This short paper reflects on a computer vision experiment examining gender representation in Chinese comics (lianhuanhua), small illustrated storybooks widely circulated in twentieth-century China for entertainment, education, and political communication. While prior scholarship has focused on close readings of individual works, the medium has rarely been analyzed at scale, leaving broader patterns of visual and textual representation underexplored. To address this gap, the study combines structured annotation with computational modelling to investigate how female figures are visually constructed and narratively positioned across panels.

The central methodological decision involved model selection and experiment design. Rather than maximizing the number of detected figures, the project prioritizes interpretive rigor and contextual accuracy. Supervised object detection was avoided due to potential gender bias and the highly stylized nature of lianhuanhua imagery; instead, image clustering was applied to annotated character regions. A multimodal annotation framework cross-references visual features with narrative text, and large language models assist in identifying gendered cues, enhancing efficiency while preserving interpretive oversight.

The process generated a structured, reproducible multimodal dataset linking image regions, textual content, and annotations. Preliminary findings reveal both visible imbalances and subtler phenomena, such as textual mentions of characters absent in illustrations, highlighting gaps between narrative acknowledgment and visual presence. An ongoing phase explores AI-assisted attribution of objects to female figures.

The study contributes to Digital Humanities by proposing a bias-aware experimental workflow, producing a reusable multimodal dataset, and demonstrating how computational analysis of textual–visual alignment can generate new insights into gendered visibility in historical media.