To preserve logos and labels in AI product photos, keep the approved product image as your source, change as little of the product as possible, and check the result against the exact package wording. If the logo or label must remain exact, retain the original product pixels or place approved artwork in a suitable layer editor. A prompt asking for preservation doesn't certify the output.
Oakgen's image editor gives you a place to try a scene change using an uploaded product photo. Start with one background edit and inspect the label before extending the workflow to a catalog. The attractive image with the wrong quantity is a rejected image.
Edit the scene around one approved product photo, then use the worksheet below to decide whether the result is usable.
Choose how much of the package you need to change
| Your job | Recommended approach | What can go wrong | Approval requirement |
|---|---|---|---|
| New scene, same product angle | Keep the product source; request a background change | The model redraws letters or changes the cap | Compare all visible label lines and package geometry |
| Exact packshot for a listing | Preserve the original product pixels with appropriate selection and layer tools | Edge cleanup or shadows damage the package boundary | Inspect product, edges, and cast shadow |
| New camera angle | Photograph that angle, then edit its surroundings | The unseen side gains invented wording | Use a real source proving the new angle |
| One incorrect label line | Restore approved artwork or source pixels in a suitable editor | A flat replacement doesn't follow the curved bottle | Review perspective, curvature, print texture, and lighting |
| New packaging concept | Generate a clearly marked concept | Someone mistakes concept text for approved copy | Keep it out of live product listings until approved |
| Product detail shot without branding | Use a crop that genuinely answers the detail question | The crop hides the feature shoppers need to inspect | Retain a separate accurate package image |
Use the first row to try a controlled scene edit in Oakgen. Choose the original-pixel route when you need exact branding and your generative results keep drifting.
Who this guide is for, and what we checked
This guide is for sellers and creative teams with an existing package: a food pouch, candle, cosmetic bottle, branded device, or other item whose name and visible specifications must match what ships. Designing a new logo is a different job.
Research date: September 30, 2026. We checked Oakgen's current editor implementation for source upload, natural-language instructions, optional reference images, model selection, and output download. Available settings depend on the selected model. This article doesn't establish a pixel-preserving mask or artwork-compositing workflow inside Oakgen.
OpenAI's current image-generation documentation acknowledges remaining limits in precise text placement and clarity, as well as consistency of brand elements across generations. That supports a review step; it doesn't establish which model will preserve your particular label best.
The fictional package and review examples below are illustrative templates, not results from a performed generation experiment. We haven't measured a pass rate, compared model scores, or produced a repaired package for this article. Oakgen is our product.
Write the label truth sheet before you generate
Your source of truth needs to survive the photo workflow. Don't ask the reviewer to remember what the package said yesterday.
Save the approved product photo and a close crop of the label in a folder named with the SKU and packaging version. Keep the flat approved artwork separately if you have it. A crop records the photographed label; the artwork records the approved design. Neither gives permission to invent a new quantity or claim.
Copy this worksheet into the brief:
SKU:
Variant:
Packaging version / approval date:
Approved full-product photo:
Approved label crop:
Approved artwork file, if available:
Brand name, exact characters:
Product name, exact characters:
Variant wording:
Quantity and unit:
Other visible wording requiring an exact match:
Logo shape and placement:
Existing symbols / badges:
Requested change:
Product details that must remain unchanged:
Intended placement and displayed size:
Reviewer:
For a fictional candle, the wording might be NORTH PINE, CEDAR, and 180 g. Record the lowercase unit and the space too. 180 ml doesn't describe the same quantity; CEDER doesn't pass because it looks close at thumbnail size.
Add symbols as visual facts. A leaf-shaped logo, a registered trademark symbol, or a circular badge needs its own inspection. A matching transcript cannot prove that the model drew the right mark.
If two references show different packaging revisions, settle the conflict before uploading them. An old logo in the detail crop and a new logo in the full photo leave the reviewer without one approved target.
Prepare a source that lets you inspect the text
Use the original photograph where possible. Avoid a marketplace screenshot with a tiny product, a heavily compressed social image, or a crop that cuts off the unit after the number.
