Practical guide · Updated September 11, 2026
GPT Image 2.5 Prompt Guide: 15 Prompts, Templates, and Fixes
A good GPT Image 2.5 prompt states the image you need, the visible details, and the limits. For edits, say what to change and what must stay the same. The 15 templates below are built for real product, portrait, poster, reference, and design tasks.
Written by the GPTImage2-5.art Editorial Team
Reviewed against OpenAI documentation on September 11, 2026. We tested five representative workflows in the GPTImage 2.5 workspace. Results may vary between runs.
Key takeaways
A strong GPT Image 2.5 prompt is a short, testable creative brief. State the finished asset, subject and action, composition, visible style and lighting, exact text, and constraints. For edits, separate what should change from what must stay unchanged.
- New image: Define the deliverable, subject, composition, lighting, materials, and exclusions.
- Image edit: Write a “Change only” instruction and a separate “Keep” list for identity, geometry, layout, lighting, and text.
- Reference images: Number each input and state what to transfer—and what not to transfer—from each one.
- Output check: Verify exact text, subject identity or product geometry, and unintended changes before revising one variable at a time.
Use generator controls for model, aspect ratio, resolution, quality, and transparency. Start with Flare for fast, everyday generation. Use Sunburst when output quality or editing precision matters more than speed.
Prompting principles follow the current OpenAI guide. Workflow and QA notes reflect five representative single-run tests completed September 11, 2026. Results can vary between runs.
Quick start
Choose your task
Start with the outcome you need. Each card opens the closest prompt template below; the model and reference notes are practical starting points, not fixed requirements.
Product photo
Ecommerce product photo
Preserve a face
Background replacement
Add exact text
Poster with exact text
Create a transparent asset
Transparent product or icon
Keep a character consistent
Character design sheet
Source check
Official guidance and practical advice
OpenAI released ChatGPT Images 2.5 on September 8, 2026. Its API offers two models. GPT Image 2.5 Flare is the smaller, faster option for most applications. GPT Image 2.5 Sunburst is the base model for workflows with demanding quality requirements. OpenAI also reports better reference preservation and more reliable multi-turn editing than Images 2.0. These are provider claims, not a promise that every prompt will work on the first try.
The official prompting guide gives the practical rule this article follows: state the image you need, then describe the subject, composition, style, and constraints. For edits, separate changes from preserved details. Refine one thing at a time. See the OpenAI image prompting guide and launch announcement.
Prompt
Subject, scene, composition, visible details, exact text, and constraints.
Tool settings
Model, quality, dimensions, output format, and background mode.
Human review
Spelling, identity, geometry, edges, factual accuracy, and rights.
60-second method
A minimal prompt formula for GPT Image 2.5
You do not need special syntax. Use short sentences or labeled sections—whichever is easier to edit. Start with this six-part brief:
Create [DELIVERABLE] for [USE].
Subject: [WHO OR WHAT + VISIBLE DETAILS].
Composition: [FRAMING, ANGLE, POSITION, SPACE].
Look: [LIGHT, COLOR, MATERIAL, MEDIUM].
Text: "[EXACT WORDS]" at [POSITION].
Constraints: keep [INVARIANTS]; no [EXCLUSIONS].For an edit, put the requested change first. Follow it with a specific keep list. For multiple images, write “reference image 1,” “reference image 2,” and so on, then define what each controls.
Prompt library
15 copy-ready GPT Image 2.5 prompts
Replace the bracketed variables and remove any line that does not apply. The settings are starting points, not universal best settings. “Reference image 1” means the first image you upload with the request. “Try this prompt” opens the workspace with the prompt and starting settings filled in; nothing is generated until you review and submit it.
Product and marketing prompts
Use these for ecommerce listings, campaign images, posters, and localized creative.
Choose this group when: Start with Flare for straightforward generation. Test Sunburst when output quality, detailed layout control, or editing precision matters more than speed.
Clean ecommerce product photo
Create a listing image with controlled geometry, materials, and negative space.
Copy-ready template
Create a photorealistic ecommerce product photograph of [PRODUCT].
Composition: one product centered, front three-quarter view, fully visible, with 12% clear margin on every side. Place it on a seamless [BACKGROUND COLOR] studio surface.
