Seeing and reading together

An image is worth 16 × 16 words

Transformers read sequences of tokens. To make one see, turn an image into a sequence: cut it into square patches, flatten each patch into a vector, and pass it through a learned linear layer. That gives one token per patch. Add position embeddings so the model knows where each patch came from, and the rest is exactly the transformer you built. That’s the Vision Transformer (ViT, 2020).