Deep Convolutional Generative Adversarial Networks (DCGAN)-Based Anime Face Generation
Xunxiong Ou · Advances in computer science research · 2024
This study delves into the realm of anime face generation with the aim of empowering individuals to create their own anime characters and easing the burden on artists.Employing Deep Convolutional Generative Adversarial Networks (DCGAN), the research focuses on generating anime face images.The DCGAN model consists of a generator and a discriminator, each designed and trained for their respective roles.The generator employs a convolutional transpose structure, while the discriminator utilizes a convolutional neural network structure.Through simultaneous training of the generator and discriminator using a diverse dataset of anime face images, a comprehensive DCGAN model is developed.Leveraging the Kaggle dataset, the study evaluates the training progress of the model through loss change curves of the generator and discriminator, alongside the final generated anime face images.Comparing different loss change curves and generated images across varying epochs and batch sizes reveals superior performance with 60 epochs compared to 30 epochs, facilitating clearer facial features.Moreover, a batch size of 32 outperforms 256, attributed to its more stable loss change curve.These findings contribute valuable insights to the domain of anime face generation research.