Performance Comparison between Generative Adversarial Networks (GAN) Variants in Generating Anime/Comic Character Images - A Preliminary Result

In recent years, the popularity and demand for digital animation, specifically anime characters, has brought significant challenges and opportunities for the world of computer graphics and artificial intelligence. This research dives deep into the comprehensive exploration of two well-known Generative Adversarial Networks (GANs)— Deep Convolutional Generative Adversarial Network (DCGAN) and CycleGAN, — with a specific focus on anime character generation. GANs, consisting of a generator and discriminator, operate in a feedback loop to create and evaluate synthetic data. This research will identify challenges within each GAN model and develop objectives to address these challenges. Both of GAN models were mainly executed in the Google Colab environment to optimize the GPU-accelerated runtime. The dataset is sourced from publicly accessible anime image repositories and both GAN models will be evaluated using Fréchet Inception Distance (FID) and Inception Score (IS). FID compares the distribution of generated images with the distribution of a set of real images (“ground truth”) while IS only evaluates the distribution of generated images. Together, they provide a comprehensive evaluation of a GAN's performance. In the realm of anime character generation, achieving authenticity and diversity is crucial.