license: cc-by-4.0
pretty_name: grow3d training data (CC-licensed Objaverse subset)
tags:
- 3d
- image-to-3d
- text-to-3d
size_categories:
- 100K<n<1M
grow3d data
Training data of grow3d (image/text-to-3D geometry): objects from the ObjaverseXL-Sketchfab part of TRELLIS-500K, restricted to assets whose Sketchfab license allows redistribution of derived data: CC0, CC-BY 4.0, CC-BY-SA 4.0. Non-commercial (CC-BY-NC*) and no-derivatives assets are excluded.
Licensing and attribution
Every file derived from an asset is distributed under that asset's license. meta/attribution.csv lists,
for every included object: sha256 (the id used in all file names), Sketchfab uid and url, the model
name, the author and author_url, and the license (cc0, by = CC-BY 4.0, by-sa = CC-BY-SA 4.0).
Assets under CC-BY-SA keep that license (ShareAlike) for their derived files. Licenses come from the
Objaverse annotations (ODC-By).
Contents
data/shard_*.tar (+ .json with the object list). Per object sha:
| path | content |
|---|---|
lat512/<sha>.npz |
TRELLIS.2 SC-VAE shape latent (f16c32) at 512^3: sparse coords (N,3) on a 32^3 grid, feats (N,32) |
lat1024/<sha>.npz |
same at 1024^3 (64^3 grid) |
cond/<sha>/<i>.webp, views.json |
12 RGBA conditioning renders (base colour, Lambert, nvdiffrast) and their cameras (TRELLIS.2 sampling) |
gt/<sha>.npz |
ground-truth mesh (validation objects only) |
Meshes are normalised to [-0.5, 0.5]^3 in Blender's Z-up frame, like TRELLIS.2. Captions are not
duplicated here: use the captions column of TRELLIS-500K.