Lighting: large softbox from the upper left, soft fill from the right, short natural contact shadow.
Materials: preserve the exact shape of the package and render [MATERIAL] with realistic texture and reflections. Keep the label flat and readable.
Constraints: no hands, props, extra products, added text, watermark, warped packaging, or cropped edges.Replace
[PRODUCT][BACKGROUND COLOR][MATERIAL]Suggested start
Start with a square output and medium quality. Use a product reference image when shape and label accuracy matter.
Why it works and QA
Why it works
The prompt fixes the view, margins, light, and material before adding style. The exclusions protect a listing-ready layout.
Check the output
- Package shape
- Label spelling
- Material texture
- Clean edge
- Contact shadow
Lifestyle product ad with copy space
Build a campaign image that leaves a usable area for later typography.
Copy-ready template
Create a photorealistic lifestyle advertisement for [PRODUCT].
Scene: [ENVIRONMENT] at [TIME OF DAY]. Place the product in the lower-right third, large enough to recognize on a phone screen. Leave the upper-left 40% visually quiet for headline and logo placement.
Lighting and palette: [LIGHTING], using [BRAND COLORS] as restrained accents. Match the product shadow and reflections to the scene.
Preserve the product's proportions, cap, label layout, and brand colors from reference image 1.
Do not add a headline, logo, badge, people, duplicate product, or decorative text.Replace
[PRODUCT][ENVIRONMENT][TIME OF DAY][LIGHTING][BRAND COLORS]Suggested start
Use a landscape output for web ads or portrait for social stories. Add the product photo as reference image 1.
Why it works and QA
Why it works
It treats copy space as part of the composition and tells the model which product details cannot move.
Check the output
- Usable copy space
- Product identity
- Matched shadow
- Mobile readability
- No accidental text
Poster with exact text
Generate a simple poster where a short headline is part of the image.
Copy-ready template
Create a vertical event poster for [EVENT TYPE].
Text, exactly as written, with no extra words:
Headline: "[EXACT HEADLINE]"
Details: "[DATE · TIME · PLACE]"
Layout: headline in the top third, details at the bottom, one clear focal image in the center. Use [TYPE STYLE] typography with clean kerning and strong contrast. Keep a clear reading order at phone size.
Visual direction: [PALETTE], [IMAGE CONCEPT].
Do not repeat, translate, paraphrase, misspell, or add text. No logos or watermark.Replace
[EVENT TYPE][EXACT HEADLINE][DATE · TIME · PLACE][TYPE STYLE][PALETTE][IMAGE CONCEPT]Suggested start
Use portrait orientation and start with medium quality. Keep the text short; finish legal copy and pixel-perfect typesetting in a design tool.
GPT Image 2.5 Flare and Sunburst · 1K · 3:4 · opaque background · one generation per model · September 11, 2026
Flare and Sunburst, same prompt
Both first-pass images rendered “MAKE SPACE” and “OCT 18 · 7 PM · HARBOR HALL” exactly once and without extra copy.
Outputs
What we observed
- Both outputs passed the exact-character and duplicate-text checks in this single run.
- Flare used separate folded-paper shapes; Sunburst organized the artwork into a tighter circular composition.
- This is a two-image comparison, not evidence that either model will always render text correctly.
Why it works, QA, and repair
Why it works
Quoted copy, named positions, and a simple hierarchy reduce ambiguity. The prompt also tells the model what not to do with the words.
Check the output
- Every character
- Text appears once
- Reading order
- Contrast
- Safe margins
Repair prompt
Change only the poster text. Replace the headline with "[CORRECT HEADLINE]" and the details with "[CORRECT DETAILS]". Keep the illustration, colors, spacing, type style, and all other layout elements unchanged. Show each line once. Add no other text.
Localized poster without a redesign
Replace the language while preserving the approved visual layout as closely as possible.
Copy-ready template
Edit reference image 1. Translate only the visible poster copy from [SOURCE LANGUAGE] to [TARGET LANGUAGE].
