A background replacement is two requests wearing one name. One asks the model to draw the picture again with a new setting behind the subject. The other asks it to leave the subject alone and change only what sits behind it. GPT Image 2.5 documents the second one explicitly, and that is the one that keeps a product’s shape, a portrait’s features and a brand’s colours intact.
Everything below comes from OpenAI’s release notes and help centre. Neither page publishes a per-image price or a usage quota.
What changed for background edits
The launch page states the capability in its developer sentence: the Images API “gives users the ability to update a single element—such as a product, background, or piece of copy—while preserving the subject, composition, and brand treatment around it.”
Read that list carefully. A background is one of three examples of a single element, next to a product and a piece of copy. There is no “replace background” button; the capability is general single-element editing.
The same page adds the mechanism: GPT Image 2.5 “is better at editing only what you’ve asked for, while keeping the rest of the details the same—even with more complex subjects and backgrounds.” Complex backgrounds were where a local edit used to leak into the subject.
Two paths for a new background
Decide which path you are on before you write a prompt, because the two fail in opposite ways.
Redraw the whole frame
You describe a complete new picture: subject plus new setting. This is generation wearing the clothes of editing. It suits a generic subject, where you do not care whether the same face comes back, and a scene that has to interact with the subject.
The cost is that the subject is re-invented. Logos drift, hair changes, and a product’s proportions move by a few percent. If any of that is a fixed asset, this is the wrong path.
Edit one region only
You keep the image and change one area. The help centre documents two ways: “Select part of the image with the selection tool, then describe the changes in chat”, or “Skip the selection tool and describe the edit directly in the conversation panel.”
It also states the rule that makes a local edit local: “If you want the edit to apply to a specific area, include that area in your prompt.” Naming the area keeps the change from spreading into the subject.
Set up the reference first
- Start from one reference image and stay in one conversation. The launch page documents that “during longer ChatGPT conversations, Images 2.5 follows specific editing instructions more reliably across multiple edits”. Restarting the chat discards that context.
- Fix the frame before the background. “You can also generate images in any aspect ratio,” chosen with the picker or stated in the prompt. Swapping a background and then changing the crop forces you to check the edit twice.
- Describe the boundary, not only the fill. Say what must not move: the subject’s edge, the light direction, the existing highlight.
Two examples: redraw vs region edit
Both prompts below are complete and copy-pastable, and neither names a real person or a real brand.
Case one: a scene swap
Redraw this image on a new background.
Subject: the same person, head and shoulders, three-quarter view, neutral expression.
New background: a plain concrete wall, even overcast daylight, no props, no signage.
Keep the subject's face, hair shape and clothing colours as in the reference.
Match the light: soft and from the upper left, so the existing highlights still fit.
Square 1:1 frame, subject centred, a 12% clear margin inside every edge.
No text, no logo, no watermark, no second person.This is the redraw path, and its expected failure is drift. Compare the face against the reference before accepting anything. If the likeness moved, switch to the second path.
Case two: a region-only edit
Edit only the background behind the subject.
Keep the subject, the pose and every edge of the subject pixel-identical.
Replace the background with a plain concrete wall, even overcast daylight.
Match the existing light direction from the upper left; do not relight the subject.
Keep the shadow under the chin exactly as it is.
Do not add props, text, signage or a second person.Here the failure runs the other way: the edit bleeds past the boundary. Inspect the outline at full zoom, then judge the whole frame at thumbnail size. If the outline is soft, the fix is a tighter region, not a stronger prompt.
What goes wrong, and how to recover
OpenAI states the limit as plainly as the capability: “Highlights are not always precise, and edits may extend beyond the area you selected.”
- Subject drift on the redraw path. The likeness or the product shape moved. Switch to a region-only edit.
- A soft edge on the local path. The edit crossed the boundary. Re-state the area and re-run.
- Lighting that no longer matches. State the light direction, and keep the existing shadow unless you mean to relight.
- Text you never asked for. Add “no text, no signage, no watermark” instead of erasing the result afterwards.
Frequently asked questions
Can it keep the subject unchanged?
That is the documented strength of the model: it is “better at editing only what you’ve asked for, while keeping the rest of the details the same.” Name the area you are changing.
Do I need the selection tool?
No. Either highlight the area with the selection tool and describe the change in chat, or describe the edit directly in the conversation panel.
Is the subject edge pixel-perfect?
Not guaranteed. OpenAI warns that highlights are not always precise and that edits may extend beyond the selected area.
Does replacing a background cost extra?
No official page publishes a per-image price or a quota. The help centre says only that ChatGPT Images “is available on all tiers”.
A background-swap checklist
Judge the result against these rows, at the size it will be seen.
| Check | What it must show | If it fails |
|---|---|---|
| Subject identity | The same face, shape and brand colours | Switch to a region-only edit |
| Boundary | A clean outline at full zoom | Narrow the area and re-run |
| Lighting | One light direction across both layers | State the direction in the prompt |
| Layout | The frame and margins unchanged | Reset the aspect ratio and re-apply |
Keep the routine repeatable
Pick the path first: a redraw when the subject is generic, a region-only edit when it is a fixed asset. Stay in one conversation, name the area you are changing, and inspect the boundary at full zoom before you judge the whole picture. Two habits finish the job: match the ratio your destination uses, covered by the sizes and ratios guide, and re-read any text the model drew. The image editing guide drills the same loop, transparent backgrounds covers the cut-out destination, and a catalogue is better served by the product photo guide. Before anything ships, commercial use and watermarking covers what the output carries.