ViT

ViT chops an image into fixed-size patches, flattens and projects each patch to a token embedding, adds positional encodings, then processes the token sequence with a standard Transformer encoder. A class token summarizes the sequence for classification, and the same backbone adapts to detection or segmentation through lightweight heads. The appeal is simplicity and scale: no hand-crafted convolutions, strong performance when pretraining on large datasets, and reliable fine-tuning across domains. ViT families vary in depth and width to match hardware budgets, and they train well with modern augmentations and regularization. In practice, ViT is a go-to backbone for vision tasks that benefit from long-range context, clean transfer, and stable optimization.