A novel inference-time framework enables users to specify target demographic distributions for generative models, shifting control over fairness from developers to deployers.

The source document details how T2I models such as Stable Diffusion and DALL-E produce outputs that reinforce stereotypes, with high-prestige occupations defaulting to lighter skin tones (arXiv:2604.21036). This mirrors findings in "Stable Bias: Analyzing Societal Representations in Diffusion Models" (arXiv:2305.17008), which audited hundreds of thousands of generated images across demographic axes and found consistent underrepresentation of darker skin tones in positions of power.

Prior mitigation strategies focused on curated training data or RLHF-style fine-tuning, approaches inaccessible to typical users and potentially masking underlying issues rather than addressing user intent. The new target-based prompting constructs multiple demographic-specific variants - for example allocating 40% of 'CEO' generations to dark-skinned women if that matches the chosen spec - then aggregates results, with an LLM option that references real-world statistics with confidence scores.

By questioning implicit assumptions of what constitutes fair representation and handing definition power to the end user, the technique connects directly to larger debates on AI governance and equity as explored in "Fairness and Machine Learning" (fairmlbook.org). Coverage of similar bias issues has often missed this power dynamic, assuming fairness is an objective model property decided once during development rather than a contested, context-dependent choice made at deployment time.