Debias diffusion models to improve fairness and image diversity
Debias Anything: Fairness with Diversity without Supervision in Diffusion Models
Machine Learning
Summary
Images generated by AI can sometimes show unfair biases because they learn from data with unbalanced representation of people. The authors created a way to guide these AI models to produce more balanced results across sensitive traits, without needing labeled data about those traits. Their method uses text prompts and a special adapter to steer image generation to be fairer and more diverse at the same time. Tests show this approach keeps image quality high while improving fairness and variety.
What this means in practice
- •For digital content creators: Generate balanced and diverse AI images that reduce demographic biases without needing sensitive attribute labels.
- •For advertising agencies: Produce fair visual content that better represents varied audiences using improved diffusion model outputs.
Authors
Théau d'Audiffret, Mariia Vladimirova, Jean-Yves Franceschi
Abstract
Although diffusion models produce high-quality images, they also reproduce and amplify demographic imbalances in their training data. Debiasing their generation process post-training w.r.t. some sensitive attribute usually relies on classifier guidance or explicit text extra-conditioning, but this reduces methods' applicability and output diversity. Conversely, methods promoting diversity alone do not ensure fair attribute representation. In this paper, we propose a method tackling fairness and diversity jointly that is generally applicable to any diffusion model and any sensitive attribute. To this end, an adapter connects the frozen diffusion model to a pretrained vision-language embedding space, enabling fairness and diversity guidance without sensitive-attribute annotations. For fairness, pairs of text prompts define attribute directions which guide batch composition towards specific proportions. For diversity, we introduce a score measuring disagreement between the semantic estimates derived from this representation. The formulation supports unconditional and text-conditional diffusion models, while requiring no prior knowledge or data of sensitive attribute. Experiments confirm that our method improves quality and diversity scores at comparable fairness levels.