MoE-GS Studio summarizes our research series on Mixture-of-Experts (MoE) for Dynamic Gaussian Splatting.
This repository provides a compact overview of our MoE-based 4DGS papers and the core algorithmic ideas behind each work.
- 2026.07: MoDE paper is accepted to IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2026.
- 2026.07: MoDE code repository is available. Please visit github.com/cvsp-lab/MoDE.
- 2026.07: MoE-GS Studio is released as a central hub for our MoE-based Dynamic Gaussian Splatting research series.
- 2026.03: MoE-GS code, paper, and project page are available. Please visit github.com/cvsp-lab/MoE-GS.
- 2026.01: MoE-GS paper is accepted to ICLR 2026.
- We add MoDE, our TPAMI 2026 work on Mixture of Deformation Experts for Dynamic Gaussian Splatting.
- We update the paper index to organize MoE-GS and MoDE as a unified MoE-based 4DGS research series.
- We add the MoDE paper and code links, along with a concise algorithm summary.
| Project | Paper | Venue | Links |
|---|---|---|---|
| MoE-GS | Mixture of Experts for Dynamic Gaussian Splatting | ICLR 2026 | Paper / Project Page / Code |
| MoDE | On the Design of Mixture-of-Experts for Dynamic Gaussian Splatting | IEEE TPAMI 2026 | Paper / Code |
MoE-GS introduces a Mixture-of-Experts framework for dynamic Gaussian splatting. It first trains multiple dynamic Gaussian experts independently and then learns a volume-aware pixel router that adaptively blends expert outputs across space, time, and viewing direction.
This design allows MoE-GS to combine heterogeneous 4DGS models without requiring them to share the same canonical Gaussian representation.
Authors: In-Hwan Jin*, Hyeongju Mun*, Joonsoo Kim, Kugjin Yun, and Kyeongbo Kong† (* equal contribution, † corresponding author)
MoDE studies multi-deformation modeling for dynamic Gaussian splatting. Instead of relying on a single deformation model, MoDE integrates multiple deformation experts into a shared canonical Gaussian representation and jointly optimizes them within a unified dynamic Gaussian pipeline.
This design enables multiple deformation priors to cooperate directly at the representation level, providing an end-to-end alternative to routing independently trained full models.
Authors: In-Hwan Jin*, Hyeongju Mun*, Joonsoo Kim, Kugjin Yun, and Kyeongbo Kong† (* equal contribution, † corresponding author)
If you find this series useful, please consider citing the relevant papers.
@inproceedings{jinmoegs2026,
title={MoE-{GS}: Mixture of Experts for Dynamic Gaussian Splatting},
author={In-Hwan Jin and Hyeongju Mun and Joonsoo Kim and Kugjin Yun and Kyeongbo Kong},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026}
}
@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}
}