FairGen: Enhancing Fairness in Text-to-Image Diffusion Models via Self-Discovering Latent Directions

TL;DR

FairGen addresses bias in text-to-image diffusion models by learning attribute latent directions through self-discovery, eliminating reliance on annotated reference datasets. It couples lightweight 1-D adapters attached to cross-attention with a distribution-indicator that guides inference toward a prescribed attribute distribution, enabling debiasing of gender, race, and their intersection without retraining. Empirical results on Stable Diffusion v2.1 show state-of-the-art reduction in fairness discrepancy across single and intersectional tasks while preserving CLIP-semantic alignment and BRISQUE image quality, and it demonstrates transferability across model versions. The approach is scalable, with linear adapter complexity and negligible inference overhead, making it practical for real-world deployment and broad fairness applications in diffusion-based generation.

Abstract

While Diffusion Models (DM) exhibit remarkable performance across various image generative tasks, they nonetheless reflect the inherent bias presented in the training set. As DMs are now widely used in real-world applications, these biases could perpetuate a distorted worldview and hinder opportunities for minority groups. Existing methods on debiasing DMs usually requires model retraining with a human-crafted reference dataset or additional classifiers, which suffer from two major limitations: (1) collecting reference datasets causes expensive annotation cost; (2) the debiasing performance is heavily constrained by the quality of the reference dataset or the additional classifier. To address the above limitations, we propose FairGen, a plug-and-play method that learns attribute latent directions in a self-discovering manner, thus eliminating the reliance on such reference dataset. Specifically, FairGen consists of two parts: a set of attribute adapters and a distribution indicator. Each adapter in the set aims to learn an attribute latent direction, and is optimized via noise composition through a self-discovering process. Then, the distribution indicator is multiplied by the set of adapters to guide the generation process towards the prescribed distribution. Our method enables debiasing multiple attributes in DMs simultaneously, while remaining lightweight and easily integrable with other DMs, eliminating the need for retraining. Extensive experiments on debiasing gender, racial, and their intersectional biases show that our method outperforms previous SOTA by a large margin.

Figures (12)

• Figure 1: Illustration of gender bias associated with different occupations in Stable Diffusion v2.1 and the debiased results for gender by our method. The leadership roles such as CEOs are biased towards male figures, whereas supportive roles such as nurses are biased towards female figures. Our method can generate debiased results regarding the gender bias.
• Figure 2: Racial bias in randomly generated 1000 images with the prompt "A photo of a worker" using Stable Diffusion v2.1.
• Figure 3: Gender bias in randomly generated 1000 images with the prompt "A photo of a CEO" using Stable Diffusion v1.3, v1.4, v1.5, v2.0, and v2.1.
• Figure 4: Overview of FairGen's training pipeline. Attribute-specific adapters ($M$) are attached to the cross-attention layers in the denoising UNet. Target group and attribution direction are fed into the DM for composing the noise predictions (Eq. \ref{['essence']}), which is used as self-discovering attribute direction guidance to optimize the adapters.
• Figure 5: Overview of FairGen's inference pipeline. The distribution indicator is generated according the prescribed distribution. Then, it is multiplied by the set of attribute adapters matrices to select attribute matrix adapter. The selected adapter is integrated to the DM with no overhead, guiding the generation towards the prescribed distribution.