[3Xin2] Poster session 1

[3Xin2-38] Investigation of a Method to Ensure Consistency of Generated Images Through Intermediate Representation Operations in U-Net

〇Ryota Yamada¹, Yoshiaki Kurosawa¹, Kazuya Mera¹, Toshiyuki Takezawa¹ (1.Hiroshima City University)

Keywords:Image Generation

Image generation has begun to attract attention rapidly since the advent of Stable Diffusion. Among them, there is a growing interest in video production using image generation. One of the major obstacles to video production using image generation is the difficulty of ensuring image consistency. Many of the methods to improve this situation involve re-training based on inputs and tasks, which may be difficult to use in combination with existing techniques. Our proposed method stores the intermediate representation during the generation of an image, and directly uses and manipulates the result during the generation of another image. The task of generating images of two different types of bottoms was performed to verify the consistency of the images. The results of visual and numerical evaluation showed that the changes in composition and background were suppressed.