Stone surface defect detection plays a critical role in industrial quality control. Traditional Few-Shot Anomaly Detection (FSAD) methods exhibit limitations in reconstruction fidelity and segmentation accuracy, thereby restricting their broader applications in stone surface inspection. Moreover, the scarcity of high-quality stone surface defect datasets poses significant challenges for research advancement in this domain. To address these issues, we propose Mask-Guided and Semantic-Guided Diffusion-based (MSD) framework, a novel system leveraging multiple representation spaces and advanced guidance mechanisms for precise and efficient detection of stone surface defects. The framework integrates pixel-space, feature-space and latent-space representations, enhanced by two meticulously designed modules: a Mask-Guided Knowledge Distillation network (MGKD) focusing on anomalous regions to improve reconstruction accuracy, and a Semantic-Guided Enhancement Network (SGEN) for preserving semantic fidelity during reconstruction. Additionally, we introduce Stone Defect Dataset (StoneDD), to the best of our knowledge, the first few-shot stone surface defect dataset specifically designed for vision-based defect detection and segmentation. In this study, comprehensive experiments are conducted to evaluate MSD against traditional detection methods. The experimental results demonstrate that MSD achieves superior performance, underlining its potential for future deployment in industrial quality inspection applications.