By Haoran Bai, Di Kang, Haoxian Zhang, Jinshan Pan, and Linchao Bao
In CVPR 2023 [Paper: https://arxiv.org/abs/2211.13874]
Rendering demos [YouTube video]

FFHQ-UV is a large-scale facial UV-texture dataset that contains over 50,000 high-quality texture UV-maps with even illuminations, neutral expressions, and cleaned facial regions, which are desired characteristics for rendering realistic 3D face models under different lighting conditions.

The dataset is derived from a large-scale face image dataset namely FFHQ, with the help of our fully automatic and robust UV-texture production pipeline. Our pipeline utilizes the recent advances in StyleGAN-based facial image editing approaches to generate multi-view normalized face images from single-image inputs. An elaborated UV-texture extraction, correction, and completion procedure is then applied to produce high-quality UV-maps from the normalized face images. Compared with existing UV-texture datasets, our dataset has more diverse and higher-quality texture maps.

[2022-07-11] A solution for using our UV-texture maps on a FLAME mesh is available [here].
[2022-07-10] A more detailed description and a new version of the RGB fitting process is available [here].
[2022-07-10] A more detailed description of the facial UV-texture dataset creation pipeline is available [here].
[2022-03-17] The source codes for adding eyeballs into head mesh are available [here].
[2022-03-16] The project details of the FFHQ-UV dataset creation pipeline are released [here].
[2022-03-16] The OneDrive download link was updated and the file structures have been reorganized.
[2022-02-28] This paper will appear in CVPR 2023.
[2022-01-19] The source codes are available, refer to [here] for quickly running.
[2022-12-16] The OneDrive download link is available.
[2022-12-16] The AWS CloudFront download link is offline.
[2022-12-06] The script for generating face images from latent codes is available.
[2022-12-02] The latent codes of the multi-view normalized face images are available.
[2022-12-02] The FFHQ-UV-Interpolate dataset is available.
[2022-12-01] The FFHQ-UV dataset is available [here].
[2022-11-28] The paper is available [here].

• Linux + Anaconda
• CUDA 10.0 + CUDNN 7.6.0
• Python 3.7
• dlib: pip install dlib
• PyTorch 1.7.1: pip install torch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2
• TensorBoard: pip install tensorboard
• TensorFlow 1.15.0: pip install tensorflow-gpu==1.15.0
• MS Face API: pip install --upgrade azure-cognitiveservices-vision-face
• Other packages: pip install tqdm scikit-image opencv-python pillow imageio matplotlib mxnet Ninja google-auth google-auth-oauthlib click requests pyspng imageio-ffmpeg==0.4.3 scikit-learn torchdiffeq==0.0.1 flask kornia==0.2.0 lmdb psutil dominate rtree
• Important: OpenCV's version needs to be higher than 4.5, otherwise it will not work well.
• PyTorch3D and Nvdiffrast:

mkdir thirdparty
cd thirdparty
git clone https://github.com/facebookresearch/iopath
git clone https://github.com/facebookresearch/fvcore
git clone https://github.com/facebookresearch/pytorch3d
git clone https://github.com/NVlabs/nvdiffrast
conda install -c bottler nvidiacub
pip install -e iopath
pip install -e fvcore
pip install -e pytorch3d
pip install -e nvdiffrast

• Please refer to this [README] to download the dataset and learn more.

• Please refer to this [README] for details of checkpoints and topology assets.

• Please refer to this [README] for details of running facial UV-texture dataset creation pipeline.
• Microsoft Face API is not accessible for new users, one can find an alternative API, or manually fill in the json file to avoid this step.
• We provide the detected facial attributes of the FFHQ dataset we used, please find details from here.

• Please refer to this [README] for details of running RGB fitting process.
• We provide a new version of RGB fitting, where the FFHQ-UV-Interpolate dataset is used and the GAN-based texture decoder only generates facial textures. More details can found here.

• Prepare a head mesh with HiFi3D++ topology, which is without eyeballs.
• Modify the configuration and then run the following script to add eyeballs into head mesh.

sh run_mesh_add_eyeball.sh  # Please refer to this script for detailed configuration

There are two ways to generate a UV-texture map from a given facial image:

1. Facial editing + texture unwrapping (Section 3.1 of the paper)
2. RGB fitting (Section 4.2 of the paper)

• The FFHQ-UV dataset is created from the FFHQ dataset in this way.
• See source codes of facial UV-texture dataset creation pipeline, which including GAN inversion, attribute detection, StyleGAN-based editing, and texture unwrapping steps.
• Advantages:
  • The generated textures are directly extracted from facial images, which are detailed and with high-quality.
• Disadvantages:
  • The GAN inversion step would be failed for some samples.
  • The attribute detection step requires the Microsoft Face API, which is no longer accessible.
  • The StyleGAN-based editing would change the ID of the faces, resulting in textures with low-fidelity.

• This is the proposed 3D face reconstruction algorithm, which uses FFHQ-UV dataset to train a nonlinear texture basis.
• See source codes of RGB fitting.
• Advantages:
  • The generated textures are fitted based on the supervision of input faces, which are with high-fidelity.
• Disadvantages:
  • The textures are generated by a GAN-based texture decoder, sometimes with less detail than texture maps in the FFHQ-UV dataset. This is mainly due to the limitations of the nonlinear texture basis design and training.
  • Future work will improve the generation capabilities of the nonlinear texture basis, in order to take full advantage of the high-quality and detailed UV texture maps in the FFHQ-UV dataset.

• We provide a solution for using our UV-texture maps on a FLAME mesh, please refer to this [README] for details.

@InProceedings{Bai_2023_CVPR,
  title={FFHQ-UV: Normalized Facial UV-Texture Dataset for 3D Face Reconstruction},
  author={Bai, Haoran and Kang, Di and Zhang, Haoxian and Pan, Jinshan and Bao, Linchao},
  booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
  month={June},
  year={2023}
}

This implementation builds upon the awesome works done by Tov et al. (e4e), Zhou et al. (DPR), Abdal et al. (StyleFlow), and Karras et al. (StyleGAN2, StyleGAN2-ADA-PyTorch, FFHQ).

This work is based on HiFi3D++ topology, and was supported by Tencent AI Lab.