Generating new Pokemon with specified type combinations, based on ACGAN DiffAugment D2DCE.
The used dataset has been uploaded to Kaggle. It contains in total 10,437 Pokemon sprites (half of them are shiny variants) in 96x96 resolution collected from 898 Pokemon in different games. Below are the collected images of Bulbaraur:
| Gen3 E | Gen3 E (Frame 2) | Gen3 FL | Gen4 DP |
|---|---|---|---|
  |
  |
  |
  |
| Gen4 HS | Gen4 HS (Frame 2) | Gen5 | Gen5 (Back) |
|---|---|---|---|
  |
  |
  |
  |
Labels that may relate to Pokemon’s appearance are extracted from PokeIndex and stored as a CSV file. Below is the sample entry of Bulbasaur:
| name | pokedex_id | type1 | type2 | primary_color | shape | legendary | mega_evolution | alolan_form | galarian_form | gigantamax |
|---|---|---|---|---|---|---|---|---|---|---|
| Bulbasaur | 1 | Grass | Poison | Green | Quadruped | FALSE | FALSE | FALSE | FALSE | FALSE |
In this repository only the type labels are used. Of course we can also try generating new Pokemon conditioned on other labels.
The notebook is self-contained and ran on Google Colab. Just modify those Colab-specific codes and Google Drive specific paths to run locally.
Sample generated images of all possible (either one or two, in total 171) type combinations are provided in gen_results/, and the used model weights is provided in checkpoints/.
For example, let’s create some fire fairies:
and some electric ice-mon:
and some dark bugs (that we programmers hate):
