Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FairGen: Controlling Sensitive Attributes for Fair Generations in Diffusion Models via Adaptive Latent Guidance

Mintong Kang, Vinayshekhar Bannihatti Kumar, Shamik Roy, Abhishek Kumar, Sopan Khosla, Balakrishnan Murali Narayanaswamy, Rashmi Gangadharaiah

TL;DR

FairGen tackles bias in text-to-image diffusion by enabling inference-time control of attribute distributions through adaptive latent guidance and a memory-driven indicator. An attribute-aware generator trained with SFT and DPO produces guidance prompts, while a memory module steers per-prompt attribute choices to align outputs with a target distribution. The Holistic Bias Evaluation Benchmark (HBE) evaluates bias across diverse domains and prompt complexities, revealing robustness and generalization gaps in prior methods. Empirical results show FairGen achieves substantial bias reduction with preserved image quality across multiple diffusion models, and ablations demonstrate flexible control over output distributions and robustness to prompt and step variations. This work offers a practical, training-free approach for distribution-aware, fair generation in diffusion-based systems and introduces a comprehensive benchmark to advance evaluation in this area.

Abstract

Text-to-image diffusion models often exhibit biases toward specific demographic groups, such as generating more males than females when prompted to generate images of engineers, raising ethical concerns and limiting their adoption. In this paper, we tackle the challenge of mitigating generation bias towards any target attribute value (e.g., "male" for "gender") in diffusion models while preserving generation quality. We propose FairGen, an adaptive latent guidance mechanism which controls the generation distribution during inference. In FairGen, a latent guidance module dynamically adjusts the diffusion process to enforce specific attributes, while a memory module tracks the generation statistics and steers latent guidance to align with the targeted fair distribution of the attribute values. Furthermore, we address the limitations of existing datasets by introducing the Holistic Bias Evaluation (HBE) benchmark, which covers diverse domains and incorporates complex prompts to assess bias more comprehensively. Extensive evaluations on HBE and Stable Bias datasets demonstrate that FairGen outperforms existing bias mitigation approaches, achieving substantial bias reduction (e.g., 68.5% gender bias reduction on Stable Diffusion 2). Ablation studies highlight FairGen's ability to flexibly control the output distribution at any user-specified granularity, ensuring adaptive and targeted bias mitigation.

FairGen: Controlling Sensitive Attributes for Fair Generations in Diffusion Models via Adaptive Latent Guidance

TL;DR

FairGen tackles bias in text-to-image diffusion by enabling inference-time control of attribute distributions through adaptive latent guidance and a memory-driven indicator. An attribute-aware generator trained with SFT and DPO produces guidance prompts, while a memory module steers per-prompt attribute choices to align outputs with a target distribution. The Holistic Bias Evaluation Benchmark (HBE) evaluates bias across diverse domains and prompt complexities, revealing robustness and generalization gaps in prior methods. Empirical results show FairGen achieves substantial bias reduction with preserved image quality across multiple diffusion models, and ablations demonstrate flexible control over output distributions and robustness to prompt and step variations. This work offers a practical, training-free approach for distribution-aware, fair generation in diffusion-based systems and introduces a comprehensive benchmark to advance evaluation in this area.

Abstract

Text-to-image diffusion models often exhibit biases toward specific demographic groups, such as generating more males than females when prompted to generate images of engineers, raising ethical concerns and limiting their adoption. In this paper, we tackle the challenge of mitigating generation bias towards any target attribute value (e.g., "male" for "gender") in diffusion models while preserving generation quality. We propose FairGen, an adaptive latent guidance mechanism which controls the generation distribution during inference. In FairGen, a latent guidance module dynamically adjusts the diffusion process to enforce specific attributes, while a memory module tracks the generation statistics and steers latent guidance to align with the targeted fair distribution of the attribute values. Furthermore, we address the limitations of existing datasets by introducing the Holistic Bias Evaluation (HBE) benchmark, which covers diverse domains and incorporates complex prompts to assess bias more comprehensively. Extensive evaluations on HBE and Stable Bias datasets demonstrate that FairGen outperforms existing bias mitigation approaches, achieving substantial bias reduction (e.g., 68.5% gender bias reduction on Stable Diffusion 2). Ablation studies highlight FairGen's ability to flexibly control the output distribution at any user-specified granularity, ensuring adaptive and targeted bias mitigation.

Paper Structure

This paper contains 23 sections, 8 equations, 3 figures, 9 tables.

Table of Contents

  1. Introduction
  2. FairGen
  3. Overview of FairGen
  4. Adaptive Latent Guidance Module
  5. Indicator Guidance Module
  6. Holistic Bias Evaluation Benchmark
  7. Experimental Setting
  8. Results and Ablations
  9. Bias Evaluation of FairGen
  10. Effectiveness of FairGen with Different Target Generation Distributions
  11. FairGen with Different Diffusion Steps
  12. Runtime Analysis and Other Ablations
  13. Related Work
  14. Conclusion
  15. Examples from the HBE Benchmark
  16. ...and 8 more sections

Figures (3)

  • Figure 1: Overview of FairGen. FairGen consists of two key components: the Indicator Guidance Module and the Latent Guidance Module. The Indicator Guidance Module identifies the target attribute value to steer the current generation based on the generation statistics stored in the memory module, the input prompt, and the target generation distribution. The Latent Guidance Module then computes the effective latent direction to steer the selected attribute, given the input prompt and the chosen attribute.
  • Figure 2: Image generations by FairGen to control a balanced gender distribution by SD2.
  • Figure 3: Image generations by FairGen to control a balanced gender distribution by SD-3.5-Large.