This repository provides the FFHQ-UV dataset, a large-scale collection of over 50,000 high-quality facial UV-texture maps. It addresses the need for normalized, evenly lit, and clean facial textures essential for realistic 3D face reconstruction, targeting researchers and developers in computer vision and graphics.

How It Works

The dataset is generated using a fully automatic pipeline that leverages StyleGAN-based facial image editing. This process involves creating multi-view normalized face images from single inputs, followed by an elaborate UV-texture extraction, correction, and completion procedure. This approach yields diverse and high-quality UV-maps, surpassing existing datasets.

Quick Start & Requirements

• Installation: Requires Linux, Anaconda, CUDA 10.0+, CUDNN 7.6.0+, Python 3.7. Specific PyTorch (1.7.1), TensorFlow (1.15.0), and other packages (dlib, azure-cognitiveservices-vision-face, kornia, etc.) are needed. PyTorch3D and Nvdiffrast require manual compilation from source.
• Dataset: Download instructions are provided via a README link.
• Setup: The setup involves installing numerous dependencies and potentially compiling complex libraries, which may take a significant amount of time.
• Links: Paper: https://arxiv.org/abs/2211.13874, Dataset: README link

Highlighted Details

• Dataset contains over 50,000 high-quality UV-texture maps.
• Includes a pipeline for generating UV-texture maps from single facial images via facial editing or RGB fitting.
• Provides solutions for integrating UV-textures with FLAME meshes and adding eyeballs to head meshes.
• Offers pre-trained checkpoints and topology assets for the creation pipeline.

Maintenance & Community

The project is associated with CVPR 2023 and builds upon foundational works from StyleGAN, e4e, DPR, and StyleFlow. Support is acknowledged from Tencent AI Lab. Further details on dataset creation and RGB fitting are available via linked READMEs.

Licensing & Compatibility

The repository's license is not explicitly stated in the provided README. However, the project acknowledges dependencies on libraries with various licenses, including those from Facebook Research and NVlabs. Compatibility for commercial use or closed-source linking would require clarification of the specific license applied to the FFHQ-UV dataset itself.

Limitations & Caveats

The facial attribute detection step requires Microsoft Face API, which is not accessible to new users, necessitating alternative solutions or manual data preparation. The StyleGAN-based editing can alter facial identity, potentially reducing texture fidelity. The RGB fitting method, while high-fidelity, may produce less detailed textures than the dataset itself due to GAN-based decoder limitations.