How Panel Layouts Define Manga: Insights from Visual Ablation Experiments
Siyuan Feng ( [email protected] ), Teruya Yoshinaga ( [email protected] ), Katsuhiko Hayashi ( [email protected] ), Koki Washio ( [email protected] ), Hidetaka Kamigaito ( [email protected] ) · 2024
Read How Panel Layouts Define Manga: Insights from Visual Ablation Experiments by Siyuan Feng ( [email protected] ), Teruya Yoshinaga ( [email protected] ), Katsuhiko Hayashi ( [email protected] ), Koki Washio ( [email protected] ), Hidetaka Kamigaito ( [email protected] ) (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-19141.txt.
Introduction
Siyuan Feng and co-authors test whether panel layouts alone can identify a manga in a 2024 paper. Manga is a global medium, yet what makes one work visually distinct is unmeasured. The team trained a deep learning model to guess a comic's title from facing-page images. Then they stripped the pages down, in ablation tests, to panel frames only.
The experiments used over ten thousand page spreads from one hundred four works across twelve genres. Heat maps showed where the model looked. Even with characters and text removed, layout carried a strong fingerprint of each work. The authors conclude that panel arrangement is not mere packaging. It encodes a creator's habits deeply enough for a machine to recognize. They offer the ablation method as a way to weigh each visual element's role in defining a style.
Catalog record
| CREATOR | Siyuan Feng ( [email protected] ), Teruya Yoshinaga ( [email protected] ), Katsuhiko Hayashi ( [email protected] ), Koki Washio ( [email protected] ), Hidetaka Kamigaito ( [email protected] ) |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 28 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2412-19141
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
How Panel Layouts Define Manga: Insights from Visual Ablation Experiments Siyuan Feng ( [email protected] ), Teruya Yoshinaga ( [email protected] ), Katsuhiko Hayashi ( [email protected] ), Koki Washio ( [email protected] ), Hidetaka Kamigaito ( [email protected] ) Abstract Manga has gained global popularity, yet how its visual elements, such as characters, text, and panel layouts, reflect the uniqueness of individual works remains underexplored. This study investigates the contribution of panel layouts to manga identity through both quantitative and qualitative analysis. We trained a deep learning model to classify manga titles based solely on facing page images, and performed ablation experiments by removing characters and text, retaining only panel frame structures. Using 10,122 images from 104 works in 12 genres in the Manga109 dataset, we demonstrate that panel layouts alone enable high-accuracy classification. Grad-CAM visualizations further reveal that the models focus on layout features such as size, spacing, and alignment. These findings suggest that panel layouts encode work-specific stylistic patterns and support visual narrative comprehension, highlighting their role as a key component of manga’s visual identity. Keywords: Artwork layout analysis; Digital comics; Comic classification; Deep learning; Visualization analysis Introduction Comic is a globally popular medium, encompassing a wide range of themes and genres. With the increasing adoption of e-books, the digital comic market has seen significant growth, and many comic works are now available in electronic formats. This shift has drawn considerable attention to multimedia processing research for digital comics, particularly Japanese comics (referred to simply as “manga” in this paper). Manga, a distinctive form of Japanese comics, is characterized by its unique panel arrangements, which, along with creative character designs, guide the reader’s eye through the story in a carefully orchestrated manner. Panel layouts play a crucial role in conveying the narrative, as emphasized by Shogakukan’s “Newcomer Manga Award Manga Creator Training Course” 1 1 1 https://shincomi.shogakukan.co.jp/training/004.html , which outlines five fundamental rules for panel layout design. Manga artists can further enhance originality by creatively adhering to or intentionally breaking these rules to suit their storytelling needs. While readers may perceive subtle layout differences, it is often difficult to determine how these reflect the style of a work. This raises an important question: To what extent do panel layouts reflect the uniqueness of a manga work, and how do they contribute to its overall identity? The significance of panel layouts has been acknowledged in multiple foundational works. For instance, Cohn (2013) emphasized the impact of page layout on reading order and compared the technical usage of panel arrangements between Japanese and Western comics. McCloud & Martin (1993) highlighted the role of panel-to-panel transitions and particularly explored how the spacing between panels (gutter) influences the flow of storytelling. Additionally, Natsume et al. (2020) delved deeply