You do realize that today’s multi-modal LLMs are actually able to understand images? They’re tokenized very much like language is.