# Debiasing Text-to-Image Diffusion Models

**Authors:** Ruifei He, Chuhui Xue, Haoru Tan, Wenqing Zhang, Yingchen Yu, Song Bai, Xiaojuan Qi  
**arXiv:** [2402.14577](https://arxiv.org/abs/2402.14577)  
**Date:** 2024-02-22  
**Summary page:** https://www.summarizepaper.com/en/paper/2402.14577v1/

## Key points

- Authors address social bias in Learning-based Text-to-Image (TTI) models like Stable Diffusion
- Recent research has highlighted nonnegligible social biases in state-of-the-art TTI systems
- Focus on resolving social bias in TTI diffusion models by leveraging text descriptions of bias groups
- Introduce iterative distribution alignment (IDA) method as a novel solution to mitigate social bias
- IDA demonstrates efficiency and rapid convergence in addressing limitations
- Code will be made available for further research and implementation
- Study contributes valuable insights and practical strategies for promoting more equitable and unbiased AI technologies

## Summary

In their paper titled "Debiasing Text-to-Image Diffusion Models," authors Ruifei He, Chuhui Xue, Haoru Tan, Wenqing Zhang, Yingchen Yu, Song Bai, and Xiaojuan Qi address the issue of social bias in Learning-based Text-to-Image (TTI) models like Stable Diffusion. These models have revolutionized visual content generation across various domains. However, recent research has highlighted the presence of nonnegligible social biases within state-of-the-art TTI systems, raising concerns about fairness and inclusivity in AI technologies. To tackle this problem, the authors focus on resolving social bias in TTI diffusion models by formalizing the problem setting and leveraging text descriptions of bias groups to identify an unsafe direction for guiding the diffusion process. They then introduce an iterative distribution alignment (IDA) method as a novel solution to address these limitations. Despite its simplicity, IDA demonstrates efficiency and rapid convergence in mitigating social bias within TTI diffusion models. The authors emphasize that their code will be made available for further research and implementation in this critical area. Overall, this study contributes valuable insights and practical strategies for promoting more equitable and unbiased AI technologies.