This project fine-tunes a pre-trained StyleGAN model (originally trained on FFHQ human faces) on a small dataset of 500 manually annotated cartoon face images extracted from webcomics. It uses the FreezeD technique, which freezes early discriminator layers during fine-tuning to preserve learned semantic features while adapting to the new cartoon domain. The model was trained for about 50k steps on a single RTX 2080ti GPU over roughly 10 hours.