Midjourney V8.2 Edit Model: Consistent Characters Without --cref
Short answer: since August 27, 2026 Midjourney keeps a character consistent through the Edit Model for V8 — up to four reference images plus a plain instruction — instead of the --cref, --oref and Retexture tools. V8.2 has been the default since late July, so stop pinning --v. The workflow that works: one canon image, one reference per job, one sentence per change.
What changed in summer 2026
- V8.2 default (late July) — aesthetics, quality and personalization tuning on top of V8.1 (April) and the V8 alpha (March).
- Edit Model for V8 (August 27) — replaces Omni Reference, Character Reference and Retexture; takes up to four references; inpaint and outpaint by instruction.
- Lightbox editor and Style Exploration (alpha, September 2) — a visual editing surface on top of the same model.
- Niji 7 stays the anime model; there is no Niji companion for V8.x.
- Plans unchanged — $10 / $30 / $60 / $120 per month.
Old parameter → new workflow
| Before | Now | Note |
|---|---|---|
| --cref <url> (Character Reference) | Edit Model with the character photo as reference 1 + 'keep the face and hair exactly' | Up to 4 references, one job each |
| --oref <url> (Omni Reference) | Edit Model with the object/scene photo + 'keep this object, change the setting to …' | Instruction replaces weight parameters |
| Retexture | Edit Model: 'keep the composition, change the material/lighting to …' | Same image, new surface |
| --v 8.1 pinned in prompts | No --v pin (V8.2 is the default) | Keep --ar and --style raw |
| Long negative lists via --no | Short positive frame + at most 2-3 exclusions | --no support on V8.2 is documented inconsistently |
| Multi-prompt weights (::) | One clear sentence per subject | Weights marked unsupported on V8.2 in the docs |
Step 1 — make the canon image
Generate or photograph one clean, front-facing, evenly lit image of the character: neutral expression, no sunglasses, no heavy shadows, plain background. This is the only image the face ever comes from. Store it with the project; every later edit references it. If you need a full character sheet, ask the Edit Model for a 360° turnaround from the canon image first, then use the turnaround frames as references for difficult angles.
Step 2 — one reference, one job
The Edit Model does not guess. Give each of the four slots a role in the instruction: "Reference 1 is the face and hair — keep them exactly. Reference 2 is the leather jacket. Reference 3 sets the colour mood only. Change the setting to a rainy Tokyo street at night, three-quarter view, 35 mm." Midjourney's own notes warn that mood boards "may need extra prompt direction", so name the mood explicitly rather than hoping the board carries it.
Step 3 — write the prompt like a brief, keep the parameters short
V8 reads natural language well. One subject, one lighting setup, one mood, correct framing, no contradictions. Keep --ar and --style raw for photoreal work; drop --v; treat --no as an experiment, because Midjourney's own pages disagree about its V8.2 support. Skip tag soups — "8k, masterpiece, trending" adds nothing on V8.
Step 4 — build the series
- Change one variable per edit: setting, outfit, expression, lens — never all four.
- Keep the canon in slot 1 for every edit; rotate slots 2-4 for outfit and mood.
- For a new angle, reference the turnaround frame closest to it in slot 2.
- Save the best output of each edit as a secondary reference for that outfit.
When to leave Midjourney
For photoreal influencers, Nano Banana Pro (up to 14 reference objects, five consistent people) and Seedream 5.0 Pro (pixel-level edits from a canon image) are equally strong and speak plain language. For video, take the same canon photo into Seedance 2.5 (@Image1 = identity only), Veo 3.1 (ingredients) or Kling 3.0 (Elements). See how to keep an AI character consistent for the cross-model playbook.
FAQ
What is the Midjourney Edit Model?
A V8 editing mode released on August 27, 2026. Instead of the old --oref (Omni Reference), --cref (Character Reference) and Retexture tools, you give the Edit Model up to four reference images plus a plain-language instruction, and it produces a new image that keeps what you asked it to keep — the face, the outfit, the object — while changing the rest.
Are --cref and --oref gone?
They are superseded. Midjourney's update notes describe the Edit Model as the replacement for Omni Reference, Character Reference and Retexture on V8. Old prompts still run on older model versions, but new character work should move to the Edit Model.
Which Midjourney version is current?
V8.2 became the default in late July 2026 (V8 alpha launched in March, V8.1 in April). Do not pin a version with --v in new prompts — Midjourney applies its current default. Niji 7 remains the anime model; there is no Niji version for V8.
Does --no still work in V8.2?
Midjourney's own documentation is inconsistent about --no on V8.2. Test it on your account before relying on it; the safer habit is to describe the frame positively and keep exclusions short.
How many references can I use?
Up to four in the Edit Model. Give each one a job in the instruction — this photo for the face, this one for the jacket, this one for the mood — and expect mood boards to need extra prompt direction, as Midjourney notes.
Is Midjourney still the best for consistent AI characters?
It is the best for stylized and cinematic consistency. For photoreal influencers, Nano Banana Pro (up to 14 reference objects) and Seedream 5.0 Pro edits from a canon image are equally strong, and for video you move the same canon photo into Seedance 2.5, Veo 3.1 or Kling 3.0.
Want the character brief written for you? GoldenPrompts builds Midjourney-ready prompts with the right --ar and --style raw, a locked identity block and a series plan from a few clicks. Free to start: 24 hours of everything, no card.