LightGenBench

A Benchmark for 3D Emission Generation

Dongchen Yang Xingguang Yan Manolis Savva

Simon Fraser University

Paper · soon arXiv · soon GitHub 🤗 Hugging Face
A collection of LightGenBench shapes in a dark room with their emission on: a glowing ring, lamps, robots, a clock, a laptop and a ray gun. The same shapes with their emission off: the room is dim and nothing glows.

Previous 3D generation works create assets that reflect light but cannot emit it. Our work generates the emission for an object, so it can light up the scene. LightGenBench holds 36,826 artist-made emissive assets in 106 categories, curated from TexVerse, each with albedo, metallic, roughness, opacity and emission.

Data representations

We cover the main families of 3D generation: the UV, voxel and multi-view domains. For each family we include its data representation: UV atlases, O-Voxels and multi-view images.

An input asset and its three data representations: a UV atlas, O-Voxels and multi-view images.

Baselines

For the UV domain, we use TEXGen-Emission; for the voxel domain, TRELLIS.2-Emission; for the multi-view domain, Hunyuan3D-Emission. Since one key challenge for emission generation is to localize the emissive regions, we also include a segmentation-based approach, SegviGen-Emission, which works on voxels.

The baseline pipelines: TEXGen-Emission on the UV atlas, TRELLIS.2-Emission and SegviGen-Emission on the O-Voxels, and Hunyuan3D-Emission on the multi-view images.

Results

Emission generation remains a challenging task: the best baseline, TRELLIS.2-Emission, reaches an IoU of only 0.271.

Input image
GT
TEXGen-Emission
TRELLIS.2-Emission
Hunyuan3D-Emission
SegviGen-Emission
Input image: a helmet.
The helmet, ground-truth emission, emission on.
The helmet, TEXGen-Emission, emission on.
The helmet, TRELLIS.2-Emission, emission on.
The helmet, Hunyuan3D-Emission, emission on.
The helmet, SegviGen-Emission, emission on.
Input image: a ring.
The ring, ground-truth emission, emission on.
The ring, TEXGen-Emission, emission on.
The ring, TRELLIS.2-Emission, emission on.
The ring, Hunyuan3D-Emission, emission on.
The ring, SegviGen-Emission, emission on.
Input image: a saber.
The saber, ground-truth emission, emission on.
The saber, TEXGen-Emission, emission on.
The saber, TRELLIS.2-Emission, emission on.
The saber, Hunyuan3D-Emission, emission on.
The saber, SegviGen-Emission, emission on.

Interactive comparison

An interactive 3D comparison on test examples: points sampled on the original asset's surface. Drag to orbit.

Video

Download

LightGenBench is available on Hugging Face, split into 36,426 train, 200 validation and 200 test shapes. Submit the short access form on the dataset page once (access is granted right away), then log in and download:

hf auth login
hf download 3dlg-hcvc/LightgenBench --repo-type dataset --local-dir lightgenbench

BibTeX

@inproceedings{lightgenbench2026,
  title  = {LightGenBench: A Benchmark for 3D Emission Generation},
  author = {Yang, Dongchen and Yan, Xingguang and Savva, Manolis},
  year   = {2026}
}

Acknowledgements

To be added.