GPT Image 2 Brand Style Reference Workflow

Brand consistency is not the same as asking GPT Image 2 to “make it look on brand.” That sentence describes an intention, but it does not define a visual system. A useful system tells the model what is fixed, what can change, and how to judge the result.

This guide presents a reference-image workflow for campaign assets, product pages, social graphics, and visual prototypes. The goal is not to force every image into the same composition. The goal is to make different images feel as if they belong to the same brand.

You can test the workflow from the Image3 homepage, open the Image3 Workspace, inspect the model capabilities on the GPT Image 2 model page, and reuse the prompt blocks in the GPT Image 2 prompt library.

Quick answer

Build a small style contract before generating a campaign:

SOURCE OF TRUTH: use the uploaded reference images for the approved product identity,
palette, material, lighting direction, and visual tone.
PRESERVE: product silhouette, label placement, brand colors, surface finish,
camera distance, and the amount of negative space.
CHANGE: only the campaign scene, prop, crop, or headline area described below.
OUTPUT: one clean commercial image with a clear focal subject.
AVOID: invented logos, changed packaging, extra products, random text, watermarks,
unapproved colors, and decorative elements that compete with the subject.

The fixed block is the reusable part. Add a short variation block for each asset. This separation makes it easier to identify whether a failure came from the brand rules or from the new scene request.

What the reference should contain

Choose references for information, not for decoration. A good source image answers a question the prompt cannot answer reliably on its own.

Do not attach ten unrelated examples and hope the model averages them correctly. Start with one source of truth. Add another only to resolve a specific missing detail. If two references disagree about color, camera angle, or product shape, decide which one is authoritative before generating.

The six locks of a visual system

Identity lock

Describe the parts that make the subject recognizable. For a bottle, that might be the shoulder shape, cap, label ratio, and transparent material. For a character, it might be facial structure, hairstyle, outfit colors, and one distinctive accessory.

Identity lock: preserve the exact bottle silhouette, short matte-black cap,
warm ivory label block, and translucent amber liquid. Do not redesign the package.

Palette lock

Name the dominant and supporting colors, but also describe their role. “Blue” is weak. “Deep navy background with a small warm-coral accent and neutral white type area” is actionable.

Palette lock: use deep navy for the background, warm coral only as a small accent,
soft ivory for the product label, and neutral gray for secondary surfaces.

Lighting lock

Lighting changes perceived brand quality. Specify direction, softness, contrast, and shadow behavior. Keep the language short enough to reuse.

Lighting lock: large soft key from upper left, gentle fill from the front,
controlled rim light on the right edge, and one natural contact shadow.

Camera lock

Choose a repeatable camera relationship: eye level, three-quarter product view, overhead flat lay, or close portrait. Include crop and subject scale when they matter.

Camera lock: three-quarter view at eye level, product occupies about 55 percent
of the frame, 4:5 composition, with clean negative space on the left.

Typography lock

If text is part of the asset, keep the request narrow. State the exact copy, placement, hierarchy, and what must remain blank for later design work. Do not ask for a whole poster and a long paragraph of copy in the same first attempt.

Typography lock: leave the upper-left quarter empty for a headline. If text is
rendered, use exactly “NIGHT ROUTINE” in two short lines; do not invent copy.

Composition lock

Composition is the bridge between separate images. Describe where the subject sits, how much breathing room it has, and which direction the visual energy moves.

Composition lock: one focal subject, calm background, diagonal movement from
lower right to upper left, and enough uncluttered space for a short caption.

A reusable prompt structure

Keep the order stable so your team can edit prompts without losing the contract.

Use the uploaded references as the source of truth for [identity details].

STYLE CONTRACT
- Palette: [dominant colors and allowed accents]
- Lighting: [direction, softness, contrast]
- Camera: [view, crop, subject scale]
- Composition: [subject placement and negative space]
- Material: [surface, texture, finish]
- Type policy: [exact text or blank area]

VARIATION
Create [asset type] for [campaign/use case]. Change only [scene/prop/crop].
Keep every item in the style contract stable.

QUALITY CHECK
No changed packaging, invented logo, extra product, random text, watermark,
or decorative object that competes with the focal subject.

The style contract should be boring. That is a feature. Creative variety belongs in the variation block, not in the rules that define the brand.

The one-variable iteration loop

The fastest way to debug a campaign is to change one major variable per pass. A first pass might change the background from studio white to a warm kitchen. A second pass might change the crop from 4:5 to 1:1. A third pass might introduce a hand or prop. If you change all three at once, a drifted label could be caused by the scene, crop, or prop and you will not know which instruction to repair.

Use this loop:

1. Generate a baseline with the style contract and no variation.
2. Add one campaign variation.
3. Compare the output against the source image and checklist.
4. Keep the variation if identity and style remain stable.
5. If something drifted, remove the last variable before rewriting the whole prompt.

This is less exciting than asking for ten final images at once, but it produces a prompt you can hand to another person and reproduce later.

How to diagnose drift

Review in this order:

Shape and identity

Is the product still the same product? Has the character’s face, hairstyle, or clothing changed? If identity drifted, strengthen the source-of-truth sentence and remove competing references.

Color and material

Did a navy surface become black? Did a translucent package become opaque? Replace broad adjectives with a material and lighting description, then compare again.

