Back to AI Research

AI Research

PAGE combines AI reading with human checks for comics research

Key Takeaways

  • The framework organizes scans, codebook-based annotations and human review to study American comics and Japanese manga at archive scale.
  • ## Turn scans into a research collection
  • [The PAGE framework](https://arxiv.org/abs/2610.13840) proposes using AI to organize and analyze comics archives.
  • It parses scans into panels, text and bibliographic records, then applies a vision-language model using a written codebook.
  • The framework grounds samples of its readings in human annotation.

Turn scans into a research collection

The PAGE framework proposes using AI to organize and analyze comics archives. It parses scans into panels, text and bibliographic records, then applies a vision-language model using a written codebook.
The framework grounds samples of its readings in human annotation. That makes the annotation procedure and its checks part of the research method, rather than treating a model's description of a page as an established historical finding.

The demonstrations cover two comics traditions

The paper demonstrates PAGE on 3,929 American comic books and 87 volumes of Japanese manga. Its studies examine political portrayal, visual form and who speaks across the collections. The framework's analysis includes controls for publisher and genre.
Those examples show how the authors apply the method to specific archives. They do not establish that a model interprets every panel correctly, or that the same codebook transfers unchanged to another period or comics tradition. The human checks remain important when reading the reported patterns.

Make interpretation traceable to pages

Franklin's interpretation is that a useful archive workflow should preserve the path from an aggregate claim back to the scans and annotations supporting it. A researcher needs to inspect what the model counted and how the codebook defined that category.
For a small evaluation, choose a sample with different layouts and have people annotate it under the same instructions. Review disagreement before scaling the process across a collection. Keep publisher and genre context visible when comparing groups.
The paper also discusses possible applications to search, preservation and access. Treat those as directions to investigate rather than completed services demonstrated for every archive. Before adoption, check scan quality, permissions and the suitability of the annotation categories. The contribution is a research framework that combines scale with inspectable human grounding; its value depends on preserving both.

Comments