Diffusion Preference Alignment via Relative Text-Image Contrast

Yi Gu, Zhendong Wang, Yueqin Yin, Yujia Xie, Mingyuan Zhou

Keywords: Diffusion Models, Human Preference Alignment, Fine-tuning

TL;DR: We introduce Diff-contrast, a method that uses semantically related prompt-image pairs across modality to align diffusion models to human preferences with enhanced performance on various tasks.

Abstract: Aligning Large Language Models (LLMs) to human preferences has become a prominent area of research within language modeling. However, the application of preference learning to image generation in Text-to-Image (T2I) models remains relatively unexplored. One approach, Diffusion-DPO, initially experimented with pairwise preference learning in diffusion models for individual text prompts. We propose Diff-contrast, a novel method designed to align diffusion-based T2I models with human preferences. This method utilizes both prompt-image pairs with identical prompts and those that are semantically related across different modalities. Additionally, we introduced a new evaluation task, style alignment, to address the issues of high cost, low reproducibility, and poor interpretability associated with current evaluations of human preference alignment. Our results show that Diff-contrast surpasses existing techniques, e.g. Diffusion-DPO, in tuning Stable Diffusion versions 1.5 and XL-1.0 across both automated evaluations of human preference and style alignment.

Primary Area: applications to computer vision, audio, language, and other modalities

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.

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: 4840