into the techniques and effects of panel layout in manga, offering insights into how layout decisions enhance the reader’s experience. While these works underscore the importance of panel layouts, they do not delve into the extent to which panel layouts contribute to defining the unique identity of a manga work. Despite these foundational studies, existing research on manga has predominantly focused on characters, text, and other visual elements, with limited exploration of panel layouts. While some studies leverage engineering approaches, such as image processing or machine learning, these often focus on practical applications rather than delving into the artistic or structural characteristics of manga. This imbalance leaves the role of panel layouts in reflecting the creative uniqueness of manga largely unexplored. Figure 1: Brief description of the multi-class manga title classification experiment of the titles of the comics. The images shown in the figure are from AisazuNihalrarenai ©Masako Yoshi. To address this gap, in this paper, we conducted visual ablation experiments. The experimental process is shown in Figure 1 . We prepared three types of images: original unprocessed facing page images, text-and-character-masked facing page images, and panel frame-only facing page images. These images were used as input for a deep learning model to perform a multi-class classification task on manga titles. The objective of this study was to analyze the changes in model performance by removing text and characters or retaining only panel layouts. If the model performs comparably well on processed images as it does on original unprocessed images, we believe this would validate the uniqueness of features such as manga panel layouts. By using panel frame-only facing page images, we aimed to evaluate the contribution of panel layout features from each work to the classification task. Subsequently, we employed Grad-CAM ( Selvaraju et al. (2019) ) to visualize the key features in the input images and qualitatively interpret the results. Finally, we conducted additional experiments, including classification of manga publishers and genres, to rule out other factors influencing the uniqueness of panel layouts in manga works. We also performed classification experiments on noisy panel frames, further demonstrating that panel layouts are unique features of each manga work. These results indicate that panel layouts possess inherent uniqueness, even in the absence of content such as characters or text. This suggests that layout structures carry stylistic information that may reflect the unique visual identity of each work. Although our approach is based on a classification task, the broader aim is to understand how layout alone encodes work-specific style and supports visual narrative interpretation. This line of inquiry is closely related to cognitive theories of attention and sequential image comprehension. Related Work Visual Analysis in Creative Works Numerous studies have used images of creative works as input for predicting associated attributes such as titles or genres. Chu & Guo (2017) fine-tuned a pretrained CNN to classify movie genres based on poster images. Wi et al. (2020) applied a similar approach and visualized the model’s focus using Grad-CAM to verify whether it extracted expected features. Godwin et al. (2018) ; Takahashi & Kondo (2023) ; Tüchler et al. (2021) examined how layout affects comprehension or user experience, including children’s book design and web interfaces. Several works also explored deep learning approaches in comic research. Laubrock & Dunst (2020) ; Augereau et al. (2018) ; Sharma & Kukreja (2024) surveyed computer vision applications in comics. Laubrock & Dubray (2019) conducted classification of comic authors and titles using CNNs on datasets like GNC ( Dunst et al. (2017) ) and Manga109 ( Fujimoto et al. (2016) ; Matsui et al. (2017) ) . Dubray & Laubrock (2019) performed speech balloon detection, while Rigaud et al. (2021) worked on text segmentation inside speech balloons. Young-Min (2019) used CNN-based classification and feature visualization. Kukreja et al. (2023) applied CNN and Random Forest models to classify manga pages. Existing Issues in Manga Research While a growing number of studies have used manga page images, most focus on engineering applications, with limited attention to the underlying visual structure of manga. Laubrock & Dubray (2019) attempted to analyze manga features through classification, but their method treated pages as undivided wholes, without isolating individual components like panel layout. Young-Min (2019) proposed a CNN-based model for author classification and visualized learned features, but their analysis focused on stylistic textures and drawing patterns, rather than structural layout. Although previous studies have explored page-level classification, none have specifically isolated panel layouts through ablation and visualization to analyze their contribution to manga