Duplex-GS: Proxy-Guided Weighted Blending for Real-Time Order-Independent Gaussian Splatting

Project Page | ArXiv

Weihang Liu*, Yuke Li, Yuxuan Li, Jingyi Yu, Xin Lou

Overview

The paper introduces Duplex-GS, a dual-hierarchy framework that integrates order-independent rendering to significantly accelerate 3D Gaussian Splatting while reducing sorting overhead without compromising quality.

Installation

1. Clone this repo:

git clone https://github.com/LiYukeee/Duplex-GS --recursive
cd Duplex-GS

2. Install dependencies

conda env create --file environment.yml
conda activate duplex

Data

First, create a data/ folder inside the project path by

mkdir data

The data structure will be organised as follows:

data/
├── dataset_name
│   ├── scene1/
│   │   ├── images
│   │   │   ├── IMG_0.jpg
│   │   │   ├── IMG_1.jpg
│   │   │   ├── ...
│   │   ├── sparse/
│   │       └──0/
│   ├── scene2/
│   │   ├── images
│   │   │   ├── IMG_0.jpg
│   │   │   ├── IMG_1.jpg
│   │   │   ├── ...
│   │   ├── sparse/
│   │       └──0/
...

Custom Data

For custom data, you should process the image sequences with Colmap to obtain the SfM points and camera poses. Then, place the results into data/ folder.

Training

Training Mipnerf-360

# This is a parallel training script. It enables `n_jobs` (default 4) processes to train at the same time.

bash mipnerf360.sh

Training other scenes

cd script
bash train_xxx.sh

This script will store the log (with running-time code) into outputs/dataset_name/scene_name/exp_name/cur_time automatically.

Evaluation

python render.py -m <path to trained model> # Generate renderings

Result

Matrix City

Method	PSNR↑	SSIM↑	LPIPS↓	#GS/#Proxy↓	FPS↑	Time [h]↓
3DGS¹	27.03	0.808	0.310	10206 K	66	2.60
Scaffold-GS	26.82	0.811	0.288	4829 K	113	2.31
Octree-GS	25.18	0.738	0.397	4082 K	128	2.61
Octree-GS*	26.61	0.812	0.292	8010 K	79	-
LC-WSR	26.34	0.782	0.340	7841 K	34	3.86
Ours (K = 5)	27.40	0.826	0.260	4165 K	119	2.21

* Officially released pretrained model.

Mip-NeRF360

Method	PSNR↑	SSIM↑	LPIPS↓	Storage↓
Mip-NeRF360¹	27.69	0.792	0.237	-
3DGS¹	27.52	0.813	0.222	750.2 MB
Mip-Splatting¹	27.61	0.816	0.215	838.4 MB
Scaffold-GS¹ (K=10)	27.73	0.812	0.226	171.0 MB
Scaffold-GS¹ (K=5)	27.74	0.811	0.230	205.3 MB
Octree-GS¹ (K=10)	27.88	0.816	0.216	142.7 MB
Octree-GS¹ (K=5)	27.73	0.813	0.227	124.5 MB
LC-WSR²	27.21	0.800	0.219	1312.0 MB
Ours (K=10)	27.90	0.813	0.216	152.2 MB
Ours (K=5)	27.74	0.802	0.218	104.4 MB

Tanks & Temples

Method	PSNR↑	SSIM↑	LPIPS↓	Storage↓
Mip-NeRF360¹	22.22	0.759	0.257	-
3DGS¹	23.57	0.845	0.180	431.4 MB
Mip-Splatting¹	23.96	0.856	0.171	500.4 MB
Scaffold-GS¹ (K=10)	24.09	0.858	0.165	147.7 MB
Scaffold-GS¹ (K=5)	24.53	0.863	0.162	177.7 MB
Octree-GS¹ (K=10)	24.60	0.864	0.157	77.8 MB
Octree-GS¹ (K=5)	24.47	0.861	0.168	73.9 MB
LC-WSR²	23.22	0.832	0.186	672.7 MB
Ours (K=10)	24.26	0.867	0.150	124.8 MB
Ours (K=5)	24.32	0.867	0.138	112.2 MB

Deep Blending

Method	PSNR↑	SSIM↑	LPIPS↓	Storage↓
Mip-NeRF360¹	29.40	0.901	0.245	-
3DGS¹	29.61	0.900	0.251	662.7 MB
Mip-Splatting¹	29.56	0.901	0.243	736.8 MB
Scaffold-GS¹ (K=10)	30.42	0.912	0.246	111.2 MB
Scaffold-GS¹ (K=5)	30.26	0.911	0.242	143.4 MB
Octree-GS¹ (K=10)	30.44	0.911	0.239	95.5 MB
Octree-GS¹ (K=5)	30.06	0.908	0.250	75.1 MB
LC-WSR²	29.90	0.901	0.243	810.7 MB
Ours (K=10)	30.30	0.910	0.249	89.7 MB
Ours (K=5)	30.26	0.909	0.236	94.7 MB

[1] Experiments are conducted using the official public repository, with no modifications except for iteration settings as described in Sec. 4.

[2] The official implementation has not been released. Experiments are conducted using our own implementation, which is available at https://github.com/LiYukeee/sort-free-gs.

Citation

If you find our work helpful, please consider citing:

@misc{liu2025duplexgsproxyguidedweightedblending,
      title={Duplex-GS: Proxy-Guided Weighted Blending for Real-Time Order-Independent Gaussian Splatting}, 
      author={Weihang Liu and Yuke Li and Yuxuan Li and Jingyi Yu and Xin Lou},
      year={2025},
      eprint={2508.03180},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2508.03180}, 
}

LICENSE

Please follow the LICENSE of 3D-GS.