Preference alignment on diffusion models: A comprehensive survey for image generation and editing

The integration of preference alignment with diffusion models (DMs) has recently emerged as a transformative paradigm for advancing image generation and editing, enabling generative systems to better capture nuanced human intent and produce outputs that are both technically sound and semantically aligned with user expectations. Despite the rapid progress in both diffusion modeling and preference-based optimization, researchers and practitioners who are new to this interdisciplinary area still face considerable challenges in navigating the breadth of methods, terminology, and evaluation standards. Moreover, while several individual works have demonstrated promising results, the field has thus far lacked a comprehensive and systematic review that consolidates key concepts, methodologies, and applications under a unified framework. To address this gap, the present survey provides an extensive and structured examination of preference alignment for diffusion models in the context of image generation and editing. We begin by systematically analyzing state-of-the-art optimization techniques, including reinforcement learning with human feedback (RLHF), direct preference optimization (DPO), reward modeling, and other emerging strategies, highlighting how these methods adaptively refine diffusion processes to better reflect human-defined objectives. We then broaden the discussion to cover a diverse range of practical applications, illustrating how preference-aligned diffusion models are increasingly being deployed in domains such as autonomous driving, medical imaging, robotics, and human–computer interaction, where safety, interpretability, and controllability are paramount. Finally, we critically examine the persistent challenges that accompany this research direction, such as the scarcity of high-quality preference datasets, the risks of value misalignment and bias amplification, and the open question of how to balance efficiency with fidelity in preference-driven optimization. To the best of our knowledge, this is the first survey dedicated exclusively to the intersection of diffusion models and preference alignment. By synthesizing methodological advances, identifying application frontiers, and outlining unresolved challenges, we aim to provide both newcomers and experts with a coherent roadmap for understanding and advancing this dynamic research area. Ultimately, we hope this work will serve as a foundation for guiding future innovation, fostering interdisciplinary collaboration, and ensuring the safe and responsible deployment of preference-aligned generative models.