Research on the Computational Analysis of Film

Measuring the Evolution of Camera Distance Across a Century of Film

One of the most fundamental ways through which film communicates is by manipulating the space between the camera and its subject. In this work, we carry out the largest analysis to date on the use of camera distance in film. We train a classifier (SigLIP 2) to label apparent shot distance, from extreme close-up to extreme long shot, and applied it to 5,205 films viewed in theaters between 1922 and 2025, including popular films (by U.S. box office), prestige films (by award nominations) and American indies.

Examples of the apparent shot distance categories we measure, from close-up to extreme long shot.

• We see significant changes in camera distance with the introduction of new technologies. Long shots declined with the arrival of sound, and medium close-ups in film more than doubled with the widespread adoption of television. We do not yet see closer shots with an uptake in smartphone viewing.
• Despite getting 3x less screentime than men, women are far more often the subject of closer shots, receiving 20% more medium close-ups. The gap persists across the whole century and in indie films too.
• Compared with live-action films matched by genre and year, animated films use more than twice as many long shots and show no gender difference in close framing, for human or non-human characters. This lack of a gender difference in animation sheds sharper light on the objectifying function of closeups in live-action films.

• We release new models for predicting apparent shot distance that others can run on their own films.
• We release predictions for all shot distances measured in all films in our collection, along with the location of shot boundaries and who is showing up on screen.

↑ top

Measuring Gender Representation in Animated Films

As a wholly created world, animation need not inherit the same biases as live action filmmaking. While our past work (below) examined gender representation exclusively in live-action films, we build models here to recognize the characters (both human and non-human) present on screen in order to carry out a similar analysis for 224 popular animated films (1937–2025).

Output of our character recognition pipeline on Frozen (2013), which identifies both speaking characters (Elsa, Anna, Kristoff) and non-speaking ones (the reindeer Sven).

• At first glance, animation mirrors live action. From 1980 to 2025, male characters get about 3× the screentime of female characters in both, and both begin moving toward parity around 2010. Many franchises give female characters more screentime with each sequel.
• Human vs. non-human. Among human characters, animated films give women and girls far more screentime than live-action films. Non-human characters (mostly animals), however, are even more male than live-action casts.
• Female villains persist. Contrary to early work on Disney's declining female antagonists, we see that animated films show no gender difference in antagonist screentime. In 21st-century live-action films, by contrast, women are cast as antagonists far less often than men.
• Women share the screen. As a minimal prerequisite to the Bechdel test, we measure how often we see female characters appearing together; again, contrary to earlier work on Disney, we do not see animated films depicting female characters in isolation (if anything, they appear together more often than women in live-action films).

• We release new models for detecting and recognizing characters in animated films, so that others can run them on their own films.
• We release predictions for where characters show up in all 224 movies in our collection.

↑ top

Evaluating Multimodal Narrative Understanding of Popular Hollywood Films

The creation of stable benchmarks built around Hollywood films is complicated by copyright protections. In this work, we address these concerns directly, by building a new collection of Hollywood films defined by two criteria: box office popularity (where we publish the first large-scale, open collection of weekly box office earnings reported by Variety magazine from 1922–1979); and likely public domain status (by researching copyright registrations and renewals in the US Catalog of Copyright Entries). We build a new multimodal multiple choice question benchmark on top of this collection that focuses on narrative elements that directly evaluate the abilities of models to inform meaningful research on film narrative. The result is 61 sound-era films from the Classical Hollywood period, which form the basis of the Classical Hollywood Narrative Benchmark (CHNB).

Historical US box office (1922–1979) visualization, estimated from weekly reports in Variety: the top movies of 1940, and the weekly earnings of Gone with the Wind across its re-releases.

• We publish the first open database of historical box office numbers. We extract 1.4M weekly box office reports for 24,000+ films from Variety (1922–1979) with a multimodal LLM. Rankings match human annotation closely (ρ = 0.961).
• We release new multimodal narrative benchmark that others can use to assess the performance of multimodal models now and into the future. The benchmark has 628 four-way multiple-choice questions on temporality, plot, character, setting, perspective, representation, symbolism and objects. We select questions to minimize the ability of frontier LLMs to answer correctly from the title and year alone.
• Models trail humans. Expert humans score 82.0%. The best system, Gemini 3 Flash on the full video with audio, scores 61.1%.
• Vision alone is not enough. Long-video models at 64 frames perform no better than the closed-book baseline. Adding the audio track lifts Gemini 3 Flash from 39.3% to 61.1% (+21.8 points), and dialogue transcripts alone reach 48.1%.
• Captioning works about as well as watching. Captioning one-minute clips and reasoning over the text (58.9%) is statistically tied with native video processing. Frame-based retrieval agents barely beat chance.

