<p>Flat tube, as a critical component of micro-channel heat exchanger, can significantly compromise the sealing integrity and thermal management system performance when surface defects occur. However, the low occurrence frequency of defective samples creates a data scarcity challenge for data-driven deep learning detection methods. To address this issue, this article proposes an improved latent diffusion model (LDM) network for generating diverse defective samples, thereby enabling effective data augmentation. First, an improved LDM is proposed which is capable of synthesizing high-fidelity defect images from noise using only mask-guided conditioning. Second, an adjustable intensity coefficient mechanism is incorporated to precisely control defect severity, further enhancing sample diversity. Experimental results demonstrate the superiority of the proposed method in low-data regimes. Compared to existing methods, our generated samples exhibit higher visual quality and structural fidelity. When used for data augmentation, the proposed method boosts defect classification accuracy by 29.5% for VGG and 14.72% for Swin-T models, respectively.</p>