Source-linked AI summary
Measuring Gender Representation in Animated Films
David Bamman, Allison Cooper, Ruby Alvarez Rubio, Reina Kushihashi, Madison Mar
TL;DR
Research on gender in animated films has largely counted cast composition rather than character visibility, despite animation’s importance in children’s enculturation. This paper builds a computational character-recognition pipeline and applies it to 224 popular animated movies, finding that animation combines live-action-like overall patterns with greater human female representation, stronger masculine skew among non-humans, persistent female antagonist representation, and more shared scenes among women and girls.
Problem
Prior large-scale studies of gender in animated films have focused mainly on cast composition rather than fine-grained visibility such as screentime.
Method
The paper trains and evaluates animated-character detection and recognition models, constructs casts and gender labels, and analyzes 224 popular films alongside live-action films.
Results
Overall female representation tracks live action, while human female characters receive greater representation, non-human characters are largely male, female antagonist representation does not decline, and women and girls more often share scenes.
Takeaways & Limitations
Animation both retains broader gender biases and creates distinct representational patterns through its treatment of human, non-human, antagonist, and socially connected female characters.
Takeaways & Limitations
Screentime measures visibility but not whether characters are active or whether more visibility is better, and gains in one area may be offset by non-human gender imbalance.
Abstract
from arXiv · showhide
Animated films--often developed with an audience of children in mind--are an important vector for enculturation, and empirical work that has examined the representation of gender at scale in these films has largely focused on counting the gender composition of the cast rather than deploying a more fine-grained instrument (such as assessing the visibility of those characters in overall screentime). In this work, we develop a computational pipeline for recognizing animated characters in these films, and use it to test several hypotheses about gender representation in a corpus of 224 popular animated movies. We find that while the overall representation of female characters in animated films largely tracks with those of live-action films (over the period 1980-2025), we see stark differences between the representation of human characters (much greater representation among women and girls) and non-humans (largely male). Contrary to past work on Disney, we do not see female characters declining in antagonist roles in animated films, and characters who are women and girls are much more likely to share scenes together than their live action contemporaneous counterparts.
1 Introduction
This study asks whether animation’s expressive freedom produces different gender-representation patterns from live-action film. It develops computational methods to measure animated characters and applies them to popular films, finding both shared and distinct patterns.
- Motivation: Animation offers a wholly created world for testing whether gender biases persist without live-action filmmaking’s formal constraints.The study compares animated representation with contemporaneous live-action films.
- Motivation: Men and male non-human characters appear roughly three times more often than female characters across many popular media forms.These inequalities matter because children learn about gender through influential media.
- Contributions: The pipeline detects and recognizes animated characters, links them to manually constructed casts and gender information, and measures representation beyond cast counts.It extends prior computational work on live-action films and cartoon imagery.
- Contributions: The study applies the pipeline to 224 popular animated films to test hypotheses about gender representation against contemporaneous live-action films.The analysis examines gender composition, screentime, roles, and character co-occurrence.
- Findings: Women and girls have greater screentime than their live-action counterparts, while non-human characters are more strongly skewed masculine.Female animated characters also increasingly share screens, and female antagonist representation does not decline as in live action.
2 Data
The corpus consists of 224 popular animated films selected from high-grossing movies between 1937 and 2025. It captures much of the documented box-office market but is concentrated in recent releases and major commercial studios.
- Corpus: 66.7% of documented animated-film box-office grosses are represented in the collection.The selection draws on Box Office Mojo and Variety historical box-office data.
- Selection: The collection selects animated films ranked among the top 50 movies by annual box office and acquires and digitizes those available.This procedure emphasizes popular theatrical releases rather than the full set of animated films.
- Corpus: 224 films spanning 1937–2025 comprise the study’s collection, with 78% released after 2000.The corpus includes Snow White and the Seven Dwarfs through films released in 2025.
- Corpus composition: The corpus includes 59 Disney, 46 DreamWorks, 26 Pixar, and 15 Illumination films, among other studios.Nearly 47% of the films belong to franchises containing at least two top-50 movies.
3 Methods
The paper develops a computational pipeline to detect and recognize animated characters, assemble cast lists, and measure their screentime for gender-representation analysis.
- Pipeline: The pipeline segments films into shots, detects character faces, links sequential detections into face tracks, and matches tracks to cast-list characters.It primarily adapts a live-action character-recognition pipeline while addressing animated films’ lack of suitable visual cast lists.
- Detection and recognition: Models are trained and evaluated on human-face, cartoon-face, and mixed datasets, including iCartoonFace and film-specific evaluation data.The detection experiments use 60,000 annotated iCartoonFace images and a manually created film evaluation set, while recognition uses nearly 400,000 images of 5,013 identities plus film data.
- Detection and recognition: RT-DETR-L without test-time augmentation is selected for detection because it offers the best overall accuracy–speed balance on annotated films.Human-face-only models struggle on animated faces; test-time augmentation improves weaker models but adds substantial prediction cost.
- Detection and recognition: Recognition experiments compare buffalo_l and DINOv2 models across pretrained, fine-tuned, and movie-track self-supervised settings.They vary DINOv2 model size, use 0% or 25% crop padding, and evaluate adaptation either without or with test-film movies.