Choose a view where the label faces the camera and glare doesn't erase characters. You should be able to read the required wording in the source before asking the model to use it. When you can't, take a clearer photograph instead of asking the model to reconstruct missing letters.
A close crop can help communicate a small label's role, but it shouldn't contradict the full product photo. In Oakgen, the main uploaded image supplies the image you want to edit; optional reference images provide additional context. Describe that context explicitly, and only add a close crop when it answers a missing question.
Don't attach a mood image containing someone else's package beside your own label without explaining the roles. If the mood image supplies the scene, say that its product and typography don't belong in the result.
For a wider campaign, use the product consistency workflow. This article's truth sheet concentrates on package wording and logo geometry rather than the complete campaign scorecard.
Once the source passes that inspection, upload it for one background change. If you don't yet have an approved real package, use the image generator for concept exploration and keep the concept outside the live-SKU workflow.
Ask for the scene change separately from preservation
Name the requested change, the facts that must survive, and the permitted camera view. A list of aesthetic adjectives doesn't explain which pixels contain product data.
Here is a copyable prompt for the fictional candle. Replace the identifiers with wording you have verified on your own source:
Use the uploaded product photograph as the source for this edit.
Change the surrounding background and supporting surface to a pale
stone tabletop with soft window light. Keep the current camera angle,
product position, and front-facing label.
Preserve the candle jar, lid, proportions, package color, label boundary,
and original logo layout. Keep every visible label character as shown,
including NORTH PINE, CEDAR, and 180 g. Add no wording, badge, symbol,
ingredient, or promotional claim.
Keep the label unobstructed and readable. No hands or props in front
of the jar. Don't rotate the product or reveal a new side.
The prompt states the intended edit. It isn't an assurance that the output will obey it.
Keep the first attempt narrow. Rotating the jar, changing its lighting, opening the lid, and putting it in a hand creates several new demands around the same label. If a result fails, you want to know which demand to remove.
For selection-led editing in another tool, Adobe documents targeted edits using selections in Photoshop. That is a tool-specific control, separate from writing “keep the label unchanged” in a prompt. Check the available controls in the editor you actually use.
Try this background-only brief with your own product, keeping the original source beside the downloaded result. The sample prompt is copyable text; the link doesn't load a prefilled project.
Check exact text and logo geometry as separate jobs
First read the label. Then inspect the design.
The text review should compare the output line by line with the truth sheet. Pay special attention to numbers, punctuation, units, repeated characters, variant names, and any visible claim. Those details often look ordinary enough to escape a quick glance.
For logo geometry, compare the outline, spacing, letter proportions, alignment, and location relative to the label boundary. A model can spell the name correctly and still draw a different wordmark. Inspect the visual symbol too, especially when a stylized letter doubles as the logo.
This is an illustrative rejection worksheet for NORTH PINE / CEDAR / 180 g, not a log of generated outputs:
| Check | Approved source | Illustrative defect | Decision |
|---|---|---|---|
| Brand spelling | NORTH PINE | NORTH PlNE with a substituted character | Reject |
| Product variant | CEDAR | CEDER | Reject |
| Quantity | 180 g | 180 ml | Reject |
| Logo placement | Centered above product name | Shifted partly outside the label | Reject or restore artwork |
| Symbol inventory | One approved leaf mark | Additional invented seal | Reject |
| Label boundary | One continuous oval | Doubled edge or torn curve | Repair only if product geometry remains accurate |
| Scene | Pale stone and window light | Slightly darker stone, product correct | Review against the creative brief |
A prettier surface doesn't compensate for a changed quantity. Keep product accuracy as a pass/fail requirement rather than averaging it with creative quality.
OCR can help flag a missing word or changed digit, but an exact text match won't establish that the visual logo matches. If OCR reads NORTH PINE while the image has the wrong spacing and symbol, the image still fails the brand check. Use automated assistance to find candidates for inspection; retain the source comparison.
Review the file at its actual delivery size
Zooming into the generation preview isn't the last check. Download the image, inspect the label crop, then view the export at the size shoppers will see.
At thumbnail size, check the recognizable mark and product silhouette. At product-page size, check whether the variant and quantity remain clear enough for that placement. At a close crop, inspect spelling and geometry. Return to the full frame afterward so the repair hasn't created an obvious patch.
