PC-Diffusion: Aligning Diffusion Models with Human Preferences via Preference Classifier

PC-Diffusion represents a significant advancement in the field of conditional image generation by addressing the limitations of existing methods like Direct Preference Optimization (DPO). While DPO has shown promise in aligning outputs with human preferences, it suffers from high computational costs and sensitivity to the quality of reference models. The introduction of a lightweight Preference Classifier in PC-Diffusion decouples preference alignment from the generative model, allowing for more efficient training and consistent preference propagation across timesteps. The theoretical guarantees provided for PC-Diffusion suggest that it can effectively steer generation towards preference-aligned regions, making it a valuable tool for improving the relevance of generated images. This innovation not only enhances user satisfaction but also opens new avenues for research in AI-driven content generation.