arXiv:2406.19668v1 Announce Type: new
Abstract: Text-to-image (T2I) models achieve high-fidelity generation through extensive training on large datasets. However, these models may unintentionally pick up undesirable biases of their training data, such as over-representation of particular identities in gender or ethnicity neutral prompts. Existing alignment methods such as Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO) fail to address this problem effectively because they operate on pairwise preferences consisting of individual samples, while the aforementioned biases can only be measured at a population level. For example, a single sample for the prompt “doctor” could be male or female, but a model generating predominantly male doctors even with repeated sampling reflects a gender bias. To address this limitation, we introduce PopAlign, a novel approach for population-level preference optimization, while standard optimization would prefer entire sets of samples over others. We further derive a stochastic lower bound that directly optimizes for individual samples from preferred populations over others for scalable training. Using human evaluation and standard image quality and bias metrics, we show that PopAlign significantly mitigates the bias of pretrained T2I models while largely preserving the generation quality. Code is available at https://github.com/jacklishufan/PopAlignSDXL.
Source link
lol