Use this approved translation exactly:
Headline: "[TRANSLATED HEADLINE]"
Details: "[TRANSLATED DETAILS]"
Keep the illustration, people, product, colors, lighting, borders, logo position, and overall layout unchanged. Preserve the existing visual hierarchy. Adjust line breaks and font size only as needed to fit the translated copy in the same text areas.
Do not translate brand names. Do not add, remove, or move visual elements.Replace
[SOURCE LANGUAGE][TARGET LANGUAGE][TRANSLATED HEADLINE][TRANSLATED DETAILS]Suggested start
Use the approved poster as reference image 1. Keep its aspect ratio and compare medium and high quality for small type.
Why it works and QA
Why it works
Providing the final translation removes a language decision. The keep list defines the edit boundary.
Check the output
- Translation accuracy
- Brand names
- Line breaks
- Unchanged layout
- Leftover source text
Portrait and controlled edit prompts
These prompts separate the requested change from the details that must remain unchanged.
Choose this group when: Use Sunburst for identity-sensitive edits, keep the original aspect ratio when possible, and change one main variable per pass.
Natural editorial portrait
Create a polished portrait without plastic skin or vague cinematic styling.
Copy-ready template
Create a photorealistic editorial portrait of [PERSON DESCRIPTION] for [USE].
Framing: [HEADSHOT / HALF BODY / FULL BODY], camera at eye level, subject looking [GAZE DIRECTION], with [POSE].
Light: soft window light from camera left with gentle falloff. Background: [BACKGROUND], slightly out of focus but recognizable.
Skin and fabric: natural pores, small tonal variation, realistic hair strands, and clear [CLOTHING MATERIAL] texture. Neutral color grade with restrained contrast.
No beauty-filter skin, waxy texture, excessive sharpening, distorted hands, extra fingers, jewelry changes, text, or watermark.Replace
[PERSON DESCRIPTION][USE][FRAMING][GAZE DIRECTION][POSE][BACKGROUND][CLOTHING MATERIAL]Suggested start
Use portrait orientation. Medium is a practical baseline; test higher quality only if fine hair or fabric remains weak.
Why it works and QA
Why it works
Visible skin, pose, and light instructions replace broad words such as 'premium' or 'cinematic.'
Check the output
- Face anatomy
- Skin texture
- Hands
- Hair edges
- Clothing details
Change a background and keep identity
Move a person into a new setting while preserving their appearance.
Copy-ready template
Edit reference image 1. Change only the background to [NEW ENVIRONMENT] at [TIME / WEATHER].
Keep the person's identity, facial structure, skin tone, age, expression, gaze, hairstyle, body shape, pose, clothing, accessories, crop, and camera angle unchanged.
Match the new environment to the existing subject with realistic light direction, color spill, depth of field, and contact shadows. Preserve natural hair edges.
Do not retouch the face, change the outfit, move the person, alter body proportions, add people, or add text.Replace
[NEW ENVIRONMENT][TIME / WEATHER]Suggested start
Use the original portrait as reference image 1. Keep the original aspect ratio and make this a single-variable edit.
GPT Image 2.5 Sunburst image edit · 1K · 9:16 · one generation · September 11, 2026
Background replacement with an identity keep list
The rooftop garden and blue-hour lighting were added while the face, hairstyle, floral qipao, pose, and high camera angle stayed visually close to the reference.
Inputs
Outputs
What we observed
- The requested environment, depth of field, color spill, and practical lights are visible.
- Identity and clothing are recognizable, but this is not a pixel-identical preservation test.
- For identity-critical work, compare facial landmarks and overlay the result with the source before approval.
Why it works, QA, and repair
Why it works
Preserving identity requires more than one constraint. The keep list names the facial, body, clothing, and camera details that matter.
Check the output
- Identity
- Expression
- Pose and crop
- Hair boundary
- Light direction
Repair prompt
The background is acceptable, but the person changed. Restore the face, expression, skin tone, hairstyle, clothing, pose, crop, and body proportions from reference image 1. Keep the current background. Change nothing else.
Change clothing from references
Replace an outfit while protecting the person, pose, and scene.
Copy-ready template
Edit reference image 1. Change only the clothing.
Use reference image 2 for the [TOP GARMENT], including its cut, color, seams, and fabric. Use reference image 3 for the [BOTTOM GARMENT]. Use reference image 4 for the [SHOES OR ACCESSORY].
