This study investigates the impact of different image augmentation techniques on the performance of an EfficientNet model for skin disease classification. Three augmentation methods, including random rotation, random zoom, and random flip, were applied to generate distinct training datasets, which were then evaluated against a control dataset without augmentation. The results revealed that the random flip technique achieved the highest performance, with an accuracy of 84% and an F1-score of 80.47%, along with superior sensitivity (80.57%) and specificity (95.85%). Random zoom also enhanced model accuracy and robustness, while random rotation showed limited improvement, particularly in sensitivity. This study provides insights into the effectiveness of different image augmentation techniques in enhancing the performance and generalizability of AI models for skin disease classification. The findings emphasize the importance of augmentation in improving model generalizability and performance in clinical applications.