Source-linked AI summary
Studying Comics at Scale with Artificial Intelligence: The PAGE Framework and Three Studies of American Comics and Japanese Manga
Emilio Ferrara
TL;DR
Comics scholarship has not studied most of the medium’s vast record as a whole, limiting archive-wide inquiry into its social and visual history. PAGE structures, annotates, grounds, and analyzes comics pages, and three studies apply it to American comics and Japanese manga to examine political portrayals, visual form, artist attribution, and speakers.
Problem
Comics scholarship has relied on selected books or small hand-coded corpora, while the surviving record raises questions about society, visual language, and whose voices comics carried.
Source: §1 · p. 2Method
PAGE parses scans into structured pages and bibliographic records, applies a written codebook through a vision-language model, grounds sampled annotations, and estimates trends with publisher, genre, and sampling held fixed.
Source: §1 · p. 2Results
Three studies apply PAGE to 3,929 American comic books and 87 manga volumes, examining presidential portrayals, page form and penciller attribution, and who speaks across traditions.
Source: §1 · p. 3Takeaways & Limitations
PAGE’s outputs include indexes searchable by depicted content, transcripts for screen readers, and evidence that can support crediting artists whose names were not printed.
Source: §1 · p. 2Human reference annotation covers speaker gender in manga, while American pages have noisy OCR and no human reference for the reported coding.
Source: Limitations · p. 19
Abstract
from arXiv · showhide
Comics are one of the great visual records of the twentieth century: hundreds of thousands of pages in which a mass readership saw its leaders, its wars and its heroes. Most of this record has never been studied as a whole; comics scholarship proceeds one book at a time. We argue that artificial intelligence, including generative models, can change this to the benefit of the medium itself. We propose PAGE, a framework that makes artificial intelligence an instrument for comics research. It parses every scan into panels, text and a bibliographic record; annotates each page with a vision-language model that follows a written codebook, as a human coder would; grounds samples of these readings in human annotation; and explains change across an archive, with estimates controlled for publisher and genre. We demonstrate the framework on 3,929 American comic books (1938--1956) and 87 volumes of Japanese manga, in three studies, each resting on a different capability. The first pairs text search with visual reading to recover how a mass medium portrayed political power. The second measures the page's visual form across an industry and recovers the hand of individual artists. The third follows who speaks across two comics traditions, checked against human annotation. Beyond research, the same instruments can make archives searchable by what is drawn, open comics to blind and low-vision readers, and help preserve fragile collections and credit uncredited artists. We outline a research agenda in which artificial intelligence strengthens the study, preservation and reach of the medium.
1 Introduction
Comics scholarship has not been able to study the medium’s vast record as a whole, leaving broad questions about society, visual language, and voice unanswered. PAGE combines structured parsing, codebook-guided model annotation, grounding, and adjusted trend analysis in three studies of American comics and Japanese manga.
1 Introduction
- Comics scholarship has relied on selected books and hand-coded corpora of a few dozen titles, while no scholar can read enough pages to answer archive-wide questions.The corpus includes 3,929 golden-age American books and 195,795 pages.Source: §1 · p. 2
- PAGE parses scans into structured pages and bibliographic records, applies a written codebook through a vision-language model, grounds sampled annotations, and estimates trends with publisher, genre, and sampling held fixed.It also produces searchable visual indexes, screen-reader-ready transcripts, and evidence that may help credit artists whose names were not printed.Source: §1 · p. 2
- The paper catalogs AI techniques and their potential uses for discovery, attribution, access, preservation, and teaching, and proposes six research-agenda lines.These are framed as uses for readers, archives, and artists.Source: §1 · pp. 2–3
- Three studies apply different capabilities to 3,929 American comic books and 87 manga volumes, examining presidential portrayals, page form and penciller attribution, and who speaks.The framework also checks annotations against independent evidence, including a second coder and human speaker-gender annotation.Source: §1 · p. 3
2 Related work
Prior work has studied comics through close readings, small annotated corpora, and computational tools for visual analysis. PAGE brings these strands together while treating model outputs as measurements that require grounding and accounting for concerns about generative AI and artists.
Comics as an object of empirical study.
- Comics research has tested hypotheses about visual language, layouts, dialogue, and representation using hand-coded studies and annotated collections.Large-scale visual analysis has also been applied to manga and comic strips (Manovich et al. 2011; Wigard et al. 2023).Source: Comics as an object of empirical study · p. 3
Computer vision and language technology for comics.
