Non-overlapping patch-wise convolution is the default image tokenizer for all state-of-the-art vision transformer (ViT) models. Even though many ViT variants have been proposed to improve its efficiency and accur