Midjourney opened its first V8.2 image edit model on August 27, 2026. The announcement calls it a test and warns of “a lot of edge cases here”.
It covers four jobs that used to need four tools and retires two older features. Everything below comes from the announcement and the two pages it links to; no page publishes a per-edit cost.
What the Edit Model is
The documentation lists three: instruction edits on an existing image; generation “using up to 4 reference images (replacing Omni Reference and Character Reference)”; and the Editor for “inpainting” and “outpainting”.
The four capabilities the announcement lists
The release note splits the same ground into four points: instruction edits, generation from up to four references, region edits and canvas expansion, and personalization, moodboards and srefs — with the caveat that the last two “may work best with some extra prompt direction in this version”.
Versions and tools it combines with
It runs on versions 8.1 and 8.2; V8.2 is the default model, per the version guide.
| Works with the Edit Model | Excluded |
|---|---|
| Image Prompts, Style References | --tile |
| Moodboards, Personalization | Remix on edited results |
--hd, up to four reference images | A fifth reference image |
Four ways to start an edit
On the website
Drag images into the prompt bar under “attach to prompt”, or click edit in the lightbox’s lower right. A third path is the Edit tab, for uploaded and external images. In Creation Actions, “Quick Edit” sends an image to the Imagine bar and “Open Editor” opens the Editor.
In Discord with --edit
Discord takes the same job as a parameter: add --edit to your prompt “then pasting your image URL. If you want to use multiple images, separate each URL with a space.” The image must be “already online”.
make the jacket red --edit https://your-public-host.example/portrait.pngInpainting and outpainting in the Editor
The Editor page states the change: “Inpainting and outpainting now use the new Edit Model for V8.X.” The tools did not change. Paint offers an Erase brush for the region to regenerate and a Restore brush to bring pixels back; Layers stacks images and regenerates “only the areas of your image with visible transparency”; Move / Resize handles scale and rotation.
Two boundaries matter: Retexture is built into the model, and the old version needs V7 or earlier. Omni Reference images “can only be opened in the Edit tab”, and --ow must be removed first.
Smart Select, in five documented steps
- Press “Smart Select”.
- Click an area to set the first positive (Include) point.
- Add positive and negative (Exclude) points to refine it.
- Add points until the mask covers what you want.
- Apply with “Erase Selection” for the inside, or “Erase Background” for everything outside.
The docs add one warning: use either button “before submitting your image edits”.
Style controls that work with an edit
Personalization, moodboards and srefs all run through the model, and the docs recommend describing the style in the prompt too, so the aesthetic has something to attach to.
The model also changes how a prompt reads: instead of describing a scene you can hand over an instruction, “similar to Conversational mode”. Both stay valid; the prompt guide covers the descriptive half.
--raw for prompt-led edits
Raw is the documented lever for precision: “When you add --raw, it reduces the extra creative flair the Midjourney usually adds, allowing your prompt text to have more influence.”
hand drawn pencil sketch of this image --sref 2708276657 --rawWorked examples: three edits from the official page
Each prompt below is the wording the official page prints under its own example image. Paste it into the Imagine bar with an image attached.
Example one: an instruction instead of a description
make this girl into a real characterNothing is masked and nothing is described twice; the sentence names the change, the model finds it. Perspective work follows the same pattern, with the printed prompts let's see this image from the front and let's see this image from behind.
Example two: composing from references
make this silver knight ride the horse of fire holding the sword of flowersUp to four reference images feed one composition — the capability that replaced Omni Reference and Character Reference. When attributes bleed between subjects, the docs fix it on the input side: “combine your characters into a single reference image.”
Example three: retexturing with a style reference
hand drawn pencil sketch of this image --sref 2708276657Retexturing is now a function of this model: the reference does the look, the sentence does the medium. Note where the style reference guide puts --sref and the parameter list puts --edit — both are suffixes on one line.
What the official pages have not published
- A per-edit GPU cost: any figure you see is someone else’s estimate.
- Reference image details beyond the count: format, file size, resolution.
- A guaranteed preservation syntax;
--rawis described as influence, not as a lock. - How
--editcombines with other parameters on one Discord line. - Whether the test ends, what it costs, which plans may use it.
One official conflict is worth knowing. The Editor page’s visibility section says Edit model creations “will automatically populate to your Create and Organize pages”, while its tool list says layered and uploaded images stay invisible until you upscale them. Read together: standard edits arrive on their own, the other two need the upscale.
Frequently asked questions
What did the Edit Model replace? Omni Reference, Character Reference and the separate Retexture tool. --oref is still listed but marked as retired in V8.X.
Does inpainting need a new tool? No. Inpainting and outpainting moved into this model; the Editor’s brushes, Smart Select and Layers drive it.
Can I still remix an edited image? Not currently: “Edit model image results are not currently compatible with Remix.” Start fresh instead.
Which image sets the frame? The first one you attach; saved aspect ratio settings do not apply. Add --ar to the prompt to change it.