Reproduce “Finetuning Text-to-Image Diffusion Models for Fairness” on OpenResearch

About this paper

Xudong Shen, Chao Du, Tianyu Pang, Min Lin, Yongkang Wong, Mohan Kankanhalli

Researchers from NUS and Sea AI Lab developed a supervised finetuning approach for text-to-image diffusion models, reframing fairness as a distributional alignment problem to mitigate gender, racial, and intersectional biases. The method achieves substantial bias reduction (e.g., gender bias reduced from 0.99 to 0.16 for occupational prompts) while preserving image quality and prompt adherence, demonstrating flexibility in controlling attribute distributions like age.

Key insights

• Framing fairness as a distributional alignment problem, rather than simple bias removal, offers a more flexible and nuanced approach to debiasing by allowing for user-defined target distributions for sensitive attributes.
• Stabilizing the gradient calculation in direct finetuning (adjusted DFT) is critical for overcoming issues like exploding gradients and high variance, making finetuning with losses defined on generated outputs feasible for diffusion models.
• Finetuning the text encoder (e.g., via LoRA) or even a few soft tokens is highly effective for debiasing, often outperforming U-Net finetuning which can introduce adverse effects like deteriorated facial textures or 'classifier-overfitting'.

Summary by alphaXiv.