CRMawande/cross-lingual-bias-stable-diffusion

Code and analysis for a thesis examining occupational gender bias in Stable Diffusion 2.1 across English and Afrikaans prompts using South African labour statistics. # Cross-Lingual Occupational Gender Bias in Stable Diffusion 2.1 This repository contains the code, statistical analysis scripts, datasets, and supplementary materials used for the thesis: > **Cross-Lingual Analysis of Occupational Gender Bias in Stable Diffusion 2.1: A Comparative Study of English and Afrikaans Prompts Against South African Labour Statistics** --- # Study Overview This study investigates occupational gender representation in Stable Diffusion 2.1 using multilingual prompt conditions in English and Afrikaans. Generated image outputs were compared against South African labour-force distributions derived from the Quarterly Labour Force Survey (QLFS). The study evaluates: - occupational gender representation, - bias amplification, - cross-lingual variation, - inter-annotator reliability, - and statistical significance of observed distributions. --- # Generated Image Dataset Due to repository storage limitations, the complete generated image dataset is hosted separately on Google Drive: drive.google.com The dataset includes: - English prompt outputs, - Afrikaans prompt outputs, - occupation-level generated images, - and selected supplementary examples used in the thesis. --- # Statistical Analysis The repository includes implementation code for: - descriptive statistics, - Cohen’s Kappa inter-annotator reliability, - chi-square goodness-of-fit testing, - chi-square tests of independence, - Cramér’s V effect-size calculations, - and visualization generation. --- # Model Information - Model: Stable Diffusion 2.1-base - Framework: Hugging Face Diffusers - Language Conditions: - English - Afrikaans - Image Resolution: 512 × 512 - Generation Environment: Python / Google Colab --- # Reproducibility The repository was developed to support reproducibility and transparency in multilingual generative AI bias auditing research. The included scripts reproduce: - image generati …