MoGaFace: Momentum-Guided and Texture-Aware Gaussian Avatars for Consistent Facial Geometry [PRCV 2026]

Project Page | Paper (arXiv)

---

Installation & Dependencies

Linux / Ubuntu

Tested with Python 3.8, PyTorch 2.4.1 + CUDA 12.1. Environment setup follows common 3D Gaussian Splatting + PyTorch3D workflows (see setup_env.sh).

git clone https://github.com/AISHIWEILAI/MoGaFace.git --recursive
cd MoGaFace

# If already cloned without --recursive:
# git submodule update --init --recursive

# Recommended: conda environment
conda create -n mogaface python=3.8
conda activate mogaface

# One-shot setup (Tsinghua pip mirror + CUDA extensions)
bash setup_env.sh

Component	Version
PyTorch	2.4.1+cu121
torchvision	0.19.1+cu121
pytorch3d	0.7.8
diff-gaussian-rasterization / simple-knn	submodules/ (git submodules)
gridencoder	gridencoder/ → submodules/torch-ngp/gridencoder (symlink)

Note: PyTorch3D may require building from source. See setup_env.sh for compiler flags (gcc, CUDA_HOME, TORCH_CUDA_ARCH_LIST).

---

FLAME Assets Preparation

MoGaFace is built on FLAME 2023. Due to the FLAME license, model files are not included in this repository. You must register and download them from the official site:

https://flame.is.tue.mpg.de/download.php

After registration, download the following resources and place them as shown (same convention as GaussianAvatars):

Resource	Official download	Target path
FLAME 2023 (w/ jaw rotation)	FLAME 2023 model	flame_model/assets/flame/flame2023.pkl
FLAME Vertex Masks	FLAME masks	flame_model/assets/flame/FLAME_masks.pkl
Landmark embedding (w/ eyes)	FLAME landmark embedding	flame_model/assets/flame/landmark_embedding_with_eyes.npy
Head template mesh	FLAME template / geometry resources	flame_model/assets/flame/head_template_mesh.obj
Mean texture (optional)	FLAME texture resources	flame_model/assets/flame/tex_mean_painted.png
MediaPipe landmark embedding	FLAME MediaPipe resource	flame_model/assets/mediapipe/mediapipe_landmark_embedding.npz

Expected layout:

flame_model/
└── assets/
    ├── flame/
    │   ├── flame2023.pkl
    │   ├── FLAME_masks.pkl
    │   ├── landmark_embedding_with_eyes.npy
    │   ├── head_template_mesh.obj
    │   └── tex_mean_painted.png          # optional
    └── mediapipe/
        └── mediapipe_landmark_embedding.npz

Important

• You need to sign up on the FLAME website and agree to the license before downloading.
• FLAME assets are for non-commercial research only; see the model license.
• Validation inference does not require face_mask.pth.

---

Usage

All scripts should be run from the project root.

Data & checkpoints are not included in this GitHub repo. Download them from Baidu Netdisk (see below) and extract to data/ and output/ before running inference.

Script	Usage
setup_env.sh	Install conda/pip dependencies and build CUDA extensions
infer_val.sh	Batch validation inference for subjects 306 and 074
render.py	Single-subject inference with custom flags

1. Data & checkpoints

Note: Due to size limits, multi-view data and pretrained checkpoints are hosted on Baidu Netdisk and are not synced with this GitHub repository. After downloading, extract the archives into the project root so that paths match the layout below.

Resource	Baidu Netdisk	Extract code
NeRSemble data (subjects 306 & 074)	Download data	xxxx
Pretrained checkpoints (306 & 074)	Download checkpoints	xxxx

After extraction, the expected layout is:

data/
├── 306_20material_all_views/
│   └── UNION20_306_EMO1234EXP234589_v16_DS4_whiteBg_staticOffset_maskBelowLine/
└── 074_20material_all_views/
    └── UNION20_074_EMO1234EXP234589_v16_DS4_whiteBg_staticOffset_maskBelowLine/

output/nersemble/
├── 306_20material_allviews_expemo/
└── 074_20material_allviews_expemo/

Subject	Data path (-s, resolved in render.py)	Model path (-m)
306	data/306_20material_all_views/UNION20_306_EMO1234EXP234589_v16_DS4_whiteBg_staticOffset_maskBelowLine	output/nersemble/306_20material_allviews_expemo
074	data/074_20material_all_views/UNION20_074_EMO1234EXP234589_v16_DS4_whiteBg_staticOffset_maskBelowLine	output/nersemble/074_20material_allviews_expemo

Replace PLACEHOLDER_DATA / PLACEHOLDER_CKPT and extract codes with the actual Baidu Netdisk links before publishing.

2. Validation inference

Both subjects:

bash infer_val.sh

Single subject:

python render.py \
  -m output/nersemble/306_20material_allviews_expemo \
  --hum_id 306 \
  --skip_train \
  --skip_test

Flag	Meaning
-m	Checkpoint directory
--hum_id	Subject ID (306 or 074)
--skip_train	Skip training cameras (run validation only)
--skip_test	Skip test cameras

Outputs (example subject 306):

{model_path}/val_8/ours_1000000/
├── renders/           # rendered images
├── gt/                # ground-truth images
├── infer_results.txt  # PSNR / SSIM / LPIPS
├── renders.mp4
└── high_renders.mp4

---

Citation

Please cite the following paper if you use this method, model, or conduct derivative research based on this project:

@inproceedings{liu2026mogaface,
  title={MoGaFace: Momentum-Guided and Texture-Aware Gaussian Avatars for Consistent Facial Geometry},
  author={Liu, Yujian and Cao, Linlang and Chen, Chuang and Geng, Fanyu and Shen, Dongxu and Cao, Peng and Xu, Shidang and Liu, Xiaoli},
  booktitle={Proceedings of the Chinese Conference on Pattern Recognition and Computer Vision (PRCV)},
  year={2026}
}

---

Acknowledgements

This project is built upon or inspired by the following open-source projects:

• 3D Gaussian Splatting
• GaussianAvatars
• TensorialGaussianAvatar
• FLAME
• SyncAnimation

We sincerely thank the authors of these projects for their contributions to the open-source community.

---

Disclaimer

By using this project, you agree to comply with all applicable laws and regulations. You must not use it to generate or disseminate harmful content. FLAME and third-party CUDA extensions are subject to their respective licenses. The developers assume no responsibility for any direct, indirect, or consequential damages arising from the use or misuse of this software.