#vision-tokens

The Image Is a Token Budget

A multimodal model does not see an image. It reads a sequence of tokens — and the number of tokens an image produces is an engineering decision, not a fixed property of the image. Most engineers discover this only when their context window fills up.