GPT Image 2 Prompt Guide (2, 2.5 and 2.6): OpenAI's Own Advice for Exact Text, Edits and Reference Images, With Examples You Can Run (2026)

OpenAI's advice for its image models is to name the subject, what the image is for, the composition and the shape, and for anything complex to organize the prompt as "scene, subject, details, and constraints, using labeled sections". Put required words in quotes with their position and typography, ask for no extra text, and for an edit say "change only X" and list what must stay fixed. In our tests, GPT Image 2.5 Flare set every word of two ads right and made the boldest YouTube thumbnail layout, one of the three most legible.

This guide covers the structure, text in the image, edits, reference images, quality and size, which GPT Image model, what to avoid, and examples you can run.

A note on who wrote this: GenMagic runs every GPT Image model. The quoted advice comes from OpenAI's image prompting guide, written for GPT Image 2.5 and 2.6 (Flare and Sunburst), read November 10, 2026, and from its earlier GPT Image prompting guide in the Cookbook for GPT Image 2; the results come from our own tests, linked where used. Every API route and its price: GPT Image API.

Table of contents

1. The structure
2. Text in the image
3. Edits
4. Reference images
5. Quality and size
6. Which GPT Image model
7. What to avoid
8. Examples you can run
9. Questions

The structure

• Say what the image is for: a poster, a product page, a slide, a thumbnail. The use decides layout and legibility.
• Name the subject, composition, shape and placement: "portrait 2:3, the title at the top, the product centered".
• For complex requests, use labeled sections: Scene: ... Subject: ... Details: ... Constraints: .... OpenAI says prose, JSON-like structures and tags all work, so pick what is easiest to read and change.
• Name materials, light, colors and medium, and ask for "photorealistic" or "real photograph" explicitly when that is the goal.
• For people, say how they are framed, where they look and what they hold.

A template to copy: " [What it is, shape]. Scene: [...]. Subject: [...]. Details: [light, materials, colors]. Text: ["exact words"], [where], [typography]. Constraints: no other text, no logos."

Text in the image

OpenAI's rules, in its words:

• "Put required wording in quotes and describe its position and typography."
• "Spell unusual words or brand names letter by letter when needed."
• "Ask for no extra text, then check spelling and legibility in the output."
• Use medium or high quality for small text, dense information or several fonts.

In our ad test, GPT Image 2.5 Flare set every word of a feed ad and a text-heavy story ad right, and added an arrow to the "Shop now" button that nobody asked for. Every model that set every word right in our tests: images with exact text.

Edits

• Say "change only X" and list what must stay fixed: identity, geometry, layout, lighting, labels.
• State exclusions: no added text, logos or watermarks.
• One change at a time: OpenAI's advice is to "pass the previous output as the next edit input, request one change, and repeat the details to preserve", and it warns that "repeated edits can still change details you intended to preserve".
• For areas that must stay pixel-identical, OpenAI says to composite the approved edit into the original instead of relying on the prompt.

In our headshot test, GPT Image 2.5 Flare kept both people's faces, freckles, glasses and beard and added nothing.

Reference images

• "Identify each input by number and purpose: subject, style, clothing, or background."
• Explain how the inputs combine and what goes where: "Image 1 is the product; place it on the table from Image 2, in the light of Image 3."
• When turning a sketch into a finished image, add "do not add new elements or text" so the model does not reinterpret it.

Quality and size

• Quality: GPT Image 2.6 offers auto, low, medium, high, xhigh and max; OpenAI's advice is to use the top settings "only when they improve an unmet quality requirement", because "a higher setting doesn't guarantee a better result for every prompt".
• Shape: custom sizes in multiples of 16, with an aspect ratio from 1:3 to 3:1 on 2.6.
• Transparent backgrounds: ask for PNG or WebP and check the alpha channel; "a drawn checkerboard is not transparency".
• Image size in our tests: GPT Image 2.5 Flare returned 1536 pixels wide for staged rooms, where most models returned 2048 or more ( virtual staging test).

Which GPT Image model

The studio shows each image's price before you generate, and a refused or failed image costs nothing.

What to avoid

• Mood words instead of specifics for wide, dark or neon scenes: give scale, atmosphere and color.
• Concept-art language for interface mockups, in OpenAI's advice, if the result should look like a shipped product.
• Trusting numbers in generated slides and charts: check every figure.
• Long chains of edits without restating what stays fixed.

Examples you can run

Each link opens the GenMagic studio with the model and the prompt ready; you see the price before you generate.

A poster with exact text (GPT Image 2.5 Flare): "Event poster, portrait 2:3. Scene: a saxophone player silhouetted against a huge orange sun over the sea. Text: the title "JAZZ BY THE SEA" in large elegant serif capitals at the top, and "Saturday, July 18 · 8 PM · Harbor Square" in small letters at the bottom. Warm retro colors. No other text, no logos." Open in the studio

Volume on a budget (GPT Image 2.5 Sunburst): "Flat vector illustration for a blog header, 16:9: a small team of three people building a giant paper rocket together in a bright workshop, limited palette of navy, coral and cream, clean shapes, no text." Open in the studio

An edit that changes only one thing (GPT Image 2.5 Flare): upload a photo of a person at their desk and write "Change only the background: replace the office behind the person with a soft, out-of-focus green park in daylight. Keep the face, hair, glasses, clothing, pose, framing and lighting on the person exactly the same. No added text."

More prompts by what you make: posters, YouTube thumbnails, headshots, logos. GPT Image without ChatGPT Plus: is GPT Image 2 free?

Prompt guides for other image models: Nano Banana, FLUX 3. For video: Seedance, Veo 3.1, Kling 3.0.

Teaching AI video or images on YouTube, in a course or in n8n templates? The creator program pays 15% in cash of what the people you bring pay in their first month.

Questions

Is there an official GPT Image 2 prompt guide? Yes: OpenAI's image prompting guide for GPT Image 2.5 and 2.6 in its API documentation, and its GPT Image prompting guide in the OpenAI Cookbook for GPT Image 2.

How do I get GPT Image to spell text correctly? Put the words in quotes, say where they go and in what typography, spell unusual words letter by letter, ask for no other text, and check the result. Use medium or high quality for small text.

How do I edit an image without changing the rest? Say "change only" the one thing, list everything that must stay fixed, make one change per request, and restate the fixed details each time.

How many reference images can GPT Image use? Several; label each by number and purpose ("Image 1 is the product, Image 2 the background") and say how they combine.

Which GPT Image model should I use? GPT Image 2.5 Flare for text, edits and legible layouts in our tests; GPT Image 2.5 Sunburst for volume at a low price; GPT Image 2.6 for OpenAI's newest settings.