You fix one hand in a Midjourney image, and the face changes too. You ask for a warmer background, and the product label you spent twenty minutes getting right turns into nonsense. If that sounds familiar, the problem is not your prompt quality. You are using the wrong edit mode for the kind of change you want.
Midjourney's V8.2 Edit Model, opened to all users on 27 August 2026 and upgraded on 24 September, folds instruction editing, inpainting, outpainting and multi-image references into one model. That is powerful, but it also means one model can now touch your whole image when you only wanted it to touch a corner. This guide shows you how to pick the right mode and write edit prompts that change only what you mean to change.
What is Midjourney's V8.2 Edit Model?
The V8.2 Edit Model is Midjourney's image editing model for versions 8.1 and 8.2. It edits an existing image from written instructions, builds new images from up to four reference images, and repaints or extends selected areas in the Editor. It replaces the older Omni Reference and Character Reference features.
According to Midjourney's official Edit Model documentation, there are four ways in. You can drag images into the Imagine bar under "Attach to prompt", click Quick Edit or Open Editor on any image you made, open the Edit tab in the left menu and upload an image, or add --edit plus image URLs on Discord.
The biggest mental shift is the prompting style. Classic Midjourney prompts describe a scene. The Edit Model also accepts direct instructions, like "make this a real person" or "show this from behind". You can mix both styles, and knowing when to use which one is most of the skill.
The 24 September update matters too. Midjourney says targeted inpainting and outpainting now change only the pixels you select, so you can make repeated edits without the image quality degrading. Before that update, stacking five or six edits often left the image soft and muddy.
Which edit mode should you use for each kind of change?
Pick the mode by asking what must stay fixed. If one region is wrong, use inpainting. If the frame is too tight, use outpainting. If you want a new image built from your own assets, use multi-image reference. If the change fits in one sentence and the whole image may shift, use instruction editing.
Here is the decision list I keep next to my screen:
- One specific area is wrong (a hand, a logo, a stain on a shirt): open the Editor, paint over only that area with the Erase brush, and describe what belongs there. Everything outside the mask stays as it is.
- The image is right but the frame is wrong (square post that needs to be a 16:9 banner): use outpainting in the Editor and set the new aspect ratio. The original pixels stay; Midjourney invents the extension.
- You want a new scene using your own material (your mascot plus your product plus a location photo): attach up to four images in the Imagine bar and describe the composition.
- You want a different look, same content (photo to pencil sketch, daylight to neon night): use an instruction edit, and add a style reference or moodboard if you have one.
- You can say it in one sentence and you accept ripple effects ("make the lighting golden hour"): use a plain instruction edit on the whole image.
The common failure is using an instruction edit for a local fix. The model reads the whole image, so "fix the left hand" can quietly re-render the face, the sleeve and the background. If you can draw a boundary around the change, draw it.
How do you write edit prompts that only change what you want?
Use a three-part edit prompt: say what to change, say what to keep, then say how it should look. Naming what must stay fixed gives the model an explicit anchor, and describing the style in words helps style references take effect. Add --raw when you want your text to dominate over Midjourney's default flair.
I call it the Change, Keep, Look pattern. It works for instruction edits and inside the Editor, and it fixes most "why did it change that?" moments.
- Change: one clear instruction, in plain verbs. "Replace the coffee cup with a glass of iced lemon tea."
- Keep: list what must survive. "Keep the woman's face, hairstyle, outfit and the window light exactly the same."
- Look: describe the finish in words, even if you also attach a style reference. Midjourney's own docs recommend describing the style in your prompt when using style references and moodboards, because it helps "activate" the aesthetic more consistently.
Two parameters do a lot of quiet work. First, --raw reduces the extra creative flair Midjourney adds, so your instruction carries more weight. Second, --ar matters more than you think: the Edit Model tries to match the aspect ratio of the first image you attach, and your default aspect ratio settings do not apply automatically.
Try this prompt (instruction edit):
Replace the paper coffee cup on the desk with a tall glass of iced lemon tea with condensation on the glass. Keep the person's face, hair, clothing, hand position, the laptop and the soft window light exactly as they are. Clean commercial lifestyle photography, natural colours, shallow depth of field. --raw --ar 4:5
Notice what is missing: no long scene description. You already have the scene. An edit prompt should read like a note to a retoucher, not a brief to a photographer.
How can you turn one image into a full campaign set?
Start from one approved hero image and branch from it. Use outpainting for wide banners, inpainting for localised versions such as price tags or seasonal props, and multi-image reference for new scenes that keep your product or character consistent. Approve each step before the next so errors do not compound.
