This is a standalone implementation of Mini-Splatting-C, designed as a plug-in module for standard (vanilla) pretrained 3DGS models. By combining Mini-Splatting simplification with ACRF-F compression, this repository provides a strong baseline for Gaussian simplification and compression.
This code has been tested with Python 3.8, torch 1.12.1, CUDA 11.6.
- Clone the repository
git clone git@github.com:fatPeter/simplify-3DGS.git && cd simplify-3DGS
- Setup python environment
conda create -n simplify_3DGS python=3.8
conda activate simplify_3DGS
pip install torch==1.12.1+cu116 torchvision==0.13.1+cu116 -f https://download.pytorch.org/whl/torch_stable.html
pip install -r requirements.txt
- Download datasets: Mip-NeRF 360, T&T+DB COLMAP.
- Download pretrained models: Pre-trained Models (14 GB).
- Simplification scripts are in
simp:
cd simp
- Simplify pretrained 3DGS models:
python simplify.py -s <dataset path> -m <model path> --simp_iterations 5_000 --sampling_factor 0.6 --eval # -i images_4/images_2
- Compression scripts are in
comp:
cd comp
- Compress simplified models:
python compress.py -s <dataset path> -m <model path>
- Decompress:
python decompress.py -s <dataset path> -m <model path>
- Final compression results of pretrained 3DGS models:
| Dataset | SSIM | PSNR | LPIPS | Size (MB) |
|---|---|---|---|---|
| Mip-NeRF 360 | 0.806 | 27.16 | 0.234 | 30.68 |
| Tanks&Temples | 0.829 | 23.32 | 0.210 | 12.02 |
| Deep Blending | 0.901 | 29.48 | 0.254 | 17.62 |
- Final compression results of Mini-Splatting2 models (
--sampling_factor 1):
| Dataset | SSIM | PSNR | LPIPS | Size (MB) |
|---|---|---|---|---|
| Mip-NeRF 360 | 0.817 | 27.36 | 0.221 | 19.40 |
| Tanks&Temples | 0.840 | 23.31 | 0.191 | 10.67 |
| Deep Blending | 0.909 | 30.00 | 0.244 | 16.73 |