Graphic Design

AI models render type now. The no-text rule stays

Every prompt in our nightly image pipeline ends with the same sentence: "No text or words in the image." We added it in the era when every model drew letters as melted plastic, and it saved more thumbnails than any style phrase we tried. This autumn the exception arrived. Ideogram shipped version 4.5 on 30 September, and EachLabs' typography roundup, updated this month, puts it at the front of the text-rendering field. A day later Recraft announced Ideogram 4.5 inside Recraft Studio alongside its own V4.1 Flash. The models can draw words now. Our rule survives anyway, and the reason why decides when yours should too.

What actually changed

Text rendering used to fail at the letter level: swapped glyphs, fused ascenders, serif noise where a G belonged. The current generation fails at the word level, and less often. Rangy's logo comparison reports wordmark accuracy in the mid-nineties for short marks on Ideogram's typography mode, and Melies' Ideogram-versus-Recraft breakdown describes the split that formed: Ideogram optimizes the words, Recraft folds typography into a full design system with brand palettes and reusable styles. Two answers to one demand, and the demand is real. Designers have asked for usable type in generated images since the first models shipped.

Mid-nineties accuracy reads like a solved problem until you count. A five-word headline at 95% per-word accuracy fails one time in four. The failure mode moved from every word to one word, and one wrong word in a headline is still a comp in the bin.

Why our no-text rule survives

Our images serve articles. The CMS renders the post title as live HTML on every card, so a generated thumbnail with baked-in type would repeat the headline twice, once editable and once frozen. Generated type cannot be translated, retyped, or reflowed when the layout changes, and it scales badly: words that read fine at 2976 pixels turn to mush on a 400-pixel card. The print lesson from our prepress piece applies twice over: words belong in the layout as live vector type, wherever a layout exists.

The rule costs nothing in our case, because the words were never the point of the image. A person doing work carries the thumbnail, and the title does the talking.

Where baked type wins

The ban was always a default, and defaults change when the tools change. Let the model draw the words when the words are the art: a poster, a logo lockup, a packaging comp, a social card with a quote. Reach for the text-first models there, keep the phrase headline-length, and proof the result letter by letter before it ships. The models earn their reputation on short marks and lose it on paragraphs, and no copywriter signs off on a comp they have not read. Baked type also belongs where no layout layer exists downstream: a gig poster goes to the printer exactly as designed, while a blog thumbnail lives inside a page that owns its own type.

One more boundary stays firm regardless of model: our nightly pipeline renders people doing work, and people plus paragraphs of generated type in one frame is where composition breaks. The text-first models handle a headline, not a newspaper page.

So the sentence stays at the end of every prompt we send tonight, and the exception sits next to it in the toolbox. The rule was never a law. It is a default that matches our output: images that serve words.

Sources

The reported facts in this post come from these outlets:

  • EachLabs ranked the text-and-typography models this month and placed Ideogram 4.5, released 30 September, at the front.
  • Recraft announced Ideogram 4.5 inside Recraft Studio on 1 October alongside its own V4.1 Flash release.
  • Rangy compared the models on logo work and reported mid-nineties wordmark accuracy for Ideogram's typography mode.
  • Melies broke down the Ideogram-versus-Recraft split: words-first optimization versus typography inside a design system.

More in Graphic Design