Skip to content
conftrace_
2026 EMNLP EMNLP 2026

MangaLens: A Multi-Model Tool for Structured Manga Parsing

AbstractManga pages combine text, layout, characters, and narrative order, making them difficult to process as plain images or text alone. We present MangaLens, an integrated tool for structured manga parsing. MangaLens combines LLM-based reasoning with segmentation-based grounding to support manga parsing subtasks such as panel structuring, speaker linking, and related cross-page reasoning tasks. The system produces page-level and book-level JSON records and visualizes subtask outputs with aligned image evidence, enabling users to verify, correct, and reuse parsing artifacts. Evaluation results show that MangaLens produces reliable structured parsing outputs across major subtasks. The tool is released at https://github.com/nict-astrec-att/MangaLens

More from EMNLP 2026

All EMNLP 2026 papers →