The analysis of rock fragment sizes plays a pivotal role in various stages of geotechnical and mining engineering. Traditional case-based methods often falter in addressing the variability in rock fragment images due to fluctuating geological conditions and imaging environments. To counter this challenge, AI-based models necessitate training on expansive and diverse data sets, encompassing a wide range of conditions and fragment types. However, the creation of such data sets is often hampered by financial and environmental constraints. This paper introduces a novel approach for rock fragment recognition, which leverages the power of AI for data augmentation, automated annotation and prediction. The methodology integrates three key components: Style-Based Generator Architecture for Generative Adversarial Networks (StyleGAN) for data augmentation, the Segment Anything Model (SAM) for automated annotation, and the multiscale version of you only look once (YOLO-MS) real-time instance segmentation model. The proposed method begins by training a new StyleGAN model using the original data set. It then delves into exploring the latent vectors of different rock fragment images to generate specific ‘style representations’. These latent vectors form the foundational elements of each image in a high-dimensional space, encapsulating critical features such as size, shape, texture, and colour variations. The next step involves conducting quality checks on the generated images to ensure their realism and suitability as training data. Subsequently, the augmented image set is annotated using modified SAM module, followed by training the YOLO-MS real-time instance segmentation model. The performance of this model, trained on the augmented data set, is compared with a model trained on the original data set, particularly in terms of generalisation abilities in new scenarios. The results exhibit a significant enhancement in accuracy, robustness, and generalisability of the model trained on the augmented data set, underscoring the potential of StyleGAN as a powerful tool for data augmentation in geotechnical engineering image analysis.