<p>The application of object detection models in surface defect detection of industrial products require a large number of labelled, high-quality defect samples for training. However, defect samples of industrial products are rare in many highly automated application scenarios, posing a major challenge to the detection task. To overcome these issues in practice, this paper proposes a fine-grained image generation method based on Stable Diffusion XL Turbo (SDXL Turbo) and a controllable light intensity augmentation method to expand the training set of the object detection model. We created two surface defect detection datasets collected from actual production lines to demonstrate the application of our method. Experimental results show that our method can significantly improve the mean Average Precision (mAP), Recall and Precision of the object detection model for surface defect detection.</p>