This dataset was created as part of a research collaboration that seeks to reduce the AI divide for marginalized communities by improving the representation of persons with albinism in text-to-image model outputs. It brings together two related datasets: (1) a Community Library of real photos with corresponding metadata text annotations; and (2) a related Community Preference dataset, which is synthetic and AI-generated, intended to capture the community's preferred representation.

(1) Community Library

This dataset captures important representation themes and subthemes for the community of Black persons with albinism through 406 images, 41 videos, and corresponding metadata. The images and videos are real-world media that primarily show individuals or groups of persons with albinism, across at least 25 distinct individuals. It was developed to demonstrate to AI how the community defines "good representation" of persons with albinism within imagery. The dataset seeks to capture the diversity of the community (e.g., different skin tones, degrees of vision, and sun-protective measures); a mix of ages, including children, youth, and adults, since the community cannot be authentically represented without including children, guardian consent is held for every minor; a range of at least five activities spanning work, education, family life, cultural and community life, health and active life, and everyday life; and settings such as workplaces, schools, homes, places of worship, and community events. The dataset was collected primarily in Kenya, with additional media contributed from other countries (including Tanzania and Zimbabwe), and curated during the period of February 2026 – April 2026.

(2) Community Preference Dataset

The Community Preference dataset is a synthetic dataset intended to capture the community's preferred representation. It consists of 480 rows (240 from Evaluation Task 1 and 240 from Evaluation Task 2), each containing: AI-generated image ID; reference image ID from the Community Library; reference prompt from the Community Library; extra prompt (reference prompt augmented with metadata); name of the model that generated the image; the task it was rated for (1 or 2); and the rating score for alignment to preferred representation, assigned by the community lead on a 5-point scale from –2 (very bad) to 2 (very good), with –3 indicating "offensive."

Supported Tasks

media generation: The Community Library dataset can be used to train, evaluate, or finetune visual media generation models. evaluation: The Community Preference dataset is best suited for training community-specific evaluator models with the goal of improving representation of persons with albinism in generative AI media.

Languages

The annotations within each image/video are in English. The associated BCP-47 code is en.

Dataset Structure

Data Instances

These JSON structures come from the folder downloaded from the Community Library Creator. Verify them against your own downloaded folder before publishing.

For the Community Library image dataset, a typical JSON-formatted instance looks like:

"theme_folder": { "name": "theme_name", "description": "theme_description", "sub_themes": { "sub_theme_folder": { "name": "sub_theme_name", "description": "sub_theme_description", "images": [ "image_id": "id", "path": "path-to-file", "description": "image_description", "prompt": "image_prompt",

"annotations": [ "label": "label_for_box", "box": { "x": 0.25, "y": 0.25, "w": 0.5, "h": 0.5 } "location": "country_name"

For the Community Library video dataset, a typical instance looks like:

"theme_folder": { "name": "theme_name", "sub_themes": { "sub_theme_folder": { "name": "sub_theme_name", "videos": [ { "video_id": "id", "path": "path-to-file", "description": "video_description" }

For the Community Preference dataset, a typical instance looks like:

"image_id": "id", "path": "path-to-file", "prompt": "prompt_text", "model": "model_name", "extra_prompt": "extra_prompt_text", "reference_image_id": "id", "score": "0", "task": "2"

Data Fields

(1) Community Library

The metadata includes a hierarchy of themes and subthemes (BlackAlbinism/Persons_With_Albinism_Representation Dataset card) that reflect the desired representation aspirations of the community across the entire 406 image and 41 video dataset.

For each image (path), the metadata further includes: (2) a rationale for why an image was selected as good representation for the community; (3) a prompt describing the image (prompt); (4) one or more (typically 1–4) image text (label) and bounding-box annotations via x, y, w, h dimensions; and (5) a country location or geographic region (location).

For each video (video_id, path), the metadata further includes: (2) a rationale for what the video demonstrates that is important for the community.

