GPT Image 2.5 Workflows: Retouch, Backgrounds, Product, Posters

A photo I took last week was badly backlit — the face was a grey smudge against a bright window. I ran it through GPT Image 2.5 expecting the usual: a brighter face, and a stranger's face. What came back was the same person, lit properly, with individual strands of hair still separated.

That is a small thing. It is also the whole argument. The jobs in this article are not new. People have wanted better photos, cleaner product shots and faster poster variants for decades. What changed is that a model can now do them without quietly rewriting the parts you did not ask it to touch — and that moves a whole category of work from "possible in theory" to "worth doing on a deadline".

This is the companion to the full tutorial on prompts, models and settings. Here we skip the parameters and go straight to four jobs, with the prompt patterns for each.

Every workflow below is the same instruction shape:

[what to change] — change only that.
Preserve: [the specific things that must not move]
Keep everything else in the image unchanged.

If you take one thing from this page, take the middle line. A preservation list is what separates an edit from a re-roll. Naming what must stay is more important than describing what should change, because the model is good at the change and historically bad at the restraint.

1. Photo retouching

The highest-volume job, and the one with the lowest barrier. Backlit shots, underexposure, muddy skin tones, cluttered backgrounds, holiday photos that were nearly good, old prints that have faded.

The reason this is worth revisiting now is not that AI can retouch — apps have done that for years — but that it can retouch this one thing without redrawing the face. That was the blocker.

Fixing light:

Correct the backlighting on the subject's face: lift the shadows, recover
skin tone, and keep the bright window exposure natural rather than blown out.
Preserve the subject's exact facial features, hair detail, expression and pose.
Do not change the background, framing, or colour of clothing.

Cleaning a background without touching the subject:

Remove the two people walking in the background on the right.
Fill the area with the same wall and floor already visible in the photo.
Preserve the subject entirely — face, hair, body, clothing, pose, and the
shadow they cast. Keep the camera angle and lighting unchanged.

Restoring an old print:

Restore this scanned photograph: remove creases, dust and colour fading,
recover detail in the faces, and keep the original grain character.
Do not modernise clothing, hairstyles or the background. Do not change
anyone's face. Preserve any handwriting or printed text exactly as it appears.

That last constraint matters more than it sounds. Old photos often have a date or a name written on them, and an unconstrained restore will happily invent new characters.

What to check before you use the result: faces at 100%, fingers, jewellery, and any text in the frame. Those are the four places artefacts show up first.

2. Background swaps

The same subject in a different place. Portraits against a new backdrop, products in a new setting, a headshot that needs a neutral wall instead of a kitchen.

The failure mode here is not a bad background — the model draws those easily. It is that the lighting on the subject stays keyed to the original scene, so the composite reads as pasted-on. The fix is to say so explicitly:

Place the subject against a soft neutral grey studio backdrop.
Preserve the subject's face, hair, clothing, pose and proportions exactly.
Relight the subject to match the new backdrop: soft frontal key, gentle
falloff, no hard shadows. Add a realistic contact shadow beneath them.
Do not change the framing or crop.

Relight the subject to match is the line that turns a cut-out into a photograph.

For a transparent cut-out instead of a new background:

Extract the product from the input image and isolate it on a fully
transparent background. Centred product, crisp silhouette, no halos or
fringing. Preserve product geometry and label legibility exactly.
Do not add a backdrop, checkerboard, scenery or shadow.

Two notes on transparent output. First, do not describe a scene anywhere in the prompt — any scene description will override the transparency and you will get a background you did not ask for. Second, if you keep editing after this, repeat the transparency requirement every single round; it drops out otherwise.

The transparent output in this generation has visibly less white fringing than before, which is what makes cut-outs usable without a manual cleanup pass.

3. Product shots and listing sets

An e-commerce listing needs the same product in six situations. Previously the constraint was never "can the model draw a nice scene" — it was that the product itself deformed, the logo turned to mush, the label text scrambled, or the material stopped looking like the material.