- Computer-vision systems detect panels, balloons, and characters, while newer tools jointly associate balloons with speakers and support translation (Sachdeva and Zisserman 2024; Li et al. 2024b; Hinami et al. 2021).General multimodal models perform near chance on choosing the panel that completes a strip (Vivoli et al. 2025), motivating checks against independent evidence.Source: Computer vision and language… · pp. 3–4
Models as measurement instruments for culture.
- Cultural analytics applies quantitative methods to texts and images, while language and vision-language models can code cultural material but may carry biases or vary with prompt wording.The paper therefore calibrates comics-reading model annotations in its Ground step.Source: Models as measurement instruments for… · p. 4
Generative AI and the people who make comics.
- Generative image tools raise concerns about artist consent, imitation, labor, and archive provenance; the paper confines proposed generative-AI uses to scholarship, access, or preservation.The authors state that the work does not endorse or enable AI artwork generation.Source: Generative AI and the people who make… · pp. 4–5
3 The PAGE framework
PAGE treats AI as a measurement instrument: it structures comic pages, codes them under a written codebook, checks samples against independent evidence, and estimates trends with design controls. Its resulting data and tools are intended for research as well as use by readers, archives, and artists.
3 The PAGE framework
- PAGE fixes model categories in a written codebook before reading pages and checks sampled codes against human annotation and independent evidence.The four steps are Parse, Annotate, Ground, and Explain.Source: §3 · p. 5, Figure 1 · p. 5
Parse.
- Parse detects panels and text, transcribes and orders lettering, and attaches bibliographic records; Manga109-s panel detection reaches box F1 0.97, versus 0.66 for gutter rules.COMICS supplies panel and text boxes and OCR; its missing bibliographic records are reconstructed, and duplicate scans of the same issue are counted once.Source: Parse · pp. 5–6
Annotate.
- Annotate records page content under predefined categories and computes CLIP image embeddings for visual similarity, retrieval, attribution, and genre classification.Pages in the speaker study are shuffled across tasks and their years withheld to prevent coding drift from aligning with publication year.Source: Annotate · p. 6
Ground.
- Ground compares annotations with a second coder, available human annotations, independent measurements, and close readings, then revises the codebook when checks reveal a need.A first coding of 32 pages led the authors to separate gender, age, and species into distinct fields.Source: Ground · p. 6
Explain.
- Explain estimates yearly change with publisher-group and genre fixed effects, bootstraps comics or volumes for intervals, and reports contrasts between named time windows.Estimates are read against the medium’s history; the authors make causal claims only where the design supports them.Source: Explain · pp. 6–7
Uses for readers, archives and artists.
- PAGE outputs can support visual discovery, artist attribution, screen-reader access, preservation, and teaching, including evidence relevant to penciller credits and transcripts for readers who cannot see the page.The paper notes that 22% of 1938–1956 American comic stories name no penciller.Source: Uses for readers, archives and artists · p. 7
Principles of practice.
- The studies require traceable page-level claims, prewritten codebooks revised only with grounding evidence, independent references where possible, deduplicated issues, license compliance, and documented model use.Source: Principles of practice · p. 7
4 Data
The paper combines a digitized American comics release, reconstructed bibliographic metadata, and a Japanese manga dataset with existing panel and speaker annotations. It also situates these data alongside a catalog of AI techniques and their potential uses.
COMICS.
- The COMICS release contains scans, panel and text boxes, and OCR for 3,929 golden-age books totaling 195,795 pages; scans are preserved losslessly and indexed by checksum.The release and Grand Comics Database index reflect volunteer work by fans and collectors.Source: COMICS · p. 7
- Table 1 organizes AI techniques by PAGE step and by their uses for readers, archives, and artists, marking which techniques the paper tests.Source: Table 1 · p. 8
Reconstructed metadata.
- Cover text recognition and indicia readings assign years to 3,034 comics, or 77% of the release, and matching links 2,215 comics to the Grand Comics Database.The matching is exact in 72% of cases when title and issue are independently known.Source: Reconstructed metadata · p. 9
Manga.
- Manga109-s supplies 87 Japanese volumes with hand-drawn panel, balloon, face, and body boxes; its catalog records decade, readership, and genre, while Manga109Dialog links balloons to speakers.Source: Manga · p. 9
COMICS.