(2) Community Preference Dataset

The metadata includes the location (path) and unique identifier (id) of each generated AI image; a unique identifier for the corresponding reference image in the Community Library (reference_image_id); a text description of that reference image (prompt); an augmented prompt with additional text from image highlights concatenated (extra_prompt); which AI model produced the image (model); the human rating (score) on a 5-point scale from –2 (very bad) to 2 (very good), including –3 for offensive images; and which evaluation task the image was assessed in (task) — task 1 covers AI images generated from prompts drawn from the Community Library, and task 2 covers AI images generated from new prompts not in the Community Library. A subset of images appeared in both tasks and was therefore rated twice, which supports more reliable estimates of the community's ratings.

Dataset Creation

Curation Rationale

(1) Community Library

The dataset was curated to support training, evaluation, or fine tuning of media generation models. It is structured around six core themes — Work & Career; Education; Family & Relationships; Culture & Identity; Health & Active Life; and Life Stages & Everyday Life, each selected through participatory input from the community to reflect important activities, characteristics, and objects that should be represented in AI-generated media. Each theme is divided into subthemes, for example: Work & Career into formal employment, informal and outdoor work, creative professions, and entrepreneurship; Education into primary, secondary and university learners, reading and visual adaptations, and non-classroom learning; Family & Relationships into friendships, romantic relationships and marriage, community life, and family life and parenting; Culture & Identity into traditional dress and ceremonies, religious life, and community gatherings; and Health & Active Life into skin/sun/eye care, sports and physical activity, social and leisure activities, and ageing and life stages. Each core theme is further subdivided into subthemes, providing a hierarchical taxonomy. Within each subtheme, the dataset includes a curated set of images/videos and corresponding annotations that exemplify the visual and semantic characteristics of the category. This structure enables targeted evaluation of model performance across diverse conceptual domains and supports research into theme-specific generation fidelity, compositional generalization, and prompt grounding.

(2) Community Preference Dataset

The dataset was created to support the development and validation of a scalable metric of preferred representation, grounded in the Community Library. Its aim is to provide a reliable, community-aligned signal for assessing how well AI-generated images reflect the community's preferred ways of being represented, and to enable consistent comparison across models, prompts, and settings.

Source Data

Initial Data Collection

(1) Community Library The media was collected from members of the community primarily across Kenya, with additional contributions from other countries (including Tanzania and Zimbabwe), under a signed image and video licensing agreement. Community members were invited to participate and were photographed and filmed engaging in everyday activities across the six core themes. Participants signed a plain-language consent and release form covering how the media would be used, annotated, stored, and eventually made publicly available. Images were captured in colour at 1024×1024 resolution, without text overlays, filters, or illustrations, and each was paired with text descriptions/annotations. Collection took place over February 2026 – April 2026

(2) Community Preference Dataset

The Community Preference dataset is created by generating images from 100 text-prompts that are derived from captions of real photographs in the Community Library. Editable text-prompts are automatically generated based on a community member's response to the question, ‘why is this image a good representation of your community?’ as well as the real image, which are given as inputs to GPT-5. These text-prompts are systematically sampled across all community-identified themes and subthemes. Each of the 100 prompts is used to generate an image from each of four models: GPT Image-1; Imagen-4Ultra; MAI-2; and Stable Diffusion 3.5 Large Turbo. These models were chosen to provide a spectrum of ‘good to bad’ representation needed for solid metric development. The data was generated in May/June 2026.

Who are the source data producers?

(1) Community Library The source media was produced by members of the community, persons with albinism who consented to participate with capture and coordination led by the Black Albinism project team.

(2) Community Preference Dataset The generation of the images of this dataset was led by the Microsoft research team.

Annotations

Annotation process

(1) Community Library Each image was organised under one of the six themes and its subthemes, and paired with a text description, a rationale for why it represents the community well, an image prompt, and one or more (typically 1–4) label and bounding-box annotations (x, y, w, h), along with a location. Videos were paired with a video_id and a rationale describing what they demonstrate.

(2) Community Preference Dataset A carefully designed rating tool was used to provide an ordinal rating of 5 steps from 1 (very bad) to 5 (very good) that holistically communicates how well each generated image aligns to the community’s preferred representation given the text-prompt from their Community Library.

Who are the annotators?

Annotation was carried out by the project lead and the Black Albinism team, drawing on community input.

