“High-Quality Synthetic Character Image Extraction via Distortion Recognition”, Tomoya Sawada, Marie Katsurai, Masashi Okubo2023-07-09 (Danbooru AI, CNN, StyleGAN anime)⁠:

Digital avatars have become indispensable in the digital age. In Japan, virtual characters using illustration-style avatars have gained popularity and are generating a large economic impact. However, the creation of such avatars is a time-consuming and costly process that requires a great deal of expertise. To support avatar creation, research of automatic generation of character design and textures for 3D models have emerged. However, deep learning-based generative models sometimes synthesize corrupted outputs. Methods to detect collapsed outputs from a generative model have not been explored, and users of the generator need to manually exclude such outputs.

In this paper, we propose a method to extract high-quality images from a set of synthetic illustrations, generated by a deep learning model [StyleGAN2], based on the degree of distortion of the images. As it is difficult to prepare real-world distorted images to train a distortion recognition model [ConvNeXt], we propose a simple procedure to create pseudo-distorted images.

Experimental results showed superior results of the proposed method in distinguishing between human-drawn images and generated images, compared to baseline methods. Furthermore, we sorted the generated images using the confidence level of the trained distortion detection model, and qualitatively confirmed that the proposed method produces results closer to human perception.

[Keywords: high-quality character image extraction, distortion recognition, deep learning]