AvatarMoE

AvatarMoE

Decomposing Non-Rigid Deformation with Part-Aware Experts for 3DGS Avatars

Hyeri Yang, Junyoung Hong, Shinwoong Kim, Kyungjae Lee Computers & Graphics, 2026

Paper (DOI) · Project Page · Code

---

Overview

AvatarMoE is a part-aware Mixture-of-Experts (MoE) framework for animatable 3D Gaussian Splatting avatars. Instead of modeling non-rigid deformation with a single global network — which produces tearing and stretching artifacts in complex poses — AvatarMoE decomposes the body into 24 joint-centric regions, each handled by a lightweight specialized expert:

• Dynamic GMM-based gating allocates expert influence according to the input pose.
• Hybrid expert architecture shares a HashGrid encoder across 24 lightweight MLP decoders.
• Intra-Expert Coherence Loss regularizes each expert's deformation field for robust OOD poses.

This repository hosts the trained checkpoints. Code, installation, and training/evaluation instructions live in the GitHub repository; the core model is in models/non_rigid.py.

Checkpoints

Each subject is a folder containing the checkpoint (ckpt<iters>.pth) and the resolved training config (.hydra/):

checkpoints/
├── zjumocap_377_mono-best/
│   ├── ckpt8000.pth
│   └── .hydra/config.yaml
├── zjumocap_386_mono-best/
├── ...
├── zjumocap_377_refine-best/
│   ├── ckpt8000.pth
│   └── .hydra/config.yaml
├── zjumocap_386_refine-best/
├── ...
├── ps_female_3-best/
│   ├── ckpt15000.pth
│   └── .hydra/config.yaml
└── ...

• ZJU-MoCap: 377, 386, 387, 392, 393, 394 (8,000 iters)
• People-Snapshot: female-3, female-4, male-3, male-4 (15,000 iters)

Usage

Checkpoints require the AvatarMoE code and its CUDA extensions — clone the repo first:

git clone --recursive https://github.com/milab-yongin/AvatarMoE.git
cd AvatarMoE
conda env create -f environment.yml
conda activate avatarmoe
pip install submodules/diff-gaussian-rasterization
pip install submodules/simple-knn

Download a checkpoint from this repo:

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="<your-username>/AvatarMoE",
    allow_patterns="checkpoints/ps_female_3-best/*",
)
print(local_dir)  # contains ckpt<iters>.pth and .hydra/config.yaml

Then evaluate / render following the repo instructions, e.g.:

python render.py mode=test dataset.test_mode=view dataset=zjumocap_377_mono

Environment

Ubuntu 22.04, Python 3.10, PyTorch 2.1.2 + CUDA 11.8, single NVIDIA RTX 4090 (24 GB). Trained with Adam and an exponential LR schedule (γ = 0.1); 8,000 iterations for ZJU-MoCap and 15,000 for People-Snapshot.

Results

On People-Snapshot, AvatarMoE reaches 32.45 PSNR — the best among the compared baselines — while remaining competitive on SSIM and LPIPS, and substantially reduces geometric tearing and flickering under OOD poses relative to the 3DGS-Avatar baseline. See the paper for full quantitative tables.

Datasets

Prepare datasets following the official protocols of 3DGS-Avatar and ARAH. SMPL models must be obtained separately from the SMPL and SMPLify project pages under their own licenses.

License

Released under the MIT License. Some third-party dependencies and submodules (SMPL, 3DGS rasterization, etc.) are subject to their own licenses.

Citation

@article{yang2026avatarmoe,
  title={AvatarMoE: Decomposing non-rigid deformation with part-aware experts for 3DGS avatars},
  author={Yang, Hyeri and Hong, Junyoung and Kim, Shinwoong and Lee, Kyungjae},
  journal={Computers \& Graphics},
  pages={104597},
  year={2026},
  publisher={Elsevier}
}

Acknowledgement

Built upon 3D Gaussian Splatting, 3DGS-Avatar, ARAH, smplx, Anim-NeRF, and GART.