Conditional AnimeGAN

PyTorch implementation of conditional Generative Adversarial Network (cGAN) for Anime face generation conditioned on eye color and hair color.

Row-1: Brown Eyes Blonde Hair
Row-2: Blue Eyes Blue Hair
Row-3: Red Eyes Green Hair
Row-4: Purple Eyes Orange Hair
Row-5: Green Eyes Purple Hair
Row-6: Aqua Eyes Pink Hair

You can download the dataset from the following repo.

Training

Download the data and place it in the data/ directory. (Optional) Run prepro.py to clean and preprocess the data. Run train.py to start training. To change the hyperparameters of the network, update the values in the param dictionary in train.py. Checkpoints will be saved by default in the checkpoint directory every 2 epochs. By deafult, GPU will be used for training if available. (Training on CPU is not recommended)

Loss Curve

Generating New Images

To generate new images run generate.py.

python3 generate.py -load_path /path/to/pth/checkpoint -num_output n -eye_color c1 -hair_color c2

• Possible colors for eyes

['yellow', 'gray', 'blue', 'brown', 'red', 'green', 'purple', 'orange',
 'black', 'aqua', 'pink', 'bicolored']

• Possible colors for hair

['gray', 'blue', 'brown', 'red', 'blonde', 'green', 'purple', 'orange',
 'black', 'aqua', 'pink', 'white']

Results

Training Data	cDCGAN after 50 epochs

Some Generated Samples:

Blue Eyes Blonde Hair

Red Eyes Blonde Hair

Green Eyes Purple Hair

Red Eyes Green Hair

Aqua Eyes Pink Hair

Red Eyes Purple Hair

References

1. Mehdi Mirza, Simon Osindero. Conditional Generative Adversarial Nets. [arxiv]
2. Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, Honglak Lee. Generative Adversarial Text to Image Synthesis. [arxiv]
3. m516825/Conditional-GAN [repo]
4. soumith/ganhacks [repo]