SmartSplat: Feature-Smart Gaussians for Scalable Compression of Ultra-High-Resolution Images
[AAAI 2026]

Linfei Li · Lin Zhang* · Zhong Wang · Ying Shen

🌐Project page | 📝Paper

Raw Image info: 16320×10848, 189 MB -> 1.99 MB (.npz)

Table of Contents

1. Installation
2. Datasets
3. Benchmarking
4. Acknowledgement

Installation

conda create -n smartsplat python==3.12
conda activate smartsplat

# install torch
pip install torch==2.5.1 torchvision==0.20.1 --index-url https://download.pytorch.org/whl/cu124

pip install setuptools==78.0.1

pip install -r requirements.txt

cd submodules/fused-ssim
pip install -e .

# Used by GaussianImage/3DGS/SmartSplat
cd ../gsplat
pip install -e .

# Used by LIG
cd ../gsplat2d
pip install -e .

Datasets

The datasets used in SmartSplat can be downloaded from the links provided below, including the tested subsets and the full versions hosted on Hugging Face or Baidu Netdisk.

Dataset	Tested	Full
DIV8K	SmartSplat-DIV8K	Full DIV8K
DIV16K	SmartSplat-DIV16K	Full DIV16K

Benchmarking

This codebase integrates multiple GS-based image representation methods, including GaussianImage, 3DGS, and LIG. All our experiments were conducted on the A800 cluster. The corresponding run scripts are provided in the slurm folder.

• LIG

data_path="data"
base_log_path="logs"
current_cr=50
python train_lig_for_eval.py \
    -d $data_path \
    --data_name  DIV16K \
    --model_name LIG \
    --compression_ratio $current_cr \
    --log_dir $base_log_path \
    --iterations 50000 \
    --save_iter_img 10000 \
    --save_imgs

• 3DGS

data_path="data"
base_log_path="logs"
current_cr=50
python train_all_for_eval.py \
    -d $data_path \
    --data_name DIV8K \
    --model_name 3DGS \
    --compression_ratio $current_cr \
    --log_dir $base_log_path \
    --iterations 50000 \
    --save_iter_img 10000

• GaussianImage (RS)

data_path="data"
base_log_path="logs"
current_cr=50
python train_all_for_eval.py \
    -d $data_path \
    --data_name  DIV16K \
    --model_name GaussianImage_RS \
    --compression_ratio $current_cr \
    --log_dir $base_log_path \
    --iterations 50000 \
    --save_iter_img 10000

• GaussianImage (Cholesky)

data_path="data"
base_log_path="logs"
current_cr=50
python train_all_for_eval.py \
    -d $data_path \
    --data_name  DIV16K \
    --model_name GaussianImage_Cholesky \
    --compression_ratio $current_cr \
    --log_dir $base_log_path \
    --iterations 50000 \
    --save_iter_img 10000

• Image-GS

The Image-GS implementation used in our codebase is built upon GaussianImage and does not incorporate the Top-K strategy, resulting in suboptimal performance. For accurate reproduction, please refer to the official implementation.

data_path="data"
base_log_path="logs"
current_cr=50
python train_all_for_eval.py \
    -d $data_path \
    --data_name  DIV8K \
    --model_name ImageGS_RS \
    --compression_ratio $current_cr \
    --log_dir $base_log_path \
    --iterations 50000 \
    --save_iter_img 10000

• SmartSplat

data_path="data"
base_log_path="logs"
current_cr=3000
python train_all_for_eval_smart.py \
    -d $data_path \
    --data_name  DIV8K \
    --model_name SmartSplat \
    --compression_ratio $current_cr \
    --log_dir $base_log_path \
    --iterations 50000 \
    --save_iter_img 10000

Acknowledgement

We thank the authors of the following repositories for their open-source code:

• Gsplat
• GaussianImage
• ImageGS
• LIG
• FusedSSIM

Citation

If you find our paper and code useful for your research, please use the following BibTeX entry.

@misc{li2025smartsplatfeaturesmartgaussiansscalable,
      title={SmartSplat: Feature-Smart Gaussians for Scalable Compression of Ultra-High-Resolution Images}, 
      author={Linfei Li and Lin Zhang and Zhong Wang and Ying Shen},
      year={2025},
      eprint={2512.20377},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2512.20377}, 
}