- Appendix A documents panel-subset page-number offsets and duplicate scans that affect studies joining parts of the COMICS release.Duplicate scans are searched only among comics in Study 1.Source: Manga · p. 9
5 Grounding the instrument before use
Before using PAGE annotations, the authors compare samples with human annotations and independent readings. The checks show strong agreement, while retrieval results favor visual search over raw OCR.
5 Grounding the instrument before use
- Before using codes to answer research questions, the authors check samples against human annotation and evidence from independent readings.Table 2 organizes the grounding evidence by study; human references are marked, while other checks compare readings made with different information or methods.Source: §5 · p. 9, Table 2 · p. 10
- The model matched human-stated gender for 95.5% of 202 manga characters and 92.7% of 454 eligible coded speaker balloons.For the balloons, the 95% confidence interval was 85–98%; 477 balloons were eligible.Source: §5 · p. 9
- Separate coding versions agreed on 99.6% of 269 American balloons (κ = 0.99) and 96.4% of 167 manga characters (κ = 0.93).On 112 random manga pages, page- and character-level speaker gender matched for 94.2% of 1,154 balloons (κ = 0.88); the methods assigned women 39.5% and 41.6% of the text, respectively.Source: §5 · p. 10
Lessons from retrieval.
- CLIP page retrieval reached macro average precision 0.54, +0.20 over raw OCR, while its advantage over dense retrieval was not distinguishable from zero.Dense retrieval with expanded queries reached 0.47; the CLIP advantage over it was +0.06 (−0.02 to 0.15). An earlier page-key alignment was wrong for 261 of 501 images and overstated image retrieval’s advantage.Source: Lessons from retrieval · p. 10
6 Study 1: how a mass medium portrayed political power
The study combines text search with visual reading to examine presidential portrayals in American comics from 1938 to 1956. Sitting presidents’ faces were often withheld, and commander-in-chief portrayals peaked during the war.
Goal and design.
- The study searched page OCR for presidential names and titles, then used a vision model to code candidate pages’ portrayals.After adding sampled pages and removing duplicate scans, the model read 1,082 pages: 658 contained a real presidential reference, yielding 787 references.Source: Goal and design · p. 11
Findings.
- Only three of 787 references mocked a real president; other mocking or negative portrayals concerned stand-ins.The three cases were a 1921 quip outside the study period, a 1948 baseball gag, and a 1950 caricature.Source: Findings · p. 11
What PAGE enables.
- Text search cannot distinguish a portrait from a silhouette, and likenesses recognizable without a name can enter the candidate set only when another presidential term appears.The authors identify image search for unnamed likenesses as a next step.Source: What PAGE enables · pp. 11–12
Findings.
- Among drawn sitting presidents, 57% had their faces withheld, versus 3% of historical presidents; commander-in-chief references reached 11% in 1942–45.The commander-in-chief share was 4% before Pearl Harbor and 3% after the war; the wartime share was clearly above the postwar share, while its interval overlapped slightly with the prewar one.Source: Findings · p. 11, Figure 2 · p. 12
7 Study 2: the visual form of the page across an industry
Across American comics from 1938–1956, page layouts shifted during and after the war, while image embeddings recovered both penciller similarity and publisher-group clustering. Whole-corpus measures extend prior small-sample studies of page design and support research on attribution.
Goal and design.
- Across 2,924 dated comics, the study measures panels, splash pages, lettering, and CLIP embeddings from a fixed sample of 27,018 pages.The estimates adjust for publisher group and genre and use comic-level bootstrap intervals.Source: Goal and design · p. 12
- Panels per page fell from 7.28 in 1940–41 to 6.02 in 1944–45, while splash pages rose by 4.0 points during the war.These estimates hold publisher and genre fixed; Figure 3 shows adjusted and raw yearly means.Source: Goal and design · p. 12, Figure 3 · p. 13
- Words per page rose from 189 in 1944–45 to 264 in 1953–54, while narration’s share of text boxes returned to its prewar level.The postwar increase was 75 words; words per page rose modestly during the war (+21).Source: Goal and design · p. 12
- Adjusted and raw curves nearly coincide, and the layout and word-count trends retain their signs and significance in robustness checks.This suggests changes in publisher and genre mix account for little of either trend; earlier hand-coded studies covered 40 books, while these estimates cover the first two decades of that span.Source: Goal and design · p. 13
- After excluding series neighbors, duplicates, reprints, and near-duplicates, same-penciller nearest neighbors occur 2.4% of the time versus 1.5% for shuffled labels.The lift is 1.6 over the shuffled baseline (p < 0.001). In the artist network, 32% of links join artists from the same publisher group, versus 11% expected by chance.Source: Goal and design · pp. 13–14
- A cover-embedding classifier names held-out cover genres correctly 70% of the time, versus 27% for the majority class and 46% for zero-shot prompts.Across dated covers, superhero covers fell from 42% in 1942–45 to 6% in 1950–56, while horror covers rose from 2% to 26%; Manga109-s panels per page fell from 5.3 in 1980s volumes to 5.0 in 2000s volumes.Source: Goal and design · p. 14
What PAGE enables.
