You ask for a new background and the bottle shape shifts. You change one color and the face comes back as a different person. That drift is why most GPT Image 2.5 prompts fail — not because the model cannot draw.
According to OpenAI’s September 8, 2026 launch, ChatGPT Images 2.5 (GPT Image 2.5) is built for sharper detail, more precise editing, stronger reference-subject preservation, and more stable multi-turn changes. On GPT Image 2.5 AI Image Generator, those gains only show up when the prompt names the asset, the change, and the lock list in the same brief.
This gpt image 2.5 prompt guide uses KEEP — Kind, Elements, Exact copy, Preserve — tested as generate, local-edit, and reference workflows on KreadoAI. For offer-led 2.0 statics, keep the GPT Image 2 ad prompt guide. This article is the 2.5 generate-and-edit path.
What GPT Image 2.5 Changes
A GPT Image 2.5 prompt is not a mood board. It is a production brief: what the file is, what the camera sees, what text must print, and what must not move.
OpenAI’s launch post lists the prompting-relevant upgrades: more natural lighting and texture, better preservation of people and products from reference photos, edits that change only the named element, and earlier edits that are more likely to survive later turns. Prompting implication: stop rewriting the whole scene every time. State the delta. Restate the lock.
| Task | Weak prompt habit | GPT Image 2.5 habit |
|---|---|---|
| New still | Style adjectives only | Name the asset, ratio, subject, light, and exclusions |
| Local edit | “Make it better” | Change only X. Preserve Y. Match lighting on the new piece |
| Reference | “Use this photo” | Assign a role: identity, product, or style |
| In-frame type | Hope the model invents copy | Quote the string, count appearances, place the zone |
| Iteration | Paste a new novel each turn | One meaningful change, then restate Preserve if drift starts |
We tested KEEP briefs on KreadoAI the same week as the OpenAI launch: a scale-locked cinematic still, a one-change identity edit, and a poster with quoted type. The model follows the lock list when the lock list is written as a list — not implied.
Why Structured Prompts Win
If your bottleneck is “pretty but unusable,” the prompt is usually missing Kind or Preserve. If your bottleneck is “new face after a background swap,” Preserve is missing identity. If type garbles, Exact copy is missing quotes and a count.
KreadoAI is the run path: pick GPT Image 2.5, paste the KEEP brief or upload a reference, set ratio, generate, then run one local edit before you jump to Smart Canvas or Multi-Size Ad Image Generator. Do not treat ChatGPT’s consumer canvas as the trafficking workflow.
The KEEP Prompt Framework
KEEP is four blocks. Skip a block only when the job does not need it. Never skip Preserve on an edit.
| Block | Write this | Weak fragment | Strong fragment |
|---|---|---|---|
| Kind | Asset type, channel, ratio | “a nice image” | “16:9 cinematic still for a campaign hero” |
| Elements | Subject, scene, materials, light, crop | “premium, cinematic” | “palm-sized astronaut, giant vintage kitchen, lace-curtain sun, copper kettle as architecture” |
| Exact copy | Quoted strings, placement, how many times | “add some text” | Headline text "SUMMER NIGHTS" once, upper third, bold sans |
| Preserve | What must not change | (omitted) | “Keep face, hair, coat, pose, crop, camera” |
Write Kind first. Style adjectives go last, and only if they name a visible treatment (marble, tungsten, grain), not a vibe.
Recommended order on GPT Image 2.5 on KreadoAI:
- Kind + ratio
- Elements (subject, scene, composition)
- Exact copy in quotes
- Constraints / Preserve
- Generate. Then one edit turn with a new Preserve list
How to Prompt Generation
Open image generation, select GPT Image 2.5, leave the KEEP brief in the prompt box, set size and resolution, then generate. Expected result: a first still that already matches ratio and subject placement, so you are not fighting a square crop later.
Template (new image):
Kind: 16:9 cinematic still for a campaign hero. Elements: palm-sized astronaut in a white suit with gold visor on a giant vintage kitchen counter; coffee mug the size of a building; wooden spoon like a mast; copper kettle steaming like weather; sourdough loaf like a hill; lace-curtain afternoon sun; flour dust in the light; camera at counter height. Exact copy: none. Preserve: no extra astronauts, no modern appliances, no watermarks, no on-image captions.
Kind + Elements first. Scale is a subject instruction, not a later crop.
If you need type in the first pass, add Exact copy in the same brief. Do not generate a blank beauty frame and then ask the model to “design an ad” in a second novel — that is how hierarchy collapses.
How to Prompt Precise Edits
Upload the still (or keep the generated frame), then write Change and Preserve as two lists. One change per turn. If a later turn drifts, paste the Preserve list again.
Template (local edit):
Change only the background from an overcast city sidewalk to a rainy neon Tokyo night street. Preserve: face, hair, trench coat, pose, crop, camera angle, expression. Relight the coat and skin to match the new neon, but do not redesign the person. Do not add extra people.
Precision editing is a lock list, not a new description of the person.
