All posts
Brand

GPT-Image 2.6 Brand Assets: A Consistency Tutorial

Use a saved product reference to guide new campaign images, then check every detail that must stay true to the real product.

STSearchFIT team15 minutes read

GPT-Image 2.6 brand assets are easier to keep visually consistent when you save an approved product photo in ChatGPT and call it back with a named reference such as @ProductPhoto. That reference can guide new campaign images, but it does not guarantee exact packaging, logo or product geometry, so review each result before publishing. For exact product details, keep the real product photo and use AI for the scene around it.

This tutorial takes you from an approved source photo to a reviewed campaign image. It covers the ChatGPT reference workflow, how to write the prompt, how to iterate without losing useful details, and when to use a compositing workflow instead. If you are planning product imagery alongside product pages and other e-commerce content, our e-commerce industry resources offer a related starting point.

What GPT-Image 2.6 brand assets can and cannot do

OpenAI announced Images 2.6 for ChatGPT, Codex and the API on November 4, 2026. In ChatGPT, the product is called Images 2.6. The API offers GPT-Image-2.6 Flare for faster iteration and GPT-Image-2.6 Sunburst for more detailed visual work and precise edits, according to OpenAI’s announcement. Those are product positioning statements, not a promise that every image will match a particular quality or generation time.

A saved reference is a way to bring an approved image back into a ChatGPT conversation by name. You can ask ChatGPT to save a photo, give it a name, then use @ followed by that name in a later prompt. OpenAI documents this process in its ChatGPT image help guide.

Think of the reference as visual guidance, not a locked product file. The model may still redraw the item. A bottle can change shape slightly, a label can become garbled, or a logo can shift. OpenAI’s image generation guide notes that image models can struggle with fine text and consistent brand elements across generations. So use a reference to help guide the work, then check the result against your approved product photography and brand rules.

Set the goal and gather what you need

For this walkthrough, the goal is to create one lifestyle image for an e-commerce campaign. The product sits in a new setting, while its recognisable shape, colour and packaging remain as close as possible to the approved source.

Before you start, gather:

  • A clear, approved product photo that shows the front of the product and its important details.
  • The product name or SKU, so you can name the saved reference clearly.
  • A short brief describing the campaign, audience, placement and desired mood.
  • A list of details that must not change, such as the product silhouette, cap colour, logo position and label wording.
  • A list of details the model may change, such as the background, lighting or surrounding props.
  • Your review criteria, including any packaging text or legal copy that must be exact.

Choose one asset to make first. A single hero image is easier to direct and assess than a request for an entire campaign system in one go. Once you have a result you can approve, use it as the starting point for further variations, while checking that each one still meets the brief.

Step 1: Choose a useful product reference photo

Select a source image that makes the product easy to understand. A front-facing, well-lit shot is a practical starting point because it clearly shows the shape, colour and label. If the product has a distinctive side, cap or control panel, consider whether your first campaign image needs to show it. A reference that hides a key feature cannot clearly guide the model on that feature.

Check the photo before uploading it:

  • Is the product in focus?
  • Can you see its outline and proportions?
  • Are the brand mark and label layout visible?
  • Does the light make the product colour look representative?
  • Are other objects covering any important details?
  • Is this an image your team is authorised to use for this purpose?

If your approved photography includes several angles, choose the one that best matches the image you want to create. For a front-on product hero, use a clear front view. For a lifestyle scene that depends on a side detail, use a source that shows that detail. You can also keep other approved views handy for a later conversation if the first image makes a feature ambiguous.

What you should see: You should be able to identify the product, its main proportions, its colour and the parts of the packaging that matter to the campaign. If the original photo is too dark, obscured or angled away from the details you need, choose another approved image before prompting.

Step 2: Save the image with a clear reference name

Open ChatGPT and upload the selected photo. Ask ChatGPT to save it as a reference and provide a short, unambiguous name. For example:

Save this approved product photo as a reference named ProductPhoto. I’ll use it as the visual source for a campaign image.

Use a name that makes sense to everyone working on the task. If you manage multiple products, a product name or SKU can help prevent mix-ups, such as CitrusCleanserFront or SKU482ProductPhoto. Avoid vague names such as image1 or new shot if you might save several references.

