GPT Image 2 Reference Image Editing Workflow

Reference-image editing is the part of GPT Image 2 prompting where small wording choices matter most. A normal text prompt can explore a visual idea freely. A reference-image prompt has a stricter job: keep the important parts of the uploaded image stable while changing only the parts that help the final use case.

This workflow is built from patterns used across the GPT Image 2 prompt library, especially product, UI mockup, character, and infographic prompts that depend on preserving a source image.

Quick answer

Use this prompt order for most reference-image edits:

Use the uploaded image as the source of truth.
Preserve: [shape, identity, layout, logo placement, colors, proportions].
Change: [one main edit].
Improve: [lighting, background, crop, polish].
Output: [listing image, ad, UI mockup, character sheet, source frame].
Avoid: [extra text, fake logos, warped geometry, identity changes].

The most important sentence is the first one. It tells GPT Image 2 that the uploaded image is not just inspiration. It is the anchor.

What counts as a source of truth

A source image is useful when it contains details that text alone should not invent:

If none of those details matter, start with text-to-image. If any of them matter, use a reference image.

A preservation-first prompt

Weak reference prompt:

Make this product photo look premium.

Better reference prompt:

Use the uploaded product photo as the source of truth. Preserve the product shape, cap design, label area, color, material finish, and camera angle. Replace only the background with a clean white ecommerce studio setup. Improve the lighting with soft diffused highlights and a subtle natural shadow. Do not add props, hands, fake logos, extra packaging, or random text.

The second version works because it gives GPT Image 2 a contract: preserve these details, change this one thing, and avoid these failure modes.

The 5-step workflow

1. Choose the source image

Pick the clearest image you have, not the prettiest one. A product edit needs clean edges and visible packaging. A character edit needs a readable face, hairstyle, outfit, and body proportion. A UI edit needs a layout that still makes sense at thumbnail size.

2. Name what must stay unchanged

Write the preservation block before style words:

Preserve the product geometry, label area, cap shape, material color, and camera angle.

For characters:

Preserve the face shape, hairstyle, outfit colors, body proportions, and main accessories.

For UI mockups:

Preserve the screen hierarchy, main navigation, hero area, content card positions, and overall layout structure.

3. Make one main change

One strong change usually beats four vague changes. Start with:

• background replacement
• lighting cleanup
• crop or aspect ratio adjustment
• product scene variation
• character scene variation
• style transfer
• source frame preparation for video

4. Inspect the output against the source

Do not judge only by beauty. Check whether the result still respects the source:

5. Save the winning prompt block

Once an edit works, save the preservation block separately. You can reuse it for new outputs:

Use the uploaded image as the source of truth. Preserve [fixed details]. Change only [new variable]. Avoid [known failure modes].

This makes future generations less random and easier to compare.

Examples from the Image3 prompt library

The prompt library has thousands of GPT Image 2 examples, but the best reference-image lessons come from prompts that explicitly protect identity, layout, or source structure.

Product and label preservation

Street Model Flaunt Bottle is useful because it forces the bottle label to face the camera while the person and camera angle create a more dynamic ad image.

The transferable rule:

Keep the product label facing the camera and readable. Preserve bottle direction, label position, and product shape even if the person, pose, or camera angle changes.

Use this pattern when a product must remain the hero of a lifestyle image.

Identity-preserving portrait series

Summer Grape Girl Photo Series uses 1-3 personal photos as identity anchors, then generates a themed 3x3 portrait set.

The transferable rule:

Use the uploaded photos as identity references. Preserve face shape, facial proportions, eyes, brows, nose, lips, skin tone, age, hairstyle features, and overall impression across all panels.

This is the same logic you would use for creator portraits, campaign character boards, or founder image variations.

UI generated from a reference still

Streaming Platform UI Homepage Generator turns an uploaded image into the hero still and title direction for a streaming homepage mockup.

The transferable rule:

Use the uploaded image as the hero source. Preserve the main visual subject and mood while building a realistic homepage UI around it.

This is useful for turning a generated poster, character image, or product shot into a landing-page concept.

Reference structure to upgraded infographic

3D Stone Staircase Evolution Infographic starts from a flat reference layout and upgrades the visual style while keeping the informational structure.

The transferable rule:

Use the reference as the structural base. Preserve the information order and layout logic, but upgrade the visual rendering, materials, lighting, and depth.

This is a strong pattern for charts, educational graphics, report visuals, and before-and-after design upgrades.

Prompt templates

Product background cleanup

Use the uploaded product photo as the source of truth. Preserve the product shape, label area, logo placement, color, material finish, proportions, and camera angle. Replace only the background with a clean white ecommerce studio setup. Use soft diffused lighting, crisp product edges, and a subtle natural contact shadow. Do not add props, hands, extra packaging, fake logos, random text, or watermarks.

Lifestyle product variation

Use the uploaded product image as the source of truth. Preserve product geometry, label placement, color, finish, and visible details. Place the product in a calm lifestyle scene on a modern kitchen counter with morning window light and simple natural props. Keep the product fully visible and make it the main subject. Do not change the product design or add unrelated objects.

