Scenario has put a large pile of AI-generated 3D assets on Hugging Face for anyone to download. The Scenario 3D Collection holds 1,342 selected AI-generated 3D examples with 1,500 output files, about 39 GB including previews and available source inputs, under a CC BY 4.0 license. Scenario announced it on X on October 23, 2026, saying every example ships with its input image, prompt and settings.
The dataset page was created on October 19 and, at the time of writing, shows roughly 700 downloads. Here is what is in it, how to use it, and where the limits are.
What is in the Scenario 3D Collection?
| Question | Answer (from the dataset card) |
|---|---|
| How many examples? | 1,342 selected examples, 1,500 output files |
| Size | About 39 GB including previews and available source inputs |
| License | CC BY 4.0, credit "Scenario" |
| Formats | 1,271 GLB, 137 Blender, 41 SPZ, 23 PLY, 21 VOX, 7 FBX |
| Metadata | One parquet file, one row per example, with inputs, outputs, preview and download URLs |
| Companion set | 393 Tencent Hunyuan examples, separate repo, own licenses, access approval needed |
The examples were generated on Scenario, a creative-AI platform. Scenario's post names Hitem3D, Meshy, Rodin, Tripo, Claude Opus 5.5 3D and GPT-6 Astra 3D as generators that run on its platform. The dataset viewer also lists other generators in its model column, including Trellis and Trellis-2, Direct3D-S2, PartCrafter, Meta's SAM 3D Objects and World Labs' Marble, and workflows such as image-to-3D, text-to-3D and image-to-splat.
That mix is the point. Most open 3D collections are scraped human-made models. This one is a record of what current generators produce, with the exact settings, so you can compare models on the same kind of input. For background on one of those generators, see our Meshy skills guide for Hermes Agent, and for how Meta's 3D model fits in, our post on Meta's GenIA and SAM 3D Objects.
How do you load it?
The dataset card gives the pattern. The metadata is the single source of truth, and the original assets are separate files you fetch only when needed.
import json
from datasets import load_dataset
examples = load_dataset("scenario-labs/3d-collection", split="train")
example = examples[0]
settings = json.loads(example["inputs"]["settings"] or "{}")
source_assets = example["inputs"]["assets"]
outputs = example["outputs"]
Key columns, per the card: inputs (prompt, effort, generation settings and source assets with IDs, relative paths and SHA-256 hashes), outputs (formats, roles, paths and hashes), preview (an embedded thumbnail for the viewer), and input_files and output_files (download URLs). inputs.settings is JSON text because parameter names vary by model, such as polygon targets, texture size or guidance scale; parse it with json.loads. Missing values are null and no defaults are inferred, so write code that tolerates missing fields.
To pull the actual files, use huggingface_hub.snapshot_download with repo_type="dataset". At 39 GB, filter by pattern if you only want GLB files or a single model's outputs. The Hunyuan companion set requires access approval and authentication.
What can you do with it?
- Compare generators on matched inputs. Because the input image, prompt and settings are included, you can look at how different tools treat similar subjects, rather than trusting a vendor's cherry-picked gallery.
- Prototype game and web scenes. GLB is a common interchange format, and most of the set is GLB. Treat the assets as placeholders or starting points, then clean up topology and textures.
- Build evaluation sets. Hashes for every input and output file make it possible to build a reproducible benchmark or to check that a download is intact.
- Train or fine-tune with care. The license is permissive for reuse with credit, but the assets come from third-party generators with their own terms. Read the model terms before using these files to train a competing generator.
- Test agent pipelines. Scenario also open-sourced a Blender plugin on October 22, which it describes as covering around 500 models in one tab and being built for humans and agents. That is Scenario's own description from its X account; we have not tested it. For a hands-on take on the image-to-3D workflow in general, see our AI image-to-3D rigged game character pipeline.
What are the limits and caveats?
Quality is not certified. The card says structural checks passed and that all 137 Blender outputs loaded with auto-execution disabled. It also says these checks do not certify rendering or animation quality, and that one Blender scene references two missing textures. Expect the usual AI-3D problems: messy meshes, baked lighting, uneven texture resolution.
Blender files can contain scripts. Blender scenes can run Python on open if auto-execution is enabled. Scenario notes it loaded the files with auto-execution disabled. Keep that setting off when you open files from any dataset, and consider inspecting them in a throwaway environment first.