OpenAI’s documented workflow is to upload an image, ask ChatGPT to save it as a reference, name it, and later call it using @ and that name. A TechRadar test also describes saving a photo under a name and invoking it in a later prompt with @ followed by the name. That demonstrates how the reference workflow can be used; it is not an e-commerce accuracy test.

What you should see: ChatGPT should acknowledge the reference or otherwise make it available in the conversation. In a later prompt, type @ and select or enter the reference name as it appears in your interface. If it does not appear, check that the name is correct and that the photo was saved as a reference rather than merely uploaded in an earlier conversation. The exact interface can change, so follow the current controls shown in ChatGPT.

Step 3: Write a brief that separates fixed details from flexible ones

Before generating an image, decide what the model should preserve and what it can change. This is the central habit for getting more useful results. A prompt that says “make it look premium” leaves many choices open. A prompt that specifies which parts are fixed and which can vary gives the model a clearer job.

Write your brief in four parts:

  1. Asset and purpose: Describe the image and where you plan to use it, such as a product hero image for a seasonal e-commerce campaign.
  2. Reference: Name the saved product photo using its @ handle.
  3. Fixed details: State the product features and brand elements that should remain unchanged.
  4. Allowed changes: Describe the setting, light, camera angle and props the model may change.

Be concrete. “Keep the product the same” is less useful than naming the features that matter: “Preserve the bottle silhouette, proportions, cap colour, label layout, colours and logo placement.” If exact wording matters, tell the model not to invent or alter it. That instruction helps communicate the requirement, but it does not make generated text reliable. You must still inspect it.

Here is a starting prompt:

Use @ProductPhoto as the source of truth for the product. Create a premium lifestyle image for an e-commerce campaign: the product sits on a pale stone counter in warm morning light, with a soft neutral kitchen background. Preserve the product’s silhouette, proportions, cap, label layout, colours and logo placement. Do not invent or alter label text. Change only the setting and lighting; keep the camera angle close to eye level. No extra products, props over the label, watermarks or added text.

You can adapt the scene to your own product and brand. If the campaign needs a coloured backdrop, a bathroom shelf or an outdoor setting, say so directly. Keep the first brief focused on one image rather than asking for several formats, multiple products, a range of props and ad copy at once.

What you should see: The prompt should make it obvious which reference to use, what the finished image is for, which features must stay, and what is allowed to change. If you cannot tell whether a particular detail is fixed or flexible from the brief, clarify it before generating.

Step 4: Generate one first-pass image

Submit the prompt in ChatGPT and assess the first output as a draft. Do not judge only by whether the overall mood feels right. Compare the product to the reference, especially if the image is intended to represent a real item customers can buy.

Look at the result at a useful size. Check the overall composition first: is the product visible, placed where it belongs and shown from an appropriate angle? Then inspect the product itself: does its outline look right, and have its colour, proportions or packaging changed? Finally, zoom in on text, marks and small components.

Record what worked and what did not. For example:

  • Keep: warm morning light, pale stone surface, useful framing.
  • Fix: background is too busy, label is partly covered, cap colour appears different.
  • Do not publish yet: small label text is not legible enough to verify.

If the scene is useful but a product detail has drifted, do not treat that as a finished asset. You can try a focused edit to see whether it improves the result, or switch to a workflow that retains the real product photo. The more critical the detail, the less sensible it is to rely on a generated redraw.

What you should see: A draft that meets the broad creative direction and gives you something specific to review. The first image is a checkpoint, not proof that the handle has preserved the product exactly.

Step 5: Review the details that matter to your brand

Use a repeatable review rather than relying on a quick visual impression. Compare the generated image with the approved reference and your product information. Make a checklist for the product category, then inspect every required point.

Review areaWhat to compare with the approved sourceWhat to do if it differs
Product shapeSilhouette, proportions and distinctive featuresReject or use the real product photo in a composite
ColourProduct, cap, packaging and label coloursTry a focused correction, then compare again
Logo and brand marksPosition, shape, spacing and orientationDo not publish a distorted or substituted mark
Label and packaging textProduct name, wording and required copyReplace with approved artwork or real product pixels
SceneProps, shadows, reflections and backgroundRequest a local scene change if the product is sound
CompositionVisibility, crop and space for any campaign layoutAsk for a framing change without changing the product

Keep a simple record of the review: the asset name, source reference, prompt, output version, issues found, edits made and approval status. That record helps your team understand which parts of the brief worked and which details need a different production method. It also makes a useful handover if another marketer or designer needs to continue the work.

