Most llms can equally engage with text in picture form as text in token form. In fact my initial research on this (later corroborated by actual published papers) indicate that this is a cheap way to save on tokens.
Oh interesting and good to know on the token savings with this technique. My test with claude had it use vision and then programmatically test different variable font input variables (mimicking the user scrub interaction) until it was able to OCR it.
I mean I can't know for sure but I'm pretty sure that by the time the upper layers of the network are reached the lower level networks have already transformed the image tiles into proper position encoded embeddings of the tokens in the words in the image.