7 万张真实面孔人脸图像数据集

Flickr-Faces-HQ（FFHQ）是一个高质量的人脸图像数据集，最初是作为生成对抗网络（GAN）的基准创建的：

基于风格的生成对抗网络生成器架构
Tero Karras (NVIDIA)、Samuli Laine (NVIDIA)、Timo Aila (NVIDIA)
https://arxiv.org/abs/1812.04948

原数据集包含 70,000 张 1024×1024 分辨率的高质量 PNG 图像，在年龄、种族和图像背景方面变化很大。它还很好地涵盖了眼镜、太阳镜、帽子等配饰。这些图像是从 Flickr 爬取的，因此继承了该网站的所有偏见，并使用 dlib 自动对齐和裁剪。只收集了许可许可下的图像。使用各种自动过滤器来修剪该集合，最后使用 Amazon Mechanical Turk 删除偶尔出现的雕像、绘画或照片。

该数据集经过处理后，格式统一为jpg，分辨率缩放到512x512，命名统一：8-8 uuid。

Flickr-Faces-HQ (FFHQ) is a high-quality face image dataset initially created as a benchmark for Generative Adversarial Networks (GANs):

*A Style-Based Generator Architecture for Generative Adversarial Networks*
Authors: Tero Karras (NVIDIA), Samuli Laine (NVIDIA), Timo Aila (NVIDIA)
https://arxiv.org/abs/1812.04948

The original dataset contains 70,000 high-quality 1024 × 1024 resolution PNG images, with considerable variation in age, ethnicity and image background. It also covers a wide range of accessories such as eyeglasses, sunglasses, hats and more. These images were crawled from Flickr, thus inheriting all biases of the platform, and were automatically aligned and cropped using dlib. Only images released under valid licenses were collected. Various automatic filters were used to prune the dataset, and Amazon Mechanical Turk was finally employed to remove occasional misclassified content including statues, paintings or non-human photographs.

The processed dataset is uniformly formatted as JPG, with all images resized to 512 × 512 resolution, and follows a unified naming convention: 8-8 uuid.

搜集汇总

背景与挑战

背景概述

该数据集是一个高质量的人脸图像集合，包含70,000张经过处理的图像，适用于生成对抗网络（GAN）的研究和开发。图像具有多样性和高质量，覆盖不同年龄、种族和背景，并包含各种配饰。