- Whole-corpus layout estimates extend hand-coded studies of a few dozen books, and image embeddings rank pages by artist to support attribution research.This offers a way to study attribution in a corpus where much work is uncredited.Source: What PAGE enables · p. 14
- Figure 4 pairs an artist-similarity network with yearly excess distances between publisher-group centroids.The network includes pencillers with at least 15 labeled pages and connects each to three nearest artists; the second panel compares observed with label-shuffled distances.Source: Figure 4 · p. 14
8 Study 3: who speaks in two comics traditions
The study codes dialogue speakers in American comics and Japanese manga, then compares women’s share of speech with their presence on the page. Women’s dialogue share is higher in the manga sample across readerships, though the samples differ by roughly half a century.
Goal and design.
- The study codes speakers, addressees, and narrative roles for American dialogue balloons, and uses Manga109Dialog speaker links to code manga characters.It analyzes 1,055 American story pages and 10,282 balloons, alongside 7,235 manga pages; human annotation supports 92.7% agreement for American speaker-gender coding.Source: Goal and design · p. 15
- Women speak 15.7% of dialogue words in American comics, and their dialogue share largely tracks their share of characters drawn.The adjusted share rose from 10% in 1942–45 to 20% in 1946–49, then settled at 18% in 1950–56; women’s share exceeded their presence in the late 1940s.Source: Goal and design · p. 15
- Women speak 40.3% of dialogue in the manga sample versus 15.7% in American comics, and the difference holds across all four manga readerships.Shares range from 34.7% in shōnen to 49.4% in shōjo; the corpora span roughly half a century, so tradition and period are confounded.Source: Goal and design · pp. 15–16
- Women’s share of dialogue largely tracks their share of characters in both samples, while women speak more than their presence predicts in shōjo manga.The adjusted gap in shōjo is 4.0 points (1.2 to 6.6).Source: Goal and design · p. 16
What PAGE enables.
- Human speaker links are available for one manga collection but not for an American archive, so human-coded manga checks model speaker codes used for American comics.This lets the study compare the two traditions using one instrument.Source: What PAGE enables · p. 17
9 A research agenda for artificial intelligence and comics
The research agenda extends PAGE’s archive-scale study to discovery, attribution, access, preservation, comparison, and shared human references. It also identifies risks and limits that constrain what the framework’s results can support.
(1) Discovery by what is drawn.
- An image search for unnamed likenesses could find presidential depictions that text search misses when no other presidential term appears on the page.The agenda proposes this as a way for teachers and historians to assemble such pages.Source: (1) Discovery by what is drawn · pp. 17–18
(2) Attribution of uncredited artists.
- Same-penciller nearest-neighbor matches occur for 2.4% of pages, 1.6 times the shuffled baseline, offering a visual signal for artist attribution.Embeddings would complement signatures, series records, and expert judgment; recurring characters and repeated art remain difficult to separate from an artist’s drawing.Source: (2) Attribution of uncredited artists · p. 18
(3) Access for readers who cannot see the page.
- Parse and Annotate already produce reading order, speaker links, and transcripts needed by screen readers and audio descriptions.The authors propose building and evaluating these tools with blind and low-vision readers.Source: (3) Access for readers who cannot see… · p. 18
(4) Preservation and restoration with provenance.
- Restoration and colorization can create readable access copies of damaged pages, provided they remain separate from the original and carry provenance labels.Archives would also need to detect synthetic pages, with experts and automated detectors working together.Source: (4) Preservation and restoration with… · p. 18
(5) A comparative science of the medium.
- Parsing and annotating multiple traditions with one instrument could make theories of comics’ visual-language dialects testable across whole traditions.The authors propose matched periods and adding European comics and webcomics to test divergence and influence.Source: (5) A comparative science of the medium · p. 18
(6) Shared instruments and human references.
