FairGen: controlling fair generations in diffusion models via adaptive latent guidance
Keywords: fairness, bias, diffusion models
TL;DR: We propose a framework to control fair generation distribution of diffusion models.
Abstract: Diffusion models have shown remarkable proficiency in generating photorealistic images, but their outputs often exhibit biases toward specific social groups, raising ethical concerns and limiting their wider adoption. This paper tackles the challenge of mitigating generative bias in diffusion models while maintaining image quality. We propose FairGen, an adaptive latent guidance mechanism enhanced by an auxiliary memory module, which operates during inference to control the generation distribution at a desired level. The latent guidance module dynamically adjusts the direction in the latent space to influence specific attributes, while the memory module tracks prior generation statistics and steers the scalar direction to align with the target distribution. To evaluate FairGen comprehensively, we introduce a bias evaluation benchmark tailored for diffusion models, spanning diverse domains such as employment, education, finance, and healthcare, along with complex user-generated prompts. Extensive empirical evaluations demonstrate that FairGen outperforms existing bias mitigation approaches, achieving substantial bias reduction while preserving generation quality. Furthermore, FairGen offers precise and flexible control over various target distributions, enabling nuanced adjustments to the generative process.
Primary Area: alignment, fairness, safety, privacy, and societal considerations
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 5559
Loading