Synthetic Faces High Quality (SFHQ) dataset

Dataset Overview

Dataset Name

• Synthetic Faces High Quality (SFHQ) dataset

Dataset Composition

• Total Images: ~425,000
• Resolution: 1024 × 1024
• Four Parts:
  • Part 1: 89,785 images sourced from Artstation‑Artistic‑face‑HQ Dataset (AAHQ), Close‑Up Humans Dataset, and UIBVFED Dataset.
  • Part 2: 91,361 images sourced from Face Synthetics Dataset and Stable Diffusion v1.4.
  • Part 3: 118,358 images generated via the StyleGAN2 mapping network.
  • Part 4: 125,754 images generated via Stable Diffusion v2.1.

Generation Process

• Inspiration Sources: paintings, 3D models, text‑to‑image generators, etc.
• Image Processing: StyleGAN2 latent‑space encoding and fine‑tuning to produce photo‑realistic images.
• Selection: Semi‑automatic and manual filtering using a visual taste approximator tool.

Additional Information

• Facial Features: 110 facial landmark points and semantic segmentation maps.
• Tools: explore_dataset.py script provided for accessing landmarks, masks, and text‑based search.
• Privacy & License: All images are synthetic; no privacy or copyright concerns.

Use Cases

• Training machine‑learning models, especially generative adversarial networks (e.g., StyleGAN).
• Provides extensive diversity across identity, ethnicity, age, pose, expression, lighting, hairstyle, and hair color.

Download

• Available on Kaggle as separate parts.