GPT Image 2.5 Image-to-Image: 4 Steps to a Usable Edit

To get GPT Image 2.5 image-to-image right, hold onto one idea: the picture you upload is the structure, the prompt is the change, and a keep list sits between them. Describe only the change and the model will happily rewrite your subject, framing, and lighting along with it. That is what most “the edit destroyed my shot” reports actually are.

Image-to-image and text-to-image are not two products, just two starting points. Text-to-image composes a frame from a sentence. Image-to-image starts from an image and edits on top of it. If a product, sketch, or finished shot has to survive, start from the image. If nothing exists yet, start from words.

Image-to-image vs text-to-image, in one line

Ask a single question: is the frame already decided?

• You already have a photo, sketch, render, or a reference you are matching, and you want a new background, material, light, or style → image-to-image
• You only have an idea, with no framing or elements chosen → text-to-image; do not force an unrelated image through an edit
• You have a finished frame and want to touch one small area (a cup, a label) → still image-to-image, controlled by a “change only this” instruction

In the creation studio, upload a reference and pick the model: Flare is faster for exploring directions round by round, Sunburst is sharper for cleaning up material, edges, and text before delivery. Both accept the same prompt structure.

GPT Image 2.5 image-to-image in four steps

1. Choose the input. Pick a source with a complete, sharp subject and no large watermark. This image caps how much the model can see; no prompt recovers a blurry original.
2. Name the change. One sentence, as specific as you can make it. “Replace the background with an overcast street” is ten times more controllable than “change the background”.
3. Write the keep list. Immediately after: keep the subject shape, logo position, packaging text, camera angle, and shadow direction. This is the brake.
4. Iterate one variable. One change per round. Compare against the source, confirm nothing drifted, then move to the next variable.

Step three is the one people skip, and it decides the outcome. A keep list is not padding; it is the boundary.

Which uploads actually work

The same prompt behaves very differently depending on the input. Check yours against this table:

Input condition	What tends to happen	What to do
Sharp subject filling 40%+ of frame	Structure holds, edits obey	Use as is
Small subject, busy background	Model reframes the shot	Crop first, then upload
Large text or watermarks	Text gets rewritten or smeared	Swap the image, or name the text in the keep list
Overexposed, blurry, heavily compressed	Material and edges get guessed	Use the original file, not a chat-app copy
Collage or several subjects	Only one of them changes	Name the target: “only the cup on the left”

Bigger is not better; cleaner is. A sharp 1K to 2K original is plenty, and the output resolution is a parameter you set separately. For frame and 1K/2K/4K choices, see size and aspect.

Prompt skeleton: change first, keep second

A reusable order that works for most edits:

• Source: what this image is and what the subject is
• Change: the one thing to alter, and what it becomes
• Location: where that thing sits, described from the viewer’s side, not the subject’s left hand
• Material and light: color, texture, and light direction for the new element
• Keep: shape, angle, proportions, logo, text, shadow
• Exclude: no new people, no added text, no added branding

The last two lines are what separate editing from generating. Generating only needs what you want. Editing also needs what you refuse.

Why one variable per round

Editing multiplies rather than adds. Change background and lighting together and a bad result tells you nothing about which one broke it. Split them and each round is diagnosable.

This discipline matters more across multiple rounds: rewrite the keep list every time instead of assuming the model remembers the previous round. The full routine is in multi-turn edits that lock the rest.

Keeping one character or product identical across several images is a different problem: fix one identity reference, then repeat the same identity line. That path is covered in reference consistency.

When an edit goes wrong

Diagnose the failure type before retrying:

• Subject distorted or proportions shifted → the keep list was too soft. Name shape, ratio, and angle explicitly and rerun; do not patch the broken output
• Background changed and the subject changed with it → the change sentence was too broad. Name the subject and add “everything else unchanged”
• Text scrambled → quote the exact copy in the prompt, state size and placement; for text-heavy layouts, leave space and set type later, as in on-image text
• Style drifted → express style through material, light ratio, and lens instead of adjectives like “more premium”
• Getting blurrier each round → you are editing the previous output. Start a new line from the cleanest original

The rule of thumb: when you cannot say what this round actually changed, go back to the last clean source. Repairing a broken frame just carries the error forward.

Image-to-image questions people actually ask

Can I upload several references at once? Yes, but assign a role to each: this one for shape, that one for lighting, the third only for palette. Saying “reference these” splits the model’s attention evenly and matches none of them.

Can an edit output transparency? Yes, state PNG explicitly and spell out edge handling in the keep list. The acceptance routine is in transparent product cutouts.

Does image-to-image keep my aspect ratio? Not automatically. Lock the frame in the parameters rather than mentioning it in the sentence.

Next

Take the image you know best and run one round: change only the background, keep subject, angle, and light. Then split “swap material” and “relight” into rounds two and three. The feel for image-to-image is built in three rounds, not by reading.