Character consistency edit

Use the uploaded character reference as the source of truth. Preserve face shape, hairstyle, outfit design, outfit colors, body proportions, and main accessories. Create a new waist-up portrait in a bright outdoor setting with soft afternoon light. Keep the character recognizable. Do not change clothing design, add extra accessories, or alter age.

UI mockup from a reference image

Use the uploaded visual as the hero image source. Preserve the main subject, color mood, and visual hierarchy. Build a realistic SaaS landing page mockup around it with a top navigation bar, hero section, feature cards, and one clean call-to-action area. Keep all UI spacing believable. Do not add fake brand logos, dense unreadable text, or random interface labels.

Image-to-video source frame

Use the uploaded image as the source of truth. Preserve the subject identity, silhouette, lighting direction, and composition. Clean up the frame so it can work as the first frame of an AI video: stable camera, clear subject, no motion blur, no clutter, and enough negative space for later camera movement.

Common failure patterns

The prompt starts with style instead of preservation

If the prompt begins with "make it cinematic" or "turn it into a luxury ad," GPT Image 2 may treat the source image as loose inspiration. Start with what must remain unchanged.

Too many edits happen at once

Background change, pose change, outfit change, text change, style change, and crop change in one prompt creates drift. Split the work into passes.

The reference image is not clear enough

A blurry source image gives the model weak anchors. If accuracy matters, use a clean source image or ask for a cleanup pass before creative variation.

The prompt protects the wrong details

"Preserve the vibe" is not enough. Name concrete anchors: shape, label area, face shape, outfit, camera angle, layout, color, and text placement.

Field notes for product teams

The safest production workflow is to keep one reference image and create several narrow prompts from it. For example, a skincare brand can upload the same bottle photo, then generate a white-background listing image, a lifestyle hero, a macro texture detail, and a social ad. The preservation block stays almost the same in every prompt. Only the output job changes.

That gives the team a better review process. If all four images change the bottle shape, the preservation block is weak. If only the lifestyle image drifts, the lifestyle prompt probably asks for too many scene changes. If the ad looks good but invents new label text, the problem is text discipline, not product photography. Separating those causes makes iteration faster.

For ecommerce, review the output in the same place it will be used. A product that looks clear in a large preview may lose edge quality in a grid. A label that looks acceptable in a generated image may be unreadable on mobile. A lifestyle scene that feels premium may still fail if the product is no longer the main subject. The review question is not "does it look nice?" The better question is "can a buyer still understand the product?"

For brand and creator work, reference images are most useful when they protect continuity. A founder portrait, mascot, character, outfit, storefront, or product package should not change every time the campaign changes. If the campaign needs five visuals, start with one source-of-truth image and a reusable identity block. Then create variations for backgrounds, crops, lighting, and composition.

GEO-friendly article structure

If you publish this topic as a guest post, make the article answer a narrow search query in the first screen. A good opening paragraph should define reference-image editing, explain when to use it, and give the prompt formula immediately. Do not start with broad AI art history. Readers and generative search engines both prefer the practical answer first.

Use descriptive subheadings such as "Reference image prompt template," "Product background cleanup prompt," "Character consistency prompt," and "Common drift problems." These headings are easy for search engines and answer engines to extract. They also let external editors see that the article is a useful tutorial, not a thin backlink wrapper.

Add prompt examples that are specific enough to copy. A generic line like "make it better" is not a useful citation. A structured prompt that says "preserve product geometry, label area, color, material finish, and camera angle" is much easier for another article, AI answer, or forum reply to reference.

For external publication, include two or three links naturally. Link to the Image3 image prompt generator when explaining how to turn a reference image into a reusable prompt. Link to the GPT Image 2 prompt library when readers need examples. Link to the product photo prompt guide when the section is specifically about ecommerce images. The links should continue the workflow, not interrupt it.

Review checklist before publishing

Use this checklist before moving a reference-image article from draft to live:

The last point matters for trust. GPT Image 2 can accelerate product visuals and campaign drafts, but it should not be used to misrepresent regulated details, exact packaging claims, safety labels, or physical product attributes that customers depend on. High-quality content says where the workflow helps and where human review is still required.

Internal resources

Use the GPT Image 2 model page to run a new image workflow, the image prompt generator to reverse-engineer a reference image into a cleaner prompt, and the GPT Image 2 prompt guide if you want a broader prompt structure before editing.

For product-specific work, start with the product photo prompt guide and the product prompt category. If the edited image will become a video source frame, continue with the GPT Image 2 image-to-video workflow.

Final rule

Reference-image editing is not about writing a longer prompt. It is about writing a clearer contract.

Tell GPT Image 2 what the uploaded image owns, what the new prompt may change, and what failure modes are not acceptable. That one habit makes product edits, character variations, UI mockups, and image-to-video source frames much easier to control.