How to prompt GPT Image 2 for posters with text

The structure

Every poster prompt that works for GPT Image 2 has six parts, in this order:

1. Format — "Editorial print poster, 2:3 portrait."
2. Headline — exact words in quotes, font style, weight, placement.
3. Image — what fills the frame, one clear subject.
4. Palette — 3–4 named colors.
5. Finish — paper feel, print style, grain, registration marks.
6. Grounding details — small captions, data marks, brand strip.

Skip any one and the model fills the gap with something generic.

A working template

Editorial print poster, 2:3 portrait. Bold sans-serif headline "YOUR HEADLINE HERE" set in condensed grotesk, top-aligned, near-black ink on warm off-white paper stock. Below: a single full-bleed image of [SUBJECT], [LIGHTING], [COMPOSITION]. Restrained palette: ivory, deep teal, near-black, one orange accent. Subtle paper grain, faint registration marks in corners. Bottom strip: small mono caption "[CAPTION]" with a thin data mark. Risograph print feel. High legibility, museum gift-shop quality.

Why each part matters

Format

GPT Image 2 defaults to square. State the ratio explicitly ("2:3 portrait," "3:4 portrait," "1:1 square"). Without it, you'll get a square poster cropped wrong.

Headline

Three rules:

• Quote the exact words. "THE CLOCK IS TICKING" not the clock is ticking.
• Name the font style. "Condensed grotesk," "thin geometric sans," "slab serif."
• Specify placement and weight. "Top-aligned, bold," "bottom-third, light italic."

Keep it under 8 words. Long headlines garble.

Image

One subject. One frame. Be specific about lighting and composition. Compare:

• "a beautiful nature scene" — fills with stock-photo mush
• "a single full-bleed cyanotype-style image of a melting Arctic ice shelf at golden hour, deep teal sea meeting pale sky, one lone polar bear silhouette mid-frame for scale" — renders

Palette

Name 3–4 specific colors. Hex codes work. Vague descriptors ("warm," "moody") don't.

Finish

This is what makes a poster look *printed* rather than *rendered*: "subtle paper grain," "faint registration marks in corners," "risograph print feel," "museum gift-shop quality."

Grounding details

Small captions, data marks, a thin brand strip — these take output from "AI art" to "designed object."

Skip the rewrite

Depikt's generator applies this exact structure automatically. Type "a poster about climate change" and get back a six-part structured prompt ready to paste into ChatGPT. The library has 50+ poster prompts you can copy directly.