• We release weekly box-office numbers for all movies in our collection over the period 1922-1979, along with a visualization to interact with it.
• We release our MCQ narrative benchmark for others to evaluate the ability of multimodal models on reason about complex narrative tasks

Example question in CHNB (His Girl Friday): “How is the camera positioned as Hildy enters the room to talk to Earl Williams?” A.) On Hildy’s side of the bars to look through the grate at Earl. B.) The camera is at eye level and moves parallel to her. C.) On the inside of the cell to look out at Hildy. D.) High angle above Hildy and Earl.

↑ top

Multimodal Conversation Structure Understanding

While multimodal models are able to transcribe speech, how well can they parse the structure of a conversation? Drawing on sociolinguistics (Goffman, Clark & Carlson), we define a task that asks, for every utterance in a video clip, who the speaker is, who the addressees and side-participants are, and which earlier utterance it replies to. We release TV-MMPC, built on 200 TVQA clips from The Big Bang Theory, Friends, House M.D. and How I Met Your Mother released by the TVQA project. It contains 4,378 speaker and reply-to annotations, 5,599 addressee labels and 3,412 side-participant labels.

The task: for each utterance in a clip, a model predicts the speaker, addressee(s), side-participant(s), and the earlier utterance it replies to. The example is from The Big Bang Theory.

• Models lean on character names. When character identities are anonymized, the best model's speaker and addressee performance drops by 65 and 52 points, which raises questions about memorization of well-known shows.
• Gendered listening roles. Across 350,842 TVQA utterances in 4 TV series, women start conversational threads in proportion to their speaking time but are 1.2× more likely than men to be cast as addressees or side-participants rather than speakers.
• Audience design. When side-participants are present (i.e., a conversation includes more people than just a speaker-addressee pair), speech shifts from a personal, transactional register to a more social and performative one.

• We release TV-MMPC, our human-annotated dataset of speakers, addressees, side-participants and reply-to links for 200 clips from four TV series (on Dataverse), along with detected faces (with predicted actor identities) for every frame.
• We release code to preprocess TVQA clips, run vision–language models (via vLLM) and Gemini on the task, and evaluate their predictions against our annotations.

↑ top

Measuring Diversity in Hollywood through the Large-Scale Computational Analysis of Film

This paper shows how a 2021 regulatory exemption to the DMCA makes large-scale computational research on film legally possible in the US. Under it, we digitized 2,307 films: the top 50 by US box office each year from 1980–2022, plus award nominees. We measure representation through facetime, using face detection and recognition to identify which actors appear on screen, linked to perceptions of gender (from Wikidata) and race/ethnicity (from a user study).

Women's share of facetime in popular films, 1980–2022. Each point is one movie; the line shows the overall trend (with 95% confidence band).

• Men get 3× more screentime than women. From 1980 to 2010, women accounted for only about 25% of all facetime on screen
• Hollywood has grown more diverse since 2010. Since 2010, women's share of screentime has risen from about 25% to about 40% (in 2022), and Black, Hispanic/Latino, East Asian and South Asian actors have all seen increasing representation over the same period.
• We see big differences based on behind-the-scenes roles. Films directed by women reach near parity for gender, while films directed by men give men 3× the screentime of women. Women directed only 4% of the films.
• Disparities in awards and leading roles. Black actors are underrepresented in award-nominated films relative to popular ones. Lead roles go to White actors and men at significantly higher rates than supporting roles (women: 24.0% of leads vs. 33.2% of non-leads).

• We release our computational pipeline to measure actor screentime in film for others to run on their own movies.
• we release all derived measured that we have calculated in our collection (including the locations of faces within films, the face track and actor they belong to, shot boundaries, and film-level metadata).

↑ top

Once More, With Feeling: Measuring Emotion of Acting Performances in Contemporary American Film

The same line of dialogue can be delivered in many different ways. Drawing on variationist sociolinguistics, this paper treats a line as a fixed variable and each actor's delivery as a choice among possible variants, which lets us study how emotion is expressed through performance rather than through the words of the script. We transcribe and align the speech in 2,283 English-language films (box office hits and award nominees from 1980–2022), label the emotion of each utterance with a speech emotion model (wav2vec2), and group equivalent lines across films (e.g., "It's so nice to meet you") to compare how the same words are performed.

• Emotion follows narrative structure. Emotionality rises over the course of a film. Joy follows a U-shaped curve that climbs at the resolution, while anger peaks about 85% of the way through.
• Genre constrains range. Thrillers, biographies and mysteries have the narrowest emotional range, while fantasy, musicals and family films have the widest.
• Words constrain range. Functional lines ("Is that okay?", "Any questions?") are delivered in narrow ways. Open-ended, evaluative lines ("You're alive", "Are you out of your mind?") vary widely in how they are performed.

↑ top