ADL-Pokify-GAN

The proposed Generative Adversarial Networks (GAN) is a conditional Generative Adversarial Networks (cGAN) based on CycleGAN. The main difference is that we will be passing the labels such that the discriminator can compare the generated pokemon with a real pokemon that belongs to the same type of animal.

We have also used StyleGAN2 as a baseline to compare with the GAN that we proposed. The following image is the results of the generated images.

Getting Started

Links to download pretrained-model and datasets

Link to dataset for CycleGAN: HERE

Link to dataset for StyleGAN: HERE

Link to models: HERE

Preparing Dataset

For cGAN, dataset.json is required, but need to delete the file if using the normal GAN.

Move the dataset to folders:

1. StyleGAN: DO NOT unzip pkmn_label.zip, move to ./stylegan/datasets.

2. CycleGAN: Unzip dataset_cyclegan.zip. MovetrainA and trainB to ./cyclegan/train_datasets. To test images, create another folder call testA.

Preparing Pre-trained Models

Move the model to folders:

1. StyleGAN: stylegan_conditional.pkl and stylegan_unconditional.pkl to ./stylegan/training_runs.

2. CycleGAN: animal2pkmn_cond, animal2pkmn_new and animal2pkmn_xavier to ./cyclegan/checkpoints.

Using the Models

Refer to respective Jupyter Notebook file. StyleGAN.ipynb and CycleGAN.ipynb.

References

Karras, T., Aittala, M., Hellsten, J., Laine, S., Lehtinen, J., & Aila, T. (2020). Training Generative Adversarial Networks with Limited Data. Proc. NeurIPS. Available at: https://github.com/NVlabs/stylegan2-ada-pytorch.

Zhu, J.-Y., Park, T., Isola, P., & Efros, A. A. (2017). Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks. Computer Vision (ICCV), 2017 IEEE International Conference On. Available at: https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix

PokeAPI, The RESTful Pokémon API. Available at: https://github.com/PokeAPI/pokeapi