At presently, Diffusion Model (DM) has achieved state-of-the-art performanceby modeling the image synthesis process through a series of denoising networkapplications. Image Restoration (IR) is to improve the subjective image qual-ity corrupted by various kinds of degradation unlike image synthesis. However,IR for worksite images is greatly challenging in the low-level vision field due tocomplicated environmental factors. To solve this problem, we propose a Sam-pling Improved Diffusion Model for Worksite Image Restoration (SIDM-WIR).It not only improves the authenticity and representation ability for the gener-ation process, but also handles complex backgrounds and diverse object typeseffectively. SIDM-WIR has three main contributions: 1) Its framework adopts aMish-based residual module, which enhances the ability to learn complex patternsof images, and allows for the presence of negative gradients to reduce overfittingrisks during model training. 2) It employs a Mixed-head Self-Attention (MSA)mechanism, which augments the correlation among input elements at each timestep, and maintains a better balance between capturing the global structuralinformation and local detailed textures of the image. 3) It ran in a new customizeddataset named Workplace for WIR specifically. Furthermore, experiments on bothtwo datasets demonstrate that SIDM-WIR achieves better results to previousmethods both quantitatively and qualitatively.