Pokemon generation using GAN.
Dataset was adapted from https://github.com/msikma/pokesprite. Used Gen_7 sprite (as gmax sprites in Gen_8 differ in size). Cropped and centered to (50,50,3).
Tensorflow 2.4.1, Numpy 1.20.3, Matplotlib 3.3.2
A 2-layer convolutional GAN. 2 layers is sufficient as Pokemon sprites are consisted of mostly simple low-level features. Generator uses Relu activation. Discriminator uses LeakyRelu activation.
Weakness: Network might generates similar images especially the texture. This network is prone to mode collapse.
| 100 epochs | 200 epochs | 400 epochs | 1000 epochs |
|---|---|---|---|
 |
 |
 |
 |
Leverage the Wasserstein distance primarily used to address mode collapse. Images generated has more variety. Both Generator and Discriminator use LeakyRelu activation. In this version, I made the network to train the discriminator more than the generator using "n_critic" following the original implementation.
| 100 epochs | 200 epochs | 400 epochs | 1000 epochs |
|---|---|---|---|
 |
 |
 |
 |
- Due to the simplicity (low-level feature and low resolution) of dataset, a small network is sufficient. I had tried more complicated networks but the results are image blobs.
- Image augmentation (in code "data_augmentation") is useful but too much of it will do more harm than good. Random brightness was not used as the background is white (will turn greyish when used).
- "random_normal_dimensions" for setting the dimension of noise is a hyperparameter. I found that adding a larger noise dimension doesn't show any performance improvement and stick with the current dimension.