NSVQ-GS: Compressing 3D Gaussian Splatting by Noise-Substituted Vector Quantization

Haishan Wang · Mohammad Vali . Arno Solin

SCIA 2025

Paper

---

This is the public codes for SCIA2025 paper: "Compressing 3D Gaussian Splatting by Noise-Substituted Vector Quantization"

Overview of Storage structure of NSVQ-GS

Repo clone

The command to clone our repo: git clone [email protected]:AaltoML/nsvqgs.git --recursive. Since the repo contains submodules, please clone it recursively, and make sure the dir submodules contains non-empty subdirs

Environment installation

The python environment is installed by conda env create -f environment.yml with the environment name nsvqgs

Notes of env

This note contains additional information about the environment installation, you can skip it if you have installed it successfully already.

This env is very similar to the environment of Inria-GaussianSplatting repo. The solutions to some installation issues maybe exist in this repo.

We succeed to install the environment by the environment.yml with cuda=11.6.2, and gcc=9.2.0. However, this environment is very sensitive to the versions of CUDA, torch and gcc. And there are possibly bugs during the installaion of submodules. If the environment.yml built by conda fails inside the submodules with cuda version incompatiblility. We recommend to install all packages without submodules, then install cuda tool one by one: mamba install -c "nvidia/label/cuda-11.6.0" cuda-toolkit mamba install -c "nvidia/label/cuda-11.6.0" cuda-nvcc Then make sure the cuda compiler is 11.6 by nvcc -V Then install all submodules by pip.

pip install submodules/diff-gaussian-rasterization
pip install submodules/simple-knn
pip install submodules/fused-ssim

Data download

All the data could be downlowded by the following command python scripts/data_download.py The script will download the main three datasets: Mip-NeRF360 (9 scenes), Tanks & Temples (2 scenes) and Deep Blending (2 scenes). Note that the data source: Mip-NeRF360 is from Mip-NeRF 360, Tanks & Temples (2 scenes) and Deep Blending are from Inria-3DGS.

Default settings

There are some default settings following Inria-3DGS:

• Data split: Every 8th image must be selected for testing, specifically those images of index i where i mod 8 ≡ 0. All other images are used for training.
• Resolution: Full-resolution images should be used up to a maximum side length of 1600px. For larger test images, downscaling is required so that the longest dimension is 1600px.

Example usage

Training

This is an example command of training NSVQ-GS (4k bitrates) model:

CUDA_VISIBLE_DEVICES=0 python train_newvq.py  --port 4060 --vq_ncls 4096 --vq_ncls_sh 1024 \
  --vq_ncls_dc 1024 --vq_start_iter 20000 --quant_params sh dc rot scale --opacity_reg \
  --lambda_reg 1e-7 --max_prune_iter 20000 --eval -s=${path_source} -m=${path_output} \
   --total_iterations 45000 --fine_tune

Here ${path_source} denotes the path of input data, ${path_output} denotes the path of all output files

Evaluation

This is an example command of rendering and evaluating the trained model

# * rendering
python render.py -m ${path_output} --iteration 45000 --skip_train --load_quant

# * metrics calculation
python metrics.py -m ${path_output}

Here ${path_output} denotes the path of trained model.

Integrated script

For convenience, we prepare a integrated script to train, render and evaluate NSVQ-GS standalone. you can run it with this command: source scripts/train.sh