Which Image Model Understands Typography?

One letter, four models: the Ag by Nano Banana 2.1, Muse, GPT Image 2.6 Flare and Sunburst, animated with Gemini Omni Flash.

Can an image model design a typeface? We gave four recent models exactly the same assignment: GPT Image 2.6 Flare and Sunburst by OpenAI, Muse Image by Meta and Nano Banana 2.1 by Google.

How we tested

A professional typographer wrote the master prompt: around 30,000 characters on construction, rhythm, optical correction and spacing. Every model received that text unchanged, with the same brief.

Then three steps. First a sentence, from text alone. That sentence became the reference for the full character set. The character set in turn became the reference for a poster. One attempt per image, nothing regenerated, nothing retouched. We ran three rounds and sharpened the prompt each time.

TextMaster prompt±30,000 characters, by a typographer

TextBriefIdentical for every model

TextInstruction per stepSentence · Character set · Poster

CombinePromptExactly the same text

Image 1SentenceText only. Defines the typeface.

Image 2Character setOwn sentence as reference

Image 3PosterOwn character set as reference

Per model:Nano Banana 2.1Muse ImageGPT Image 2.6 FlareGPT Image 2.6 Sunburst1 attempt per image · 3 rounds

The test structure, identical for all four models.

What we saw

Nano Banana 2.1 and Muse are clearly trained on professional typography. Their specimens look like what a type foundry publishes: careful proportions, calm spacing, credible figures.

OpenAI tries to be creative, but doesn't understand it. Flare and Sunburst reach for striking shapes, such as stencil cuts and flared stems, that read as effects rather than type design.

None of the models uses a reference well. Muse switched from a sans to a classic serif between the sentence and the character set. Other models replaced straight quotes with curly ones or added text nobody asked for. And real creativity was nowhere to be found: the letters are tidy, but rarely surprising.

Conclusion

Nano Banana 2.1 is the best. Not the boldest model, but the most professional and the most reliable. If you want to explore a typeface, test a direction or visualise a specimen, that is where to start.

It does not replace a type designer. A model delivers an image of letters, not a usable font. Curious what AI can do for your brand identity? A first intake is free.

Character-set specimen by Nano Banana 2.1: hero glyphs on the left, uppercase, lowercase, numerals and punctuation on a grid — Nano Banana 2.1: character set.

Poster for NEW FORMS on a concrete wall, set in the typeface by Nano Banana 2.1 — Nano Banana 2.1: poster.

Character-set specimen by Muse Image: hero glyphs on the left, uppercase, lowercase, numerals and punctuation on a grid — Muse Image: character set.

Poster for NEW FORMS on a concrete wall, set in the typeface by Muse Image — Muse Image: poster.

Character-set specimen by GPT Image 2.6 Flare: hero glyphs on the left, uppercase, lowercase, numerals and punctuation on a grid — GPT Image 2.6 Flare: character set.

Poster for NEW FORMS on a concrete wall, set in the typeface by GPT Image 2.6 Flare — GPT Image 2.6 Flare: poster.

Character-set specimen by GPT Image 2.6 Sunburst: hero glyphs on the left, uppercase, lowercase, numerals and punctuation on a grid — GPT Image 2.6 Sunburst: character set.

Poster for NEW FORMS on a concrete wall, set in the typeface by GPT Image 2.6 Sunburst — GPT Image 2.6 Sunburst: poster.

How we tested

What we saw

Conclusion

What do you want to change?