Camera and scale

Did the product become tiny because the scene request was too ambitious? Restate the subject scale and crop. Keep the background simple until the camera relationship is stable.

Text and marks

Did the model invent a logo or alter a label? Request a blank label area or attach a clean label reference. For important legal or packaging copy, treat the generated text as a draft and verify it manually.

Negative space

Did props invade the area reserved for a headline? Move the space rule earlier in the prompt and describe it as a composition requirement, not as an afterthought.

Three campaign variations

The same contract can support different destinations without making every image look identical. Think of the contract as the brand layer and the variation as the channel layer. A marketplace listing may need a neutral background and a large product silhouette. A social post may need motion, asymmetry, and space for a short hook. A landing-page hero may need a wider crop and a calmer visual hierarchy. The subject can remain stable while the surrounding composition changes.

Keep a small destination table next to the prompt library:

This table helps a team avoid a common mistake: rewriting the entire prompt for each channel. Rewriting everything makes a successful result impossible to reproduce and makes it hard to tell whether a change improved the image or merely changed the style.

Product detail page

Use the uploaded product reference as the source of truth. Preserve the silhouette,
label proportions, cap, material, and warm ivory palette. Create a clean 4:5 product
detail image on a soft neutral background. Keep the product large and centered with
one natural contact shadow. Leave the label text unchanged and do not add props,
hands, extra packages, or invented logos.

Social launch image

Keep the approved product identity, deep navy and coral palette, soft upper-left
light, and three-quarter camera lock. Create a vertical social launch image with a
small coral geometric accent behind the product and generous empty space in the
upper-left for a headline. Show one product only. Do not render extra copy.

Lifestyle scene

Preserve the product shape, label area, ivory-and-navy palette, and soft commercial
lighting from the references. Place the product on a pale stone bathroom counter
with a restrained morning atmosphere. Change only the environment. Keep the camera
at eye level, product at 55 percent of the frame, and the upper-left area clear.
Avoid unrelated toiletries, hands, fake text, and decorative clutter.

When to use multiple references

Multiple references are useful when each one owns a different fact. For example, a front product image can own geometry, a close-up can own material, and an approved campaign image can own lighting. Tell GPT Image 2 which role each reference plays.

Reference 1 controls product shape and label proportions.
Reference 2 controls the matte ceramic material and edge highlights.
Reference 3 controls the quiet navy studio mood and negative space.
Do not blend their products or copy any unapproved text.

Role labels prevent the model from treating every image as an equal collage. If the result still blends details incorrectly, reduce the set to the one reference that contains the most important constraint.

A practical review scorecard

Before approving an image, score each item as pass, revise, or unknown:

Do not approve an image because it looks attractive in isolation. A beautiful image that breaks the product shape or leaves no room for the headline is a failed campaign asset.

What this workflow cannot guarantee

Reference prompts improve control; they do not turn generation into a deterministic renderer. GPT Image 2 can still misread small labels, alter a hand, simplify a texture, or make an unexpected composition choice. For high-stakes packaging, legal copy, faces, and final campaign claims, keep a human review step.

The workflow also does not replace a real brand guide. If your team has approved color values, logo rules, accessibility requirements, or legal restrictions, place those rules in the project brief and use the generated image as a draft until it passes your normal review.

Make the workflow usable by a team

A prompt becomes valuable when someone else can use it without asking the original creator what every sentence means. Store five things together: the approved reference images, the fixed style contract, the variation prompt, the destination and crop, and the review result. Give each asset a simple version name such as brand-style-v03-social-4x5. When an image fails, record the failed lock instead of replacing the whole prompt with a vague note like “make it better.”

Also record what the model was allowed to change. A reviewer should be able to distinguish an intentional variation from an accidental drift. “Background may change; packaging may not” is more useful than a long paragraph about overall quality. Keep the record close to the prompt so a future creator does not have to reconstruct the decision from a finished image.

Keep approved and experimental references in separate folders. An experimental image can be useful for exploration, but it should not silently become the source of truth for the next campaign. The same rule applies to generated outputs: mark the image that passed review, not the image that merely received the most likes in an internal chat. This small amount of labeling prevents style drift from accumulating across a large asset library.

For a handoff, write a one-line acceptance rule beside the prompt: “The product shape, label area, navy palette, upper-left light, and headline space must match the contract; the environment may change.” That sentence gives a reviewer a fast way to reject a visually attractive output that is still wrong for production. It also gives the next creator a clear starting point for the next controlled variation.

Final checklist

Before producing a batch, confirm:

• One approved source image is designated as the source of truth.
• Every reference has a clear role.
• The style contract lists identity, palette, lighting, camera, composition, and type policy.
• The variation block changes one major variable.
• The first output is checked against the source, not just against personal taste.
• Drift is fixed by removing or isolating the latest variable.
• Important text, logos, packaging, and claims receive human review.
• The winning prompt is saved with the asset and its intended destination.

Once the contract is stable, open the Image3 Workspace, choose the GPT Image 2 text-to-image workflow, attach the approved references, and test one controlled variation at a time. For more model-specific prompt patterns, continue with the GPT Image 2 model guide and the prompt library. If you need a deeper editing foundation, read the GPT Image 2 reference image editing workflow; for commerce examples, compare the GPT Image 2 product photo prompt guide.