Fit the garments naturally to the pose. Preserve the person’s identity, face, hair, body shape, hands, stance, background, camera, lighting, and crop from reference image 1.
Keep correct layering at the neck, waist, wrists, hair, and hands. Do not blend garments or add accessories.Replace
[TOP GARMENT][BOTTOM GARMENT][SHOES OR ACCESSORY]Suggested start
Number the person photo and each clothing reference in upload order. Start with one garment if exact control matters.
Why it works, QA, and repair
Why it works
Each input has a clearly defined role in this example, and the prompt calls out the boundaries where clothing edits often fail.
Check the output
- Identity
- Garment match
- Sleeves and waist
- Hands
- Natural fabric folds
Repair prompt
Keep the person and current scene unchanged. Correct only [FAILED GARMENT]. Match reference image [NUMBER] more closely in color, cut, fabric, seams, and length. Repair the boundary at [WRIST / WAIST / NECK / HAIR]. Do not change the face, pose, body, other garments, or background.
Remove or replace one object
Make a local scene edit while retaining composition and lighting.
Copy-ready template
Edit reference image 1. Remove [OBJECT TO REMOVE] from [LOCATION]. Replace it with [NEW OBJECT], positioned in the same physical area.
Match the scene's perspective, scale, focus, light direction, reflections, occlusion, and surface contact. Reconstruct any background that the old object covered.
Keep every person, face, hand, product, piece of furniture, architectural line, color, crop, and camera angle unchanged.
Do not move nearby objects, widen the frame, restyle the scene, or add text.Replace
[OBJECT TO REMOVE][LOCATION][NEW OBJECT]Suggested start
Use the original image as reference image 1. In ChatGPT, a selection can guide the area, but inspect beyond its boundary.
Why it works and QA
Why it works
The prompt explains both the replacement and the physical cues needed to make it belong in the scene.
Check the output
- Old object fully removed
- Background reconstruction
- Scale
- Occlusion
- Unrequested changes
Reference image and consistency prompts
Use numbered inputs and repeat invariant details when you need consistency across images.
Choose this group when: Give each reference a clear role. Start with Sunburst when product geometry, identity, or character continuity must be accurate.
Turn a sketch into a finished image
Keep a rough composition while adding materials, depth, and light.
Copy-ready template
Use reference image 1 as the composition sketch. Create a finished [OUTPUT TYPE] of [SUBJECT / SCENE].
Preserve the sketch's object positions, relative scale, camera angle, horizon, major silhouettes, and negative space.
Resolve the sketch into these visible materials: [MATERIALS]. Set the scene in [ENVIRONMENT] with [LIGHTING] and [PALETTE]. Add realistic depth, surface contact, and shadows.
Do not add new focal objects, change the viewpoint, crop the composition, add text, or copy any signature from the sketch.Replace
[OUTPUT TYPE][SUBJECT / SCENE][MATERIALS][ENVIRONMENT][LIGHTING][PALETTE]Suggested start
Use a clear sketch when composition and perspective must be preserved. Use a mood board when palette, texture, or visual direction matters more. Keep the source aspect ratio when its layout matters.
Why it works and QA
Why it works
It states what the sketch controls and supplies the material and lighting information the sketch cannot contain.
Check the output
- Composition match
- Camera angle
- Object count
- Material logic
- No invented focal object
Combine three reference images
Combine a subject, product, and visual direction in one controlled scene.
Copy-ready template
Create a [OUTPUT TYPE] using three references.
Reference image 1 controls the person's identity, face, hair, and body proportions.
Reference image 2 controls the product's exact shape, label layout, colors, and materials.
Reference image 3 controls only the palette, lighting, and background mood. Do not copy its people or objects.
Scene: [PERSON ACTION AND PRODUCT POSITION]. Composition: [FRAMING AND CAMERA].
Make the contact, gaze, hand position, shadows, and reflections physically believable. Do not merge identities, redesign the product, copy text from reference 3, or add extra objects.Replace
[OUTPUT TYPE][PERSON ACTION AND PRODUCT POSITION][FRAMING AND CAMERA]Suggested start
Upload references in the numbered order used in the prompt. Start with Sunburst when accurate reference matching is the main goal.