Don't describe tiny text as “verified exact” if you can't read it in either the source or export. Choose an accurate close-up or a better source when the placement requires that information. Blurring required information doesn't turn an inaccurate packshot into an accurate one.
Upscaling is a separate operation. If you use an image upscaler, repeat the text and visual checks on the enlarged file; additional pixels don't certify the wording. Compression and later crops also deserve a final look, especially when the unit sits close to an edge.
Repair, composite, or keep the original?
Use a local repair only when the surrounding package is already correct and the correction is small enough to inspect completely. Return to the approved source for a retry rather than using a faulty package as the next reference.
Retry the scene edit if the input was unclear, the camera change was unnecessary, or the prompt mixed product and scene instructions. Simplify the demand and retain the failed attempt so you can compare what changed.
Restore approved artwork or source pixels when the text keeps changing but the package geometry is sound. Use a suitable layer editor. A flat label needs the right perspective, curvature, edge treatment, and lighting to belong on a photographed bottle. Overlaying a correct PNG on a distorted jar doesn't repair the jar.
Keep the original photograph when the altered view exposes packaging you haven't photographed, several label regions fail, or the repair would mean rebuilding the product. An accurate source on a plain background is a useful listing asset.
For grouped packaging, inspect every visible unit independently. The product-bundle photography guide addresses arranging a set; one correct hero label doesn't approve all the smaller packages behind it.
Record the decision in the delivery note:
Asset ID / final filename:
SKU and packaging version:
Source used:
Requested change:
Label text: PASS / FAIL / UNREADABLE
Logo geometry and placement: PASS / FAIL
Package shape and quantity: PASS / FAIL
Repair method, if any:
Approved placement:
Reviewer and approval date:
Avoid a vague approved filename with no approved placement. A detail crop that omits the label may suit a material feature section while the main listing still needs the accurate full package.
Mistakes that keep the label problem alive
Rewriting “exact” ten different ways can become an expensive substitute for fixing the input. If the source has unreadable lettering, improve the source. If the shot requires an unseen side, photograph it. If the same line keeps failing, restore the approved artwork or retain the original photo.
Another trap is approving by resemblance: the brand name looks right, so nobody checks the unit. Review each recorded field independently. The controlled image-editing benchmark gives you a wider method for evaluating preservation and requested edits separately.
Keep offer copy outside this task. Sale prices, expiration dates, and campaign headlines change more often than the photographed package. Add approved copy through an editing workflow that keeps it editable rather than requesting new promotional wording inside the label.
Frequently asked questions
Can AI keep every word on my product label exact?
A generative edit can still change individual characters, quantities, or logo geometry. Compare every required detail against an approved source; use the original product pixels or approved artwork when exact reproduction is essential.
Why did the quantity change even though the package looks right?
A plausible-looking package can contain incorrect text. Treat quantities and units as independent checks rather than accepting them because the bottle and background look convincing.
Does a higher-quality setting solve label errors?
It may improve image detail, but it doesn't certify spelling or product accuracy. Approve the wording separately and inspect the final exported file.
Should I mask the product before editing the background?
In an editor with suitable selection and layer controls, protect the product and edit the background. A preservation prompt alone isn't a pixel lock, and annotations don't prove that the model leaves the marked area untouched.
When should I composite the original label?
Use approved artwork or an original label crop when text remains wrong and the package geometry supports an accurate placement. Use a suitable layer editor and review perspective, curvature, lighting, and edges afterward.
Can OCR approve the label for me?
OCR can help flag a missing word or changed number, but a text match doesn't establish that the logo, spacing, symbol, or curved label looks correct. Keep a human visual check.
Can I use a generated label for packaging print?
Treat a generated label as a design concept. Final packaging needs approved artwork, confirmed wording, and the printer's production specifications; a polished product mockup doesn't establish print readiness.
Make one accurate asset before expanding the shot list. Open Oakgen's image editor, upload the approved source, request one scene change, and complete the label truth sheet against the downloaded result. If your next step is motion, take that approved still into the product video workflow.
