TruncGradGS

Synthetic dynamic benchmark released with TruncGradGS: Improved 3D Gaussian Splatting via Truncated Gradient Updates (Pacific Graphics 2026).

The benchmark is six path-traced scenes with large motion: fluids, foliage, night lighting, particles, and caustics. Each scene is 300 frames (10 s at 30 fps) at 1600×900. Cameras are arranged in rigs of several circular arrays. Evaluation holds out 1–4 cameras per scene; the remaining cameras are the training views.

Five scenes are in this repo as Parquet files. The underwater scene (1_4_underwater) is coming soon.

Project page: https://dubiouscactus.github.io/trunc-grad-gs/
Paper: https://arxiv.org/abs/2609.03534

Scenes

Folder / Parquet	Scene	Cameras	Rig	Test Views	Depth	Status
1_1_neon	Neon city	42	6 × 7	3	EXR	Released
1_4_underwater	Underwater	36	6 × 6	2	EXR	Coming soon
4_5_tree	Windy tree	50	5 × 10	3	EXR	Released
alley	Alley	42	6 × 7	4	EXR	Released
bouncy_balls	Bouncy balls	40	4 × 10	4	EXR	Released
cup	Water cup	35	5 × 7	1	EXR	Released

License

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International License.

Acknowledgement

Scenes by Adam Handschuh, funded by Dolby Laboratories, Inc. TruncGradGS is supported by Research Ireland (22/FFP-P/11522). Cite the paper below when you use the data.

Citation

@inproceedings{morales2026truncgradgs,
  title         = {TruncGradGS: Improved 3D Gaussian Splatting via Truncated Gradient Updates},
  author        = {Morales, Th{\'e}o and Le-Pham, Nhat-Quynh and Atkins, Robin and Hua, Binh-Son},
  booktitle     = {Proceedings of Pacific Graphics 2026},
  year          = {2026},
  publisher     = {The Eurographics Association},
  eprint        = {2609.03534},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV}
}