How to Write AI Image Prompts That Render Text Correctly

What actually changed with text rendering

For years, text inside AI-generated images was unreliable. Garbled letters, made-up characters, hybrid hallucinations. This was the running joke of the industry, and the main reason these tools were unusable for branding, posters, or anything with copy.

GPT Image 2 fixed most of it. Independent reviews of the April 21, 2026 launch put text accuracy above 95% across Latin, Chinese, Japanese, Korean, Hindi, Bengali, and Arabic scripts. PetaPixel called it the first AI image model where text rendering "actually works for production." OpenAI's announcement specifically highlighted text rendering as one of the model's three core upgrades.

But "above 95%" still leaves room for errors. And if you don't follow the protocol below, your hit rate drops fast. Here's what works.

Rule 1: Wrap exact copy in quotes

Never write: a poster that says data night

Always write: a poster with the headline "DATA NIGHT" displayed prominently

Quotation marks tell the model: this is the literal string. Not a description. Not a theme. Render exactly these characters.

Rule 2: Specify font style, weight, color, placement

The model has to make four decisions about every text element. Don't leave them to chance:

• Style: sans-serif, serif, display serif, condensed sans, monospace, handwritten, italic
• Weight: thin, regular, bold, black
• Color: name a hex code (#0F1729) or descriptive color (off-white, deep navy)
• Placement: upper third, centered, lower-left corner, vertical along right edge

Example: Subtitle "MARCH 14" in thin sans-serif, off-white, lower third, letter-spacing 0.2em

Rule 3: Use the "verbatim" trigger

For accuracy-critical text — brand names, dates, prices, taglines — append this phrase:

Verbatim text — no extra characters, no substitutions, no duplicate text, no text artifacts.

This works because GPT Image 2's reasoning layer attends to explicit constraints. The phrase has become a common pattern in the prompt engineering community for a reason — it lowers the error rate noticeably on critical text.

Rule 4: Spell out unusual brand names

If your brand name has unusual letter combinations or made-up words, spell it letter by letter on first reference:

Logo "WANDER" (W-A-N-D-E-R) in bold sans-serif, white

Overkill for common words, but lifesaving for brand names the model might "correct" to a real word.

Rule 5: Don't ask for paragraphs

Image models are not document layout engines. Asking for "a paragraph of body text below the headline" produces gibberish that looks like text but isn't.

If you need real paragraph text in a layout, use the right tool — Figma, Canva, or InDesign — and use the AI for the visual elements only.

What still doesn't work reliably

Even with the protocol, three things remain inconsistent:

1. Existing brand logos — the model can't reproduce specific real-world logos (Nike swoosh, etc.) accurately. Composite them in Figma.
2. Long-form text — paragraphs, lists, or any text over ~10 words. Generate the visual, add text after.
3. Mixed languages — mixing scripts (Latin + Hindi + Chinese in one image) often produces errors in one or more scripts, though GPT Image 2 does better at this than any previous model.

Putting it together

Here's a complete text-heavy prompt that follows the full protocol:

Minimalist event poster, 2:3 portrait. Headline "FUTURE STACK" in bold condensed sans-serif, ALL CAPS, off-white (#F7F3EC), upper third. Subtitle "MARCH 14 — 16 • 2026" in thin sans-serif, off-white, below headline, letter-spacing 0.2em. Bottom-right corner: small attribution "Conference 2026" in tiny gray sans-serif. Background: deep navy (#0F1A2E). Generous negative space. Verbatim text — no extra characters, no substitutions, no duplicate text, no text artifacts.

Three text elements, each with its own font weight, color, and placement. Hex codes for color lock. The verbatim trigger at the end.

The shortcut

Manually applying this protocol on every prompt is tedious. Depikt bakes the entire text-rendering protocol into every poster, social graphic, and typography-heavy prompt it generates. Paste your rough idea, get back a prompt with quotes, placements, weights, and the verbatim trigger already in place.