The instruction pattern is four parts: preserve, adjust, compose, exclude.

Using the uploaded product photo as the only reference, produce a listing
image at 4:5.

Preserve: the product's shape, colour, model, logo, buttons and port
positions exactly as they appear in the reference.
Adjust: background to warm off-white, product on a pale wooden surface,
light from upper left, soft shadows.
Compose: product occupies about two thirds of the frame, leave headline
space at the top.
Do not add: accessories, feature callouts, or promotional text that does
not exist in the original.

That last line is the one people forget, and it is why listing images come back with invented badges and made-up feature labels.

For a series, change one variable per round. Lock the layout first, then vary only the setting:

Same product, same framing, same lighting. Change only the surface to
dark slate and the background to a deep neutral grey.

Run that pattern five times and you have a coherent set instead of five unrelated pictures.

4. Poster and layout work

This is where text rendering matters, and where the honest position differs from the marketing one. Chinese and other non-Latin layouts already rendered well in the previous generation. What is new is that editing text inside a dense layout no longer disturbs the rest of it — you can change one headline in a poster carrying hundreds of characters and everything else stays put.

Write every required string explicitly, one per line:

Event poster, vertical 4:5, clean editorial layout.
Text (exact, verbatim, each on its own line):
"SPRING SESSIONS"
"Saturday 14 March · 7pm"
"Doors at 6:30 · Limited seating"
Fine print, smaller, at the bottom: "Programme subject to change"
Typography: bold sans-serif headline, high contrast, generous margins.
Ensure each line appears once and is fully legible.
No watermarks, no logos, no additional text.

Ensure each line appears once prevents the duplicated-headline failure. No additional text prevents the model from inventing plausible-looking filler.

Editing a finished poster:

Change the headline from "SPRING SESSIONS" to "SPRING SESSIONS 2027".
Keep the typeface, size, colour, position and every other text block
exactly as they are.

Before it goes to print: read the small type character by character. Headlines come out reliably; fine print is where errors still hide.

Two more things it turns out to be good at

Sprite sheets and simple animation. Ask for a numbered sequence and you get frames that hold their character across the set:

Create a sprite sheet of [character] performing [action], 12 frames,
left to right, consistent character design and lighting across every frame.
Plain background. No text.

Stitch those into a GIF and you have a short loop. The same trick works for material studies — one logo rendered in eight finishes, assembled into a switching animation.

Batch layouts from a single reference. You can hand over one product photo and ask for a whole identity system derived from it. The critical instruction is the output format, because the default failure is one collage instead of eight separate files:

Generate 8 separate 16:9 images, output individually.
Do not combine them into a grid, contact sheet, collage or presentation board.
Each image is one complete piece.

Without that block you will get a nine-panel grid every time.

Style consistency across a batch

Once a look works, you can carry it forward. State the lock explicitly rather than relying on the model to infer it:

Keep this exact style — same palette, same line weight, same lighting,
same level of detail. Now produce the same treatment for [next subject].

This holds up well for illustration and icon sets. It holds up less well for a specific character across separate images; independent testing found character consistency improved less than the other areas, so a recognisable person or mascot across four panels will still need the full description restated each time and some outputs discarded.

5. Personal styling and wardrobe

This job did not exist as a product because the service cost too much. Hiring someone to design a few outfits around your age, job, build and the rooms you actually walk into is not something most people will pay for. Once a model can hold a real person's face and proportions steady, the economics change.

The lock line does the work:

Keep my exact face, likeness and identity unchanged: do not alter features,
age, or skin tone.

Restyling without moving the person:

Restyle my hair into a voluminous retro look and my clothing into bold
period fashion, while keeping my exact pose and background.
Keep my exact face, likeness and identity unchanged: do not alter features,
age, or skin tone. Saturated palette, soft glow, period grain. Output 4:5.

