BaMSGAN: Self-Attention Generative Adversarial Network with Blur and Memory for Anime Face Generation

In this paper, we propose a novel network, self-attention generative adversarial network with blur and memory (BaMSGAN), for generating anime faces with improved clarity and faster convergence while retaining the capacity for continuous learning. Traditional self-attention generative adversarial networks (SAGANs) produce anime faces of higher quality compared to deep convolutional generative adversarial networks (DCGANs); however, some edges remain blurry and distorted, and the generation speed is sluggish. Additionally, common issues hinder the model’s ability to learn continuously. To address these challenges, we introduce a blurring preprocessing step on a portion of the training dataset, which is then fed to the discriminator as fake data to encourage the model to avoid blurry edges. Furthermore, we incorporate regulation into the optimizer to mitigate mode collapse. Additionally, memory data stored in the memory repository is presented to the model every epoch to alleviate catastrophic forgetting, thereby enhancing performance throughout the training process. Experimental results demonstrate that BaMSGAN outperforms prior work in anime face generation, significantly reducing distortion rates and accelerating shape convergence.