1. Evaluation

No ground truth that would enable fair comparisons across GANs → hard to evaluate GANs

1-1. Metrics

• Fidelity
  • Quality of images
• Diversity
  • Variety of images

1-2. Comparing Images

• Pixel Distance → simple, unreliable
• Feature Distance → Using high level features

3. Feature Extraction

• Classifier → Feature Extractor
• Last pooling layer of a classifier is most used

3-1. Embeddings

• Comparing embeddings of fake images and real images to get the feature distance

4. Evaluation Metrics

4-1. Frechet Inception Distance (FID)

Frechet distance