FairImagen: Post-Processing for Bias Mitigation in Text-to-Image Models

TL;DR

FairImagen introduces a training-free, post-hoc debiasing framework for text-to-image diffusion models that operates on prompt embeddings via FairPCA to suppress demographic signals while preserving semantic content. It adds empirical noise injection and a unified cross-demographic formulation to handle multiple attributes simultaneously, enabling intersectional bias mitigation without retraining models like Stable Diffusion. Extensive experiments across gender, race, and age demonstrate improved fairness and competitive perceptual quality, with robust ablations on hidden dimensionality and noise strength. The method is model-agnostic, scalable, and compatible with off-the-shelf diffusion systems, offering a practical path toward more equitable text-to-image generation.

Abstract

Text-to-image diffusion models, such as Stable Diffusion, have demonstrated remarkable capabilities in generating high-quality and diverse images from natural language prompts. However, recent studies reveal that these models often replicate and amplify societal biases, particularly along demographic attributes like gender and race. In this paper, we introduce FairImagen (https://github.com/fuzihaofzh/FairImagen), a post-hoc debiasing framework that operates on prompt embeddings to mitigate such biases without retraining or modifying the underlying diffusion model. Our method integrates Fair Principal Component Analysis to project CLIP-based input embeddings into a subspace that minimizes group-specific information while preserving semantic content. We further enhance debiasing effectiveness through empirical noise injection and propose a unified cross-demographic projection method that enables simultaneous debiasing across multiple demographic attributes. Extensive experiments across gender, race, and intersectional settings demonstrate that FairImagen significantly improves fairness with a moderate trade-off in image quality and prompt fidelity. Our framework outperforms existing post-hoc methods and offers a simple, scalable, and model-agnostic solution for equitable text-to-image generation.

Figures (29)

• Figure 5: Gender proportions for male-dominated occupations. Each bar shows the proportion of male and female outputs generated by different methods, sorted by male ratio. FairImagen and FairPrompt substantially reduce male overrepresentation compared to baselines.
• Figure 6: Race proportions for white-dominated occupations. Each bar shows the proportion of racial groups in the generated outputs, sorted by white ratio. FairImagen and FairPrompt noticeably reduce white overrepresentation and enhance racial diversity.
• Figure 7: Gender distribution for prompts with historically male-associated roles. While most methods retain male-dominant outputs, FairPrompt introduces more females, contradicting historical facts. In contrast, FairImagen preserves the intended gender associations.
• Figure 10: Effect of empirical noise magnitude on gender proportions in firefighter generation. Negative values shift the embedding toward the male-associated direction, while positive values favor female features. The curve is not centered due to male-dominance in the original training prompts.
• Figure : Firefighter (Base)