VoroTracing

Differentiable Voronoi Ray Tracing Beyond Rasterization Speeds

Paper · Code · Project Page · Web Viewer

Pretrained VoroTracing scenes for the seven Mip-NeRF 360 scenes used in the paper.

Contents

checkpoints/mipnerf360/<scene>/model.pt      # trained scene, loadable with the VoroTracing code
checkpoints/mipnerf360/<scene>/config.yaml   # training config used for that scene
mipnerf360/<scene>.foam                      # export for the web viewer

Scenes: bicycle, bonsai, counter, garden, kitchen, room, stump. Each checkpoint holds 2M Voronoi cells and is about 3.4 GB.

Usage

Set up the code and place the Mip-NeRF 360 dataset under data/mipnerf360/ in the repository root. Then, from the repository root, download all seven checkpoints (~23.5 GB, without the .foam viewer files):

hf download bertaveira/vorotracing --include "checkpoints/*" --local-dir .

or a single scene:

hf download bertaveira/vorotracing --include "checkpoints/mipnerf360/garden/*" --local-dir .

Evaluate:

uv run python benchmark_eval.py --base-dir checkpoints/mipnerf360

Or open the scene in the local viewer:

uv run python viewer.py --ckpt-dir checkpoints/mipnerf360/garden/model.pt

Results

Test-set metrics from benchmark_eval.py with default settings (fp16) on an RTX 5090:

Scene	PSNR	SSIM	LPIPS	FPS
bicycle	24.28	0.7107	0.2975	559
bonsai	33.38	0.9378	0.2194	559
counter	29.45	0.8908	0.2428	589
garden	26.97	0.8251	0.1660	826
kitchen	31.52	0.9176	0.1600	595
room	31.21	0.9145	0.2638	585
stump	25.93	0.7324	0.2926	649
Mean	28.96	0.8470	0.2346	623

Citation

@article{taveira2026vorotracing,
  author  = {Taveira, Bernardo and Lindstr{\"o}m, Carl and Johnander, Joakim and Kahl, Fredrik},
  title   = {Differentiable Voronoi Ray Tracing Beyond Rasterization Speeds},
  journal = {arXiv preprint arXiv:2608.17682},
  year    = {2026}
}