Outfits for a specific occasion:

Using the uploaded photo as a face reference, create a realistic
three-quarter portrait. Occasion: [business meeting / casual dinner /
conference talk]. Season: [autumn].
Clothing, hair and overall styling should suit that occasion and read as
appropriate for my age and build.
Keep my facial features recognisably mine.
No costume-like exaggeration — real, wearable combinations.

One technique matters more than people expect on period looks: state the country and city. Without a geographic constraint the model defaults to an American version of any decade. "Shanghai, 1987" and "the 1980s" produce different photographs.

Worth saying plainly: a convincing period portrait can be mistaken for a genuine archival photo. Do not present one as a real record.

6. Ad variants from one winning image

When an ad is performing, the team does not want a hundred different images. They want the product and the subject held constant while the background, scene or visual direction changes, so they can test thirty variants against each other.

Two-round structure. Lock the layout and copy first:

Create a realistic billboard mockup of the product in a highway scene at sunset.
Billboard text (exact, verbatim, no extra characters): "Fresh and clean"
Typography: bold sans-serif, high contrast, centred, clean kerning.
Ensure text appears once and is perfectly legible. No watermarks, no logos.

Then change one condition per round:

Make it a winter evening with snowfall.

The billboard layout, the copy and the product all survive the weather change. That is the whole workflow — round one sets the format, every round after moves exactly one variable.

The older way of working was closer to "draw a new one each time". This is closer to "keep working on the asset you have", and that difference is what decides whether it fits into a real campaign pipeline.

How to tell a good result from a nearly-good one

The four jobs above share a check list, and running it takes about thirty seconds:

Faces at 100%. Eyes, teeth, the hairline. Artefacts show here first.
Hands and fingers. Count them.
Text, character by character. Headlines are reliable; fine print is not.
Product geometry. Compare against the reference — is the silhouette identical, or has it subtly softened?
Edges on cut-outs. Drop the transparent version onto a dark background; fringing that is invisible on white shows up instantly.
The thing you asked to preserve. This is the one people skip. Open the original next to the result and check the preservation list you wrote.

If a result fails on point six, do not iterate on it — go back a step and rewrite the constraint. Iterating on a drifted image compounds the drift.

Where this does not help

Worth saying plainly, because the four sections above are all upside:

Perfect region isolation. Some pixels outside your marked area still shift. It is much better, not absolute.
Long edit chains on faces. Artefacts accumulate. If you are eight rounds deep on a portrait, compare against the original rather than against round seven.
Fine print. Still fallible. Proof it.
A character across many images. The weakest of the consistency improvements.

None of these makes the workflows unusable. They just mean you should keep a check step, which is the difference between a tool and a toy.

FAQ

Do I need a ChatGPT subscription for any of this? Not for the workflows on this page. This site runs GPT Image 2.5 directly with no account required to start. A ChatGPT account is only needed for the in-app features — Sketch, Templates and Comments.

Which model should each job use? Flare for everything while you explore. Switch to Sunburst for the final pass of retouching and product shots, where fine detail matters. Poster and layout work benefits from Sunburst earlier, because text rendering is the thing you least want to re-roll.

Can I use the results commercially? Commercial use follows OpenAI's terms for the underlying model. Note that output carries C2PA metadata and an invisible watermark identifying it as AI generated, which some clients and platforms detect.

Why did my transparent background come back with a scene in it? Almost always because the prompt described a scene somewhere. Any scene description overrides the transparency setting. Strip every environmental adjective and try again.

How many reference images can I use? Up to 16 in one request. Assign each an explicit job or the model will average them.

My edit changed something I told it to keep. Rewrite the preservation list to name that specific thing, and add the catch-all line. Region isolation is much better in this generation but not absolute — naming the thing explicitly is what fixes it.

Getting access

You do not need a ChatGPT subscription or an API key to try any of the above — the generator on this site runs GPT Image 2.5 directly and the first render is free. If you would rather go deeper on the controls first, the tutorial covers prompts, model choice, quality tiers and troubleshooting.

image25.online is an independent third-party platform and is not affiliated with OpenAI. Site credits are separate from any official account.