Exploring the influence of generator channel number on the quality of anime-style portrait generation based on DCGAN

Abstract

In the realm of contemporary image synthesis, this research delves into a crucial objective: exploring the connection between the quantity of generator channels and the production of anime-style portraits through Deep Convolutional Generative Adversarial Networks (DCGAN). Employing an extensive dataset of anime faces encompassing diverse artistic styles, this study systematically examines the nuanced interplay between architectural parameters and the fidelity and intricacy of the generated images. By employing the Frechet Inception Distance (FID) as a metric for image quality, this investigation contributes significantly to the field by enhancing the understanding of how the number of generator channels impacts the ultimate quality of anime-style portraits. The DCGAN framework, and in particular its variants, is the backbone of this investigation. The generator and discriminator components are involved in adversarial training, a competitive process that improves image quality through iterations. The findings reveal a non-linear relationship between the number of generator channels and image quality. While increasing the number of channels initially improves image quality and decreases the FID value, exceeding the optimal threshold leads to diminishing returns and image quality degradation. The intricate interplay between structure selection and image quality is further confirmed by the dynamics of the generator and discriminator loss functions. By elucidating the trade-off between complexity and image fidelity, this study contributes to the advancement of image synthesis techniques and encourages future exploration of architectural nuances in the field of artistic image generation.