PERSE: Personalized 3D Generative Avatars from a Single Portrait

Hyunsoo Cha  Inhee Lee  Hanbyul Joo

Seoul National University

CVPR 2025

teaser_pp.mp4

TL;DR: Given a reference portrait image input, PERSE generates an animatable 3D personalized avatar with disentangled and editable control over various facial attributes.

• 2025.07.28: We release code for inference and training.
• 2025.02.27: 🎉 PERSE got accepted into CVPR 2025!
• 2024.12.31: 👏 AK and MrNeRF introduced our work! Thank you for sharing our paper.
• 2024.12.30: 🎉 We release paper and project page.

• Released the paper and project page on 2024.12.30.
• Release training, inference code.
• Release pretrained weights of avatar model.
• Release sample of datasets.

Our code supports preprocessing, training, and inference on NVIDIA GPUs with at least 48 GB of memory, such as A6000.

Our code has been tested with CUDA 11.8 on Ubuntu 22.04. Functionality in other environments is not guaranteed. We recommend installing and running it via Docker.

(1) Clone our repository

git clone https://github.com/snuvclab/perse.git

(2) Update WORK_DIR

Before running the installation script, please update the WORK_DIR in

scripts/common.sh

. WORK_DIR is the location where the repository was cloned.

(3) Run install script. We use conda.

# in project root.
mkdir -p submodules data data/datasets data/experiments
cd scripts
bash 1_install_conda_portrait_editor.sh
bash 2_install_conda_portrait_champ.sh
bash 3_install_conda_guidance.sh
bash 4_install_conda_perse_avatar.sh

NOTE: In our preprocessing code and training of PERSE, we rely on the face segmentation model from Sapiens (Khirodkar et al.). However, unfortunately, we are unable to provide the face segmentation checkpoint. We recommend that you refer to How to create a face segmentation checkpoint from Sapiens to finetune the face segmentation model and use that.

# in project root.
cd scripts/run
bash 1_generate_prompt.sh
bash 2_edit_portrait.sh
bash 3_normal_portrait.sh
bash 4_generate_guidance.sh
bash 5_preprocess_guidance.sh
bash 6_portrait_champ.sh

# in project root.
cd scripts/run
bash 9_train_perse_avatar_model.sh

# in project root.
cd scripts/run
bash 12_test_perse_avatar_model.sh

Codes are available only for non-commercial research purposes.

If you have any questions about the code, suggestions for improvement, or any other inquiries, please do not hesitate to contact the first author, Hyunsoo Cha (729steven@gmail.com or 243stephen@snu.ac.kr).

If you find our work useful, please cite our paper:

@InProceedings{Cha_2025_CVPR,
    author    = {Cha, Hyunsoo and Lee, Inhee and Joo, Hanbyul},
    title     = {PERSE: Personalized 3D Generative Avatars from A Single Portrait},
    booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)},
    month     = {June},
    year      = {2025},
    pages     = {15953-15962}
}