1Hong Kong University of Science and Technology (Guangzhou)
2Vivo, Hangzhou
3Hong Kong University of Science and Technology
2Vivo, Hangzhou
3Hong Kong University of Science and Technology
- [2026.01] 🔥 Our inference code and models are released.
- [2025.10] 🔥 High-Quality Face Dataset Released! We release 2,104 high-quality face images with detailed captions.
- [2025.07] Our paper has been accepted by ACM MM 2025.
- [2024.10] This repo is created.
We present a high-quality face dataset for authentic blind face restoration. Due to proprietary constraints, the released dataset is limited to samples collected from Unsplash and is made available under a non-commercial license. The dataset consists of 2,104 high-resolution images, divided into male and female categories.
| Dataset Content | Download Link | File Size |
|---|---|---|
| AuthFace-HQ (Full) | Google Drive | - |
| Category | Count | Content | Description Format |
|---|---|---|---|
| Woman | 1,198 | High-Quality Face Images + Qwen Captions | .png image + .json metadata |
| Man | 906 | High-Quality Face Images + Qwen Captions | .png image + .json metadata |
| Total | 2,104 | - | - |
The dataset is organized as follows. For each subset, we also provide .json files (generated by Qwen) containing detailed textual descriptions or attributes corresponding to the images.
AuthFace-HQ dataset examples.
Click to view Dataset Directory Structure
Dataset_Root
├── woman (1198 images + json captions)
│ ├── woman_11010_qwen.json <-- Textual description
│ ├── woman_11010.png <-- Image file
│ ├── woman_10566.png
│ └── ...
│
├── man (906 images + json captions)
│ ├── man_14935_qwen.json <-- Textual description
│ ├── man_14935.png <-- Image file
│ ├── man_11609.png
│ └── ...
In the project root, follow these steps:
conda create -n authface python==3.10
conda activate authfacepip install -r requriment.txt
huggingface-cli download Ethanliang99/AuthFace --local-dir /path/to/save --repo-type modelsh ./option/test/test_base.shNotes for ./option/test/test_base.sh
- Update the path-related arguments (e.g.,
--pretrained_model_name_or_path,--validate_image_path,--output_dir) to your local paths. --promptand--minor_color_fix_strengthcan be customized, but the paper uses the default values in the script.
If you find our dataset or code useful, please cite our paper:
@inproceedings{liang2025authface,
title={AuthFace: Towards Authentic Blind Face Restoration with Face-oriented Generative Diffusion Prior},
author={Liang, Guoqiang and Fan, Qingnan and Fu, Bingtao and Chen, Jinwei and Gu, Hong and Wang, Lin},
booktitle={Proceedings of the 33rd ACM International Conference on Multimedia},
year={2025}
}