GPT Image 2.5 Sunburst image edit · 1K · 3:4 · three references, followed by one repair edit · September 11, 2026
A real reference-role failure and targeted repair
The first output preserved the person and introduced the amber bottle, but it missed most of the woodland palette and lighting. The repair prompt addressed those two problems. The repaired output restored the woodland palette and dappled light while largely preserving the approved subject and product composition.
Inputs
Outputs
What we observed
- The first pass shows why assigning reference roles does not guarantee that every role receives equal weight.
- The repair restored green foliage, depth, dappled light, color spill, and amber reflections. The approved subject and product composition remained consistent.
- The bottle shape remained recognizable, but its exact geometry and original display scale were not preserved perfectly.
Why it works, QA, and repair
Why it works
The role map reduces the risk that the style reference becomes an unintended source of people, objects, or text.
Check the output
- Person identity
- Product geometry
- Correct reference roles
- Hands and contact
- No source leakage
Repair prompt
Correct only the reference-role error. Use reference image [NUMBER] only for [ROLE]. Remove any [PERSON / OBJECT / TEXT / SHAPE] copied from that reference. Preserve the approved identity, product, composition, lighting, and background from the current image.
Consistent character design sheet
Define a reusable character before producing scenes or story frames.
Copy-ready template
Create a clean character design sheet for an original [CHARACTER ROLE].
Invariant design: [AGE RANGE], [BODY PROPORTIONS], [FACE SHAPE], [HAIR], [OUTFIT], [SIGNATURE ACCESSORY], and [COLOR PALETTE].
Layout: four full-body views in one row—front, left profile, back, and right three-quarter view—plus three small expression studies below. Keep the same face, proportions, outfit construction, colors, and accessory in every view.
Use a plain neutral background, even studio light, and generous spacing. No perspective drama, cropped feet, labels, logos, alternate outfits, extra accessories, or watermark.Replace
[CHARACTER ROLE][AGE RANGE][BODY PROPORTIONS][FACE SHAPE][HAIR][OUTFIT][SIGNATURE ACCESSORY][COLOR PALETTE]Suggested start
Use landscape orientation and medium or high quality. Approve this sheet before generating scene variations.
GPT Image 2.5 Sunburst text-to-image · 1K · 16:9 · one generation · September 11, 2026
Character sheet consistency check
The output produced four full-body views and three expression studies with stable hair, jacket, trousers, shoes, pouch, and palette.
Outputs
What we observed
- All four full-body figures are uncropped, and the front, profile, back, and three-quarter views are distinct.
- The orange pouch and teal jacket remain consistent, although the strap position varies between viewpoints.
- The sheet is suitable as a master reference, but downstream scene consistency still needs separate testing.
Why it works and QA
Why it works
The invariant list turns a character idea into a visual specification that can be repeated later.
Check the output
- Same face in every view
- Proportions
- Outfit construction
- Accessory
- Feet fully visible
Four-panel storyboard or anime keyframe
Create a short visual sequence with a clear action and stable character design.
Copy-ready template
Create a four-panel storyboard for [SCENE]. Use reference image 1 as the approved character design.
Panel 1: [ESTABLISHING ACTION].
Panel 2: [SECOND ACTION].
Panel 3: [TURNING POINT].
Panel 4: [FINAL ACTION].
Keep the character's face, hair, body proportions, outfit, accessory, and color palette consistent across all panels. Keep the environment and time of day continuous. Vary camera distance only where specified.
Use clear panel borders and readable visual staging. No captions, speech bubbles, extra characters, costume changes, duplicated limbs, or watermark.Replace
[SCENE][ESTABLISHING ACTION][SECOND ACTION][TURNING POINT][FINAL ACTION]Suggested start
Use an approved character sheet as reference image 1. Generate a wide image or build one panel at a time when consistency matters more than speed.
Why it works and QA
Why it works
Panel-by-panel actions reduce narrative ambiguity, while the invariant list supports continuity.
Check the output
- Character continuity
- Action order
- Environment continuity
- Limbs
- Panel separation
Design and creative prompts
Use these for information design, transparent production assets, and controlled style transfer.
