LightGenBench: A Benchmark for 3D Emission Generation

     

LightGenBench is a dataset of emissive 3D objects for training and evaluating emission texture generation. Its key characteristics:

1. Scale and source: 36,826 emissive shapes from TexVerse (Sketchfab models), each with albedo, metallic, roughness, opacity and an emission map. It is curated from the 859k TexVerse models.
2. Three representations: every shape comes as a 512×512 UV atlas, 256³ sparse voxels and six orthographic 512×512 views, the inputs of UV-, voxel- and multiview-based texture generators.
3. Splits: train 36,426, val 200, test 200 shapes; val and test are category-stratified.

Questions or problems: open a discussion on this repository.

Download

The repository holds about 117 GB:

README.md          this card
splits.json        the split of each shape: {"train": [uuid, ...], "val": [...], "test": [...]}
metadata.parquet   one row per shape: uuid, split, shard, category, license, author, author_username, source_url
checksums.sha256   sha256 of every other file, to verify a download
assets/            the images this card shows
data/              <split>/<kind>/<kind>-<shard>.tar, kind: atlas, voxels, multiview, thumbnail

The shapes are sorted based on uuids and cut into shards of 1,000 shapes: train has 37 shards (the last holds 426), val and test have one for each (200 per shard), 39 shards in all. A shard is stored as four tars: the three representations (UV atlas, O-Voxels, multiview images) and the Sketchfab thumbnail; the four tars of a shard hold the files of the same uuids. For example:

data/train/atlas/atlas-00007.tar          <uuid>/atlas.npz
data/train/voxels/voxels-00007.tar        <uuid>/emission_voxels.vxz, <uuid>/pbr_voxels.vxz
data/train/multiview/multiview-00007.tar  <uuid>/multiview/000_albedo.png ... 005_alpha.png, transforms.json
data/train/thumbnail/thumbnail-00007.tar  <uuid>/thumbnail.png

metadata.parquet gives each shape's shard; there are 156 tars in all. Each atlas.npz is a deflate-compressed .npz, which np.load reads as usual.

# the root files, then one kind of file for every split
hf download 3dlg-hcvc/LightgenBench README.md splits.json metadata.parquet checksums.sha256 assets/teaser.jpg --repo-type dataset --local-dir lightgenbench
hf download 3dlg-hcvc/LightgenBench --repo-type dataset --include "data/*/voxels/*" --local-dir lightgenbench
# or everything
hf download 3dlg-hcvc/LightgenBench --repo-type dataset --local-dir lightgenbench
# check, then unpack every tar in place: lightgenbench/<uuid>/<file>, next to splits.json
cd lightgenbench && sha256sum -c --ignore-missing checksums.sha256
for t in data/*/*/*.tar; do tar -xf "$t"; done   # rm -r data/ afterwards to free the tar space

Dataset structure

After unpacking, every shape is one directory named by its TexVerse uuid, and splits.json lists the uuids of each split in train/val/test:

<uuid>/
  atlas.npz
  emission_voxels.vxz
  pbr_voxels.vxz
  multiview/         00N_{albedo,mr,normal,pos,emission,alpha}.png (N = 0..5), transforms.json
  thumbnail.png

Splits

split	shapes
train	36,426
val	200
test	200

Val and test hold 200 shapes each, drawn category-stratified at random; train is every other released shape.

The validation split picks checkpoints; the test set produces published numbers. Read a split as json.load(open("splits.json"))["train"] and a shape's files as lightgenbench/<uuid>/<file>.

Preprocessing

Every representation is built from the .glb file, normalized to [−1, 1]. Emission is the material's emissive texture, or its emissive factor as a color when it has no texture.

Representations

atlas.npz

The UV atlas of one shape: a single .npz holding eight 512×512 maps over the same UV layout, each stored as an array whose dtype and channel count are listed below. color and emission_color hold linear RGB values.

key	dtype	shape	content
occupancy	bool	512×512×1	texel covered by the UV layout
position	uint16	512×512×3	position in [−1, 1] frame
objnormal	uint16	512×512×3	object-space normal
color	uint8	512×512×3	base color
metal	uint8	512×512×1	metallic
rough	uint8	512×512×1	roughness
emission_color	uint8	512×512×3	emission
alpha	uint8	512×512×1	opacity

emission_voxels.vxz, pbr_voxels.vxz

Sparse voxels on a 256³ grid over [−0.5, 0.5]³, in the O-Voxel format of TRELLIS.2, whose o_voxel package reads them. o_voxel.io.read_vxz(path) returns: an int32 tensor of shape N×3 holding each stored voxel's grid index (0–255 on each axis), where N is the number of voxels stored for that shape, and a dict of per-voxel attributes, each a uint8 tensor with one row per voxel. emissive and base_color are linear RGB. The two files carry different attributes over the same list of voxels in the same order. Attributes:

file	attribute	dtype	shape
emission_voxels.vxz	emissive	uint8	N×3
pbr_voxels.vxz	base_color	uint8	N×3
	metallic, roughness, alpha	uint8	N×1

multiview/

Six orthographic 512×512 views (front, left, back, right, top, bottom), rendered with the six fixed cameras of Hunyuan3D-2.1's training example, each with six maps: albedo, mr (metallic-roughness), normal, pos, emission, alpha. The material maps (albedo, mr, emission, alpha) hold linear bytes; normal and pos are geometry maps. The object mask is mr red channel == 255. transforms.json holds the six camera frames.

thumbnail.png

The TexVerse preview images: 36,824 hold JPEG data and 2 hold PNG data. 35,304 are 1920×1080 RGB; the other 1,522 are smaller, down to 256×144, and 277 of those are grayscale.

License

Every shape keeps the license of its source model on Sketchfab (via TexVerse); metadata.parquet gives each shape's license, author and source_url. Credit the authors and filter on the license column for your use: NonCommercial licenses allow non-commercial use only, ShareAlike licenses require the same license on derived work, and NoDerivs licenses do not allow sharing adapted material.

• CC BY: 34,319
• CC BY-NC: 1,341
• CC BY-NC-SA: 459
• CC BY-NC-ND: 381
• CC BY-SA: 264
• CC BY-ND: 49
• CC0: 13

BibTeX

@inproceedings{lightgenbench2026,
  title  = {LightGenBench: A Benchmark for 3D Emission Generation},
  author = {},
  year   = {2026}
}