GANkémon

This project aims at the synthetic generation of new Pokemon (GANkemon). The network used is a Progressive Generative Adversarial Network.

Here are some examples of our results:

Datasets

• Pokemon - Image dataset or Alternative Link

• The Complete Pokemon Images Data Set

• Pokemon Images Dataset

• Game Assets Dump

• Bulbapedia - scraped top 10 images from bulbapedia for every Pokemon and manually cleaned the Dataset

some dirty ones

• 7,000 Labeled Pokemon

• Pokemon Generation One

Final Datasets

Base Dataset:

• Pokemon - Image dataset

2D Assets:

• The Complete Pokemon Images Data Set

• Pokemon Images Dataset

• Game Assets Dump

All Assets:

• Pokemon - Image dataset

• The Complete Pokemon Images Data Set

• Pokemon Images Dataset

• Game Assets Dump

Quality Scores

Sources:

• Medium: GAN — How to measure GAN performance?

Scores:

• Inception Score (IS) - measures quality based on quality of generations and their diversity

• Fréchet Inception Distance (FID)

  • uses the Inception Network to extract features and calucates FID based on them
  • is sensitive to mode collapse
  • more robust to noise than IS
  • better measurement for image diversity
  • FID between training and test set should be zero, since both real images (not valid for batches of train)
  • Lower FID values mean better image quality and diversity

• Wasserstein Distance

• SSIM Metric

• Precision?

• Recall?

• F1-Score?