Flickr-Faces-HQ (FFHQ)

Dataset Overview

Name: Flickr‑Faces‑HQ Dataset (FFHQ)

Description: FFHQ is a high‑quality face image dataset containing 70,000 PNG images at 1024×1024 resolution. The dataset exhibits significant diversity in age, race, and background, and includes accessories such as glasses, sunglasses, and hats. Images are sourced from Flickr and have been automatically aligned and cropped.

Usage: Primarily intended for research on Generative Adversarial Networks (GANs); not for development or improvement of facial recognition technologies.

Dataset Content

• Number of Images: 70,000
• Image Format: PNG
• Resolution: 1024×1024
• Dataset Size: 2.56 TB

Dataset Structure

• Root Folder: ffhq-dataset
• Subfolders and Contents:
  • ffhq-dataset-v2.json: Metadata (including copyright information, URLs, etc.) – 255 MB
  • images1024x1024: Aligned and cropped 1024×1024 images – 89.1 GB
  • thumbnails128x128: 128×128 thumbnails – 1.95 GB
  • in-the-wild-images: Original Flickr images – 955 GB
  • tfrecords: Multi‑resolution data for StyleGAN and StyleGAN2 – 273 GB
  • zips: ZIP archives of each folder's contents – 1.28 TB

Dataset Usage

• Download Script: download_ffhq.py script is provided for automated download and verification.
• Training & Validation: First 60,000 images are used for training; the remaining 10,000 for validation.

Copyright & License

• Image Licenses: Various Creative Commons licenses that permit free use, redistribution, and adaptation, with some requiring attribution and indication of changes.
• Dataset License: Released by NVIDIA Corporation under Creative Commons BY‑NC‑SA 4.0, allowing non‑commercial use, redistribution, and adaptation provided the original paper is cited and changes are noted; derivative works must use the same license.

Privacy Protection

• The dataset only includes photos whose authors have explicitly permitted free use and redistribution.
• Mechanisms are provided for users to check whether their photos are included and request removal.