How to Edit Reference Images with GPT Image 2.5

GPT Image 2.5 reference editing works best when the instruction is precise about two things: what must change and what must remain stable.

Quick Answer

For a controlled reference edit:

1. start with a clear source image
2. name the identity or product elements to protect
3. request one main change
4. describe the new scene or styling
5. keep framing and realism requirements explicit
6. review the result against the source

The goal is not to make a completely new image. The goal is to create a useful variation without losing the approved parts.

What Makes a Good Reference Image

Choose a source image with:

• a clear face when identity matters
• enough light to see important details
• limited visual clutter
• a readable product shape when a product is included
• a composition that can reasonably support the requested edit

A poor source image creates uncertainty before the model sees the instruction. Start with the clearest approved input available.

Use a Protect-and-Change Structure

A reliable edit prompt has two sections.

Protect

State what should remain stable:

• same person and facial identity
• same hairstyle and general age
• same product shape and branding
• same camera angle
• same pose or hand position
• same realistic visual style

Change

State what should be different:

• background
• wardrobe
• lighting
• product placement
• facial expression
• camera crop

Example:

> Keep the same person, facial identity, hairstyle, and natural expression. Change the background to a bright home office with soft window light. Replace the jacket with a neutral blue shirt. Keep the camera angle and mid-torso framing. Make the result look like a realistic creator-style product image.

This is easier to follow than “make this person look like a creator in an office.”

Make One Main Change First

Start with one large edit. For example, change only the background. Then review the result before asking for new wardrobe or product placement.

This staged process gives you better control because you can see where the image changed.

Use a larger batch only after the base edit works. If the first prompt changes identity, wardrobe, lighting, expression, and product position at once, it becomes difficult to diagnose the failure.

Protect Products Explicitly

Product edits need extra care. Mention:

• product type
• shape
• color
• label orientation when it matters
• how the avatar holds or uses it
• whether it should remain the same size

Example:

> Keep the same reusable bottle, including its shape, lid, and blue color. Move it from the right hand to the desk in front of the avatar. Keep the bottle fully visible and do not add extra products.

This instruction defines the product as an approved object instead of a generic prop.

Use Composition Language That Helps Publishing

Reference edits should account for the next use.

Ask for:

• portrait framing for short-form content
• clear space for a hook or caption
• product visibility at a useful scale
• a simple background when the image will carry text
• a consistent crop across a character batch

Good composition reduces work later. You should not need to crop away the product or cover the face with the first caption.

Review Against the Source

Use a short review checklist:

• Is the person still recognizable?
• Did the requested change happen?
• Did any protected feature change unexpectedly?
• Is the product still accurate and visible?
• Does the framing match the intended social format?
• Would this result be safe to save as a reusable character?

Save the output only when it passes the checklist. If identity or product integrity is weak, keep the source image and revise the instruction instead of treating the output as final.

When to Use Quality Levels

Use lower quality while testing whether the requested change is possible. Once the edit direction works, use high quality for the final asset.

This keeps exploration efficient and protects the quality of approved outputs.

Common Reference-Editing Patterns

Scene change

Keep the person and wardrobe. Move the image from a studio to a home office.

Wardrobe change

Keep identity and pose. Change the outfit to match the campaign palette.

Product placement change

Keep the person and background. Move the approved product into a clear hand or desk position.

Campaign variation

Keep the character stable. Create a new scene for a different audience or product angle.

Common Mistakes

Asking for a new scene without protecting identity

The model may change the person more than expected.

Changing several important objects at once

Competing instructions reduce control.

Ignoring the source composition

Some edits require a new crop or a better source image.

Saving the first acceptable result

Review identity, product, and reuse value before saving.

Final Take

GPT Image 2.5 reference editing is a controlled variation process. Protect the approved elements, request one clear change, and review the result against the source before adding it to a reusable workflow.