Midjourney Edit Model: Instruction Edits vs --cref

The Midjourney Edit Model opened to everyone on August 27, 2026. This is not another “V8.2 looks nicer” note. It folds instruction edits, up to four reference images, inpainting, and canvas expansion into the same V8.2 model, and it explicitly replaces Omni Reference, --cref, and the old Retexture tool. If you still lock faces with the portrait and —cref guide, move that workflow into the official editor.

What the Midjourney Edit Model actually shipped

The official Edit Model update lists four capabilities:

1. Editing images with instructions
2. Generating images from other images, up to four references at once, described as replacing omni-reference
3. Inpainting or outpainting (expand the canvas)
4. Personalization, moodboards, and —sref, with a note that moodboards / srefs work better in this build when you add extra prompt direction

The Version article adds: V8.2 has been the default since July 24, 2026, and the Edit Model also replaces Character Reference and Retexture. Pan, Zoom Out, and the Editor on V8.1 / V8.2 now run through the Edit Model. That is a replacement, not an optional add-on.

This site is an independent community resource, not official docs. Dates, flags, and entry points come from Midjourney’s pages; the rest maps them onto the jobs people actually do: lock a face, change a costume, hold a prop, expand a frame.

Four doors: drag-and-drop, lightbox, Edit tab, Discord —edit

Officially there are four ways in. Use one until it is muscle memory:

Door	Use it when
Drag onto attach to prompt	You are generating a new still with 1–4 references
edit in the lightbox	You just made the image and want an instruction or a local fix
Left-hand Edit tab	You are uploading an external file, or leaving the light editor
Discord --edit url...	You already have a URL and work in the bot

The UI is changing on both midjourney.com and alpha.midjourney.com; alpha will keep experimenting. The Discord --edit line comes from the same update post. The parameter-list page still describes --oref in a V7 context—do not treat that table as the V8 front door.

Practical order: attach one key portrait first and confirm the face still holds, then add images two through four (wardrobe, setting, prop). Dumping four vague files at once is how attributes bleed across characters.

How four references replace —cref and Omni Reference

Both Character Reference and Omni Reference now open with When using V8.X, use the Edit Model instead. Omni still notes that V7 --oref takes a single image; the Edit Model raises the cap to four, so you can hold a person, an object, a vehicle, or a non-human creature in the same job.

Migrate from the older on-site workflow with this map:

Old habit (V7 / --cref)	V8.2 Edit Model
--cref key-art-URL --cw value	Drag the key still onto attach to prompt
A second wardrobe image	Attach it as image two; spell out “wear this outfit”
--oref for an object or vehicle	Hang the object / setting on images three and four—do not feed four faces
Change only the action line for a series	Keep the references fixed; change action, camera, and light in text

Copy-ready Discord skeleton (swap in your URLs):

same young woman as the reference, now sitting at a rain-soaked cafe window,
85mm editorial portrait, cool window light, sharp eyes
--edit https://example.com/key-portrait.png --ar 4:5 --stylize 140 --v 8

On the web you will more often drag than type --edit. Either way, do not skip the text prompt. Omni docs already required words plus a reference; the Edit Model update says moodboards / srefs want direction too. Attach files with no “same face, swap this coat, night street behind her,” and the model will invent a plot.

If your gallery still holds Omni Reference jobs, the Editor article says those opens currently live on the Edit tab, and you should strip Omni Reference and --ow before submitting.

Instruction edits, inpainting, and canvas expansion

Treat “edit” as two layers:

• Instruction edits: the whole frame stays; one sentence changes wardrobe, light, props, or weather. Best when the face is right and the coat is wrong.
• Inpainting / outpainting: redraw a selection, or grow the canvas. On V8.X, inpainting, outpainting, Pan, and Zoom Out all use the Edit Model. The standalone Retexture tool, if you still need it, means dropping the default version back to V7 or earlier.

Write “keep this / change only that”:

keep the same face, hair, and camera distance,
change only the coat to a charcoal wool overcoat,
add faint street neon in the background, photorealistic

If you want less automatic beautification, --raw / --style raw still exist as general controls. The Edit Model post does not introduce RAW as a new dedicated switch—do not chase it as a fresh flag.

The Version article still warns that inpainting or outpainting an HD still returns SD, then you upscale again. Check resolution before you hand off a series.

Using —sref and personalization without mixing jobs

The Edit Model can stack personalization, moodboards, and --sref. Official copy also says this build wants extra prompt direction. In practice:

• --sref / moodboard own the grade (film still, fashion editorial, neon night)
• Images attached to the prompt own identity and objects (who, what they wear, what they hold)
• The text prompt owns what happens in this shot

Do not ask one file to do all three. If a style reference pulls the face off-model, drop --sref, keep the key portrait plus the action line, then add the grade back. Personalization got stronger in the V8.2 update; reuse the same profile while you migrate instead of rating a new one from scratch.

This is a first public test. Odd edges are expected. If a result is wrong, use the lightbox sad face or pass the jobid / URL through official channels. Fast-follow patches are already promised—do not treat today’s miss as a permanent hole.

Suggested path: draft here, edit on the official site

The Edit Model lives on the official web app and Discord, not on this site. Splitting the two stages is more reliable:

1. On the generate page, run a full prompt until you have a key still. Prompt habits still match how to write better Midjourney prompts.
2. Save that keeper as the identity anchor. For multi-scene work, add wardrobe, prop, and setting stills—four images max.
3. In the official Editor, attach to prompt or --edit. Write only the change; do not rewrite the whole appearance paragraph.
4. If the face or object drifts, remove references before you add adjectives.
5. Recheck official plan and license terms before commercial use. This site is not affiliated with Midjourney, Inc.

Bottom line: lock identity with references, not a legacy flag

The Midjourney Edit Model makes “change this image” and “use several images as references” default V8.2 behavior. What you actually change is the path: on V8, stop treating --cref / --oref as the main consistency tools. Hold identity with up to four references, change clothes and setting with one instruction, and use the Editor for local fixes and canvas. Get the still right here, then migrate in the official editor—that tracks the current default better than stacking weights on a retired flag.

midjourney AI Art (midjourney.wang) is an independent community resource for tutorials, prompts, and generation guidance. It is not affiliated with Midjourney, Inc.