License layers. CC BY 4.0 covers Scenario's release and requires you to credit "Scenario". It does not tell you whether the underlying generator's terms restrict particular uses of outputs, and the card does not show those terms. For the Hunyuan companion, each row's license field names the license that applies, and the card says terms differ by model. If you plan a commercial product, read both layers.
Selected, not random. The card calls the examples "selected", so this is a curated slice. It is not a fair sample of what each model produces on average, and it will flatter the generators somewhat.
Hosting. Preview thumbnails in the viewer use expiring URLs, and file links point to the Hub, so keep your own copy of anything you rely on.
A practical first hour with the dataset
You do not need to download 39 GB to learn something. A sensible order is below.
- Read the metadata only. Load the parquet rows with the
datasetslibrary and look at themodelandworkflowcolumns. Count how many examples each generator contributes, so you know which comparisons the data can support and which it cannot. - Pick a narrow slice. Filter to one workflow, such as image-to-3D, and to GLB outputs. Download a few dozen files rather than the whole repository, using the file URLs in
output_filesor a pattern filter insnapshot_download. - Check integrity. Each input and output carries a SHA-256 hash in the metadata. Verify a handful after download so you know the files are what the card says they are.
- Open them somewhere safe. View GLB files in Blender or a web viewer. For
.blendfiles, keep Python auto-execution off and use a disposable machine or container first. - Score what you see. Write down the things that matter for your use: silhouette fidelity to the input image, texture sharpness, mesh cleanliness, polygon count, and whether the object is watertight if you plan to print it.
- Re-run one input yourself. Because the prompt, settings and source image are included, you can feed the same input to a generator you pay for and compare the result with the stored output.
How to compare generators with it fairly
The matched inputs and settings make this the most useful part of the release, but a few cautions keep the comparison honest.
- Control for the input. Only compare outputs that started from the same or very similar source image and prompt. The dataset has many image-to-3D rows and fewer text-to-3D ones, so text-driven comparisons will be thin.
- Control for settings. Fields like target polygon count, texture resolution and PBR on or off change the look and file size. The
inputs.settingsJSON lets you filter for rows that match. - Remember it is curated. The card says the examples were selected, so you are comparing best-of picks, not average output. A generator with a few brilliant selected examples and a bad typical result would look good here.
- Do not read counts as quality. If one generator contributes many more rows, that reflects what Scenario chose to include, not which generator is better.
- Record versions. The viewer shows several versions of some generators, such as multiple Meshy and Tripo releases. Version differences can matter more than the vendor difference.
Quick answers to questions people will ask
Is it really free to use? The main collection is CC BY 4.0: you may copy, adapt and use it commercially if you credit "Scenario". The Hunyuan set has different terms.
Do I have to credit Scenario in a game? Yes, attribution is the condition of the license. Put it in your credits or asset list, and keep a note of which files you used.
Can I fine-tune a 3D model on it? The license on Scenario's side is permissive, but the files were produced by third-party generators. The card does not show those generators' terms, so check them first if training is the goal.
Why does a Blender file matter? The card lists 137 Blender outputs, which are scene files rather than a single mesh. That is closer to what an artist would edit, but it is also why the file-safety note above applies.
Does the dataset prove 3D generation is production ready? No. It shows what the platform's chosen outputs look like and documents how they were made. Production readiness depends on topology, rigging, and how much cleanup your pipeline can afford, none of which the structural checks cover.
How does this fit the 3D generation moment?
Image-to-3D has moved from demos to pipelines. We have covered GPT-6 Astra's 3D and spatial reasoning community builds and World Labs' Atlas world model, and the common gap is data: people want to see outputs, not marketing reels. A dataset that ships inputs, settings and hashes is a better yardstick than a leaderboard screenshot.
It is also a promotional asset. Scenario sells a platform that runs these generators, and the collection is a showcase of that platform. That does not make the release less useful, but it explains why the sample leans toward what the platform supports, and why you should treat it as a starting library rather than a neutral benchmark.
Bottom line
If you build with 3D, this is a free, well-documented set of AI-generated assets with the provenance most galleries lack. Download a slice, check the licenses, open the Blender files safely, and judge the quality yourself. If you evaluate 3D generators, the matched inputs and settings are the most valuable part.
Related reading
- Meshy skills for Hermes Agent: install guide
- AI image-to-3D rigged game character pipeline
- GPT-6 Astra 3D and spatial reasoning community builds
- World Labs Atlas multimodal world model
- Meta GenIA and SAM 3D Objects
- Hugging Face Carbon-X: an open 14B DNA chat model
Details are from the dataset card and Scenario's posts and accurate as of October 23, 2026.