"AI can't spell" is out of date
It was true, it isn't really anymore, and the advice built on it — keep words short, avoid text entirely — is now costing people designs they could have had.
STRENGTHH. Two H's, printed, on somebody's chest.
That failure is real and it still happens. What's changed is how often, and what you should do about it — and most of the advice in circulation is answering the question as it stood two or three years ago.
Why did image models used to fail at text?
Because they were drawing letter-shaped things rather than writing.
Early image models learned what text looks like — the visual texture of letterforms in a particular style — without any representation of words as sequences of characters. The output was convincing-looking type that said nothing, because nothing in the process was tracking spelling.
That produced the era everyone remembers: signs full of plausible gibberish, and advice like keep it under three words or add the text yourself afterwards.
Is it still true?
Largely not, and this is worth stating plainly because the old advice is everywhere.
Current models spell short words reliably and handle longer ones without much trouble. Our own testing found the received wisdom — a maximum of two words, ten letters each — to be well out of date, and that's a limit we'd previously written into our own documentation and had to correct.
Deliberately deformed lettering works too: bootleg-style distortion, where the letterforms are mangled on purpose, comes out well rather than turning into mush. The capability moved and the lore didn't.
So what still goes wrong?
Chance, mostly, and it clusters at the end of words.
A doubled final letter — the STRENGTHH above — is the characteristic surviving failure. So are
occasional dropped letters in longer strings, and inconsistencies when the same word appears twice
in one image. These are stochastic: the same request rendered again usually comes out correct.
That distinction matters more than the failure rate, because it determines the fix.
What's the fix?
Draw it again. Don't edit the prompt.
This is the single most useful thing to know about text failures, and it's counterintuitive. A spelling error is a bad roll, not an instruction problem — so rewriting your request in response to it is treating a random event as a systematic one. You add a clause, the next draw happens to be fine, and now that clause lives in your prompt permanently with a false claim to credit.
Repeat that for a month and you have a long, brittle prompt where most of the lines are superstition. Diagnosing which layer actually failed is the general version of this, and the chance layer's correct response is always another draw.
Should I avoid text on a design, then?
No — that advice was overcorrection when it was current and it's simply wrong now.
Text is one of the most powerful things you can put on a garment, and a great deal of what people actually want to wear is words. A shirt that's one line and nothing else is a legitimate and often better design than an illustration.
What the old constraint did was push people toward pictures they didn't want. That's a real cost, and it was being paid for a limitation that had already lifted.
What's the one thing I do have to do?
Read the words before ordering. Every time.
This is a genuine residue and it isn't going away: no automatic step knows what you meant to
say. A design can be beautifully rendered, correctly composed, perfectly printed and still say
STRENGTHH, and nothing in the pipeline can flag that because nothing else knows the intended
word.
It takes two seconds and it's the only manual check worth insisting on.
Does the style affect it?
Somewhat — heavily deformed lettering is harder to proofread than to produce.
Bootleg distortion, blackletter, chrome with deep bevels and anything with letters interlocking all render fine but make errors harder for you to spot. When the letterforms are wild, read the word letter by letter rather than glancing at its shape.
Which is the reverse of the old worry: the risk isn't that the model can't do it, it's that you won't notice when it didn't.
How do I get one made?
Give the words at JustOG — exactly as they should read — and pick a direction. Read the text on the preview, drag the crop frame, see it composited on the real garment, and it's made to order and shipped.
Designs other people have published are in the shop.
The model can spell. It just can't know that you didn't want two H's.