README.md

---
pretty_name: LightGenBench
license: other
license_name: mixed-cc
viewer: false
tags:
- 3d
- pbr
- emission
- texture
- uv-atlas
- voxels
- multiview
size_categories:
- 10K<n<100K
---

# LightGenBench (preliminary, private)

This repo holds every shape of the released dataset described under "The data", packed as tar
shards per representation.

## Repository layout

```
README.md          this card
splits.json        the split of each shape
metadata.parquet   one row per shape: uuid, split, shard, category, license, author, author_username, source_url
checksums.sha256   sha256 of every file except this card
data/              <split>/<representation>/<representation>-<shard>.tar, shard in 5 digits (data/train/voxels/voxels-00000.tar)
```

Each split's uuids are sorted and cut into shards of 1,000 shapes. A tar holds one
representation's files for one shard (atlas: `atlas.npz`; voxels: the two `.vxz` files;
multiview: `multiview/`; thumbnail: `thumbnail.png`) at paths `<uuid>/<file>`, as in the tree
under "The data". Shard k holds the same shapes in every representation. The `atlas.npz` files
here are stored compressed; the arrays are the same and `np.load` reads them the same way.

## Download and unpack

Downloading needs `hf auth login` with an account that can read `3dlg-hcvc`.

```bash
# the root files, then one representation for every split
hf download 3dlg-hcvc/LightgenBench README.md splits.json metadata.parquet checksums.sha256 --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 and splits.json into ../root
cd lightgenbench && sha256sum -c --ignore-missing checksums.sha256
mkdir -p ../root && for t in data/*/*/*.tar; do tar -xf "$t" -C ../root; done && cp splits.json ../root/
```

## The data

Emissive 3D shapes from TexVerse with their material and emission maps in three
representations: a UV atlas, sparse voxels, and six rendered views. One directory per shape,
named by its TexVerse uuid.

```
README.md
splits.json          {"train": [uuid, ...], "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 are 200 shapes each. Every shape of the earlier paper's validation split and test
set that passed stays in val and test respectively (170 and 178); the rest were drawn
category-stratified at random with seed 42. Train is every other released shape. 232 pass shapes
were left out of train: 91 outside the bake pass (they failed the bake's alignment gate, so they
have no views) and 141 without a thumbnail. None were in val or test.

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 `<root>/<uuid>/<file>`.

### `atlas.npz`

512×512 maps. Material values are linear, stored as `clip(v, 0, 1) · 255`.

| key | dtype | shape | content |
|---|---|---|---|
| `occupancy` | bool | 512×512×1 | texel covered by the UV layout |
| `position` | uint16 | 512×512×3 | position in the canonical [−1, 1] frame, stored as `(v · 0.5 + 0.5) · 65535` |
| `objnormal` | uint16 | 512×512×3 | object-space normal, same encoding |
| `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, both files on the same voxels. Read with
`o_voxel.io.read_vxz(path)`, which returns the coordinates (int32, N×3) and a dict of attributes.
The voxel frame and the atlas frame hold the same normalized object, centered the same way; the
voxel frame is half the scale (a point p in the atlas frame is at p / 2 in the voxel frame).
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, right, back, left, top, bottom; the Hunyuan3D-Paint
training cameras), 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 in Hunyuan3D-Paint's encoding. Outside the silhouette every
map is gray 127.
The object mask is `mr` red channel == 255 (`alpha` is material opacity, not the mask).
`transforms.json` holds the six camera frames.

### `thumbnail.png`

The TexVerse preview image as shipped, under a `.png` name: 36,824 hold JPEG data and 2 hold
PNG data. 35,304 (both PNG among them) are 1920×1080 RGB; the other 1,522 are smaller, down to
256×144, and 277 of those are grayscale. Decode by content (PIL does) and convert to RGB:
`Image.open(path).convert("RGB")`.

### Emission rule

A material's emission is its emissive texture alone (sRGB-decoded to linear) when it has one,
else its emissive factor color, else black. `emissiveFactor` does not scale a texture, and
`KHR_materials_emissive_strength` is not applied.

### How the shapes were selected

1. **Metadata filtering** of the TexVerse GLBs: 83,124 pass (declared emission and a full PBR
   material; 103,762 declare emission).
2. **Heuristic filtering**: the zero-emission check passes 77,569; the copy check (emission is not
   a copy of the base color) passes 73,432.
3. **Duplicate check**: 69,942 (copies with the same geometry, albedo and emission removed).
4. **Data preparation**: atlas and voxels for 70,996 of the 73,432, views for 70,988.
5. **VLM filtering**: 37,058 pass of the 69,942 (a vision-language model judges eight review
   images per shape; shapes whose baked emission map is all zero are rejected).
6. **Released**: 36,826, the pass shapes with every file above.


## Licenses

Every shape keeps the license of its source model on Sketchfab (via TexVerse). Credit the
authors. NoDerivs licenses do not allow sharing adapted material, and NonCommercial licenses
allow non-commercial use only; a public release must filter on the `license` column.

| license | shapes |
|---|---|
| CC Attribution | 34,319 |
| CC Attribution-NonCommercial | 1,341 |
| CC Attribution-NonCommercial-ShareAlike | 459 |
| CC Attribution-NonCommercial-NoDerivs | 381 |
| CC Attribution-ShareAlike | 264 |
| CC Attribution-NoDerivs | 49 |
| CC0 Public Domain | 13 |