[1J5-OS-10c] OS-10

[1J5-OS-10c-01] Generation of person images from a set of lines using ChatGPT

〇Tsubasa Odani¹, Wataru Sunayama², Shun Hattori² (1. Graduate School of Engineering, The University of Shiga Prefecture, 2. School of Engineering, The University of Shiga Prefecture)

Keywords:ChatGPT, Large Language Models, Estimation of person characteristics, Generation of person images

Generative AI has been attracting attention in recent years, and chatbots and smart speakers are two of the ways in which it is being used, and the opportunities for their use are increasing. Chatbots have pre-defined personalities that cannot be changed by the user. By providing a person's lines and other relevant information as input, the output can reflect the way the person speaks and thinks, allowing for user-preferred dialogue. Many studies have focused on the way a person speaks using dialogue information. However, they do not focus on the generation of person images, and the content of the learned person images cannot be confirmed. In this study, we propose a system that extracts lines that reflect a person's characteristics using ChatGPT and adds an explanation of the situation in which the lines are used, thereby generating a portrait of a person from a smaller number of lines, consisting of the person's actions and the ''way of thinking about things,'' ''policy,'' and ''action policy''' that form the basis of the lines. Through experiments, we have changed the method for extracting lines. Through experiments, we compared the estimation results with different methods of extracting lines, and evaluated the quality of the estimation results.