README.md

---
license: cc-by-nc-sa-4.0
---
# Dataset Card for Arc2Face

<div align="center">

[**Project Page**](https://arc2face.github.io/) **|** [**Paper (ArXiv)**](https://arxiv.org/abs/2403.11641)

</div>

This is the dataset used in [Arc2Face: A Foundation Model for ID-Consistent Human Faces](https://arc2face.github.io/) (ECCV 2024).

## Dataset Summary

This dataset consists of approximately 21M facial images from 1M identities at a resolution of 448×448. It was produced by upsampling 50% of the images from the [WebFace42M database](https://openaccess.thecvf.com/content/CVPR2021/html/Zhu_WebFace260M_A_Benchmark_Unveiling_the_Power_of_Million-Scale_Deep_Face_CVPR_2021_paper.html) (originally at 112×112 resolution) using a state-of-the-art blind face restoration [network](https://github.com/TencentARC/GFPGAN). This dataset was used to train the identity-conditioned generative face model presented in [Arc2Face](https://arxiv.org/abs/2403.11641).

## Tasks
The Arc2Face model is based on Stable Diffusion v1.5 and is designed for generating images at 512×512 pixels. To accommodate the requirements of large diffusion models, Arc2Face introduces a refined version of the WebFace42M dataset. Although the original database is intended for Face Recognition (FR) training, the restored dataset provided here is designed for training generative models. Its large number of IDs and considerable intra-class variability make it particularly helpful for ID-conditioned generation.

Please note that the original WebFace42M dataset contains images tailored to extreme conditions for FR robustness. Despite post-restoration filtering, the restored dataset may still include some poor quality 448×448 images. Moreover, all images are limited to tightly cropped facial areas. Therefore, it is suggested to use this dataset in combination with other high-quality datasets (e.g., FFHQ) when training face models, as described in the [paper](https://arxiv.org/abs/2403.11641).

## Dataset Structure

The dataset consists of 35 zip files split into 5 groups (7 zip files per group). Each zip file is approximately 30GB in size.

You can download the zip files from this repository manually or using python (e.g., for the first zip):

```python
from huggingface_hub import hf_hub_download

hf_hub_download(repo_id="FoivosPar/Arc2Face", filename="0/0_0.zip", local_dir="./Arc2Face_data", repo_type="dataset")
```

And unzip them: 
```bash 
unzip 0/0_0.zip -d ./Arc2Face_448x448
```

After unzipping, the dataset structure will be:

```
Arc2Face_448x448
       └── IDs
            └── images
```

Please note that due to the large dataset size, downloading and unzipping may take many hours to complete.

## License

The dataset is made available under [Creative Commons BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/) license. You can use, redistribute, and adapt it for **non-commercial** purposes, as long as you:
- give appropriate credit by citing our paper
- indicate any changes that you have made
- distribute any derivatives under the same license.

https://creativecommons.org/licenses/by-nc-sa/4.0/

## Citation

If you use this dataset in your research, please cite our paper:
```bibtex
@inproceedings{paraperas2024arc2face,
      title={Arc2Face: A Foundation Model for ID-Consistent Human Faces}, 
      author={Paraperas Papantoniou, Foivos and Lattas, Alexandros and Moschoglou, Stylianos and Deng, Jiankang and Kainz, Bernhard and Zafeiriou, Stefanos},
      booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
      year={2024}
}
```
as well as the original dataset paper:
```bibtex
@inproceedings{zhu2021webface260m,
      title={WebFace260M: A Benchmark Unveiling the Power of Million-scale Deep Face Recognition},
      author={Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Jiwen Lu, Dalong Du, Jie Zhou},
      booktitle={IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
      year={2021}  
}
```