Do not assume that text is correct because it looks plausible at a glance. OpenAI’s documentation warns that fine text and consistent brand details can be challenging. Inspect wording closely, and compare any legally required copy with the approved source. If you cannot confidently verify a mark or text, treat it as unapproved.

How to check it worked: A reviewer can compare the image against the source photo and confirm each required detail. If any critical item is unclear, changed or obscured, the check has not passed.

Step 6: Iterate one change at a time

When an output is close, ask for one specific change per turn. This makes it easier to see whether the edit fixed the issue without disturbing everything else. OpenAI’s image prompting guide recommends saying what should change and what should stay fixed, then refining one thing at a time.

For example, if the product looks right but the background is too dark, try:

Change only the background to a warm cream studio. Keep the product, label, logo, proportions and camera angle unchanged.

If the product is sound but the framing is not, try:

Keep the product and scene unchanged. Reframe the composition to leave more open space on the left for campaign layout.

Avoid combining several corrections into a vague request such as “make it brighter, fix the bottle, change the label and improve the layout”. If the next image changes in several ways, it may be difficult to identify what caused the improvement or created a new problem.

A local correction is worth trying when the change is genuinely local, such as adjusting a background or removing a distracting prop. If the model continues to alter a brand-critical feature, stop iterating on a redraw and use the real product photo in a composite. Repeatedly prompting a generated image to reproduce exact product pixels is not a reliable substitute for using the original image.

Step 7: Make variations without losing the approved direction

Once you have a reviewed image, you may want more than one creative variation. Create them one at a time and describe the new variable clearly. For example, keep the product and camera angle fixed while changing the setting from a kitchen counter to a neutral studio. Or keep the setting fixed while requesting a different crop for a placement.

Do not assume that an approved output automatically makes every later variation accurate. Check each new image against the original product reference, not only against the previous AI-generated image. If one generated version contains a subtle change, using it as the sole source for another round may carry that change forward.

For every variation, note:

  • Which source reference you used.
  • Which single creative change you requested.
  • Which product details you checked.
  • Whether the asset passed review or needs another method.

This is especially useful when several people create campaign assets. A short record helps a teammate reproduce the direction without treating a prompt as a guarantee that the product will remain unchanged.

Step 8: Keep real product pixels when accuracy is essential

When the image must show exact packaging, a precise logo or fine print, the safest production choice is often to retain the real product photography. Use AI to generate or change the background, scene or supporting elements, then composite the approved product image into that scene using your usual design workflow.

This is a risk-control recommendation, not a special guarantee or feature of GPT-Image 2.6. It separates the parts that can be creatively generated from the parts that must remain accurate. Your designer can position and mask the actual product image, then check shadows, scale, lighting and edges so it sits naturally in the new setting.

A practical split is:

  • AI-generated: background, room, surface, general lighting direction or non-critical props.
  • Real product photography or approved artwork: packaging, logo, product geometry, label text and regulated claims.
  • Human review: final alignment, visual quality, brand compliance and approval to publish.

If you use this method, keep the approved source file and the final composite together with the campaign record. That makes it easier to verify what is real product photography and what has been generated. It also gives your team a clear route when a generated result is attractive but does not pass a product accuracy check.

Step 9: Know the difference between ChatGPT references and API inputs

The @ProductPhoto handle described here is a ChatGPT reference workflow. Do not assume it is an API parameter or that you can use the same @ syntax in an API request. The OpenAI API image generation documentation describes providing image files as inputs for generation and editing. In an API workflow, send the relevant image file as documented rather than relying on a saved ChatGPT handle.

This distinction matters if your team is building a repeatable creative pipeline. A marketer can use a saved reference in ChatGPT for hands-on exploration. A developer building an API workflow should follow the API’s image input and editing instructions, and decide how approved source files will be selected and supplied for each request.

