To address the issues of data scarcity and class imbalance in crack image datasets, a crack image generation and dataset augmentation method based on low-rank adaptation (LoRA) fine-tuning of stable diffusion models was proposed. By freezing the backbone of stable diffusion and inserting low-rank adaptation matrices into the attention layers of the U-Net model to fine-tune the attention weights, efficient transfer and precise modeling of crack semantic features were achieved. Comparative experiments with mainstream generative models (DCGAN, WGAN-GP, and StyleGAN) and the original stable diffusion demonstrated that the proposed method achieved superior crack clarity, texture fidelity, and background consistency, with significant improvements in multiple image quality metrics. When combined with the DeepCrack dataset for mixed training, the generated images significantly improve the segmentation performance of U-Net, TransUNet, and MobileViT. In particular, on TransUNet, precision, recall, F1-score, and IoU are improved by 5.9, 7.2, 6.4, and 5.6 percentage points, respectively. The proposed method effectively generated crack images with realistic structures and diverse morphologies, enhancing the robustness and generalization ability of segmentation models, and demonstrated strong potential in scenarios with limited data, difficult annotation, and high-risk environments.