AI image generator with readable text: which Sume models to try

For text inside an image, three families in Sume's image catalog have a vendor statement about it: Nano Banana (Google says its Gemini 3 image models render legible, stylized text), FLUX.2 (Black Forest Labs calls [flex] strongest at text and fine detail), and ChatGPT Image 2.5, whose quality tiers Sume's docs point to for dense text. Whether a given model spells your exact string right is a test you run, not a promise.

Vendor statements are from their own pages and Sume's from the Image API docs and Image 1.0, read 2026-09-29.

What do the vendors say?

How should I prompt for exact text?

Black Forest Labs' guide says to put the exact string in quotation marks, say where it sits, describe the lettering style and use a hex code for a brand color. The same habits cost nothing on any model.

Which quality setting should I use for dense text?

openai/gpt-image-2.5 takes quality of auto, low, medium, high, xhigh or max, and omitted quality defaults to high. Note that auto reserves max, so an auto request reserves the largest amount. A short label can use a lower tier; a menu or packaging panel may need more.

curl -X POST https://api.sume.com/v1/images \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-image-2.5",
    "prompt": "A bakery menu board. Heading: FRESH TODAY. Three lines: Croissant, Baguette, Rye.",
    "quality": "high",
    "aspect_ratio": "3:4"
  }'

How do I check the result?

Read every character. Misspellings are the failure that matters, and they show only when you look. Generate a few with n, keep the correct one, and only then move to production sizes.