[ICLR 2026] CompMarkGS: Robust Watermarking for Compressed 3D Gaussian Splatting

Sumin In¹, Youngdong Jang¹, Utae Jeong¹, MinHyuk Jang¹, Hyeongcheol Park¹,
Eunbyung Park², Sangpil Kim<sup>1*

1</sup>Korea University,   ²Yonsei University

Installation

1. Unzip the files

cd submodules
unzip diff-gaussian-rasterization.zip
unzip simple-knn.zip
cd ..

2. Install the environment

conda env create --file ./setup/environment.yml
conda activate CompMarkGS

3. For compressed extraction, clone ContextGS into the ./attacks folder.
   For compressed extraction, please clone ContextGS into ./attacks and prepare pretrained ContextGS checkpoints in advance.

Data

1. Download datasets
   NeRF Synthetic & LLFF
   MipNeRF360
2. Configure the data structure 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/
 ...

Training

We provide the following scripts in ./scripts:

• embed_watermark.sh: watermark embedding (training)
• extract_watermark.sh: watermark extraction/evaluation (uncompressed)
• extract_compressed_watermark.sh: watermark extraction/evaluation (compressed)

Run with GPU id:

bash scripts/embed_watermark.sh 0
bash scripts/extract_watermark.sh 0
bash scripts/extract_compressed_watermark.sh 0

Citation

@inproceedings{
  in2026compmarkgs,
  title={CompMark{GS}: Robust Watermarking for Compressed 3D Gaussian Splatting},
  author={Sumin In and Youngdong Jang and Utae Jeong and MinHyuk Jang and Hyeongcheol Park and Eunbyung Park and Sangpil Kim},
  booktitle={The Fourteenth International Conference on Learning Representations},
  year={2026},
  url={https://openreview.net/forum?id=NXQvejGBFx}
}

Acknowledgement

We thank all authors from HAC, ContextGS, Scaffold-GS and 3D-GS for excellent works.