<p>Precision agriculture can be achieved through AI-driven recognition of plant leaf diseases. Still, in practice, recognition accuracy is often limited by insufficient data and a lack of dataset diversity, especially in mango cultivars. To overcome this issue, this work presents a new high-resolution mango leaf disease dataset, MangoLeafDS2025, and proposes LeafDiffusion, a modified latent-diffusion model that generates high-quality plant leaf images, serving as an effective data augmentation framework for enhancing plant leaf disease images. The proposed framework compares traditional data augmentation, StyleGAN3, a CycleGAN-based LeafGAN, and LeafDiffusion using 5-fold cross-validation across various Bayesian-optimised deep learning classifiers. The color-space-sensitive learning was added in the form of RGB, HSV, YCbCr, and L*a*b* color space, with the latter color space yielding modest performance improvement for subset of classifiers. LeafDiffusion showed higher image quality and a wider variety of datasets, as measured by a Fréchet Inception Distance of 8.35, and much higher classification accuracy, with DenseNet-161 achieving the highest overall results. Statistical significance was assessed using paired t-tests and Wilcoxon signed-rank tests, and expert annotation, a visual Turing test, and Grad-CAM analysis were used to verify the realism and interpretability of the generated samples. Altogether, the findings show that the proposed modified latent-diffusion model-based augmentation achieves gains comparable to those of traditional methods and GAN-based approaches. Overall, the proposed LeafDiffusion augmentation offers a scalable, generalizable solution for plant leaf disease detection.</p>