Compressing Double-Phase Holograms
using 2D Gaussians

EUROGRAPHICS 2026

Xiaoyue Fan¹, Yicheng Zhan¹, Amrita Mazumdar², Kaan Akşit¹

¹ University College London (UCL)   ² NVIDIA

Getting Started

This project is based on the gsplat submodule, licensed under the MIT License.

cd Gaussgram-main
git clone https://github.com/XingtongGe/gsplat.git
cd gsplat
pip install .[dev]
cd ../
pip install -r requirements.txt

---

Compressed Representation

Each bash script corresponds to a specific sample hologram. Two factorization options are available:

Directory	Factorization
./scripts/gaussianimage_cholesky/	Cholesky
./scripts/gaussianimage_rs/	RS

All sample holograms are licensed for both commercial and non-commercial use.

Example — Police Dog hologram with patch size 64:

# Cholesky factorization
bash scripts/gaussianimage_cholesky/dog_64.sh dataset/samplehologram

# RS factorization
bash scripts/gaussianimage_rs/dog_64.sh dataset/samplehologram

Dataset Structure

The default setting crops four neighbouring patches from the sample hologram in ./dataset/samplehologram/. After cropping, each patch and its decomposed components are sequentially numbered:

dataset/samplehologram/
├── patch_001    # cropped patch from sample hologram
├── patch_002    # high-value component, vertical decomposition
├── patch_003    # high-value component, horizontal decomposition
├── patch_004    # low-value component, vertical decomposition
├── patch_005    # low-value component, horizontal decomposition
├── patch_006    # next neighbouring patch
└── ...

---

Output Structure

After training, results are saved under ./checkpoints/:

checkpoints/
└── samplehologram/
    ├── GaussianImage_Cholesky_70000_200/
    │   ├── dog_patch_64_001/
    │   ├── dog_patch_64_002/
    │   ├── ...
    │   ├── dog_patch_64_020/
    │   ├── results/        ← compressed patches + simulated reconstructions
    │   └── stat_plots/     ← parameter changes throughout training
    ├── GaussianImage_RS_70000_200/
    └── quality_plots/      ← PSNR, SSIM vs compression ratio plots

Example: Police Dog · patch size 64 · Cholesky · 200 Gaussians · 70 000 epochs ./checkpoints/samplehologram/GaussianImage_Cholesky_70000_200/results

---

Citation

@inproceedings{fan2026comdph2dgs,
 booktitle = {Eurographics 2026 - Posters},
 editor = {Gerrits, Tim and Teschner, Matthias},
 title = {{Compressing Double-Phase Holograms using 2D Gaussians}},
 author = {Fan, Xiaoyue  and Zhan, Yicheng  and Mazumdar, Amrita  and Akşit, Kaan },
 year = {2026},
 publisher = {The Eurographics Association},
 ISSN = {1017-4656},
 ISBN = {978-3-03868-300-1},
 DOI = {10.2312/egp.20261010}
 }