Scenario has published scenario-labs/3d-collection on Hugging Face: 1,342 AI-generated 3D examples, about 39 GB with previews and source inputs, under CC BY 4.0 with attribution to “Scenario”.
It was created on 19 October 2026 and last touched the same day. It has since been climbing — likes rose from 118 to 199 in the day to 25 October, the largest likes gain of any dataset in that window, which is how it reached us.
The licence is the first reason to care. The second is that this is the best-documented creative dataset we have read this month, and the documentation is checkable rather than reassuring.
What is actually in it
We counted the repository’s files against the README’s claims rather than taking either on trust. Under assets/:
| Format | Count |
|---|---|
| GLB | 1,271 |
Blender (.blend) | 137 |
| SPZ (Gaussian splat) | 41 |
| PLY | 23 |
| VOX | 21 |
| FBX | 7 |
Those are the README’s figures exactly. There are also 1,342 preview JPEGs — one per example, matching the stated example count, across 1,342 asset directories.
A separate inputs/ tree holds 852 files: 778 PNG and 47 JPEG source images, plus a handful of 3D files people supplied as starting points — 8 GLB, 7 Blender, 2 PLY, 2 glTF, 2 OBJ and one USDZ. There are 830 input directories against 1,342 examples, so roughly 512 examples were generated without a source asset and about 830 were conditioned on an image or a model.
If you count file extensions across the whole repository you get slightly higher numbers than the README gives — 1,279 GLB rather than 1,271, 144 Blender rather than 137. That is not an error in either direction. The README is counting outputs; the extra files are user-supplied inputs that happen to be the same formats. Split by directory, the output counts match to the file.
The paperwork
The metadata is a single Parquet file, data/train-00000-of-00001.parquet, one row per example. What each row carries is the part worth reading twice:
inputs: structured prompt, effort, generation settings and source assets (IDs, relative paths and SHA-256 hashes).outputs: structured output formats, roles, IDs, relative paths and SHA-256 hashes.
A SHA-256 for every input and every output. That means you can verify that the file you downloaded is the file the row describes, and that the source image named in a row is the source image on disk. Very little published creative data lets you do that.
The settings field is JSON text rather than typed columns, and the README says why: parameter names and value types vary between the models used, so they preserved the original values instead of flattening them. It adds, specifically, that missing values appear as null and no defaults are inferred. That is a small sentence with real consequences — an inferred default is a fabricated record of how an asset was made.
They published the validation, including the failure
Most dataset cards assert integrity. This one reports a test:
Outputs: 1,271 GLB, 137 Blender, 41 SPZ, 23 PLY, 21 VOX and 7 FBX. Structural checks passed; all 137 Blender outputs loaded with auto-execution disabled. One Blender scene references two missing textures. These checks do not certify rendering or animation quality.
Three things there. The checks are named. The one defect across 1,342 examples is disclosed rather than quietly fixed or left to be discovered. And the scope of the claim is bounded — structural validity is not “this will render nicely”.
The phrase “with auto-execution disabled” is the detail a working artist should notice. A .blend file can carry Python that runs when the scene opens. Validating 137 downloaded Blender files with auto-execution off is both the correct way to test them and a quiet instruction about how to open them. If you pull these into a studio, open them the same way.
The licence split is the thing Gradium got wrong
A further 393 examples made with Tencent Hunyuan models are not in this repository. They are published separately as scenario-labs/3d-collection-hunyuan, which we checked: it is tagged license:other, it is gated behind access approval, and it holds 1,144 files.
That is the right way round, and it is the exact opposite of the pattern we documented at Gradium yesterday, where a family was presented as Apache-2.0 and only one of its seven models actually was. Here the differently-licensed outputs have been lifted out, labelled honestly as “other”, and put behind a gate, so the CC BY 4.0 on the main collection means what it says for all 1,342 examples in it.
There is also a LICENSE file at the repository root, not merely a tag. It reads:
Except where otherwise stated, this collection is licensed under the Creative Commons Attribution 4.0 International license (CC BY 4.0). Attribution: Scenario.
We checked that the file and the tag agree. They do. That sounds like a low bar until you have seen how often it is missed.
What you can do with it
CC BY 4.0 permits commercial use and modification with attribution. So: game assets, backgrounds, previs, kitbashing, look development, and training — a permissively licensed 3D corpus with per-asset provenance is unusual enough that the training use is probably the most significant one.
The README’s own caveat is the one to carry: all assets are AI-generated. These are model outputs, selected by Scenario, not scans or hand-modelled work, and the copyright status of AI-generated output remains unsettled in several jurisdictions regardless of what licence sits on top. The attribution requirement is straightforward; the question of what exactly Scenario is licensing is not one a dataset card can settle.
The 41 SPZ files are worth a separate look if you have not handled Gaussian splats before — they sit oddly alongside meshes in the same collection, and the two formats do not substitute for each other.
Sources
- scenario-labs/3d-collection — Hugging Face
- scenario-labs/3d-collection-hunyuan — Hugging Face
- CC BY 4.0 licence deed
- Scenario