The API documentation also describes markup-guided edits, where an annotation layer can indicate the area to change. This may be useful when you want to direct an edit to a particular region, but it does not remove the need to inspect the edited image. Confirm that the requested region changed and that other important parts remained acceptable.

OpenAI positions Flare for faster iteration and Sunburst for more detailed visual work and precise edits. Treat that as a starting point for choosing what to test, not as a rule that one option will always suit a particular brand. If you use the API, test with representative products and compare the output against the same review checklist.

Step 10: Test the workflow before scaling it

Before building a large asset library around saved references, run a small review with a few representative products. Include products with different packaging shapes, colours and text layouts. The aim is not to prove that every output will be accurate. It is to learn where the workflow helps and where your team still needs real photography, a designer or a different production method.

For each test, keep the source photo, prompt, output and review notes together. Record the errors that matter to your business, not just whether a marketer liked the image. For example, note a changed label, altered cap colour or covered product name. Then decide which tasks can use generated imagery and which require the original product image.

A simple internal test table can help:

Test itemRecord
Product and reference nameWhich approved source did you use?
Intended assetWhat campaign image or placement was requested?
Preservation requirementsWhich details had to stay accurate?
Review resultWhat matched, drifted or became hard to verify?
Production decisionUse, revise, composite or reject?

If you find that the product changes frequently, narrow the task. Generate a background only, use fewer product transformations, or compose the final image from real product photography. The point of a test is to help set sensible review rules before a team starts producing at a larger scale.

API costs and output formats to keep in mind

If you use the API, there is no single fixed price per image in the cited model rate card. The GPT-Image-2.6 Flare model page lists rates per 1 million tokens: text input at $5.00, cached text input at $1.25, image input at $8.00, cached image input at $2.00, and image output at $30.00. These are token rates, not a flat per-image charge. Total cost depends on the input and output tokens used, and settings such as quality and image size can affect output use.

The API image guide documents quality settings from low through max, alongside a separate complexity setting that controls how much the model plans before rendering. It also describes GIF generation and SVG output support. Availability of a format does not mean it is the right choice for every product asset. Test the required output format and inspect the result before adding it to a live campaign.

If you are estimating API spend, use the current model page and the actual request settings for your workflow. Do not turn an example calculation into a universal per-image price. ChatGPT and API access also have different workflows; check OpenAI’s current product information for the options available to your team.

Troubleshooting common reference and consistency problems

The reference name does not work. Check the name you saved and how it is entered in the current ChatGPT interface. Make sure you saved the image as a reference, rather than assuming that any earlier upload will be available by handle in a later prompt. If it remains unavailable, upload the source again and follow the current reference-saving instructions.

The product looks similar but the label has changed. Similarity is not the same as accuracy. Compare the label to the approved product photo. Try one focused edit if appropriate, but use approved artwork or the original product pixels when exact wording matters.

The output changes the product when you only asked to change the scene. State which features must stay fixed and which region should change. Make one edit at a time. If the product still drifts, stop asking for a full redraw and use a composite that retains the actual product photo.

The logo is close but not correct. Do not publish a near-match as an approved logo. Use the brand’s approved logo file in the final design, or retain real product photography that already contains the correct mark.

A second variation loses details from the first. Check the variation against the original approved reference, not only the first generated output. Each generation can introduce changes, so review every version independently.

You need a dependable set of campaign assets. Define a repeatable prompt and review process, but do not treat the prompt itself as a quality control system. A human reviewer still needs to approve the product, text, logo, scene and final crop.

A final pre-publication check

Before you publish, confirm that the image has passed a human review against the approved product reference. Check the product’s outline, colour, packaging and visible features. Verify the logo and label wording against approved sources, and make sure any required copy is legible and correct. Confirm that the crop suits the placement and that no prop hides an important detail.

Then record the final asset and its approval status. If the product’s exact appearance is critical and the image redraw does not pass, use a composite with real product pixels instead. For related guidance on making e-commerce content useful to both shoppers and AI search, see our DTC e-commerce AEO playbook. For more practical marketing guidance, browse our resources.

Get SearchFIT Weekly

What we are learning about getting brands cited in AI search, in your inbox every week. Unsubscribe any time.