Here is a realistic workflow for a marketer who needs five assets for an autumn promotion from one product shot.
Step 1: Lock the hero. Generate or upload your best product image in a 1:1 or 4:5 frame. This is your master. Do not edit the master directly; duplicate your work from it.
Step 2: Make the banner. Open the hero in the Editor, change the aspect ratio to 16:9 and let outpainting extend the scene. Add a short instruction for the new space, such as "extend the wooden table and add soft autumn leaves on the right, leave empty space on the left for text".
Step 3: Localise with inpainting. For a Mid-Autumn version, mask only the empty corner of the table and ask for "a small plate of mooncakes and a paper lantern, same lighting". The product is outside the mask, so it cannot drift.
Step 4: Build new scenes with references. Attach the product image, a location photo and your brand mascot (three of the four slots) and write a composition prompt: "the mascot holding this bottle at this rooftop at dusk". If features of two characters start mixing, Midjourney's docs suggest combining your characters into a single reference image.
Step 5: Restyle for social. Run an instruction edit on the banner, for example "hand-drawn ink illustration of this image", with a style reference and the style also described in words.
Because each asset branches from an approved parent, a bad result costs you one step, not the whole set.
What are the common mistakes and limits of the Edit Model?
Most failures come from scope and compatibility. Instruction edits ripple across the whole image, edit results cannot go into Remix, the model does not support --tile, defaults like aspect ratio are not inherited, and old Omni Reference images need their --ow parameter removed. The interface on alpha is still changing week to week.
Here are the gotchas worth knowing before you burn GPU time:
- Remix does not work on edit results. Midjourney's documentation states that Edit Model image results are not currently compatible with Remix. Plan to re-edit instead of remixing.
- No --tile with the Edit Model. If you need seamless patterns, generate them in standard V8.1 or V8.2 mode. The 24 September update improved
--tileseams there, not inside edits. - Old Omni Reference images need cleanup. According to a SketchTo walkthrough of the Editor docs, gallery images made with Omni Reference open only in the Edit tab, and you must remove the reference and its
--owparameter before submitting edits. - Style references need words. Moodboards and srefs "may work best with some extra prompt direction in this version", per the official announcement. A bare sref with no style description often under-delivers.
- Editor work does not appear in your gallery until you upscale it. If an edit seems to vanish, check whether you upscaled it.
- Do not chase one region with instructions. If three instruction edits in a row keep breaking something else, stop and switch to a mask.
Be honest with yourself about one more limit: some fixes are not creative problems. Removing a stray object or resizing for a fixed slot is often faster in a dedicated tool or a photo editor than in any prompt.
Try it now: a 15-minute edit drill
Pick one image you already like, then run three edits on copies of it: one inpaint, one outpaint and one instruction edit using the Change, Keep, Look pattern. Compare which mode kept the important details intact. Fifteen minutes of this teaches you more than an hour of re-rolling full prompts.
Use this template for the inpaint step. Mask only the area you want to replace, then paste it into the prompt field and fill in the brackets.
Try this prompt (inpaint inside the Editor):
Change: replace the masked area with [describe the new object or detail].
Keep: match the existing lighting direction, colour temperature, shadows and camera angle of the rest of the image.
Look: [one line on finish, for example "clean commercial product photography, soft daylight"]. --raw
After the drill, write down one rule for yourself, such as "anything with a face gets a mask" or "banners always start from outpainting". That single rule will save you more re-rolls than any prompt library.
If you want more on image workflows, our guide to ChatGPT Images 2.5 Sketch and Comments covers a different approach to targeted edits that pairs well with this one.
The real upgrade in V8.2 is not one feature. It is that you no longer switch tools to switch strategy. Decide what must stay fixed, choose the mode that protects it, and write prompts like a retoucher. We understand AI. We understand you better. With UD by your side, AI doesn't feel cold.
Reviewed by the UD AI team. Feature details reflect Midjourney V8.1 and V8.2 as documented in October 2026. Sources: Midjourney: Edit Model for V8 (27 August 2026), Midjourney: Edit updates (24 September 2026), Midjourney Docs: Edit Model.
How strong are your AI skills, really?
Knowing which edit mode to use is the kind of judgement that separates casual users from power users. UD's free AI IQ Test checks where your AI know-how stands in a few minutes, and if you want to build these skills into a repeatable team workflow, we'll walk you through every step, from tool setup to process design and rollout.
