Deep Learning has greatly improved medical image diagnosis, especially in MRI scans, by enabling accurate classification and early detection. However, training such models well needs a large amount of high-quality medical images. Collecting this data is often difficult due to high costs, privacy concerns, irregular patterns and the low number of cases for rare diseases. Traditional data augmentation methods do not create new data distributions and generative approaches are often either too costly or require large datasets. This study presents DreamDiffGAN, a hybrid framework that combines a fine-tuned Stable Diffusion model (DreamBooth) with a GAN-based discriminator to generate anatomically accurate synthetic MRI images. By merging the generative strength of Stable Diffusion with a discriminator that checks image realism, the model produces high-quality images that closely resemble real MRI scans, even with limited training data. The proposed method increases dataset diversity, lowers the risk of overfitting and improves brain tumour classification using less data. Experiments show that DreamDiffGAN performs better than traditional augmentation and standalone DreamBooth methods. It achieves a classification accuracy of 96.67% and a recall of 93.33% using a ResNet-18 classifier, significantly reducing false negatives, which is vital in clinical use. This framework offers a scalable solution for data-limited medical imaging with potential applications extending beyond MRI to other modalities.