Data report: Can an AI image generator edit any part of a product photo without changing the product? A reproducible benchmark of localized edits—background swaps, prop removal, label/logo preservation, colorway changes, and campaign-format variants—using Lamina versus Midjourney-style generation for ecommerce creative.
AI can localize ecommerce photo edits, but no supplied evidence proves any tool preserves every product detail. Use a reproducible, SKU-level fidelity gate before publishing.
Lamina Team
Product Team @ Lamina
Can an AI image generator change part of a product photo without altering the product itself?
No. Current evidence does not justify treating any generative image editor as invariant across every part of a product. You can swap a background, remove a prop, change a defined color area, or make a new format, yet logos, label text, material grain, hardware, stitching, proportions, and exact color still need publication-critical review.
That line matters on a PDP. The edit has failed if the background works but the bottle label picks up a character, the zipper shifts, or the fabric weave goes soft. For strict SKU-identity work, the documented safer pattern keeps the original product crop and generates the surrounding floor, shadow, lighting, and background rather than regenerating the product body.
Is Lamina better than Midjourney for SKU-preserving product-photo edits?
No defensible winner emerges from the supplied evidence. It contains no independent, controlled Lamina-versus-Midjourney results across the five requested ecommerce edit types. That gap is the report’s central finding: capability pages and polished examples are not approval-rate evidence.
Lamina says its brand kit can lock product references and score outputs against a brand kit; its developer documentation says the same brief, brand, and seed produce the same output. Useful controls for a repeatable test arm. They do not promise pixel-perfect preservation. Midjourney officially offers Vary Region, which regenerates a selected area while leaving the rest of the image untouched, though its own documentation does not establish unchanged product identity in every edit.
Log the model path with every result. Midjourney’s Editor documentation says V8.1 images can enter the Editor, while Editor functionality currently uses V6.1; naming only the initial model leaves out a variable that can alter the output.
Evidence review for ecommerce teams deciding whether localized AI edits are safe to publish on SKU-specific catalog and campaign assets.
Controlled Lamina-versus-Midjourney comparison available in the supplied evidence
over Evidence set reviewed through July 2026
Definition of a shippable localized edit
over For every output evaluated
Workflow for strict SKU-identity work
over Compliance-sensitive catalog and PDP use
| Metric | Value | Source |
|---|---|---|
| Virtual-model generations evaluated in Photoroom’s vendor-produced editing benchmark | 4,250 | photoroom.comas of 2026-07-06 |
| Best reported full-product-fidelity pass rate in that benchmark; treat an unreviewed output as a review candidate, not a publishable asset | 29.0% | photoroom.comas of 2026-07-06 |
| Reported full-product-fidelity pass rate after Photoroom’s Fidelity Layer; even an added fidelity mechanism did not make preservation automatic | 38.2% | photoroom.comas of 2026-07-06 |
| Localized edit families a useful benchmark should separate: background, object, text, style transfer, and format-specific work | 5 | wavespeed.aias of 2026-07-09 |
Why score product fidelity separately from edit completion?
Score product fidelity on its own because an image can finish the requested background swap or prop removal while quietly changing the SKU. The vendor benchmark figures above are not a Lamina-versus-Midjourney comparison. They do make the point: visual plausibility is a poor stand-in for checking protected regions.
Make the final call conjunctive, not aesthetic. An output is shippable only if the requested change happened and every protected element remains acceptable; either miss fails the asset. That keeps a handsome campaign frame with wrong packaging text or altered product geometry out of the catalog.
Why do logos and labels need their own acceptance metric?
Give logos and labels a hard gate. Generative systems can imitate a mark and still produce an image that passes at a glance. Runflow co-founder and CEO Ricardo Ghekiere calls out the operational issue directly: brand logos on garments are a known failure point, so a generic realism score cannot cover logo exactness.
AI image generators get brand logos wrong on garments.
Why is localized editing safer than regenerating a product image from a fresh prompt?
Localized editing is safer because it limits the generated area and gives reviewers a defined product region to protect. Daniel Geyne’s observation lands squarely in ecommerce production: ask for one small visual change through a wholly new prompt and the original SKU identity becomes part of the gamble.
Use masking as a constraint, never proof. Midjourney says its selected Vary Region area is regenerated, so a benchmark should inspect the requested edit area and the supposedly untouched product area rather than assume selection boundaries guarantee fidelity.
When a creative director requests a minor change—perhaps a different color for a jacket, a specific product on a table, or a less distracting background—regenerating the entire image via a new prompt is a gamble.
How do you run a reproducible benchmark for localized ecommerce edits?
Use rights-cleared SKUs, declare editable and protected regions before generation, then blind-score outputs against the original product image. Attractive examples will not tell you whether a tool keeps the product intact. Every arm needs the same source image, edit brief, output count, format, and evaluation rule.
Reproducible localized-edit benchmark protocol
Build a SKU set that exposes failure modes
Use 30–50 rights-cleared products across reflective, transparent, textured, apparel, packaging, and small-text items. Keep the original source asset as ground truth for every output. A smooth box will not expose the mistakes that show up on a foil pouch or woven garment.
Mark allowed and protected regions before you generate
Set out which pixels may change for each task. Protect the product body, logos, labels, material texture, hardware, stitching, and proportions unless the test explicitly allows a masked colorway change; keep separate task families for background swaps, prop removal, label/logo preservation, limited colorway edits, and campaign-format variants.
Keep generation conditions fixed
Give both arms the same edit brief, source image, requested output count, and aspect ratio. Record the complete model and editing path. For Lamina, log the brand configuration and seed, since its documentation describes repeatable output from identical brief, brand, and seed inputs. Log Midjourney’s generation and Editor versions separately.
Blind-score the requested change and the protected product
Use at least three raters for every output. Track task-completion rate, product-fidelity pass rate, label/logo exactness pass rate, color error, protected-region similarity, campaign-format compliance, median iterations, time, and cost.
Publish disqualifying failures alongside the passes
A shippable pass requires the requested edit to succeed with no material product drift. Break out failures by task family—especially small text, reflective surfaces, and branded packaging—so a buyer sees where a tool fits instead of receiving one blended score.
What should ecommerce teams do before publishing AI-edited product imagery?
Use AI generation for new concepts, complex styling, campaign formats, and localized scene changes. Put every SKU-specific output through human product-fidelity approval before it reaches a PDP or paid placement. The most reliable production setup keeps the source product where identity must stay exact, then generates the creative world around it.
Treat format expansion as a separate test, not a resize. A 1:1, 4:5, 9:16, or 16:9 campaign variant can meet framing requirements while clipping the product, shifting its scale, or introducing drift near the crop. The practical call is not Lamina versus Midjourney in the abstract; it is whether a documented tool-and-workflow combination clears your protected-SKU gate across the formats and edits you actually publish.
Methodology
Original Lamina experiment run 2026-08-02. Hypothesis: With identical source images, edit instructions, output size, and generation budget, Lamina’s localized-edit workflow will preserve the untouched product more faithfully than a Midjourney-style image-to-image workflow, while achieving similar or better instruction completion across background swaps, prop removal, label/logo preservation, colorway changes, and campaign-format variants.. Measured 3 variant(s) for cost and latency on the Lamina image engine; numbers cited here are our own measurements.
