Gaussian Splat Objects Dataset

Objects for experiments as 3D Gaussian splats, scanned and generated. Free to download and use.

Marcel Padilla
ETH Zürich
GitHub Data on GitHub 📑 Cite

Viewer

Drag to orbit, scroll to zoom, right-drag to pan. On a touch screen a sideways swipe turns the object and two fingers zoom, while an up or down swipe scrolls the page; full screen hands the object every gesture.

Three ways of making a splat, in one place. Every object says which one it is and what license it carries, and I know exactly where each one came from.

PyPI package

$ pip install splatset
Show what you can do with it
import splatset s = splatset.load("plant") # downloads once, caches, verifies, decodes s.positions # (113648, 3) float32 xyz s.colors # (113648, 4) uint8 RGBA # no URLs anywhere: say what you want and it is fetched splatset.load("lucy", lod="10k") # one object, small tier splatset.load_random(source_method="capture") # any real scan splatset.find(source_method="mesh2splat", lod="10k") # every 10k mesh object for s in splatset.load_all(category="object"): # loop the whole set print(s.id, len(s))

Filtering (splatset.find(source_method="capture"), splatset.find(tags=["plant"]), splatset.find(open_license=True)) runs offline against a bundled inventory; only the splat bytes are downloaded. The objects made from meshes ship at 10k / 100k / 500k / 1M gaussians, and every call takes lod= (or "min" / "max"). There is a command line too: python -m splatset list --source mesh2splat --lod 10k. Full docs on GitHub.

Gallery

Filter by source below. The objects from geometry ship at four levels of detail; pick the one to download in each row, and note that their colour encodes it (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

Pick the kinds you want. The zip holds the splats, their meta.json, and a LICENSES.txt with the terms for exactly what you took.

Loading the gallery…
Filter by rights
⬇ Whole repository (.zip)

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

The license depends on how an object was made (it is shown under each Source in the table, and per object in its meta.json):

  • Scanned objects are CC-BY-4.0. Use them for anything, including commercially. You must credit the author.
  • Geometry and rendered objects inherit the license of the mesh they came from. Converting a mesh to splats does not launder its license. Of the 41, eight are CC0 or public domain and ask for nothing (Spot and Bob by Keenan Crane, the Utah teapot, the primitives); nineteen are CC-BY and want their own author credited rather than Marcel Padilla; six are the Stanford subset below; and eight are NonCommercial. The rendered bunny is CC-BY-3.0, from the Stanford Bunny STL by Makerbot via Wikimedia Commons; credit that original author as well.
  • The Stanford subset is free to use, but is not Creative Commons. stanford_bunny, armadillo, dragon, happy, lucy and xyzrgb_dragon come from the Stanford 3D Scanning Repository, which permits use including commercially and in derivative works, but asks that you acknowledge the source and do not misrepresent the data. Credit Stanford University Computer Graphics Laboratory (and XYZ RGB Inc. for the XYZ RGB dragon).
  • Some mesh-derived objects are NonCommercial. Eight objects from the odedstein-meshes collection (cow, house, tower, well, mushroom, strawberry, brucewick, falconstatue) came from meshes licensed CC BY-NC, so you may not use them commercially, and falconstatue is additionally ShareAlike. They are unticked by default in the download tool above. Everything else from that collection is plain CC-BY or CC0 and free for any use with credit.
  • Generated objects are provided as-is, with no warranty of rights. They were invented by an AI image-to-3D model whose training data is undisclosed, so their copyright status cannot be warranted and they are not under CC-BY. You are responsible for clearing any rights for your use.
  • The loader code is 0BSD. No attribution required.

Credit line you can paste for a CC-BY object: Gaussian splat by Marcel Padilla, from github.com/marcelpadilla/splats, licensed CC-BY-4.0. For an object derived from someone else's mesh, credit that author too: each row's licence carries the exact wording, and it is in the object's meta.json as attribution. The license covers the capture, render or generation and the reconstruction, not the design of any object depicted.