The Chrysanthemum (Chrysanthemum morifolium), an economically and aesthetically valuable ornamental plant, is highly vulnerable to leaf diseases that reduce yield and quality. Traditional management approaches often fall short due to unpredictable disease prevalence and limited data availability. While deep learning (DL) models show promise for automated disease detection, their effectiveness is constrained by scarce and imbalanced datasets. To address this challenge, our proposed Diff-AS, a novel diffusion-augmented strategy that leverages Stable Diffusion models fine-tuned with DreamBooth for generating realistic synthetic chrysanthemum leaf images. Our pipeline integrates generative augmentation with an EfficientNetB0-based classification framework, supported by additional loss functions (prior, instance, and L1) to enhance fidelity during denoising. Experimental results demonstrate that incorporating synthetic images with original samples significantly improves performance, achieving 94.88% accuracy, 96 % precision, 91 % recall, and 93% F1-score. These findings highlight the potential of diffusion-driven augmentation to overcome data scarcity and imbalance, offering a scalable and robust solution for plant disease classification in real-world, resource-constrained environments.