Practical prompt writing

GPT Image 2.5 Prompting Guide

Learn how to structure prompts, preserve important details in edits, guide reference images, and improve one change at a time.

Start with the formula Try GPT Image 2.5

Winter evening billboard refined with a focused GPT Image 2.5 prompt

Edit the uploaded billboard image to show the same scene during a quiet winter evening. Change only the season, weather, and time of day. Add light snowfall, fresh snow on nearby surfaces, cool blue evening light, and warm streetlight reflections. Keep the product, billboard design, exact wording, typography, colors, layout, camera angle, and composition unchanged.

Subject•Composition•Lighting•Constraints

A reusable structure

The GPT Image 2.5 Prompt Formula

Treat the prompt as a short creative brief. Add only the parts your image needs, and keep the most important visual requirements easy to find.

Asset

Name what you are making and where it will be used.

Subject

Describe the main person, product, object, or idea.

Scene

Set the location, action, time, and surrounding details.

Composition

Choose framing, camera angle, hierarchy, and negative space.

Visual direction

Describe lighting, color, materials, texture, and medium.

Text

Put exact copy in quotes and define its placement and style.

Constraints

State what must stay fixed and what must not appear.

Copy this template

Create a [type of asset] for [use]. Show [subject] in [scene]. Use [composition], [lighting], [materials], and [style]. Include "[exact text]" with [placement and frequency]. Preserve [fixed details]. Do not add [exclusions].

Choose the Prompt Template for Your Task

Start with a generation template for a new visual. Use the editing template when an uploaded image already contains details you want to keep.

Text to image

Create a new image

Use this structure when you are starting from an idea rather than an existing image.

Create a [type of asset] featuring [subject] in [setting]. Use [composition], [lighting], and [visual style]. Include the exact text "[copy]" in [location]. Avoid [unwanted elements].

Image editing

Edit an uploaded image

Name the change first, then state which approved details should remain unchanged.

Using the uploaded image, change only [target element]. Keep [identity, pose, composition, lighting, and background] unchanged. Match the existing [style, perspective, and texture].

How to Write Better GPT Image 2.5 Prompts

These rules cover prompt structure, readable text, reference images, focused edits, consistent results, and useful iteration.

Define the result

Start with the asset you need: a product photo, poster, diagram, interface, transparent cutout, or focused edit.

Describe visible details

Use concrete lighting, materials, colors, framing, texture, and actions instead of vague quality words.

Make composition explicit

State subject placement, camera distance, angle, scale, hierarchy, and where open space should remain.

Quote exact text

Place required wording in quotation marks. Say where it belongs, how often it appears, and whether extra text is allowed.

Separate change and preserve

For edits, list the requested change separately from the identity, geometry, layout, lighting, or labels that must stay the same.

Assign reference roles

Identify each reference by number and explain whether it controls the subject, product, clothing, style, layout, or background.

Keep settings separate

Choose model, quality, size, background, and output format in the generator controls instead of hiding them in the prompt.

Revise one thing at a time

Use the approved output as the next input, make one focused adjustment, and repeat critical preservation rules.

GPT Image 2.5 Prompt Examples

Each example demonstrates one prompt pattern. Adapt the subject and details while keeping the structure that makes the request clear.

GPT Image 2.5 Flare

Photorealistic portrait

Describe the subject, framing, light, lens cues, texture, and the retouching you do not want.

View prompt

Create a photorealistic editorial portrait of an elderly sailor repairing a fishing net on a small wooden boat. Show him from the waist up at eye level in soft coastal daylight. Use a 50mm lens, shallow depth of field, realistic skin and fabric texture, subtle film grain, and restrained natural colors. Keep the moment candid and unposed. Do not add text, logos, heavy retouching, or watermarks.

GPT Image 2.5 Sunburst

Process infographic

Name the process, audience, components, direction of flow, and information that must remain readable.

View prompt

Create a polished vertical infographic titled How an Automatic Coffee Machine Works. Show a cutaway machine and explain the process in six numbered steps, from grinding the beans to pouring espresso into the cup. Label the bean hopper, grinder, brewing chamber, water tank, boiler, pump, outlet, and cup. Use clear arrows, readable text, warm cream, dark brown, and muted orange.

GPT Image 2.5 Flare

Text in an ad

Quote approved copy, define its hierarchy, and reject extra words, logos, and watermarks.

View prompt

Create a vertical streetwear campaign advertisement featuring three young adult friends outside a contemporary city art space. Use natural poses, soft late-afternoon light, realistic skin texture, subtle film grain, and premium fashion photography. Add the exact headline "YOURS TO CREATE." once at the top in bold uppercase sans-serif type. Do not include other words, logos, or watermarks.