Choose this group when: Define the finished asset first. Verify text and facts manually, and inspect transparent files over both dark and light surfaces.
Readable infographic or presentation slide
Turn supplied facts into a visual with a defined reading order.
Copy-ready template
Create a 16:9 presentation slide titled "[EXACT TITLE]" about [TOPIC].
Use only these facts:
1. [FACT ONE]
2. [FACT TWO]
3. [FACT THREE]
Layout: title at top, one large central diagram, three short supporting labels, and a source note area at bottom. Reading order must be obvious from left to right. Use [PALETTE], simple vector shapes, consistent line weight, and high contrast.
Do not invent data, add statistics, repeat labels, use paragraphs, or add decorative text. Spell every supplied word exactly.Replace
[EXACT TITLE][TOPIC][FACT ONE][FACT TWO][FACT THREE][PALETTE]Suggested start
Use landscape orientation and compare medium with high quality. Verify every fact and finish dense typography in a slide tool.
Why it works and QA
Why it works
Restricting the prompt to supplied facts reduces unsupported additions. The layout establishes hierarchy before visual styling.
Check the output
- Fact accuracy
- Spelling
- Reading order
- Contrast
- Source area
Transparent product or icon asset
Produce an isolated asset with a real alpha channel for later composition.
Copy-ready template
Create a single isolated [ASSET] on a fully transparent background.
View: [ANGLE], centered with generous padding and the complete silhouette visible. Render [MATERIAL / STYLE] with clean details and controlled edge softness.
Keep the alpha edge clean around thin parts, hair, glass, holes, and soft shadows. Include only [ALLOWED SHADOW].
No solid backdrop, scenery, floor, checkerboard pattern, frame, text, logo, watermark, cropped edge, white fringe, or black halo.Replace
[ASSET][ANGLE][MATERIAL / STYLE][ALLOWED SHADOW]Suggested start
In the API, set background to transparent and use PNG or WebP. A checkerboard drawn into the image is not transparency.
GPT Image 2.5 Sunburst image edit · 1K · 1:1 · transparent background · one generation · September 11, 2026
Transparent product extraction with an alpha check
The generated PNG contained an RGBA alpha channel. The stone block, orange panel, floor, and backdrop were removed, leaving the complete bottle and dropper.
Inputs
Outputs
What we observed
- The returned file was inspected as RGBA before conversion to an alpha-preserving WebP.
- The silhouette is complete and the transparent glass remains readable on the dark page surface.
- A faint warm fringe is visible along part of the bottle edge, so production compositing would still benefit from a manual edge pass.
Why it works, QA, and repair
Why it works
The prompt asks for isolation, while the output setting creates the transparent channel. Both are needed.
Check the output
- Real alpha channel
- Thin edges
- Glass or hair
- No halo
- Full silhouette
Repair prompt
Keep the asset's approved shape, material, color, angle, and scale. Remove the visible background and checkerboard. Restore a real transparent alpha channel. Clean the edge around [PROBLEM AREA] without creating a white or black halo. Do not restyle the asset.
Controlled visual style transfer
Apply visual traits from a reference without copying its subject or layout.
Copy-ready template
Create a new image of [NEW SUBJECT] in [NEW SCENE].
Use reference image 1 only for these visual qualities: [PALETTE], [TEXTURE OR MEDIUM], [LINE / SHAPE LANGUAGE], and [LIGHTING]. Do not copy its subject, characters, composition, text, logo, or distinctive objects.
Composition: [FRAMING AND SUBJECT POSITION]. Make the new subject original and clearly separate from the reference content.
No watermark, signature, copied typography, extra text, or recognizable protected character.Replace
[NEW SUBJECT][NEW SCENE][PALETTE][TEXTURE OR MEDIUM][LINE / SHAPE LANGUAGE][LIGHTING][FRAMING AND SUBJECT POSITION]Suggested start
Use the style image as reference image 1. If you need a specific person or product too, assign that input a separate numbered role.
Why it works and QA
Why it works
It extracts visible style properties instead of using a creator's name or asking for a copy of the source image.
