GPT Image 2 Reference Editing Prompts for Controlled Changes

Reference editing is most reliable when the prompt defines an edit boundary: what must remain fixed, what may change, and how the result will be judged. OpenAI’s official GPT Image prompting guide describes GPT Image 2 as supporting high-fidelity image inputs, identity-sensitive edits, reliable text rendering, compositing, and production workflows. Those capabilities still benefit from a disciplined brief.

GPT Image 2 availability in Genzy depends on the image providers enabled for the workspace. Use this guide with the reference-editing interface and controls available to you.

Use a fixed-change-review structure

Write three blocks. Fixed: identity, geometry, branding, crop, or approved style. Change: one background, object, color, pose, or layout request. Review: the visible conditions that prove the edit worked.

Example: “Preserve the exact bottle silhouette, cap, cream label, logo, existing text, and camera angle from the reference. Replace only the background with a dark green stone bathroom scene in soft morning light. Keep the label unobstructed, add a realistic grounded shadow, and leave the upper-left area quiet for copy.”

Background replacement prompt

“Use the supplied portrait as the fixed subject. Preserve facial identity, age, hairstyle, skin texture, clothing, pose, and crop. Replace only the studio background with a softly blurred independent bookstore, warm practical lights, and natural depth. Match edge light and color temperature to the new environment. Do not add people, signs, or generated text.”

The prompt includes light matching because a pasted-looking edge can undermine an otherwise accurate edit.

Object removal prompt

“Remove only the red cup from the lower-right corner. Reconstruct the wooden table surface, grain direction, contact shadows, and background behind it. Preserve every other object, the framing, lighting, color, and subject exactly.”

Name the replacement surface. “Remove the cup” explains the deletion but not what should appear in the empty region.

Add one object prompt

“Add a closed navy notebook on the empty table area to the left of the laptop. Match the existing camera perspective, warm window light, shadow softness, and realistic scale. Preserve the person, hands, laptop, desk layout, and all existing text. Do not cover any current object.”

Give the new object a location and physical relationship to approved elements.

Controlled color change prompt

“Preserve the exact jacket construction, seams, fabric texture, folds, lighting, person, pose, and background. Change only the jacket color from beige to deep forest green. Keep natural highlights and shadows consistent with the original fabric.”

When changing color, protect material behavior so the edit does not become a new garment.

Layout expansion prompt

“Extend the existing square product image into a wide 16:9 website hero. Keep the approved product at the same visual size in the right third. Continue the wall, surface, shadow, and light naturally into the new left area. The left 45 percent should remain low detail and suitable for a dark headline. Do not stretch or redesign the product.”

This describes outpainting as a layout task rather than simply asking for more canvas.

Identity-preserving scene change

“Keep the same person from the reference with unchanged facial structure, age, eye spacing, blunt black bob, silver ear cuff, body build, and rust-red jacket. Move the character to a rainy elevated train platform at blue hour. Use a medium shot and cool overhead light. Change only location and lighting; preserve identity, wardrobe, and natural body proportions.”

For larger pose or angle changes, create an approved intermediate view rather than expecting one portrait to define unseen anatomy.

Style transfer without content drift

“Render the supplied street photograph as a restrained gouache editorial illustration. Preserve the exact building arrangement, street direction, number and position of people, vehicle shapes, and overall crop. Use visible matte brush texture, a limited navy, cream, and rust palette, simplified detail, and no new signs or objects.”

Separate content constraints from style attributes. Otherwise the model may treat “make it illustrated” as permission to redesign the scene.

Edit text as a dedicated pass

If a poster headline or product label must change, state the exact old and new text, location, line break, alignment, contrast, and which surrounding design elements remain fixed. Review every character at full size. For legal copy, ingredients, or dense typography, preserve the original artwork or replace it in a design tool after the image edit.

Use transparent output for reusable assets

OpenAI’s guide lists transparent-background output as a GPT Image 2 workflow for reusable objects such as product cutouts, logos, stickers, and presentation graphics. Ask for a clean isolated edge, complete object, no shadow unless needed, and transparent background. Inspect semi-transparent materials, hair, and glass carefully.

Compare the result against the fixed block

Create a checklist from the prompt before generating. Review identity, geometry, text, crop, object count, light direction, edges, shadows, and the requested change. If one detail fails, repeat the fixed block and revise only that failure. Do not introduce a new style during repair.

A controlled GPT Image 2 edit is not a request to “make this better.” It is a small production change with protected inputs, a defined edit region, and measurable acceptance criteria.