On the Design of Mixture-of-Experts for Dynamic Gaussian Splatting

Official Repository for MoDE

IEEE TPAMI 2026

In-Hwan Jin^1* · Hyeongju Mun^1* · Joonsoo Kim² · Kugjin Yun² · Kyeongbo Kong^1†
1 Pusan National University ² Electronics and Telecommunications Research Institute
^* Equal contribution     ^† Corresponding author

Summary: This repository provides the official implementation of MoDE, a Mixture of Deformation Experts framework for dynamic Gaussian Splatting. MoDE jointly optimizes multiple deformation experts on a shared canonical Gaussian representation.

MoE-GS Series: MoDE is part of MoE-GS Studio, a research series on Mixture-of-Experts architectures for Dynamic Gaussian Splatting.

🚧 TODO List

• Grid4D based Code Release
• E-D3DGS based Code Release
• 4DGaussians based Code Release (Coming Soon)

Contents

1. Setup
2. Preprocess Datasets
3. Training and Rendering
4. Related Projects
5. BibTeX

Setup

Download Repository and Thirdparty Modules

git clone https://github.com/cvsp-lab/MoDE.git

Environment Setup

Install the required packages and submodules:

pip install -r requirements.txt
pip install -e submodules/diff-gaussian-rasterization/
pip install -e submodules/simple-knn/

Preprocess Datasets

Neural 3D Video Dataset

For the Neural 3D Video dataset, extract frames and reorganize the scene directory:

python script/pre_n3v.py --videopath <dataset>/<scene>

Downsample dense point clouds:

python script/downsample_point.py \
    <location>/<scene>/colmap/dense/workspace/fused.ply \
    <location>/<scene>/points3D_downsample.ply

Training and Rendering

MoDE with 4DGaussians and E-D3DGS

Train on the Neural 3D Video dataset:

python train_emb.py \
    -s <N3V_DATASET_ROOT>/coffee_martini \
    --expname <SAVE_PATH> \
    --configs "arguments_MoDE/dynerf/config_rot_0/coffee_martini.py"

Render:

python render_emb.py --skip_test --skip_train \
    --model_path <SAVE_PATH> \
    --configs "arguments_MoDE/emb/config_0/coffee_martini.py" \
    --iteration 30000

MoDE with 4DGaussians and Grid4D

Train on the Neural 3D Video dataset:

python train_hash.py \
    -s <N3V_DATASET_ROOT>/coffee_martini \
    --expname <SAVE_PATH> \
    --configs "arguments_MoDE/dynerf/config_rot_0/coffee_martini.py"

Render:

python render_hash.py --skip_train --skip_test \
    --model_path <SAVE_PATH> \
    --configs "arguments_MoDE/hash/config_0/coffee_martini.py" \
    --iteration 30000

Related Projects

• MoE-GS Studio: Overview of our MoE-based 4DGS research series.
• MoE-GS: Mixture of Experts for Dynamic Gaussian Splatting.

BibTeX

@article{jinmode2026,
    title={On the Design of Mixture-of-Experts for Dynamic Gaussian Splatting},
    author={In-Hwan Jin and Hyeongju Mun and Joonsoo Kim and Kugjin Yun and Kyeongbo Kong},
    journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
    year={2026},
    note={Accepted}
}