Vision Transformer (ViT)

A Vision Transformer, or ViT, adapts the transformer architecture to images. It divides an image into fixed-size patches, converts each patch into an embedding, adds positional information, and processes the resulting sequence with self-attention. With sufficient data and compute, ViTs can perform strongly on image classification and form the visual backbone of multimodal, detection, and segmentation systems. Unlike convolutional networks, they do not build the same local image assumptions directly into every layer, which can make pretraining especially important. Practical variants use hierarchical features, smaller patches, distillation, or hybrid convolutional components to improve data efficiency, resolution handling, and deployment speed.