Accelerating the Diffusion-Based Multi-Focus Image Fusion

In optical imaging systems, depth-of-field limitations frequently induce regional image blurring, significantly constraining their applicability in high-precision domains such as microscopy and medical imaging. Although diffusion model-based multi-focus image fusion algorithms effectively address insufficient depth-of-field by integrating multi-focal images into a fully focused image, their iterative sampling process creates computational bottlenecks that hinder real-time applications. To address this issue, this paper introduces the Respace Sampling technique, which samples the original time steps sequence to obtain a sampling subsequence. This enables the model to perform sampling using a sequence with shorter timesteps, thereby improving the sampling speed and enhancing the utilization efficiency of the model. Through numerous experimental validations, this technique can increase the image generation speed by approximately 10 to 100 times while ensuring the quality of the generated images, greatly promoting the practical application of the multi-focus image fusion algorithm based on the diffusion model.