README.md

DASH: 4D Hash Encoding with Self-Supervised Decomposition

for Real-Time Dynamic Scene Rendering

Jie Chen, Zhangchi Hu, Peixi Wu, Huyue Zhu,
Hebei Li, Xiaoyan Sun
University of Science and Technology of China
ICCV 2025

Quick Start

Dataset Preparation

To train DASH, you should download the following dataset:

• Neural 3D Video Dataset
• Technicolor dataset

We follows 4D-GS for preprocessing the Neural 3D Video dataset, and STGS for the Technicolor dataset. Thanks very much for their excellent work.

Installation

git clone https://github.com/chenj02/DASH.git
cd DASH

conda env create -f environment.yaml
conda activate DASH

pip install -e ./submodules/diff=gaussian-rasterization
pip install -e ./submodules/simple-knn

Training

bash train.sh

or

CUDA_VISIBLE_DEVICES=0 python train.py -s <input path> \
                --model_path <output path> \
                --conf <config path> \
                --resolution 1 # for Technicolor dataset

Render

bash render.sh

or

CUDA_VISIBLE_DEVICES=0 python render.py -s <input path> \
                --skip_train \
                --model_path <output path> \
                --conf <config path> \
                --resolution 1 # for Technicolor dataset

Evaluation

python metrics.py -m <output path>

Citation

If you find our work useful, please cite:

@inproceedings{chen2025dash,
  title={DASH: 4D Hash Encoding with Self-Supervised Decomposition for Real-Time Dynamic Scene Rendering},
  author={Chen, Jie and Hu, Zhangchi and Wu, Peixi and Zhu, Huyue and Li, Hebei and Sun, Xiaoyan},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={26349--26359},
  year={2025}
}

Acknowledgements

Our code is based on 4D-GS and Grid4D. We thank the authors for their excellent work!