The training procedure for image training involves data augmentation or masking, an interactive element and an attempt to minimize the difference between the clean and corrupted versions of the image. The transformer represents the image as non-overlapping patches to make it easier to mask parts for training purposes.