GPT Image 2.5 Sunburst

Mobile UI preview

Describe the product as a working interface with real sections, spacing, hierarchy, and usable controls.

View prompt

Create a realistic mobile app interface for discovering a local farmers market. Show Market Day, a Portland location selector, search, a featured market, vendor cards, today's specials, hours, and bottom navigation. Use a warm white background, forest green accents, soft terracotta highlights, clean typography, practical spacing, and readable text.

GPT Image 2.5 Sunburst

Identity-preserving edit

Say exactly what changes, then state which identity, pose, background, camera, and lighting details should remain unchanged.

View prompt

Edit the person in image 1 using the clothing references. Replace only the jacket with the navy tailored wool coat from image 2 and the pants with the charcoal wide-leg trousers from image 3. Preserve the person's face, identity, skin tone, hairstyle, expression, body proportions, hands, pose, background, framing, camera position, and lighting.

GPT Image 2.5 Flare

Refine one change at a time

After a strong first result, request one narrow change and keep the approved composition and text.

View prompt

Focused editing

Say What Changes and What Stays

A useful editing prompt has two clear parts. First define the edit target. Then state which approved details should remain stable.

Change

Name one edit target

Change only the jacket from black leather to a navy wool coat with a clean tailored fit.

Preserve

State what should remain

Keep the face, expression, pose, body shape, hands, background, framing, camera angle, and lighting unchanged.

Using multiple reference images

Label inputs by number: "Use image 1 for the person's identity, image 2 for the jacket, and image 3 for the color palette. Do not transfer the background from images 2 or 3."

Choose Flare for faster exploration

Use Flare for early drafts, prompt tests, and quick creative variations when speed matters most.

Choose Sunburst for detailed work

Use Sunburst when you need richer visual detail, closer reference handling, or a more focused image edit.

Prompt for a transparent background

Describe an isolated subject with clean edges and no surrounding scene. Choose a transparent background in the generator and use PNG or WebP when you need alpha transparency.

Turn a weak prompt into a useful brief

Weak

Make a product photo of this bottle.

Better

Create a front-facing studio photo of the uploaded perfume bottle on dark marble. Use soft side lighting and a subtle reflection. Preserve the bottle shape, label, logo, and cap design.

Common GPT Image 2.5 Prompting Mistakes

Most weak results come from unclear priorities, missing preservation rules, or too many changes in one request.

Starting with style only

Give the image a job, subject, and composition before adding mood or style words.

Combining conflicting directions

Choose one clear visual direction instead of mixing incompatible media and lighting.

Leaving edits open-ended

Use "change only" and follow it with a separate list of details to preserve.

Giving references no role

Label each input and state exactly what should transfer from it.

Requesting exact text loosely

Quote the copy, define placement and hierarchy, reject extra text, and check the result.

Changing everything at once

Revise one meaningful variable per turn so you can see what improved or drifted.

Prompt Checklist Before You Generate

The intended asset and audience are clear.
The subject, action, and scene are visible and specific.
Composition, lighting, materials, and style do not conflict.
Required text is quoted with placement and hierarchy.
Each reference image has one clearly stated role.
Edit targets and preservation rules are separated.
Model and output settings are selected in the generator.
Check the result for accurate text, consistent identity, correct layout, and unwanted changes.

Ready to try your prompt?

Put these techniques into practice. Start with a clear description, add references when needed, and refine one detail at a time.

Generate an image

GPT Image 2.5 Prompting FAQ

Quick answers about prompt length, editing, reference images, exact text, and model choice.

Try GPT Image 2.5

How should I structure a GPT Image 2.5 prompt?

Start with the asset, subject, scene, composition, visual direction, exact text, references, and constraints. For an edit, separate what should change from what must stay the same.

Do GPT Image 2.5 prompts need to be long?

No. Use the shortest prompt that makes the result clear. Add labeled sections when the request includes several subjects, references, text requirements, or preservation rules.

How do I write a GPT Image 2.5 editing prompt?

Name the edit target first. Then list the identity, geometry, composition, lighting, text, labels, and surrounding details that must remain unchanged.

How should I use multiple reference images?

Number the references and assign each one a role, such as subject identity, product, clothing, composition, lighting, style, or background.

How do I get readable text in an image?

Put the exact wording in quotes, specify placement and typography, say how many times it should appear, ask for no extra text, and verify spelling after generation.

Should I choose Flare or Sunburst?

Start with Flare when speed matters. Use Sunburst when a demanding generation or edit needs more detail or control, then compare both with the same prompt and settings.