Check the output
- New subject is original
- Requested visual traits
- No copied text
- No source composition
- No signature
Adapt the template
How to write a better prompt without making it longer
Name the finished asset first
“Create an image” leaves too many decisions open. Say “ecommerce product photo,” “vertical event poster,” “four-panel storyboard,” or “transparent app icon.” The asset type suggests a useful structure, but you should still state the layout.
Replace mood words with visible evidence
Instead of “make it premium,” describe a dark stone surface, one soft side light, tight reflections, a restrained palette, and generous negative space. Camera terms can help describe appearance, but they do not guarantee a physically exact lens simulation.
Treat text as data
Put the exact copy in quotation marks. State its position, hierarchy, and type style. Ask for no extra words. Then inspect every character. Small legal lines, dense tables, and final brand typography still belong in a design tool.
Give each reference a clear role
Do not say “use these images.” Say which image controls identity, product geometry, clothing, pose, composition, palette, or texture. If a style reference contains a person or headline, explicitly tell the model not to copy them.
Models and settings
Choose Flare, Sunburst, quality, size, and background
If speed is the main requirement, OpenAI recommends starting with Flare. If a complex workflow has demanding quality requirements, start with Sunburst. Use the same prompt, references, dimensions, format, and quality setting for both models. Then compare the outputs against the same checklist.
| Where you work | What you usually control | What not to assume |
|---|---|---|
| ChatGPT Images | Conversational generation, uploads, editing, selections, and product-level controls. | Do not assume the interface exposes the underlying API model or every API parameter. |
| OpenAI API | Model ID, quality, dimensions, background, output format, prompt, and inputs. | Request settings belong in parameters, not only in prompt prose. |
| Third-party tools | Only the models and controls that the product has exposed. | A model label does not prove that every native control is available. |
| Control | Official API options | Practical rule |
|---|---|---|
| Model alias | gpt-image-2.5-flare gpt-image-2.5-sunburst | Use an undated alias to access OpenAI's current version of the model. Recheck important workflows after the alias is updated. |
| Pinned snapshot | gpt-image-2.5-flare-2026-09-08 gpt-image-2.5-sunburst-2026-09-08 | Pin a dated snapshot when you need a stable model version for repeatable testing. Flare and Sunburst remain different models, so compare them with the same review checklist. |
| Quality | auto, low, medium, high, xhigh, max | Start at a clear baseline. A higher setting does not guarantee a better result for every prompt. |
| Size | Common sizes include 1024×1024, 1536×1024, 1024×1536, 2048×2048, 2048×1152, 3840×2160, and 2160×3840. Valid custom sizes are also supported. | Set dimensions in the tool or API. As checked on September 11, 2026, outputs with more than 3,686,400 total pixels—the number in a 2560×1440 image—are experimental. Writing “4K” in the prompt is not a size setting. |
| Background | auto, opaque, transparent | For transparency, use PNG or WebP and inspect the real alpha channel. |
API details can change. Confirm the current values in the official image prompting documentation before shipping a production workflow.
As checked on September 11, 2026, both models use the same API token rates. Text input costs $5 per million tokens, or $1.25 for cached input. Image input costs $8 per million tokens, or $2 for cached input. Image output costs $30 per million tokens. These are token rates, not a fixed price per image. Check the Flare model page or Sunburst model page for current rates.
Failure fixes
What to do when the first image is wrong
Diagnose the visible error. Do not rewrite the entire prompt unless the composition is unusable. Use the last acceptable image as the input for the next edit. Change only one main element.
| Problem | Next instruction | Check |
|---|---|---|
| Identity drift | Restore the face, age, expression, skin tone, hair, and proportions from reference image 1. Keep the approved background. | Compare eyes, jaw, hairline, and age cues side by side. |
| Misspelled text | Change only the text. Quote the corrected copy. Keep layout, artwork, and colors unchanged. | Read every character at 100% zoom. |
| Wrong object count | Keep the current scene. Show exactly [NUMBER] [OBJECTS], each separate and fully visible. Remove all extras. | Count repeated small objects and reflections. |
| Product label warped | Restore package geometry and label layout from reference image 1. Keep the scene, camera, and light. | Inspect corners, cap, logo baseline, and tiny type. |
| Fake transparency | Request an isolated asset and set the background parameter to transparent. Preserve PNG or WebP alpha. | Open over dark and light backgrounds; inspect hair, glass, and shadow. |
| Edit changed too much | Return to the last approved output. Restate the keep list and ask for one smaller edit. | Overlay old and new images when precision matters. |
Know when prompting is no longer the right tool
OpenAI notes that repeated edits can change details you meant to preserve. If a logo, label, face region, or layout must remain pixel-identical, composite the approved edit into the original. The ChatGPT editor selection is guidance, not a hard boundary; edits can extend outside it.
