. Fine for a still image; useless the moment an object has to be placed, rotated or articulated in a simulator.
• Objects go unnamed. A scene arriving as Object001 through Object184 cannot be segmented automatically. Someone labels it by hand, per asset, forever.
• Materials follow different conventions. Mixed PBR and legacy shader setups mean appearance varies with the importer, not with the object - so a vision model learns the pipeline instead of the thing.

None of these are visible in a thumbnail. They surface after the licence is signed and the assets are in the pipeline, which is why the cleaning cost is almost always discovered late.

One studio. One standard. 90,000+ models.

3D Molier International is not an aggregator. Every model in our catalog - vehicles, aircraft, naval, medical and anatomy, animals, characters, architecture, industrial equipment and more - is produced in-house, by the same studio, to the same production standard, for over 20 years.

• Real-world scale on every model - no rescaling pass needed before training
• Clean topology across the entire catalog - no per-source cleanup step
• Complete UV unwraps and PBR materials as standard, not the exception
• Consistent metadata - category, tags and format structured catalog-wide

That consistency is the product. It's what lets you license at scale without paying a hidden cleaning tax on every asset.

Who checks it, and against what

The usual objection to any large 3D catalog is that volume and consistency pull against each other. They do - unless the checking is structural rather than occasional.

Every model, without exception, goes through our own QA before it is accepted. Not a sample. Not the ones flagged as risky. Every one.

• A technical reviewer checks what a pipeline breaks on: topology, real-world scale, object and material naming, pivots, UVs, stray geometry.
• A visual reviewer checks what a render exposes: silhouette, proportion, material response, whether the thing reads as the object it claims to be.
• A project manager owns the model end to end and decides whether it is finished.

Part of the production capacity is contracted rather than salaried, and that changes nothing about the output: a contributed model is not accepted because someone made it, it is accepted because it passed the same review as everything else. The standard is enforced at the gate, not requested at the brief.

This is the structural difference from licensing a stock marketplace's catalog. A marketplace licenses contributions - assets it did not produce, accepted under a listing policy rather than a production standard, from thousands of independent authors who each have their own conventions. It can guarantee that a file exists and is legally clear. It cannot guarantee that two files were built the same way, because it never built either of them. We built all of ours, and we checked all of them.

What the catalog holds, by category

Every supplier of 3D training data publishes one aggregate number. Here is the actual breakdown, counted from the live catalog on this site rather than estimated. Each row links to the category so you can inspect the models before you ask for a quote.

These are product pages, not marketplace listings. One model offered for 3ds Max, for Maya and in three colours is five listings on a marketplace and one page here, with every version listed on it. Licensed volume is counted in source models, so ask for the listing count if that is the number your procurement needs.

Ground truth comes with the geometry

Synthetic data is only worth generating if the labels are trustworthy. In an aggregated corpus the labels have to be reconstructed - measured, guessed or annotated by hand - because the source assets never agreed on anything. Here they come out of how the models were built in the first place.

• Real-world scale gives you metric labels. Object dimensions, distances and physical plausibility in a simulator are correct without a calibration pass, because nothing needs rescaling to begin with.
• Named objects and materials give you segmentation. Per-object and per-material masks can be rendered directly. No manual annotation pass, no labelling vendor.
• Consistent pivots give you pose and articulation. Objects can be placed, rotated and grasped programmatically, which is what makes large-scale scene randomisation possible at all.
• Clean topology gives you stable geometry labels. Depth, normals and surface curvature stay well defined under subdivision and decimation, so the same asset serves both a high-fidelity render and a real-time simulation.
• Catalog-wide metadata gives you queryable structure. Category, tags and format are structured the same way across the catalog, so a subset can be specified as a query rather than assembled by hand.

This is the practical difference between licensing assets and licensing training data. The geometry is the same either way; what you are paying for is that the labels come with it rather than after it.

How licensing works

• 1. Tell us your dataset requirements - category, volume, format, metadata needs
• 2. We curate or produce to match - draw from 90,000+ existing models, or fill gaps with custom production to the same standard
• 3. Licensed delivery - structured, categorized, ready for your pipeline; full-catalog or custom-subset, exclusive and non-exclusive terms on request

Pricing is enterprise / custom - built around your dataset scope, not a one-size price list.

What the licence covers

Two forms, the same catalog behind both. The difference is not what you may do with the data - it is whether anyone else may hold it at the same time.

What no licence covers

• Redistributing the source assets. The geometry itself may not be resold, sublicensed or published as downloadable assets, in any format, including as part of another dataset.
• Passing the files to parties not named in the agreement. Subsidiaries and research partners are covered when they are written into it, not by default.
• Claiming authorship of the catalog. What you build from the data is yours; the models themselves stay ours.

Everything else is a term of the specific agreement rather than a fixed rule: how long the licence runs, which subset it covers, which entities it names, what happens to models we publish during the term, and how delivery is scoped. Those are the things we agree in writing before anything is transferred - ask and we will send the full text.

Let's talk about your dataset

Tell us what you're training and what you need. We'll scope a dataset from our catalog or propose custom production to fill the gap.