- Shared codebooks, annotations, and human references would let future studies report model error against common standards.The authors call for reference sets covering narrative role, depiction, and speaker identity, plus codebook replication across model families.Source: (6) Shared instruments and human… · p. 18
Risks and how the framework measures them.
- Model errors can recur across sessions, and training bias, prompt wording, style imitation, and synthetic pages pose risks to research and archives.The paper’s proposed uses are analytic or access-oriented, not generation in an artist’s style; attribution remains a human-expert decision.Source: Risks and how the framework measures… · p. 19
Limitations.
- The manga comparison is noncausal because the traditions span different decades and Manga109-s is curated; penciller matching cannot exclude recurring cross-series features.Other limits include single-model-family annotation, incomplete human references, noisy American speaker data, and a public-domain COMICS selection with noisy dates and links.Source: Limitations · p. 19
10 Conclusions
PAGE applies structured parsing, codebook-guided vision-language annotation, human checks, and controlled estimates to comics archives. Across three studies, it reports findings on presidential depictions, wartime page form, and women’s share of dialogue relative to their presence.
10 Conclusions
- Across three studies of 3,929 American comic books and 87 manga volumes, PAGE finds results on presidential depictions, wartime page form, and women’s dialogue relative to their presence.The framework parses pages, annotates them under written codebooks, grounds samples in human annotation and independent checks, and controls archive estimates for publisher and genre.Source: §10 · p. 19
- The studies find that golden-age comics withheld the sitting president’s face, wartime American pages had fewer and larger panels, and postwar pages had more words.In both traditions, women’s share of dialogue largely tracks their share of the characters drawn.Source: §10 · p. 19
- PAGE’s instruments can support search by depicted content, access for blind and low-vision readers, preservation of fragile collections, and efforts to credit artists.The conclusion frames AI as extending study of the medium when grounded in human annotation and used with creators’ consent and credit.Source: §10 · pp. 19–20
Declarations
The declarations describe data and code availability, while the appendix notes corrections and exclusions used in the studies.
Declarations
- COMICS, Manga109-s, Manga109Dialog, and the GCD are available through their respective authors, licenses, or website; code, codebooks, annotations, and derived measures are to be released publicly.Page images will not be redistributed; the declarations also identify the annotation model, author review, and no competing interests or funding.Source: Declarations · p. 20
Page numbering in the panel-annotation subset of COMICS.
- Of 501 matched panel-annotation images, 261 had a different page key, usually one page higher, so joining subset images to OCR by key can pair the wrong text.The article uses the full archive for its results.Source: Page numbering in the panel-annotation… · p. 20
Duplicate scans.
- Four duplicate issue pairs were found among the 684 comics read for Study 1, and each issue was counted once in every study.Repeated art across series and repeated advertisements were excluded from the Study 2 penciller-matching test.Source: Duplicate scans · pp. 20–21
Appendix B Gallery of pages
The appendix gallery presents unaltered COMICS scans and licensed manga pages, with examples from the paper’s three studies.
Appendix B Gallery of pages
- Gallery images use public-domain COMICS scans from the Digital Comic Museum and licensed Manga109-s pages, with manga speakers from Manga109Dialog.The COMICS scans are reproduced unaltered and identified by page key.Source: Appendix B · p. 21
Study 1
- The presidential examples include an unnamed Roosevelt identified by pince-nez and features, an unnamed neutrality-pledging president, and wartime scenes with government and military figures.The gallery also shows Churchill with the president after a foiled assassination.Source: Appendix B · pp. 22–23
Study 2
- Typical-page examples compare two story pages from each of three windows, selected nearest the window’s median panels and words per page.The reported medians are 8 panels and 142 words in 1940–41, 7 and 157 in 1944–45, and 7 and 195 in 1953–54.Source: Figure B3 · p. 24
- The penciller-matching gallery pairs pages from different series by the same credited artist after excluding reprints and repeated pages.The listed artists are Bernard Dibble, Matt Baker, Ralph Mayo, and Jack Kamen.Source: Figure B4 · p. 25
Study 3
- The gallery shows American story pages with model speaker codes and manga pages whose balloons are outlined by annotator-linked speaker gender.The manga examples show one page per readership with both women and men speaking, selected for the most linked balloons among pages coded with the page-level procedure.Source: Figure B5 · p. 26, Figure B6 · p. 27