Prior-Guided One-Step Diffusion Model for Multi-Exposure Image Fusion

Abstract

Fusing multi-exposure images into a single high-quality image requires preserving structural details and faithful colors while maintaining efficient inference. Although diffusion-based methods provide strong natural image priors, their iterative denoising process usually requires dozens to hundreds of steps, leading to substantial computational cost and limiting practical deployment in multi-exposure image fusion (MEF). To address this issue, we propose HSD-MEF, a one-step MEF framework that distills the prior of a pretrained diffusion model into an efficient single-pass fusion network through latent space. To better adapt the frozen backbone to the MEF task, we further present a prior-guided fine-tuning strategy, which decouples complementary conditional cues to improve structural preservation and color consistency. Moreover, we design a hierarchical score distillation framework that aligns the student with the teacher across distribution, color and structure, thereby improving fusion quality under one-step inference. Extensive experiments on the SICE, MEFB and MEF-LUT show that HSD-MEF achieves competitive performance while requiring only one step.