Fine-Tuning StyleGAN for CryptoPunk-Style Image Generation: A Study in Style Transfer and Generative Adversarial Networks

Seong Woo Ahn, Alexandre Petit, Victor Osso

Abstract:

This report explores the application of StyleGAN and StyleGAN2 architectures to develop a CryptoPunk-style image generator, blending artistic creativity with cutting-edge generative adversarial networks (GANs). The project aimedto fine-tune StyleGAN models on a dataset of 10,000 pixelated CryptoPunk images, leveraging transfer learningto create personalized cartoon versions of user-provided
inputs. Despite challenges such as low-resolution training data and limited computational resources, our experiments successfully demonstrated the feasibility of generating CryptoPunk-style outputs by fine-tuning a pre-trained StyleGAN model. Through trials of layer swapping and style mixing, the project also provided insights into the mechanisms behind style transfer and advanced GAN techniques. While the final implementation faced limitations, the study highlights the potential for further optimization and application of StyleGAN architectures in creative AI projects.