How to Edit Reference Images With GPT Image 2.5

To edit a reference image with GPT Image 2.5, supply a clear source, identify what must stay, and request one main change. Compare the result with the original before using it in a campaign.

ReelsFarm's GPT Image 2.5 option uses Sunburst. Reference edits can change unintended details, so preservation instructions are review criteria rather than a guarantee.

1. Select the Reference for the Details You Need

Choose an image with a clear subject and visible features. If you need the same character, the face should be readable. If you need product accuracy, the product should be visible rather than hidden behind a hand.

Save the original separately. Do not replace it with the first edited output because later trials may need a clean source.

2. Write the Keep Instructions

Name the important fixed details. For a character, that could include facial structure, hairstyle, age range, clothing, and pose. For a product, it could include shape, label layout, cap, color, and count.

Example: “Keep the person, expression, clothing, bottle shape, and label layout from the reference.” Include only details that are genuinely required so the request does not contradict itself.

3. Name One Main Change

Follow with the allowed change: “Replace the room background with a bright kitchen. Keep the original camera angle and framing.”

A scene edit and a wardrobe edit are easier to inspect separately. If you request both, establish which details can change and review each one after generation.

4. Explain the Final Use

A social image may need clear space for copy. A product crop may need the complete package in frame.

Example: “Use a vertical composition. Keep the bottle fully visible in the lower-left area and leave uncluttered space on the right for a short headline.” Then verify that the chosen application can produce the intended layout.

5. Run an Initial Edit

Use the supported reference workflow and quality settings exposed in the application. The provider may offer additional options that are not present in the app.

Inspect the first result before adding further changes. A higher quality setting cannot substitute for a clear source or resolve incompatible instructions.

6. Compare the Result Side by Side

Check the face, product, hands, shadows, framing, and any generated text. Compare protected details directly with the original, not from memory.

If an important detail changed, return to the original and narrow the request. Repeated edits from a drifting result can carry errors forward.

Three Example Edit Briefs

Background change

“Keep the subject, clothing, expression, and pose. Change only the background to a small home office with soft window light.”

Product placement

“Keep the character and room. Place one supplied product on the table beside the subject. Preserve its shape and label. Do not add extra packages.”

Campaign crop

“Preserve the character and product. Adjust the framing to leave clear space for copy. Do not change the expression, package design, or product count.”

These are starting briefs. Use only reference combinations supported by your workflow and inspect the actual outputs.

Save a Reusable Edit Record

Store the source, prompt, model, settings, approved output, and review notes. Use the editing mistakes checklist when a protected detail keeps changing.

FAQ

Can I preserve a character across scene changes?

You can request it through reference editing, but inspect each result for identity drift. Keep a clean source reference for future edits.

Should I ask for several changes in one prompt?

Start with one main change when you need to diagnose consistency. Add further changes after the first edit passes review.