When a Picture Has Words In It

Someone attached a picture — a card with Russian text on it — and asked, in Russian, for the same card in Turkish. What came back was two images, both billed, on a different background, with lettering that was not words in any language. The reply described them in Turkish as being "from the link below".

One report, three separate faults. None of them would have produced this alone.

The cheap lane cannot spell

Image tools ask for "medium" quality by default, and medium with no preference for quality resolves down the cost-ordered branch of the router, which lands on the cheapest active image model. That model is a diffusion model. Diffusion models are very good at pictures and famously bad at letters, and worse at letters outside the Latin alphabet.

That is a fine default for "a fox in a forest" and a bad one for a poster, an invitation, a menu or a card, where the text is the deliverable. Requests like that now leave the cheap lane. A prompt is read for the signs of it — quoted wording, lines of capitals, bullet lists, a bounded list of words for the kinds of things that are mostly text, in several alphabets — and the tool also has an explicit flag the model can set when it knows. Negations are stripped first, so "no text" does not trip it. Either signal sends the call to the quality tier. Which tier of picture you get is still your choice; this only decides which model is allowed to try.

A model that could not see the picture

The second fault is worse, because it is the one that produced garbage instead of merely mediocre lettering. The detectors that recognise "make me an image" and "change this image" were written for English and Turkish. A Russian sentence matched neither, so the turn was never marked as one that needed to see, and it went to a text model with no vision. That model called the image tool with the sentence as its prompt. It was drawing from scratch something it had never looked at.

Russian patterns were added for both. More importantly, so was a guard for the phrase that caused this: "the same one", "aynısını", "такую же", "like the one I sent". When a request points at a picture that is attached, forcing a fresh generation is not a near miss. It is wrong, and it is billed.

The tool was describing itself incorrectly

The third fault is embarrassing in the plainest way. The image tool's own description still said that the editing tool was "not yet wired" and that the result came back as a link. Neither had been true for some time. So the model avoided the edit tool it should have used, and then narrated an inline picture as a link.

The descriptions were rewritten — what each tool is actually for, that "the same card in another language" is an edit, that the result is a picture and never a link, that the model cannot see it, and that the description should be in the user's own language. Every image result now carries a hint about how to talk about it.

Three fixes, all pinned by tests that replay the original turns verbatim, including the near misses that must keep going to the cheap lane.