MoE-GS Studio

MoE-GS Studio is a research hub for Mixture-of-Experts (MoE) architectures for Dynamic Gaussian Splatting.
This repository organizes our research exploring how expert specialization and routing mechanisms can improve dynamic 3D scene reconstruction.

Our work investigates how multiple dynamic Gaussian splatting models can be combined through Mixture-of-Experts frameworks to better handle diverse motion patterns and scene dynamics.

---

Research Index

• (NEW) MoDE (Mixture of Deformation Experts) is our latest research project in the MoE-GS Studio series, extending the Mixture-of-Experts paradigm for dynamic Gaussian splatting.
• MoE-GS (ICLR 2026) introduces a Mixture-of-Experts framework with a pixel-wise routing mechanism that adaptively selects and combines multiple dynamic Gaussian splatting models. Code Paper ProjectPage

MoDE: Mixture of Deformation Experts

Instead of relying on a single deformation model, MoDE introduces multiple deformation experts, each specializing in different motion behaviors.
A routing mechanism dynamically selects or combines experts depending on the spatial and temporal characteristics of the scene.

This design enables the model to better handle:

• complex object motion
• non-rigid deformation
• heterogeneous dynamic regions

MoDE builds upon insights from our earlier work MoE-GS, which demonstrated the benefits of combining multiple dynamic Gaussian splatting models.

1. Virtual Environment Setup

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

---

2. Data Preprocessing

# automatically extract the frames and reorganize them
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

---

3. MoDE - 4DGaussians w/ E-D3DGS (Train & Render)

# 4D Gaussian + E-D3DGS (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"

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

---

4. MoDE - 4DGaussians w/ Grid4D (Train & Render)

# 4D Gaussian + Grid4D (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"

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