Image models went from producing nonsense lettering to producing readable short phrases, and long text still degrades. The reason is what letters are compared with everything else in a picture.
Why letters are a different kind of object
Most image content tolerates variation. A tree drawn slightly differently is still a correct tree, and no viewer checks it against a reference.
Letters do not tolerate variation. A character is either the right shape or it is a different character, or nothing.
The model is generating pixels that resemble training data, which is a good strategy for foliage and a poor one for a symbol system with exact forms.
How the improvement happened
Better text rendering came from training on far more images containing legible text with the text recorded in the caption.
That gave the model an association between requested words and the corresponding letterforms, rather than a general sense that signs contain marks.
Larger models with stronger internal representations of language also carry the spelling more reliably from prompt to picture, which is why the capability tracked general model scale.
Why length is the breaking point
Short text is common in the training data and appears at large sizes: signs, logos, headlines and packaging.
Body text appears smaller, is rarely transcribed in captions, and offers the model no reliable signal about what the individual words are.
Beyond a few words, output degrades into letterlike shapes with plausible rhythm and no content, which is the model reproducing the texture of text rather than text.
What else breaks alongside the letters
Typography involves consistent spacing, alignment and a single typeface applied throughout, all of which are global properties of an image.
Generation works locally, so a word can be spelled correctly while its letters drift in weight, width or baseline across the phrase.
Curved surfaces, perspective and reflections compound this, since each requires the letterforms to distort in a coordinated way that local plausibility does not produce.
How production work routes around it
Design teams generate the image without text and set the type afterwards in a layout tool, which takes minutes and is entirely reliable.
Where text must appear on a surface within the scene, the usual approach is to generate a blank surface and composite the type onto it.
The persistence of that workflow, long after text rendering visibly improved, indicates how little tolerance commercial work has for a spelling error that reaches print.