How Pixar tracks and improves character diversity

This article is also available in: French

In the last few years, Pixar Animation Studios took several steps towards improved diversity. This endeavour has been visible both on screen (with animated features such as Soul, Turing Red) and behind the scenes (Turning Red by Domee Shi, for example, is the first Pixar movie solely directed by a woman).

At SIGGRAPH 2022, Pixar explained that this evolution can also be found at the heart of the creative process, including when it comes to background characters and crowds, thanks to a dedicated tool and a long thought process.

Soul and unconscious bias

While working on Soul, the creative team wanted to portray New York City in an accurate manner. To achieve this goal, a character pool was created with 80% BIPOC (Black, indigenous and people of color) and 50% male/female. This way, randomised crowds would achieve the representation the team was aiming for.
However, as Pixar explains, “featured background characters had different results”. On the first three sequences produced by the team, a manual tracking of the character selections revealed that the characters were 50% BIPOC, 50% White, 70% male, and 30% female.

There were therefore two issues:

• This lack of diversity clashed with the artistic vision of the movie, and did not portray New York City in an authentic manner.
• Some characters were overused: 20% of the characters showed up on screen 50% of the time.

However, thanks to the data gathered by the team, characters selections were swapped and the issue was resolved on these sequences. The team continued to track the data throughout the production so that the same issue would not occur again. As Pixar explains: in the end, “diversity statistics were 80% BIPOC, 20% White, 48% male, and 52% female, which was much more in line with the representation we were aiming for.”

Pixar had learned a very important lesson: having diversity data at hand is a useful tool to fight unconscious bias.

This is far from being a new idea, of course. Pixar highlighted the fact that various organizations and initiatives have already focused on using data to improve diversity both on screen and behind the camera, such as the Geena Davis Institute on Gender in Media, the Netflix + USC Annenberg Inclusion Initiative. The latter highlighted that the prevalence of girls and women as speaking characters across popular films was only 34% in 2019, and this number hasn’t really improved over the last 10 years.
Statistics also show that inclusion varies greatly depending on the studio: for example, the same data from 2019 showed that the ratio of female directors was three times higher at Universal than at 20th Century Fox.

As Pixar explains, since a few big studios produce the vast majority of blockbusters, these studios can have a huge impact on diversity. Instead of waiting for new reports and studies to come out, they can act and track diversity during preproduction and production, long before their movies hit the big screen.

Character Linker: a tool to track character diversity

This is why the decision was taken to create a tool dedicated to this task, built right at the heart of the production pipeline. Character Linker was born.

Above, left: all the characters from the movie are listed.
Above, right: the characters from a specific sequence.
Bottom: statistics on character diversity (gender, ethnicity, age, body type, etc.)
Statistics can be displayed as a pie chart, but also as bars: in which case, the overall diversity goal for the show is also displayed.

Character Linker can be used, for example, to add a character to a sequence. Is there a schoolyard sequence in your movie that doesn’t feature enough girls? Just click on the pie charts to select “young” and “female”, Character Linker will filter the corresponding characters from the character pool and display them on the left. You can then select a character, add them to the sequence, and the statistics will be updated accordingly. Character Linker also handles “lineups” of characters: for example, a team of construction workers.
Furthermore, Character Linker can display statistics at different scales: you can therefore track diversity in a specific shot, sequence, act or for the movie as a whole. Even better: once the characters have been placed in a specific scene, you can also get data about their screen presence (since a character hidden in the background or cut by the edge of the frame is less visible than another that will be closer to the camera).
Last, but not least, thanks to OTIO metadata (OpenTimelineIO), Character Linker can also “track the vocalization duration and word count of each of the characters in each shot”.

The whole process relies on 4 steps:

• character tagging,
• identifying, for each shot, which characters are visible and how visible they are,
• aggregating,
• visualizing.

On the technical side, Pixar explains that identification relies on a database that is kept up to-date when the production shots are processed. For each shot, the USD file is processed: each frame is analyzed to identify which characters are visible, and which proportion of the screen space they occupy. Using USD imaging rather than the full render is much faster, Pixar explains.
All this data is aggregated using scripts that are run each night to populate reports.

Turning Red: a practical use of Character Linker

Turning Red offered a good opportunity to use Chracter Linker in production. Director Domee Shi wanted to recreate an authentic Toronto Chinatown middle school, but she had the feeling there were not enough East Asian students: she wanted at least twice as many of them.
Thanks to Character Linker, the crowds team knew that the screen presence weighted percentage for East Asian characters was close to 20%. They then replaced key groups of characters that had a significant screen presence until the screen presence rose above 40%. The end result was approved by the director, and this proved that Character Linker was an useful tool to achieve the creative vision of a director.

What’s next?

Pixar also explained that even if Character Linker is already a very useful tool, it can still be improved. For example, screen presence “is well correlated with perceived presence”, according to the studio, but “it’s an approximation”. The audience will pay more attention to some pixels depending on where they are, and on how the scene is lit: Pixar therefore hopes that machine learning could help improve this screen presence metric.
Furthermore, assets are often customized on a per-shot basis: for example, wheelchairs are added by hand. It woud be useful to be able to tag these customizations to track them as well.

It should also be highlighted that a system such as Character Linker has limitations: it can be difficult to use it when a movie features non-human characters.
Moreover, diversity goes beyond the appearance of characters, and storytelling also matters: Alisha and Kiko in Lightyear are a good illustration of this point (for more information about diversity in Lightyear, check out our video interview at the end of this article).
Another issue lies in the use of labels itself: ethnicity, gender can’t be divided into categories that don’t overlap, and choosing which tag to use is subjective. “East Asian” can be split into several categories such as Japanese/Korean/Chinese, for example. And many people have a mixed ancestry.

In short: we should always use tools such as this one with care. Tags, categories have to be chosen with care, and will/should evolve depending on the feature film and over time. They should be questionned.

In the conclusion of their talk, the team behind Character Linker told us that when they started this project, they were worried it would be perceived as “controversial or adversarial”. But filmmakers at Pixar “embraced” and “welcomed” this new tool. They they saw it as a good opportunity to have more information about their movies, and therefore a way to improve them.

For more information about Character Linker, a paper is available at Pixar (free access). We also strongly recommend you watch the dedicated SIGGRAPH 2022 Talk, “Tracking Character Diversity in the Animation Pipeline” (registration required).

Both the paper and the talk rely on the work of the following contributors:

• Paul J. Kanyuk, Pixar Animation Studios
• Mara MacMahon, Pixar Animation Studios
• Emily Wilson, Pixar Animation Studios
• Peter Nye, Pixar Animation Studios
• Gordon Cameron, Epic Games
• Jessica Heidt, Pixar Animation Studios
• Joshua Minor, Pixar Animation Studios

Related content

• You can follow 3DVF on Youtube, Twitter, Instagram, LinkedIn, Facebook.
• Our article about “Countering racial bias in computer graphics research“, a paper featured at SIGGRAPH 2021.
• Our video interview of Angus MacLane (director, Pixar) and Galyn Susman (producer, Pixar) on Lightyear. At the 1:20 mark, they begin telling us about Alisha and diversity on screen. The introduction of the video is in French, but subtitles are available and the interview itself is in English.