GPT Image 2.5 Multi-Turn Edits: Change One Thing, Lock the Rest

Multi-turn editing in GPT Image 2.5 is not “write a longer prompt again.” It is one change per round, paired with an explicit lock list. On Create, explore in Flare and finish in Sunburst—faster than blind rerolls, steadier than adjective stacking.

What multi-turn editing actually solves

The first pass is often “almost right”: subject and camera are fine; you only need a new background, light, or headline. Rewriting the whole prompt invites a full reinterpretation—face, bottle shape, and crop move together.

Goals of a multi-turn loop:

1. Traceable changes — know which turn introduced drift
2. Stable identity — people, products, and approved type stay put
3. Controlled cost — fewer wasted generations in the wrong stage

It complements reference consistency: references lock who / which SKU; change-and-preserve locks what this step is allowed to touch.

Change and preserve must travel as a pair

Public image-prompting guidance separates the requested change from what must remain. On image25.xin that means two lines:

List	What to write	Example
Change	The single allowed delta	Change only the background to a rainy Tokyo street at night
Preserve	Everything that must not move	Face, hair, pose, camera angle, crop, expression, light direction

Weak:

Make the background nicer, keep the person roughly the same

Strong:

Change only the background to a rainy Tokyo street at night.
Preserve exact face, hairstyle, pose, camera angle, crop, expression, and lighting direction.
Do not redraw the person. No extra text, no watermark.

Rule of thumb: “change X” and “Y must not move” appear together. Naming only the change invites a full reshoot.

Four-step loop you can copy

1. Baseline — Build a passable draft in Flare with the four-layer prompt structure; attach references when identity matters.
2. Name this turn’s change — Background or wardrobe or light or camera or one on-image line.
3. Paste the preserve list — Restate subject, crop, light, and text that already passed; do not rely on “same as last.”
4. Inspect before stacking — Did the change land? Are locks intact? Fix this turn before adding another ask.

Keep composition hunting on Flare; move to Sunburst when faces, product edges, or labels must survive consecutive edits (Flare vs Sunburst).

Paste-ready local-edit templates

Swap background (lock the person)

Change only the background to soft morning cafe window light and wooden table.
Preserve the woman's exact face, hair, clothing, pose, camera height, and crop.
Keep natural skin texture. No extra text, no logo.

Swap product scene (lock geometry)

Change only the scene to an oak desk with a laptop slightly out of focus.
Preserve the same matte black earbuds geometry, material, proportions, and logo placement.
Keep the product centered. No readable text invented on the case.

Change one headline only

Follow on-image text rules, then:

Change only the poster headline to exactly once: "CITY LIGHTS".
Preserve layout, subject, lighting, and all other lettering absence.
No subtitle, no watermark, no extra words.

Why one variable per turn is faster

Change weather, wardrobe, and camera together and you cannot tell which clause failed. Split the work:

1. Freeze the subject line + prior output (or the same reference)
2. Swap only background → confirm identity
3. Swap only light → check materials still match
4. Then adjust type or accessories

Each step has its own acceptance check—ideal for series KVs, PDP sets, and poster polish.

Common failure modes

• Naming a background change without locking face, pose, and crop → person redrawn
• Three asks in one round → impossible to debug
• “Keep the style” instead of concrete locks → vibe stays, identity leaves
• Burning Sunburst early without a preserve list
• Needing pixel-identical regions yet stacking more prompt gambles (composite instead)
• Editing type, logo, and background in the same turn → everything drifts

60-second acceptance checklist

After each generation:

1. Was exactly one change requested—and did it land?
2. Are locks still true? (face, geometry, crop, light direction)
3. Any stray type, watermark, or extra limbs?
4. Right model for the stage? (Flare draft / Sunburst finish)
5. Ready for the next turn? If not, fix this round first

How this fits the rest of the site

1. Unstable new frames → prompt structure
2. First edit round not working yet → image-to-image in four steps
3. Can’t recognize the same person/SKU → reference consistency
4. Labels and poster type wobble → on-image text
5. Model routing → Flare vs Sunburst
6. Practice → Create; onboarding → getting started

For public edit patterns, see the OpenAI image prompting guide (site commentary, not an official doc).

Next step

Open Create: Flare a person or product draft, then change only the background with a full preserve list. When that passes, switch to Sunburst for one detail polish. If both turns still read as the same subject, the multi-turn loop is working.