[COGS’25 – 2nd Place, ICCV 2025 Workshop] Post-Hoc Compression Algorithm 🚀

Runner-up (2nd place) at the COGS’25 workshop, ICCV 2025 🥈

Our compression method is post-hoc and takes inspiration from gradient-based pruning in deep neural networks, specifically informative gradient–based pruning 🧠✂️.

You can download compressed ckpts here.

Overview 📌

Starting from a pretrained Gaussian Splatting model (checkpoint provided), the algorithm alternates between two phases:

1. Gradient accumulation over the training set.
2. Pruning + fine-tuning to recover quality after each pruning step.

At each pruning step, we remove the Gaussians with the smallest gradient magnitudes.
The gradient of a Gaussian is defined as the sum of the norms of the gradients of all its parameters.

We iterate this process—pruning a small fraction each time—until a stopping criterion is met ✅.

During fine-tuning, we apply quantization ⚙️:

• Spherical harmonics coefficients → 8-bit unsigned integers
• All other parameters → 16-bit float

Finally, we compress the parameters with LZMA 💾.

How to Run 🚀

Download and install the Vanilla 3DGS repo along with the pre-trained ckpts. Run train.py. Be sure to point to the right pre-trained dir. There is also an example of compatibility script to run it with Flash-GS (file flashGS_example.py).

Computational Details 🧮

• Gradient accumulation requires a single pass over the training set and typically takes only a few seconds.
• Fine-tuning runs 1,000 iterations per pruning step, and 5,000 iterations at the final stage.
• In total, the algorithm performs about 30,000 iterations.
• Each pruning step removes 5% of the remaining Gaussians.

Results

Results obtained with vanilla-3dgs implementation

Scene	PSNR	SSIM	LPIPS	NUM GAUSSIANS	FPS	SIZE (MB)
bicycle	25.22	0.755	0.248	579283	171	21.66
bonsai	32.14	0.939	0.215	296058	205	10.56
counter	28.98	0.902	0.220	203121	175	8.04
flowers	21.55	0.596	0.368	491931	234	20.21
garden	27.37	0.861	0.125	1130273	143	44.51
kitchen	30.93	0.919	0.150	263772	185	10.39
room	31.42	0.913	0.244	166795	195	5.62
stump	26.66	0.767	0.260	327342	262	13.05
treehill	22.27	0.590	0.457	94202	304	3.44
AVG	27.39	0.805	0.254	394753	208	15.28
drjohnson	28.99	0.897	0.276	165140	325	6.17
playroom	30.22	0.910	0.252	381647	251	13.93
AVG	29.61	0.903	0.264	273394	288	10.05
train	21.79	0.764	0.296	87527	415	3.38
truck	25.04	0.872	0.165	361868	243	12.64
AVG	23.42	0.818	0.230	224698	329	8.01

Acknowledgments

This work was built upon two 3DGS compression techniques.
If you liked the project, please cite:

@inproceedings{ali2025elmgs,
  title     = {Elmgs: Enhancing memory and computation scalability through compression for 3D Gaussian splatting},
  author    = {Ali, Muhammad Salman and Bae, Sung-Ho and Tartaglione, Enzo},
  booktitle = {2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  pages     = {2591--2600},
  year      = {2025},
  organization = {IEEE}
}

@article{di2025gode,
  title   = {Gode: Gaussians on demand for progressive level of detail and scalable compression},
  author  = {Di Sario, Francesco and Renzulli, Riccardo and Grangetto, Marco and Sugimoto, Akihiro and Tartaglione, Enzo},
  journal = {arXiv preprint arXiv:2501.13558},
  year    = {2025}
}