How to edit ecommerce product images with AI without changing the SKU: a Lamina workflow for accurate, on-brand backgrounds, retouching, channel crops, and campaign creative

Edit ecommerce product images with AI without changing the SKU: treat the approved source product as an immutable layer, and generate only the scene around it. The package, silhouette, variant color, material, labels, logo position, visible claims, included components, and scale stay fixed. Background, surface, ambient light, canvas, crop, props, and composition are the creative layer.

This is a production rule, not prompt decoration. One approved packshot can feed a white-background PDP asset, a 4:5 social crop, and a lifestyle campaign composition, yet each is its own output class with channel-specific requirements. Lamina’s create, track, evaluate, and distribute model puts the work in a useful order: brief the asset, generate candidates, score them against brand and SKU requirements, then send only approved files to the destination your team uses.

What did the Lamina workflow experiment actually show?

The experiment measured all three tested Lamina workflows at $0.04 per generation. The protected-product masked workflow ran at about 37 seconds per generation—roughly 17 seconds slower than the reference-guided unmasked version and about 11 seconds slower than campaign-first compositing. Budget that masked path for catalog accuracy; speed alone is the wrong thing to optimize.

The test covered three variants: a SKU-locked masked control, a reference-guided workflow without a protected product mask, and campaign-first compositing with explicit negative constraints. It supplied no outcomes for SKU preservation, mask compliance, product similarity, background adherence, crop compliance, approval rate, reruns, or failure rate. This is operational data, not a quality verdict: it establishes a per-generation cost and one-test latency, excluding human review, revisions, and media spend. It does not prove any variant won on product fidelity.

What belongs in a SKU truth pack before AI editing?

A SKU truth pack needs the untouched, full-resolution approved source image, the SKU and variant identifier, intended channel, approved brand kit, and a written list of product facts that cannot move. Write the non-negotiables plainly: package geometry, front-facing artwork, visible text, logo coordinates, colorway, finish, included items, pack count, and scale.

Split locked facts from permitted changes before anyone writes a scene prompt. Cleanup, surface replacement, shadow repair, relighting, canvas extension, and channel crops are generally editable; changing package copy, variant color, claims, accessories, or product construction is a new verified-product request rather than an image-editing instruction. That line stops an attractive campaign output from quietly becoming an inaccurate catalog asset.

Which ecommerce product edits need closer human review?

Reflective metal, transparent packaging, intricate fabric texture, exact color, small text, food surfaces, and luxury finishes need tighter human review. Small visual drift is easy to miss and expensive to publish, especially for hero SKUs and product-detail pages where shoppers zoom into the file. Treat them as a higher review tier.

Review truth, not whether the image merely looks polished. Check whether a highlight altered the apparent finish, transparent edges swallowed package details, tiny type was rewritten, or a prop hid a required marking. If the scene works but a local edge or shadow fails, repair that area or composite the protected product over the approved scene rather than regenerating the whole asset.

What is a product-safe prompt for background, crop, and campaign edits?

A product-safe prompt identifies the uploaded approved SKU as the fixed reference, lists the exact attributes to preserve, and permits generation only for the scene and format. Use this template: Use the approved uploaded [SKU] image as the fixed product reference. Create a [channel and aspect ratio] asset with [approved background or scene]. Preserve exactly the product silhouette and proportions, SKU color, material and texture, packaging structure, all visible label text, logo placement, visible claims, included parts, camera angle, and natural contact shadow. Change only the background, surface, ambient lighting, canvas, crop, and composition.

Close with explicit exclusions: Do not invent, remove, rewrite, recolor, resize, or obscure logos, text, claims, certifications, accessories, features, pack count, badges, ratings, or discounts. Reject the result if any locked attribute differs from the source. That hard stop matters; a model can make a convincing image that is still wrong for a purchasable SKU.

How should ecommerce teams publish AI-edited product images safely?

Publish AI-edited product images only after a named reviewer checks the SKU facts, brand treatment, and channel crop against the original source. Make the approval gate stricter for PDP mains and hero products than secondary campaign placements. Every file still needs a traceable decision.

Keep the immutable source, approved brief, output, and evaluation record together. Your team can then reuse an approved visual system across product launches without losing proof of what was protected, what changed, and why a particular asset was allowed into the catalog.