Skin cancer is a complex ailment that encompasses several types, each requiring distinct treatment approaches. In recent years, machine learning methodologies have been utilized to develop automated systems for dermatological diagnosis, including picture categorization and segmentation approaches. Timely identification of skin cancer might enhance recovery prospects and avert 30–50% of cancer-related deaths. Convolutional neural networks (CNNs) have become prevalent in cancer diagnosis; yet, the Vision Transformer (ViT) encounters challenges such as inadequate model generalization, suboptimal detection accuracy, and insufficient labeled data for training. However, the proposed model indicates that dependence on CNNs is unwarranted and that a pure transformer applied directly to sequences of picture patches may perform remarkably well in image classification tasks. This study establishes a multi-class skin disease classification model utilizing a vision transformer (ViT) that surpasses existing advanced approaches. Furthermore, our model offers comparisons with several current and contemporary models that are widely used. The ISIC 2019 dataset shows that the model obtains an accuracy of 99.45%, precision of 99.45%, F1 score of 99.44%, and recall of 99.44%. The suggested approach employs data augmentation strategies to mitigate imbalances in class, utilizing multi-scale and overlapping sliding windows, as well as multi-scale patch processing, to enhance classification accuracy.