Conditional Anime Generation using conditional GAN.
Different from vanilla GAN, the generator and discriminator in conditional GAN (abbr. CGAN) will be given a specific condition (could be a one-hot vector indicating classes or a word-embedding).
The discriminator considers the following cases:
- Real distribution, correct conditon -> positive
- Real distribution, wrong condition -> negative
- Fake distribution, real condition -> negative
Our training process roughly follows the paper: https://arxiv.org/abs/1605.05396, only that the text descriptions are replaced with a one-hot class vector.
| Fixing noise |
|---|
 |
| Changing eye color |
|---|
 |
 |
 |
| Change hair color |
|---|
 |
 |
 |
| Condition | Generated |
|---|---|
| blonde hair, purple eyes |  |
| blue hair, red eyes |  |
| white hair, green eyes |  |