Verified against Midjourney · 2026-07-25
Fix or change one part of a generated image with Vary (Region)
A targeted-editing workflow using Midjourney's Vary (Region) tool to regenerate just one selected area of an already-generated image — the wrong hand, an unwanted background object, a piece of clothing — while leaving everything else in the frame untouched.
The prompt
Ready to copy — highlighted parts are example details you can swap.
IMAGE BEING EDITED an upscaled portrait of a chef standing in a restaurant kitchen, plating a dish — the already-generated (and typically already-upscaled) image this edit applies to. EXACT REGION TO SELECT the chef's left hand and the plating tool it's holding — describe this precisely enough that whoever is doing the selection (drawing the mask in Midjourney's editor) knows exactly where the boundary should sit; a selection that's too tight leaves a visible seam around the edit, and one that's too loose regenerates more of the image than intended. WHAT IS WRONG WITH THAT REGION NOW the hand has six visibly malformed fingers and the tool it holds looks like a warped, unreadable shape WHAT SHOULD REPLACE IT a correctly-formed hand with five fingers holding a plating spoon at a natural angle over the dish — describe only the replacement content for the selected region; do not re-describe the rest of the image, since Vary (Region) already holds everything outside the selection fixed and re-describing it risks introducing a subtly different interpretation of the untouched area into the edit prompt. REGION-EDIT PROMPT a correctly-formed hand with five fingers holding a plating spoon at a natural angle over the dish, matching the warm kitchen light from overhead, slightly steamy atmosphere, same photoreal rendering as the rest of the frame of the surrounding image exactly, seamlessly blending at the selection edge. WHY THIS NEEDS TO MATCH THE SURROUNDING IMAGE warm kitchen light from overhead, slightly steamy atmosphere, same photoreal rendering as the rest of the frame — restate the specific lighting direction, color temperature, and rendering style already present in the untouched parts of the image, since the region-edit prompt is evaluated somewhat independently and will not automatically infer the surrounding image's exact lighting and style unless told explicitly. VARIATION STRENGTH strong, since a subtle variation of a badly malformed hand tends to just produce a slightly different malformed hand — a subtle setting keeps the edit close to the original region's composition while changing the specific problem; a strong setting gives Midjourney more freedom to substantially reimagine the selected area, useful when the problem is severe enough that a close variation of it would just reproduce the same flaw. OUTPUT A version of the original image identical outside the selected region, with the selected region regenerated according to the edit prompt above, blended at the selection boundary rather than pasted with a visible hard edge. IF THE EDIT DOESN'T BLEND CLEANLY Re-select a slightly larger region that includes a small margin of the surrounding, already-correct area rather than only the exact flawed pixels — giving Midjourney a little context from the correct surrounding area inside the selection itself, not just outside it, is usually what fixes a visible seam that a tightly-fitted selection produced.
Customize
Optional — swap in your own details for the highlighted parts above.
Why this works
Vary (Region) holds everything outside a drawn selection mathematically fixed and only regenerates pixels inside the boundary, which is a fundamentally different operation from re-running the whole prompt and hoping the next generation happens to fix one specific flaw while keeping everything else that was already good — a full re-roll risks losing a composition, expression, or lighting result that took several attempts to land, purely to fix one unrelated malformed hand. Region editing decouples "fix this one thing" from "keep everything else," which is the entire reason it exists as a distinct tool rather than users simply re-rolling the full generation more times. The instruction to describe only the replacement content, not the rest of the image, inside the region-edit prompt matters because the untouched area outside the selection is not being regenerated at all — it is preserved directly from the source image. Re-describing it inside the edit prompt does nothing to protect it and instead risks confusing the region-specific instruction with irrelevant detail about pixels that were never going to change anyway, diluting the actual instruction that matters: what goes inside the selected boundary. Explicitly restating the surrounding lighting, color temperature, and rendering style inside the edit prompt — rather than assuming the model will infer it from context — addresses a real limitation of how region editing evaluates its prompt: the model generating new content for the selected area is working from the text description given for that edit plus some visual context at the boundary, but it has no guaranteed strong signal that the new content must match the specific light source and style of the rest of the image unless that match is stated as an explicit requirement; without it, a regenerated hand can come back correctly formed but lit from a subtly different angle than the rest of the photograph, which reads as an even more obvious edit than the original flaw did. The guidance to choose strength based on the severity of the original flaw, and to widen the selection with a small margin of correct surrounding area when a blend looks seamed, both target the same underlying mechanical reality: a subtle-strength edit stays close to the original region's existing (flawed) composition, so it is the wrong choice when that composition itself is the problem, and a tightly-fitted selection gives the regeneration no correctly-rendered context inside its own boundary to blend from, so a small margin of already-correct pixels inside the mask gives it something reliable to match against at the seam.
Verified against
Midjourney v7 · 2026-07-25
Changelog
- 2026-07-25 — Initial publish, verified against Midjourney v7 Vary (Region) for targeted local edits.
Need this built into your business?
If a prompt isn't enough — what Scult builds, built and maintained for you — that's Scult's day job.
EXPLORE WHAT SCULT BUILDS
