README.md

May 19, 2026 · View on GitHub

Structure-Aware Fine-Grained Gaussian Splatting for Expressive Avatar Reconstruction, arXiv

1. Directory Structure

Please organize your project directory as follows. In particular, make sure the model files are correctly placed under the common directory:

SFGS
${ROOT}
|-- main
|-- common
|   |-- utils/human_model_files
|       |-- smplx/SMPLX_FEMALE.npz
|       |-- smplx/SMPLX_MALE.npz
|       |-- smplx/SMPLX_NEUTRAL.npz
|       |-- smplx/MANO_LEFT.pkl             
|       |-- smplx/MANO_RIGHT.pkl     
|       |-- smplx/MANO_SMPLX_vertex_ids.pkl
|       |-- smplx/SMPL-X__FLAME_vertex_ids.npy
|       |-- smplx/smplx_flip_correspondences.npz
|       |-- flame/flame_dynamic_embedding.npy
|       |-- flame/FLAME_FEMALE.pkl
|       |-- flame/FLAME_MALE.pkl
|       |-- flame/FLAME_NEUTRAL.pkl
|       |-- flame/flame_static_embedding.pkl
|       |-- flame/FLAME_texture.npz
|-- data
|   |-- XHumans
|       |-- data/00028
|       |-- data/00034
|       |-- data/00087
|-- tools
|-- output

Model Downloads

  • SMPL-X: Version 1.1
  • FLAME: Version 2020

2. XHumans Data Preparation

Download the dataset from the following link,extract together with files of the same name:

Firstly：

Secondly：

3. Training

Navigate to the main directory and run the following command (taking subject 00028 as an example):

CUDA_VISIBLE_DEVICES=0 python train.py --subject_id 00028

The trained checkpoints will be saved to:

output/model/00028

4. Visualization and Animation

4.1 Neutral Pose Visualization

To render a rotating avatar in the neutral pose:

python get_neutral_pose.py --subject_id 00028 --test_epoch 20

The results will be saved under:

./main/neutral_pose

4.2 Motion-Driven Animation

To animate the avatar using motion parameters:

python animation.py --subject_id 00028 --test_epoch 20 --motion_path $PATH
  • $PATH should contain the SMPL-X parameters used to drive the avatar.

To render animation with a rotating camera view:

python animate_view_rot.py --subject_id 00028 --test_epoch 20 --motion_path $PATH

5. Testing and Evaluation

Rendering Results

python test.py --subject_id 00028 --test_epoch 20

The rendered results will be saved to:

output/result/00028

Quantitative Evaluation

Navigate to the tools directory and run:

python eval_xhumans.py --output_path ../output/result/00028 --subject_id 00028

Citation

If you find this work useful, please consider citing:

@misc{su2026structureawarefinegrainedgaussiansplatting,
  title={Structure-Aware Fine-Grained Gaussian Splatting for Expressive Avatar Reconstruction}, 
  author={Yuze Su and Hongsong Wang and Jie Gui and Liang Wang},
  year={2026},
  eprint={2604.09324},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2604.09324}, 
}