GPT Image 2.5 prompt — swap the outfit, preserve the person exactly

The longest preservation list in the guide, and every item on it earns its place.

The same woman in the same pose and setting, wearing the jacket, tank top and boots from the references
The original photograph of a woman standing in a museum
InputResult

Also attached

  • A jacket supplied as a clothing reference
  • A tank top supplied as a clothing reference
  • Boots supplied as a clothing reference
Settings the Change what someone is wearing without changing their face prompt ran at
ModeModelQualitySizeRatioCredits
Image to ImageThinkStandard1K2:310 cr

Published by OpenAI as an example for this model — not generated on this site. Model: gpt-image-2.5-sunburst. Source: OpenAI The published example ran at 1024x1536, a legal size we do not currently offer; the frame here is the same shape at our nearest size.

One photograph of a person plus three clothing references, and only the clothes change. The prompt names nine things that must stay fixed — face, features, skin tone, body shape, pose, identity, likeness, expression, hairstyle, proportions — and then nine more it must not add. That is not over-writing. Each clause is closing a specific way this edit is known to drift.

In full

The prompt, nothing cut

Edit the image to dress the woman using the provided clothing images. Do not change her face, facial features, skin tone, body shape, pose, or identity in any way. Preserve her exact likeness, expression, hairstyle, and proportions. Replace only the clothing, fitting the garments naturally to her existing pose and body geometry with realistic fabric behavior. Match lighting, shadows, and color temperature to the original photo so the outfit integrates photorealistically, without looking pasted on. Do not change the background, camera angle, framing, or image quality, and do not add accessories, text, logos, or watermarks.
629 characters

Run it here

This exact prompt, at these exact settings

The mode and all five settings are locked to this template, so there is nothing to line up by hand. Press Run this prompt above and the panel opens filled in — prompt, mode, quality, size and frame — and your first image needs no account. If your plan sits below what this ran at, the line under the panel says so.

PNG, JPEG or WebP · up to 25 MB

Up to 16 images. The first four are included; each one after that costs 3 credits.

0 / 4000
Keep intact — tick anything the model must not touch

We turn your ticks into OpenAI's own "change only X" phrasing, then re-render from your original file — not from the last output.

No card, no waitlist. Credits only start after the free one.

This example was published at Think · Standard · 1K. A free run is Fast · Standard · 1K, so expect less fine detail and softer small text. The composition is the same; the finish is not.

3 levers

What is safe to change

Most libraries publish this list and stop, which is why so many copied prompts come back worse than the original: the swappable nouns are the safe half. The three sections under it are the other half — the clauses that are doing the work, the ones that break it if you touch them, and why it is written in this order.

  • The identity list

    Face, features, skin tone, body shape, pose, identity, likeness, expression, hairstyle, proportions. This block transfers verbatim to any edit where a person must stay themselves — a background swap, a scene change, a lighting change.

  • The integration clause

    "Match lighting, shadows, and color temperature to the original photo so the outfit integrates photorealistically, without looking pasted on." This is what separates an edit from a collage, and it is the sentence people leave out.

  • The garment references

    Three separate images rather than a written description. Describing a jacket gets you a jacket; attaching one gets you that jacket. The endpoint takes up to sixteen references in this mode.

Three ways to break it

Each one is a real failure with a reason attached. A rule without a reason is not usable.

  1. 01Trimming the preservation list

    Every clause maps to something that actually drifts. Drop "skin tone" and it shifts with the new garment's colour cast; drop "proportions" and the body subtly resizes to suit the clothes. It reads as over-writing because it is a list of nine separate lessons.

  2. 02Leaving out the lighting-match clause

    Without it you get the right clothes lit for the reference photograph rather than for this one. That is exactly the pasted-on look, and it is the most common reason this edit gets rejected.

  3. 03Expecting a garment to be reproduced exactly

    Preservation here is aimed at the person, not the product. The clothing will be close, not pixel-exact — pattern placement, seam detail and hardware all move. For catalogue accuracy, composite rather than generate.

Why it is written this way

Inside the Change what someone is wearing without changing their face prompt

When OpenAI announced GPT Image 2.5 the quote it chose to lead with was about the model understanding what *not* to change. This prompt is that claim written out in full, and it is the best available demonstration of how much preservation language a real editing job needs.

The instinct on reading it is that it is over-specified. Nine things preserved, nine things prohibited, for one change. But go through them and each one is a distinct failure that someone has seen. Skin tone shifts because a new garment introduces a colour cast the model balances against. Body shape changes because clothes imply a body and the model reconciles toward the implication. Expression drifts because faces get resampled. Hairstyle changes because hair and collars occupy the same region and the model is redrawing that region anyway. Not one of those clauses is decoration.

The clause worth copying above all others is the integration sentence: match lighting, shadows and colour temperature so the outfit does not look pasted on. Every reference garment arrives with its own lighting attached — it was photographed somewhere, under something. Without an explicit instruction the model has no reason to prefer this photograph's light over the reference's, and the result is the composite look that makes people say an image is obviously edited. They are usually reacting to a lighting mismatch rather than to a geometry error.

The final prohibition list closes the other flank. Background, camera angle, framing, image quality — those are the properties of the photograph rather than of the person, and they drift for a different reason: the model is regenerating the whole frame, so everything in it is up for renegotiation unless pinned. And "do not add accessories" is there because fashion imagery is full of accessories, so a model dressing someone will helpfully supply a bag.

One honest limit. All of this preservation points at the person. The garments themselves are interpreted rather than reproduced — near enough for a styling concept or a look test, not near enough for a catalogue where the product must be exact.

4 questions

Change what someone is wearing without changing their face — common questions

Is the preservation list really that long in practice?
Yes. Each clause corresponds to something that drifts in this specific edit. Trimming it is how people end up with a different-looking person in the right clothes.
Why does my result look pasted on?
Almost always the missing lighting-match clause. The reference garment brings its own light with it, and without an instruction the model has no reason to prefer the original photograph's.
How many clothing references can I attach?
The endpoint takes up to sixteen images in this mode. Three or four is typical — past that the model starts blending rather than assembling.
Will the garment be reproduced exactly?
No. Close, not exact. Pattern placement and hardware move. For product accuracy, composite the real garment rather than generating it.

Written and maintained by Andy SwiftPublished Last updated

Your first GPT Image 2.5 image is free

Generate my first image