SeqAvatar is a human avatar method based on sequential motion context. Details are described in our paper.

📍 If there are any bugs in our code, please feel free to raise your issues.

⭐️ Update:

• [2025/10/11] Results, including optimized models and rendered images, are now available at this link.

    # clone this repository
    git clone https://github.com/zezeaaa/SeqAvatar.git --recursive 
    # create environment
    conda env create -f environment.yaml
    conda activate seqavatar

1.Download SMPLIFY_CODE_V2.ZIP from SMPLify website. We use smplify_public/code/models/basicModel_neutral_lbs_10_207_0_v1.0.0.pkl inside for neutral gender.

2.Follow this page to remove Chumpy objects from the SMPL model.

3.Copy and rename the modified pkl file to ./smpl_model/models/SMPL_NEUTRAL.pkl

1.Download models_smplx_v1_1.zip from SMPLX website.

2.Copy SMPLX_NEUTRAL.pkl, SMPLX_FEMALE.pkl, and SMPLX_MALE.pkl.pkl to ./smpl_model/models/

The folder structure should look like

./
├── ...
└── smpl_model/
    ├── models/
        ├── SMPL_NEUTRAL.pkl
        ├── SMPLX_NEUTRAL.pkl
        ├── SMPLX_FEMALE.pkl
        └── SMPLX_MALE.pkl.pkl
    ├── smpl/
    └── smplx/

Download the dataset in data/ folder from this link. Extract the data to path/to/I3D-Human/.

The folder structure should look like

<path/to/I3D-Human/>
├── ID1_1-novelpose/
    ├── images/
    ├── masks/
    ├── ...
├── ID1_1-novelview/
├── ID1_1-train/
├── ID1_2-novelpose/
├── ID1_2-novelview/
├── ID1_2-train/
├── ...

Apply and download the dataset from DNA-Rendering website. Extract the data to path/to/DNA-Rendering/. We use 6 sequences for our experiments: 0007_04, 0019_10, 0044_11, 0051_09, 0206_04, and 0813_05. Extract the data to path/to/DNA-Rendering/.

The folder structure should look like

<path/to/DNA-Rendering/>
├── 0007_04/
    ├── 0007_04.smc
    └── 0007_04_annots.smc
├── 0019_10/
├── ...

Then run the following script to extract necessary data for experiments.

python process_dnarendering.py --dna_root path/to/DNA-Rendering/ --output_root path/to/DNA-Rendering/ 

After processing, the folder structure should look like

path/to/DNA-Rendering/
├── 0007_04/
    ├── bkgd_masks/
    ├── cameras/
    ├── images/
    ├── model/
    ├── 0007_04.smc
    └── 0007_04_annots.smc
├── 0019_10/
├── ...

Download the dataset from this link. We use 6 sequences for our experiments: CoreView_377, CoreView_386, CoreView_387, CoreView_392, CoreView_393, and CoreView_394.

Extract the dataset to <path/to/ZJU-MoCap/>, the folder structure should look like

<path/to/ZJU-MoCap/>
├── CoreView_377/
    ├── Camera_B1/
    ├── Camera_B2/
    └── ...
├── CoreView_386/
├── CoreView_387/
├── ...

Training and evaluation on I3D-Human:

# set data_path as <path/to/I3D-Human/>
bash scripts/exps_i3dhuman.sh

Training and evaluation on DNA-Rendering:

# set data_path as <path/to/DNA-Rendering/>
bash scripts/exps_dnarendering.sh

Training and evaluation on ZJU-MoCap:

# set data_path as <path/to/ZJU-MoCap/>
bash scripts/exps_zjumocap.sh

Results (including optimized models and rendered images) of the current version can be downloaded from this link.

If you find our work useful in your research please consider citing our paper:

@inproceedings{xu2025seqavatar,
  title={Sequential Gaussian Avatars with Hierarchical Motion Context},
  author={Xu, Wangze and Zhan, Yifan and Zhong, Zhihang and Sun, Xiao},
  booktitle={International Conference on Computer Vision (ICCV)},
  year={2025},
}

Our code is heavily based on GauHuman, Dyco, and 3DGS-Avatar. We would like to thank the authors for their great work and generously providing source codes, which inspired our work and helped us a lot in the implementation.