A StyleGAN3-based facial image synthesis database for rare genetic disease phenotypes

• Overview
• Key Features
• Model Architecture
• Synthetic Image Gallery
• Interactive Web Gallery
• Getting Started
  • 1. Environment Setup
  • 2. Data Preprocessing
  • 3. Model Training
  • 4. Generate Synthetic Images
• License

This repository provides a class-conditional StyleGAN3 model trained to generate synthetic facial images representing patients with rare genetic diseases. The generated images capture disease-specific facial phenotype patterns (facial gestalt) and serve multiple purposes in computational biology and medical AI research.

Primary Use Cases:

• Visualization of disease-specific facial patterns learned from real patient data
• Generation of synthetic training/evaluation data for rare disease facial recognition models
• Educational materials and clinical teaching resources
• Data augmentation for downstream phenotype analysis tasks
• Privacy-preserving research by providing synthetic alternatives to real patient images

ℹ️ Important Note: This repository focuses on synthetic face generation using StyleGAN3. It does not perform disease diagnosis. All generated images are fully synthetic and do not represent real patients.

• ✨ Class-conditional generation across multiple rare genetic disease categories
• 🎯 Alias-free architecture (StyleGAN3) for improved geometric consistency
• 🖼️ High-quality 224×224 resolution synthetic faces compatible with standard computer vision pipelines
• 🎮 Controllable generation with latent space manipulation capabilities
• 🌐 Interactive web gallery for exploring generated samples by disease type, demographics, and other attributes

The model is based on StyleGAN3 with the following key characteristics:

• Alias-free generator ensuring better pose consistency and geometric transformations
• Conditional generation supporting 10+ disease classes (configurable)
• Training data: Preprocessed facial images from genetic disease patient databases
• Resolution: 224×224 pixels (can be adjusted)

We provide sample synthetic faces generated by our trained StyleGAN3 model across different disease categories:

Explore our complete synthetic image database through an interactive web interface:

🌐 Phenotype-Disease Synthetic Image Database

• Browse by disease type: Filter synthetic faces by specific genetic disease categories
• Demographic filtering: Search by age range, gender, and ethnicity
• Metadata annotations: View generation parameters and conditional attributes
• Download capability: Export images for research and educational use
• High-quality samples: Access almost hundreds of synthetic faces per disease class

⚠️ Disclaimer: All images in the gallery are fully synthetic and generated by StyleGAN3. They do not represent real patients and are intended solely for research and educational purposes. Please respect ethical guidelines when using these materials.

# Clone the repository git clone https://github.com/WGLab/PDI-DB.git cd PDI-DB # Create a new conda environment conda create -n PDI-DB python=3.8 # Activate environment conda activate PDI-DB # Install PyTorch with CUDA support pip install torch==2.7.1 torchvision==0.22.1 torchaudio==2.7.1 --index-url https://download.pytorch.org/whl/cu118 # Install other dependencies pip install click pillow scipy numpy requests tqdm ninja matplotlib imageio pip install imgui glfw pyopengl imageio-ffmpeg pyspng

Before training, you need to prepare your facial phenotype images in the correct format.

Organize your raw images in the following structure:

raw_data/ ├── disease_0/ │ ├── patient_001.jpg │ ├── patient_002.jpg │ └── ... ├── disease_1/ │ ├── patient_001.jpg │ ├── patient_002.jpg │ └── ... └── disease_N/ ├── patient_001.jpg └── ... 

📊 Note: Training Dataset Information
The synthetic images showcased in this repository were generated using a model trained on 2,357 facial images across 11 categories:

• Disease images (10 categories, 1,857 images): Cornelia de Lange Syndrome (447), Williams-Beuren Syndrome (273), Noonan Syndrome (217), Kabuki Syndrome (206), KBG Syndrome (179), Angelman Syndrome (160), Rubinstein-Taybi Syndrome (116), Smith-Magenis Syndrome (102), Nicolaides-Baraitser Syndrome (97), and 22q11.2 Deletion Syndrome (60). All disease images are sourced from the GestaltMatcher Database (GMDB).
• Healthy control images (500 images): Selected from the FFHQ dataset with age < 20 years.

You can also train the model on your own custom disease categories and sample sizes by following the data preprocessing steps below.

Manual Preprocessing:

1. Face Alignment: Align all faces to a standard pose using landmark detection
2. Resolution: Resize images to 224×224 pixels
3. Format: Convert to PNG or JPG format
4. Quality Control: Remove low-quality or ambiguous images

python dataset_tool.py \ --source=raw_data/ \ --dest=datasets/phenotype_disease_224x224.zip \ --resolution=224x224

The final dataset should be a ZIP file containing:

phenotype_disease_224x224.zip ├── 00000/ │ ├── img00000000.png │ ├── img00000001.png │ └── ... ├── 00001/ │ └── ... └── dataset.json # Metadata with class labels 

Example dataset.json format:

{ "labels": [ ["img00000000.png", 0], ["img00000001.png", 0], ["img00000002.png", 1] ] }

Where the second value in each pair is the disease class index (0, 1, 2, ..., N).

Train a class-conditional StyleGAN3 model on your phenotype dataset:

python train.py \ --outdir=./training-runs \ --data=./datasets/phenotype_disease_224x224.zip \ --cfg=stylegan3-t \ --gpus=8 \ --batch=32 \ --cond=True \ --gamma=2 \ --mirror=1 \ --kimg=5000

python train.py \ --outdir=./training-runs \ --data=./datasets/phenotype_disease_224x224.zip \ --cfg=stylegan3-t \ --gpus=8 \ --batch=32 \ --cond=True \ --gamma=2 \ --mirror=1 \ --kimg=5000 \ --snap=20

Training outputs are saved in timestamped directories:

training-runs/ └── 00000-stylegan3-t-phenotype_disease_224x224-gpus8-batch32-gamma2/ ├── network-snapshot-000000.pkl # Initial model ├── network-snapshot-000100.pkl # Checkpoint at 100k images ├── fakes000000.png # Sample images at start ├── fakes000100.png # Sample images at 100k ├── training_stats.jsonl # Training metrics └── metric-fid50k_full.jsonl # FID scores (if enabled) 

Key files to monitor:

• fakes*.png: Visual quality of generated images
• training_stats.jsonl: Loss values and training metrics
• network-snapshot-*.pkl: Model checkpoints

Once training is complete, use the trained model to generate synthetic phenotype images.

python gen_images.py \ --network=./training-runs/00000-stylegan3-t-.../network-snapshot-005000.pkl \ --outdir=./generated_images \ --samples=100 \ --class=0 \ --trunc=1.0

The --trunc parameter controls the trade-off between quality and diversity:

• --trunc=0.5: Higher quality, less diversity (more typical faces)
• --trunc=0.7: Balanced quality and diversity
• --trunc=1.0: Maximum diversity, lower average quality

Examples:

# High quality, low diversity python gen_images.py --network=model.pkl --outdir=out_quality --samples=100 --class=0 --trunc=0.5 # Balanced python gen_images.py --network=model.pkl --outdir=out_balanced --samples=100 --class=0 --trunc=0.7 # High diversity python gen_images.py --network=model.pkl --outdir=out_diverse --samples=100 --class=0 --trunc=1.0

This project is licensed under the MIT License.