Production workflow
Make a consistent series, not just one good image
- 1
Approve a master reference
Choose one product image, character sheet, or layout that already passes your checks.
- 2
Write an invariant list
Record the face, proportions, colors, packaging, outfit, camera language, and other details that must not drift.
- 3
Change one variable
Change only the scene, pose, size, language, or weather in each subsequent request.
- 4
Repeat critical constraints
Do not rely on 'same as before' for the details that define identity or brand.
- 5
Run the same QA
Check every image for text, geometry, identity, hands, edges, and series consistency.
- 6
Measure accepted output
For API work, track latency, retries, failures, and cost per accepted image—not just cost per request.
Migrating an older prompt?
Save the old model, settings, references, and accepted output as a baseline. Remove repeated adjectives and turn hidden intent into visible requirements. Split edits into change and keep sections, then assign each reference a role. Test the new model with the original prompt first. Otherwise, you will not know whether the model or your rewrite caused the difference.
FAQ
GPT Image 2.5 prompt questions
Do GPT Image 2.5 prompts need to be long?+
No. A short prompt is enough when the task is simple. Add detail when you need to control composition, text, references, or protected elements. Clear requirements matter more than length.
Should I use GPT Image 2.5 Flare or Sunburst?+
OpenAI positions Flare for speed and most applications, and Sunburst for demanding quality requirements. Test both with the same prompt, references, dimensions, and quality setting. Keep the faster option only if it passes your quality checks.
Can I get a transparent background by writing it in the prompt?+
The prompt should request an isolated subject, but API users must also set the background parameter to transparent and return PNG or WebP. Always inspect the decoded alpha channel. A checkerboard inside the image is not transparency.
Does writing 4K in the prompt create a 4K image?+
No. Output dimensions are a product or API setting. In the API, choose a supported size or valid custom resolution. Use the prompt for the image content, not as a substitute for request parameters.
Why did an edit change something I asked it to keep?+
Image editing is probabilistic. Repeat a specific keep list, make one change at a time, and compare against the last approved image. If a region must remain pixel-identical, use compositing instead of relying on prompting alone.
Can I use GPT Image 2.5 output commercially?+
Images created through GPTImage 2.5 may be used personally or commercially, subject to the GPTImage 2.5 Terms of Service, applicable law, and third-party rights. If you use ChatGPT or the OpenAI API directly, the applicable OpenAI consumer or business terms govern instead. Outputs may not be unique or copyrightable, and they may raise third-party-rights issues. Review each output before use.
Evidence
Sources and testing note
This guide uses official documentation for model names, prompting principles, API controls, editing limits, and content terms. We tested five representative workflows with GPT Image 2.5 at 1K on September 11, 2026. The examples include one failed first pass and its separately labeled repair. We converted the original PNG outputs to WebP without changing their dimensions or image content. Each result comes from a single run and does not represent a statistically controlled benchmark or guarantee identical results.
- OpenAI image prompting guide
- Introducing ChatGPT Images 2.5
- Images in ChatGPT help article
- OpenAI Terms of Use — individuals outside the EEA, Switzerland, and UK
- OpenAI Europe Terms of Use — EEA, Switzerland, and UK
- OpenAI Services Agreement for APIs and businesses
- GPTImage 2.5 Terms of Service
GPTImage 2.5 is an independent product. It is not affiliated with, endorsed by, or sponsored by OpenAI. OpenAI and GPT are marks of their respective owners.
Your next image
Start with one clear task and one review checklist.
Copy the closest template, replace its bracketed variables, and remove anything you do not need. You can also explore the GPTImage 2.5 workspace and current model options.