Gaussian Splat Objects Dataset

Objects for experiments as 3D Gaussian splats, scanned and generated. July 2026.

Marcel Padilla
ETH Zürich
GitHub Data on GitHub 📑 Cite

You can interact with the viewer (drag, zoom, etc).

The dataset contains Gaussian splatted objects that are...

PyPI package

$ pip install splatset
Show what you can do with it
import splatset s = splatset.load("plant") # downloads once, caches, decodes s.positions # (113648, 3) float32 xyz s.colors # (113648, 4) uint8 RGBA # no URLs anywhere: ask for what you want and it is fetched splatset.load("lucy", lod="10k") # pick a resolution splatset.find(source_method="capture") # search the set splatset.load_all(category="object") # loop over all of it

Objects made from meshes come in 10k / 100k / 500k / 1M gaussians. Full docs on GitHub.

Gallery

Filter by source below. The objects from geometry come in four levels of detail. Pick one in each row. Their colour shows which (10k red, 100k yellow, 500k green, 1M blue). Every row carries its own license. Click a thumbnail to open that object in the viewer.

Download

Make your selection and download the zip with auxillary files (meta.json, LICENSES.txt).

Loading the gallery…
Rights
⬇ Whole repository (.zip)

To download everything, 2.7 GB, run download_all_splats.py.

Citing

If these are useful in your work, a citation is appreciated (and required by the data license): 

Loading citation…

License

per object, most are free to use with credit

Every object carries its own license, shown on its row in the gallery and in its meta.json. It depends on how the object was made.

  • Scanned: CC-BY-4.0. Use them for anything, credit Marcel Padilla.
  • From geometry: the license of the source mesh. Most are CC-BY or CC0, eight are NonCommercial, and six come from the Stanford 3D Scanning Repository. Credit the mesh author, not me.
  • Generated: rights unclear. An AI model invented them, and nobody knows what it was trained on, so nobody can promise these are free of copyright held by someone else. Checking that your use is in the clear is on you.
  • The code: 0BSD. No attribution required.

The credit line for a single object is already written in its meta.json, under attribution.