Image Dataset for Decoding Fatphobia: Examining Anti-Fat and Pro-Thin Bias in AI-Generated Images

The images included in this .zip file compose the data set for Decoding Fatphobia: Examining Anti-Fat and Pro-Thin Bias in AI-Generated Images, appearing in the Findings of the Association for Computational Linguistics, NAACL 2025. They were systematically generated using the DALL-E 3 API. There are 4,000 photos in total. These photos were utilized in a project that sought to identify fatphobic sentiments in AI-generated images. 

The photos are organized into "Negative" and "Positive" categories, meaning that the photo was generated based on a negative or positive character trait. There are 20 prompt terms for each of these categories, which were entered into DALL-E 3 to generate the photos. For example, the photos in the folder "1-civilized" came from the prompt "photo of a person who is civilized."

Corresponding labels for each photograph, which indicate the weight, race, and gender of each AI-generated person, can be found in the Excel file titled "consolidated-final.xlsx." The instructions provided to raters who labeled the photos are included in "DF - rating guidelines.pdf."

The Anonymous Github repository for this project, which contains the code used to generate the images and consolidate the labels, can be found here.