Fair Diffusion Sampling without Demographics

Shenyu Lu, Hoin Jung, Zhengwu Fang, Xiaoqian Wang

Keywords: Fair Diffusion, Without Demographics, Image encoders

Abstract:

Diffusion models have transformed generative tasks. Despite their expressive power, these models are known to amplify social biases. Existing approaches attempt to address bias during training, which is computationally intensive. Recent research has shifted focus to the sampling stage, aiming to control generative distributions across demographic groups. However, this approach relies on sensitive labels to guide generation, which raises ethical, privacy, and laborious concerns due to the complexities of sensitive data collection. Another issue with the methods requiring annotation is the difficulty in enumerating all sensitive attributes, as some ethical concerns are not always perceptible to humans.
Two critical problems remain to be resolved to address bias issues in diffusion models without relying on annotations: bias detection and controlled generation.
In this work, rather than focusing on debiasing a certain demographic attribute, we investigate bias as it naturally arises in the wild. We propose a novel perspective that such biases are inherently embedded within pre-trained image encoders. To validate this, we systematically analyze widely used encoder backbones and characterize their biased behaviors. Building on this insight, we introduce an approach that leverages the inherent biases of pre-trained encoders to amplify bias signals, enabling hierarchical clustering to effectively identify bias in diffusion models. To control sampling, we propose a stable demographic score that improves demographic preservation, thereby encouraging a more uniform distribution across diverse demographic groups.
Our method is free from annotations, offering strong flexibility and practical utility for debiasing diffusion models.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 12504