leaderj1001 / minimal-stylegan

minimal-stylegan (A Style-Based Generator Architecture for Generative Adversarial Networks)

• Paper Link
• Author: Tero Karras, Samuli Laine, Timo Aila
• Organization: NVIDIA
• Reference Code URL

Method

1. prepare dataset (FFHG Dataset)
   - Use 0-2999 photos (Total: 3000)
   

python prepare.py

2. train

python main.py

3. eval

python eval.py

Results

• Batch_size = 30
• Loss = 'wgan-gp'
• Training Time: 75 Hours, (80000 Iterations)
• 1 GPU, image size: 128x128