A simple Tensorflow implementation of Generative Adversarial Networks, focusing on generating anime faces.
AniGAN leverages the capabilities of Generative Adversarial Networks (GANs) to produce Anime faces. This project was born out of a passion for both deep learning and anime. It showcases the potential of GANs in creating stunning, high-quality images and provides a foundation for further exploration and improvement.
Some images generated by AniGAN
The model was trained on a dataset containing around 63k anime faces for 100 epochs.
Note -: The model's training was constrained to only 100 epochs because of resource limitations. I believe that training for a longer period will produce more refined results.
- The GAN has two neural networks, the 'generator' and the 'discriminator'.
- The generator takes in a random vector which then uses transposed convolutions to generate an image out of it.
- The discriminator is a Convolutional network that then classifies whether an image is real or fake. It takes in samples of images from the dataset and also images generated by the generator.
- Both networks try to improve each other's performance through backpropagation.
- It can help artists and designers come up with unique character concepts quickly.
- Design custom merchandise, such as posters, prints, and apparel, featuring anime-style artwork.
- We have the flexibility to train this identical model on various datasets, such as human faces, enabling versatile applications like data augmentation and style transfer.
And lastly, I would like to say one more thing - Training GANs is really hard!!!