A patch is a token, and that is the entire adaptation.

Demonstration

The pipeline, in full:

image  224 × 224 × 3
  ↓  cut into 16 × 16 patches
patches  196 of them, each 16 × 16 × 3
  ↓  flatten → 768 numbers, multiply by a learned matrix
embeddings  196 × d
  ↓  add a positional embedding to each
tokens  196 × d
  ↓  standard transformer encoder
representation

The minimalism is the argument. There is no vision-specific machinery anywhere in it — no kernels, no pooling, no notion of adjacency. If it works, it works because attention over patches suffices, not because image priors were built into the layers.

Note what the patch projection actually is: a convolution with kernel size 16 and stride 16. Every implementation does it that way, which is a quiet admission that one convolution survived.

The notation comes from here and you will type it for the rest of your career. ViT-B/16 is Base-sized with 16 × 16 patches; ViT-L/14 is Large with 14. When a modern model card says So400m or ViT-7B/16, that is this convention still running.

The consequence that reshaped the field is not the accuracy. It is that once an image is a sequence of tokens, it concatenates with text tokens and one model trains over both. Every vision-language model in the Marginalia depends on that single representational choice.

Corollary

The classification token is a convention, not a necessity. ViT prepends a learnable token and reads the answer off it; plenty of later models average the patch tokens instead and do as well. What matters is that the encoder produced one vector per patch — everything downstream is a choice about how to read them.