Reference-led prompts use the same Preserve logic. Assign each upload a role:
Image 1 = product geometry and label. Image 2 = lighting mood only. Keep Image 1’s bottle shape. Restyle the scene from Image 2. Do not invent a new pack.
If identity must stay pixel-close and the model still drifts after two restated locks, stop prompting and composite in Smart Canvas. Prompting is not a substitute for a flatten when legal needs an identical label.
Ratio and Output Specs
Set ratio in the KreadoAI generator, then repeat the same ratio in Kind so the prompt and the control agree.
| Output | Ratio | Prompt Kind line | Watch |
|---|---|---|---|
| Feed / PDP square | 1:1 | “1:1 product still for Meta feed” | Keep headline out of center product |
| Stories / Reels | 9:16 | “9:16 story ad, CTA in bottom 12%” | Top 12% safe for UI chrome |
| Blog / landscape hero | 16:9 | “16:9 campaign still” | Leave a quiet third for type |
| After export | — | — | Continue in Multi-Size or Smart Canvas |
GPT Image 2.5 does not remove QA. After generate, check spelling, duplicate words, extra logos, and whether the pack still matches the reference. OpenAI’s own launch framing is “more reliable,” not “skip review.”
Copy-Paste Prompt Examples
Three KEEP briefs we ran for this article. Swap the nouns; keep the lock language.
Generate (scale still):
Kind: 16:9 cinematic still for a campaign hero. Elements: palm-sized astronaut, giant vintage kitchen, mug like a building, copper kettle, sourdough hill, lace-curtain sun, counter-height camera. Exact copy: none. Constraints: no extra astronauts, no modern appliances, no watermarks.
Local edit:
Change only the background to a rainy neon Tokyo night street. Preserve face, hair, trench coat, pose, crop, camera. Relight to match neon. Do not redesign the person.
Event poster with type:
Kind: 16:9 event poster still. Elements: tropical rooftop night, infinity pool, palm silhouettes, warm string lights, Miami skyline, teal and gold. Exact copy: headline "SUMMER NIGHTS" once, upper third; date "AUG 22" once; venue "ROOFTOP, MIAMI" once; three details "LIVE DJ SET" / "SUNSET TO LATE" / "OPEN AIR"; no other words. Typography: bold geometric sans, high contrast, no warped letters. Constraints: no extra logos, no QR, no watermarks.
Quote the string, count it, place it. Mood adjectives do not typeset a headline.
Common Prompting Mistakes
These are the failures we keep seeing on first-pass GPT Image 2.5 prompts:
- Style soup — “cinematic, premium, aesthetic” with no Kind. The model invents a poster when you needed a pack shot.
- Edit without Preserve — “change the background” with no lock on face, outfit, or bottle. You get a new person.
- Three changes in one turn — color, camera, and headline together. Drift stacks. Split the turns.
- Unquoted type — the model paraphrases your offer. Put trafficking strings in quotes and say “exactly once.”
- Reference with no role — two uploads and “make it like these.” Say which image owns identity vs style.
- Ignoring ratio — 1:1 generate, then a 9:16 crop that beheads the product. Write Kind to the channel you will traffic.
If GPT Image 2 already ships your 2.0 campaign, do not migrate the brief blindly. 2.5 rewards shorter deltas and longer lock lists. Keep the 2.0 OCAL ads templates for offer-first statics; use KEEP when the job is generate, then local edit, then reference lock.
FAQ
How do I write a GPT Image 2.5 prompt?
Use KEEP. State Kind (asset + ratio), Elements (subject, scene, light), Exact copy in quotes, and Preserve for anything that must not move. Generate on GPT Image 2.5, then run one local edit if needed.
What is the difference between a generate prompt and an edit prompt?
A generate prompt describes the full still. An edit prompt names only the change, then lists Preserve. Rewriting the whole scene on an edit turn is how geometry and faces drift.
Can GPT Image 2.5 keep a product or person from a reference?
Yes, when you assign the reference a role and list the lock: face, outfit, bottle shape, label. Restate Preserve if a later edit drifts. This is the reference path on KreadoAI, not a ChatGPT-only trick.
How do I prompt GPT Image 2.5 image editing?
Write “Change only [X]. Preserve [Y].” Match lighting on the new element. Do not add a second change until the first pass QA’s.
Should I put headlines in the first generate?
If the asset is an ad, yes — Exact copy in the first KEEP brief. If the asset is a pack shot for later layout, leave Exact copy empty and typeset in Smart Canvas.
GPT Image 2.5 vs GPT Image 2 for prompts?
2.5 prompts should separate Change from Preserve and lean on reference roles. 2.0 ad briefs still work with OCAL (offer first). On KreadoAI both models stay selectable; pick 2.5 for tighter local edits and lock-from-reference stills.
Do I need API settings in the prompt?
No. On KreadoAI, ratio and resolution are generator controls. Repeat the ratio in Kind so the prose matches the UI. Do not paste developer parameters into the prompt box.