- Detection and recognition: DINOv2 vitl14 with 25% crop padding and track-based adaptation is selected for subsequent experiments despite its computational cost.The paper reports that padding, larger DINOv2 models, and fine-tuning on tracks generally improve recognition accuracy.
- Pipeline: Animated-film cast lists are manually built because IMDb generally lists voice actors and omits non-speaking characters’ animated appearances.The pipeline uses annotated character images to create visual representations for matching detected faces.
4 Analysis
The analysis measures animated-film screentime by character category, gender, and narrative role, then compares these patterns with contemporaneous live-action films. Animated films give human female characters more representation than live action, while non-human characters are more male-skewed; female animated characters do not show a meaningful antagonist disadvantage and more often co-occur.
- 4.1 Gender representation: Animated films give human characters substantially greater female representation than comparable live-action films, while non-human characters are even more disproportionately male.Animals dominate the non-human category.
- 4.2 Protagonists and antagonists: The study measures protagonist and antagonist roles across all characters because antagonists can have non-human or boundary-crossing forms.Protagonists are characters whose goals or arc drive the story; antagonists primarily oppose them or drive the central conflict.
- 4.2 Protagonists and antagonists: Female animated characters do not differ meaningfully from male animated characters in the share of protagonist-and-antagonist screentime devoted to antagonists.The animated-film confidence intervals are large, whereas live-action women are cast significantly less often as antagonists than men across the 21st century.
- 4.3 Female character co-occurrence: Since the 1990s, animated films generally depict female human characters with greater co-occurrence than contemporaneous live-action films.Co-occurrence counts another female character in the same or subsequent shot, reducing isolation and capturing shot/reverse-shot patterns.
5 Conclusion
The conclusion finds that animation both departs from and preserves live-action gender patterns: it gives human women and girls greater representation but intensifies gender imbalance among non-human characters. The paper also emphasizes that visibility and role measures do not capture the nature or activity of representation.
- 5 Conclusion: Animated films represent women and girls at greater rates than live action but gender non-human characters male at roughly four times the rate of women.The conclusion describes this as animation both resisting and embracing androcentrism.
- 5 Conclusion: The pipeline supports both larger-scale trend analysis and fine-grained identification of which characters are visible and when.The released data include frame-level character locations for the 224-film collection.
- 5 Conclusion: The analysis does not characterize whether visible characters are active or passive, so more screentime should not automatically be treated as better representation.The authors identify the nature of visibility as a subject for further work.
6 Note on AI Usage
The paper reports extensive use of Claude Code for coding assistance while stating that the resulting scripts are publicly available for inspection and reproduction. It distinguishes this assistance from the paper’s writing and manual annotations.
- 6 Note on AI Usage: Claude Code was used extensively for coding assistance, with the resulting computational scripts made publicly available for inspection.The authors state that no AI was used in writing the paper or creating manual annotations.
A Author Contributions
Author contributions span annotation, statistical analysis, computational methodology, and writing and theoretical framing. The listed contributors are assigned to specific stages of the project.
- A Author Contributions: RAR, RK, and MM contributed animated character detection, recognition, and character clustering annotation.
- A Author Contributions: DB and RK annotated character categories, while RAR and DB annotated protagonist and antagonist roles.
- A Author Contributions: DB handled statistical analysis and computational methodology, while DB and AC handled writing and theoretical framing.
B Detection
Table 3 reports AP@0.5 with 95% bootstrap confidence intervals on iCartoonFace, alongside inference throughput.
- Table 3 evaluates detection using AP@0.5 on iCartoonFace.
- Results include 95% bootstrap confidence intervals.
- The table also reports inference throughput in images per second.
C Recognition
Table 4 reports Rank@1 identification accuracy on iCartoonFace under crop 0% and crop 25% conditions, with inference throughput.
- Table 4 evaluates identification using Rank@1 accuracy on iCartoonFace.
- Results are reported for crop 0% and crop 25% conditions.
- The table includes 95% bootstrap confidence intervals and inference throughput in images per second.
D Manual clustering
Manual clustering uses an annotation interface in which annotators review candidate image clusters and build a cast list of entities.
- Manual clustering: The annotation interface presents candidate clusters in a left panel, each containing 12 images.
- Manual clustering: Annotators create a cast list of entities in the right panel during the annotation process.
- Manual clustering: Annotators can deselect images from clusters before finalizing entities, as illustrated by one deselected Gulliver image.
E Franchises
The franchise analysis examines female-character representation across franchises and uses plot summaries with cast lists to identify narrative roles and character prominence.
- Franchises: Figure 8 shows female-character representation across franchises containing at least 3 movies in the collection.
- Franchises: A film narrative analyst identifies protagonists and antagonists from each movie’s title, plot summary, and complete cast list.
- Franchises: Protagonists are main characters whose goals and character arcs drive the narrative.
- Franchises: Antagonists primarily oppose protagonists or drive the central conflict against them.
- Franchises: The procedure also classifies identified characters as major or minor according to their narrative centrality and role duration.
- Franchises: Only cast-list entries are selected, and roles remain empty when the plot summary does not support confident identification.