Midjourney Now Protects Unselected Pixels. Sellers Still Need a Clean Master

Repeated AI edits can damage a product image even when each instruction sounds harmless. Remove a prop, widen the background, soften a shadow and fix a small mark, and the final frame may look polished while the label, chain, clasp or engraving has quietly drifted from the real SKU.

Midjourney announced a useful change on 24 September 2026: targeted inpainting and outpainting edits with its V8 Edit Model now change only the pixels a user selects. The company says this allows repeated edits without degrading the rest of the image. On 2 October, it also said editor results now keep the base image’s aspect ratio, so an accidental aspect-ratio parameter should not stretch the product.

These are practical safeguards for ecommerce work, but they do not turn Midjourney into a product-accuracy system. They reduce the area exposed to generation. Sellers still have to decide what is safe to regenerate and what must remain factual.

What changed in the editor

Midjourney’s current Editor accepts both Midjourney creations and external images. Its Erase brush and Smart Select tools let a user define an area for regeneration. Smart Select can be refined with include and exclude points, then applied either to the selected object or everything around it.

The September update changes the behaviour after that selection is submitted. Midjourney now says only selected pixels change during a targeted edit. That is materially different from an edit model that rebuilds the full frame while merely trying to preserve unmentioned details.

The distinction matters because a product can deteriorate through an edit chain. A necklace may survive the first background change, lose one link after a second pass and acquire a slightly different clasp after the third. If unselected pixels genuinely remain fixed, local edits can be stacked without repeatedly asking the model to reinterpret the entire necklace.

The safest ecommerce use is editing around the SKU

This update is strongest when the product is already correct and the surrounding presentation is not. A seller might remove a stand beside a ring, repair a dirty patch of paper behind a bottle, replace a small prop, or extend an empty background to make space for advertising copy.

For a square marketplace image that also needs a 4:5 social crop, outpainting can add canvas around the approved product instead of stretching it. Midjourney’s 2 October aspect-ratio fix adds another guardrail by keeping the base image’s proportions when a stray parameter appears in the prompt.

The same principle applies to catalogue cleanup. If the jewellery, packaging and printed label are already accurate, keep them outside the editable region. Give the model the smallest possible job rather than selecting the whole frame and asking it to understand what should stay unchanged.

A repeatable local-edit workflow

Keep an untouched master before opening the editor. Work from a duplicate, and never replace the source photograph with an AI-edited derivative.

Make one narrowly defined selection. Use the Erase brush for simple shapes or Smart Select when the boundary is more complex. Add exclude points around the product if the automatic mask creeps onto a chain, stone, logo or polished edge. Apply the mask before submitting the prompt, as Midjourney’s documentation requires.

Describe only the replacement area. A useful instruction is specific about surface, light and unwanted objects: “Continue the same warm ivory paper background, matching the existing light direction and grain; no new props, text or product changes.” Avoid restating the entire product because that invites unnecessary interpretation.

Review the mask boundary at full size. Hairline chains, reflective edges, transparent glass and soft shadows are easy to include accidentally. If the selection touches the product, assume those pixels can change and compare them with the master.

Approve each step before starting the next. Save a clean checkpoint after a successful edit, then use that version for the following local change. If a pass fails, return to the last approved checkpoint instead of repairing a repair.

Finally, export the marketplace and campaign crops from the approved file. Keep the real product master, the final edited asset and the prompt or mask notes together so another team member can repeat the process without treating the AI output as new product truth.

Unselected pixels are not the same as an accurate result

Midjourney’s claim is about where pixels change, not whether the regenerated area is commercially correct. Inside the mask, the model can still invent geometry, materials, shadows and reflections.

That makes some edits poor candidates even with precise selection. Do not regenerate an engraving and assume the characters will remain exact. Do not replace a clasp, stone setting, barcode, certification mark, ingredient panel or safety label with a plausible-looking version. If those details need correction, use the original product evidence and conventional retouching or rebuild the controlled text and artwork.

Shadows also need judgement. A new background may look convincing while the contact shadow implies that the item floats or sits at a different angle. Polished metal can preserve its outline yet reflect a scene that does not match the new environment. Product fidelity includes physical coherence, not only unchanged pixels.

There is still a resolution trade-off

Midjourney’s current version documentation notes that inpainting or outpainting an HD V8.1 or V8.2 image downscales the result to standard definition. The image can be upscaled afterwards, but sellers should not confuse restored output size with restored source detail.

Inspect fine engraving, chain links, texture, stitching and small label text after the full edit and upscale sequence. If the job depends on keeping the original high-resolution product pixels, a conventional layer-and-mask workflow may remain safer.

Where this update genuinely helps

Midjourney’s update makes the editor more useful for campaign adaptation than it was at launch. Sellers can now make several local changes around an approved product without accepting whole-image regeneration on every pass. The aspect-ratio protection also removes one avoidable way to distort a product while preparing new formats.

The best use is conservative: preserve the photographed SKU, edit the environment in small regions, check every boundary and keep a clean master. Treat “only selected pixels change” as a valuable control, not permission to regenerate factual product details.

Sources

Midjourney, “Edit updates, thumbnail previews, and more,” published 24 September 2026: https://updates.midjourney.com/edit-updates-thumbnail-previews-and-more/

Midjourney, “Alpha Changelog - 10/1/26,” published 2 October 2026: https://updates.midjourney.com/alpha-changelog-10-1-26/

Midjourney Documentation, “Editor,” reviewed 2 October 2026: https://docs.midjourney.com/hc/en-us/articles/32764383466893-Editor

Midjourney Documentation, “Version,” reviewed 2 October 2026: https://docs.midjourney.com/hc/en-us/articles/32199405667853-Version