Anime Style Image Generation Based on StyleGAN3

Abstract

With the rise of anime culture and the development of computer vision technology, automatically generating images with a specific comic style has become a research hotspot. The animation industry is in urgent need of high-quality and diverse comic-style images, but traditional hand- drawing methods are inefficient. This paper chooses to use the existing Style-Based Generative Adversarial Network（StyleGAN3） model because of its excellent image generation capabilities and high-resolution output. Compared with the earlier StyleGAN2, StyleGAN3 eliminates aliasing in generated images. StyleGAN3 can achieve more natural refinement and generate more realistic and stylized images. This paper adjusts the latent space vector of StyleGAN3 to achieve precise control of anime style features, and to achieve the purpose of anime-style image generation based on StyleGAN3. The output results show that the generated images not only accurately retain the key style elements of the animation, but also show a high degree of diversity and authenticity, which fully verifies the effectiveness and feasibility of the method proposed in this paper. This research result provides strong technical support for animation creation, game design, and other fields, greatly enriches image resources, and significantly improves creation efficiency. It has important application value and practical significance.