CryptoGANs is a deep learning project that uses two different architectures for the generator and discriminator models to regenerate Cryptopunks art and perform a comparative analysis of them. This project was created as a final project for the course Digital Image Processing CS-4055.
Cryptopunks is a collection of 10,000 unique 8-bit characters, each with their own distinct traits and characteristics. The goal of this project is to regenerate these Cryptopunks characters with GANs and create new and unique punk inspired art.
Convolutional networks generate outcomes that are significantly superior to those of plain feed forward networks. Compared to images generated using feed forward networks, those generated using convolutional networks are significantly sharper and more defined. Convolutional Networks also resulted in a significantly lower fake score.
To create even better photos of CryptoPunks in the future, we might investigate the usage of different architectures like Stacked GANs or Wasserstein GANs (WGANs). We might also try utilizing varia- tional auto encoders (VAE) to create images of other distinctive figures or things. To further enhance the findings acquired, we might further look at the usage of other optimization techniques or data augmentation approaches.