Prompting cheatsheet

What GPT Image 2 does better than other models, and the prompt patterns that consistently work.

1. Pixel-perfect text rendering

GPT Image 2 is the first widely-available model that nails small text. Quote what you want using double quotes: that reads "BLUE BOTTLE". It works best with 1–5 words. Don't ask for paragraphs.

A coffee shop sign that reads "BLUE BOTTLE",
with a menu board underneath listing
"espresso", "cappuccino", "cortado".

For Chinese / Japanese / Korean, specify the typeface: "in 思源黑体 Heavy".

2. Multi-element composition

Unlike SDXL which collapses on dense prompts, GPT Image 2 handles 200-word prompts that describe scene + subject + lighting + camera + palette in one go.

A 3D isometric illustration of a digital workspace.
A glowing laptop in the center with a 3D speech bubble
coming out of the screen saying "AI IS HERE" in bold,
clean 3D letters. Soft purple-pink gradient background,
floating geometric shapes, low-poly style. 1:1.

3. Native 2K resolution

Don't upscale separately. Ask for it: "8K detail" or"ultra-sharp focus on...". Specify the aspect ratio explicitly:1:1 / 2:3 / 16:9 / 9:16 / 21:9.

4. Cross-image consistency

For series, anchor with consistent character / object descriptors. Reuse exact phrasing across prompts. Specify outfit, hair, distinguishing marks. Don't describe a person as "woman" — say "a woman in her early 30s with shoulder-length wavy auburn hair and a small mole below her left eye".

5. Reasoning before generating

GPT Image 2 has "O-series" reasoning baked in. Give the model the goal, not just keywords:

❌ safety poster
✅ a marketing poster that emphasizes safety in a warm, non-fearful way

Reusable structure

The prompt structure that works for almost everything:

A [STYLE] [SUBJECT] [in/on SETTING],
under [LIGHTING],
[CAMERA ANGLE / DEPTH OF FIELD],
the text "[QUOTED WORDS]" rendered cleanly,
[ASPECT RATIO].

Open the Prompt Lab to remix any of these atoms and run instantly.

Common mistakes

Negative prompts. GPT Image 2 doesn't support them. Don't write "no people, no text". Just describe what you DO want; the model omits the rest.
Style references by artist name. Will trigger filters. Use descriptive adjectives instead: "watercolor children's book illustration" beats"in the style of Beatrix Potter".
Asking for "perfect". Wastes tokens. Specify what perfect means:"ultra-sharp focus on the label, soft bokeh background".

Found a pattern that works? Open a PR with the prompt + a preview image.