[1Win4] Poster session 1

[1Win4-74] Manga109Script: A Script Dataset for Manga Layout Generation Models

〇Kengo Watanabe¹, Yuro Okada¹, Shun Kitagawa¹, Inaba Michimasa², Atsushi Hashimoto^3,1, Satoshi Kurihara¹ (1.Keio University, 2.The University of Electro-Communications, 3.OMRON SINIC X Corporation)

Keywords:LLM, layout generation, manga

In creative endeavors, a well-designed layout plays a crucial role. An effective layout enhances visual appeal and draws readers’ attention. While this principle applies to media, manga features a unique layout reflecting its narrative. Manga layouts are established in preliminary drafts called “name,” based on scripts and storytelling. Although layout generation models have evolved in domains such as Web UIs, where visual clarity is prioritized, script-based layout generation had not been explored due to data scarcity. In this study, we created script data for approximately 20,000 pages from the Manga109 dataset and built the Manga109Script dataset linking scripts to manga layouts. Using this dataset, we developed manga layout generation models that directly take scripts as input using several LLMs as base models. We evaluated these models, comparing existing generators and assessing the impact of script input to establish a baseline for script-based layout generation (Script2Layout).