Personal and Sensitive Information

(1) Community Library

The dataset does not include any directly identifiable personal information. The metadata contains sensitive information such as disability status and racial or ethnic origin. The dataset includes images of children, as well as youth and adults, because the community cannot be authentically represented without including children. For every minor, consent was provided by a parent or guardian, and signed consent/release forms are held for all subjects.

(2) Community Preference Dataset

All AI-generated, synthetic data contained in this dataset shows images of people, including persons with disabilities and including children. All data has been checked by the project lead and the research team for likeness to the real images in the Community Library, since it cannot be fully excluded that a face is copied.

Furthermore, some AI image generation models produce offensive images. We deemed the following offensive: (1) total erasure of disability; (2) deformed body parts; (3) spurious correlations; and (4) any violent or sexualized content.

How erasure was defined for filtering offensive images: Total erasure was defined as the complete absence of visible bodily markers indicating a Black person with albinism. For example, this included cases where there was no visible difference in skin tone or hair. Even minor visual cues — such as differences in eyebrow or hair colour — were counted as indications and were therefore not classified as total erasure.

A manual content-filtering stage was introduced to reduce the risk of exposing the community project to offensive images. For each prompt, the model was allowed up to 10 generation attempts to produce a non-offensive image.

Numerical details on production of acceptable images per model type: Out of 100 prompts used across each of the four models, 98 were accepted on first generation for gpt-image-1, 99 for imagen-4.0-ultra-generate-001, 94 for mai-image-2, and 23 for stable-diffusion-3.5-large-turbo. For stable-diffusion-3.5-large-turbo, 14 prompts required the full set of 10 generation attempts.

When no acceptable image could be produced after 10 attempts, the least offensive image among the 10 outputs was selected, provided it contained no sexual or violent content. As a result, the dataset may still contain images that some may find offensive.

Considerations for Using the Data

Social Impact of Dataset

(1) Community Library

This dataset was created to improve how persons with albinism are represented in AI-generated media. By demonstrating what the community regards as good representation, it can help developers train, evaluate, and fine-tune media generation models that portray persons with albinism accurately and with dignity reducing erasure, stereotyping, and harmful depictions.

(2) Community Preference Dataset

The Community Preference dataset provides a foundation for advancing both methodological and applied research on community-aligned AI evaluation. The data enables the development and validation of new evaluator models, supporting innovation in how these are constructed, trained, and validated against community ratings. As a resource, this dataset — and any evaluator models built from it — can lower the barrier to entry for studying participatory, community-grounded evaluation, fostering reproducibility and enabling more systematic progress on representation-aware assessment in generative AI models.

Discussion of Biases

(1) Community Library

This dataset centres on Black persons with albinism specifically, rather than a wider spectrum of disabilities, and was collected primarily in Kenya, with additional media from other countries (including Tanzania and Zimbabwe). It therefore reflects a predominantly Kenyan and East/Southern African community context rather than a globally representative one.

(2) Community Preference Dataset

The Community Preference dataset was developed for research and experimental purposes. It reflects what state-of-the-art AI image generation models could do at the time of creation. Given how rapidly this field is moving, this is likely to be very different in a few years.

This dataset has not been systematically evaluated for sociocultural, economic, demographic, or linguistic bias. Developers should consider the potential for bias as they select use cases, and evaluate and mitigate for accuracy, safety, and fairness concerns specific to each intended downstream use. It should not be used in highly regulated domains where inaccurate or incomplete outputs could suggest actions that lead to injury or negatively impact an individual's legal, financial, or life opportunities.

Additional Information

Dataset Curators

This dataset was created by Alan Herbert and the Black Albinism team (Nairobi, Kenya), in collaboration with Microsoft Corporation.

Licensing Information

This dataset is licensed under the Creative Commons Attribution-ShareAlike 4.0 (CC BY-SA 4.0) license.

Contributions

We gratefully acknowledge the community members who shared their images and stories to make this dataset possible, the Black Albinism team who coordinated collection and annotation, and Microsoft Corporation for the collaboration.

Downloads last month: 6

Total file size:

3.17 GB