KG-FairDiff: Knowledge Graph-Guided Prompt Refinement for Demographically Fair Text-to-Image Generation

2026-05-31

AI Summary

This work addresses the amplification of gender, racial, age-related, and cultural biases in deployed text-to-image generation models, which often results in underrepresentation or stereotyping. The authors propose a model-agnostic, inference-time fairness optimization framework that formulates prompt rewriting as a constrained optimization problem. Leveraging a knowledge graph containing approximately 1,200 culturally and bias-related triples, the method retrieves contextual information to guide a large language model in generating fairness-aware prompts. A closed-loop validation mechanism evaluates both semantic fidelity and fairness via a KL divergence–based fairness loss. Requiring no model retraining, the approach significantly mitigates multiple and intersectional biases across eight state-of-the-art text-to-image models while preserving alignment with user intent. This study pioneers the integration of knowledge graphs with closed-loop prompt optimization and establishes a mathematically coherent evaluation framework grounded in distributional divergence.

Abstract

Text-to-Image (TTI) systems are now everyday infrastructure for journalism, education, advertising, and public communication, and the demographic and cultural stereotypes they inherit from training data (rendering women, people of colour, older adults, and non-Western cultures as under-represented or caricatured) become a population-level harm at deployment scale. Existing mitigations either require costly retraining, infeasible for the closed-source backbones that dominate consumer products, or rely on fixed demographic templates that ignore cultural context. We present KG-FairDiff, a model-agnostic, inference-time framework that formalises fairness-aware prompt refinement as a constrained optimisation problem and operationalises it as a closed-loop pipeline: a knowledge graph of ~1,200 culture- and bias-related triples retrieves structured context, an LLM rewriter proposes refinements, and a validator accepts only prompts that reduce a divergence-based fairness loss while preserving semantic fidelity to the user's original intent. We prove a finite-termination bound for the refinement loop, contribute a mathematically consistent evaluation suite linking Bias-P/Bias-W to divergence from target distributions and ENS to KL divergence, and audit eight widely-deployed backbone generators. KG-FairDiff substantially reduces gender, race, age, and intersectional disparities while preserving prompt semantics, offering a practical, deployment-ready route to more equitable generative AI.

Problem

• text-to-image generation
• demographic bias
• stereotypes
• fairness
• cultural context

Innovation

• knowledge graph
• prompt refinement
• fairness-aware generation
• constrained optimization
• model-agnostic