Midjourney V8 Character Consistency Without --cref

Midjourney's own compatibility chart has a column that quietly ends an era. Next to "Character Reference," under V8.1 and V8.2, there is no checkmark. There is a no-symbol^[1]. The same row shows a checkmark under V6. Character Weight, the --cw dial that decided how much of a face carried over, is marked the same way.

So if you upgraded to V8 and your --cref prompts stopped behaving, nothing is broken on your end. Midjourney character consistency on V8 is a different job than it was on V6, and the flag most tutorials still teach was retired two versions ago.

TL;DR

• Midjourney's official feature chart marks Character Reference and Character Weight as unsupported on V8.1 and V8.2. Both are supported on V6^[1].
• The Character Reference doc itself now redirects you elsewhere: "When using V7, use Omni Reference instead"^[2].
• Omni Reference is not the V8 answer either. It is documented as compatible with V7, and adding an omni image "will automatically run the prompt in V7"^[3].
• What V8.2 did invest in is Personalization. The release notes describe it as an update focused on aesthetics, image quality, and Personalization, which "understands your aesthetic tastes much better"^[1].
• Style Reference, Style Weight, Image Prompts, Image Weight and Seed all remain supported on V8^[1]. Consistency on V8 comes from stacking those, not from one flag.
• A widely shared V8.2 workflow does exactly that, then hands the character off to a second model to lock the sheet before any video is made^[4].

What the official chart actually says

Midjourney publishes a feature compatibility table comparing V6, V7, and V8.1/V8.2. It is the least ambiguous document the company has on this subject, and it is worth reading literally rather than through the lens of older tutorials.

Four rows matter here:

Two things follow. The --cref and --cw pair is a V6-era feature, already gone by V7. And Omni Reference, the thing that replaced it, carries an asterisk on V8: the chart says "Uses V7," and the Omni Reference doc confirms that supplying an omni image runs the prompt in V7^[3]. You do not get a V8 render with an omni reference attached. You get a V7 render.

That single detail is what trips people up. They attach a reference, the job succeeds, the output looks slightly off from the V8 work they were doing an hour ago, and they blame the prompt. The version silently changed underneath them.

V8.2 became the default version on July 24, 2026, and V8.1 shipped on April 14, 2026^[1]. Any tutorial written before those dates that teaches --cref was accurate when published and is not accurate now.

Why the flag disappeared instead of improving

The V6 approach treated a character as a bundle of visual attributes to be copied at an adjustable strength. That is what --cw was: turn it to 0 and you kept mostly the face, turn it to 100 and you dragged hair and clothing along too^[2].

The approach had a ceiling. Midjourney's own guidance told you to start from an image Midjourney itself generated, warned that photos of real people "typically won't look exactly like them," and cautioned that details like freckles or logos "might not come out exactly right"^[2]. It was a copy mechanism fighting a model that was never copying in the first place.

Omni Reference generalised the idea to characters, objects, vehicles and creatures, and cost twice the GPU time of a regular V7 image^[3]. Then V8 arrived and put its weight behind Personalization instead, where the model learns your taste from ranked images rather than reading one reference at generation time.

I have no inside knowledge of why the decision went that way, and Midjourney has not published a rationale I could retrieve. What the documentation supports is narrower and enough to act on: the flag is gone on V8, and the replacement is a different mechanism with different ergonomics.

What still works for Midjourney character consistency on V8

The chart's checkmark column is more useful than its no-symbol column. On V8.1 and V8.2 you still have Personalization, Style Reference, Style Reference Codes, Style Weight, Image Prompts, Image Weight, Multi-Prompting, Seed, Chaos, Stylize, Raw and Tile^[1].

Four of those do real work on consistency.

Personalization is the closest thing V8 has to a character system, and it is where V8.2 put its effort. Profiles built for V8.2 draw on "a much larger and improved pool of images"^[1]. A profile code is a stable modifier you can carry across every prompt in a set.

Style Reference holds the rendering language steady. It transfers palette, linework and texture from a reference image without importing its subject, and Style Weight controls how literally. This is what stops a set of images from looking like ten unrelated renders even when the figure changes pose.

Image Prompts with Image Weight put an actual picture into the prompt and let you decide how much it outweighs the words.

Seed is the underrated one. The chart marks Seed on V8 as "99% identical" rather than a plain checkmark^[1], which is an honest way of saying reproducible but not bit-exact. Same seed plus same prompt gives you a repeatable starting point to iterate against.

None of these is a face lock. Used together with a written description you keep constant, they get you a coherent set, which is what most projects actually need.

The three-model workflow people actually run

A V8.2 workflow posted on August 8, 2026 by Kōda, a creative technologist, shows how this plays out in practice. It has drawn 57.5K views, 719 likes and 481 bookmarks^[4].

The stack is Midjourney v8.2, GPT Image 2, and Seedance 2.5. The Midjourney prompt is short enough to quote in full:

anime girl --ar 2:3 --profile geutxpd x2o8dm8

Two words of subject, an aspect ratio, and a Personalization profile. No character reference, because there is none to use. The consistency lives in the profile code.

The second step is the interesting one. Rather than asking Midjourney to produce a turnaround sheet, the character is handed to GPT Image 2 with an instruction to "create a 2:1 premium cinematic CHARACTER IDENTITY BOARD using Image A as the sole character and style reference"^[4]. Then Seedance 2.5 animates from that.

That division of labour is not arbitrary. Midjourney is an aesthetics engine with a strong opinion; it is not the tool you reach for when an image must contain small legible labels in a fixed layout. An identity board is exactly that kind of image. Using an instruction-following model for the board and Midjourney for the art works around each model's real limits instead of arguing with them.

The author describes the piece as built from "one character reference and one environment reference"^[4] — reference in the everyday sense of a source image carried through the pipeline, not the retired --cref parameter.

Running the same stack without Discord

Every model in that workflow is available here, which means you can run the pattern without a Midjourney subscription, a Discord account, or three separate billing relationships.

Midjourney v8.2 and v8.1 run from a normal form on seedance2.so/midjourney, with the native flags as real controls: aspect ratio, stylize, chaos, weird, quality, raw, draft, HD, seed, style reference and style weight. One run returns up to four images and is billed once, whether four come back or content review trims it to fewer.

Speed is the only billing dimension. Relax, fast and turbo change queue priority and price, not the rendered image, so exploration costs less than finishing does.

For the board step, GPT Image 2 sits at seedance2.so/text-to-image and costs 1 credit for a 1K image. For the animation step, Seedance 2.5 is at seedance2.so/seedance-25. Current credit costs for every tier are on seedance2.so/pricing.

Two caveats worth stating plainly. Personalization profile codes are tied to a Midjourney account, so a --profile code from your own account is not something a third-party form can mint for you. And because V8 has no character reference, no platform can offer you one on V8 — anyone advertising --cref on a V8 model is selling a flag the model rejects.

FAQ

Does --cref work on Midjourney V8?

No. Midjourney's compatibility chart marks Character Reference as unsupported on both V8.1 and V8.2, and as supported on V6^[1].

Is Omni Reference the V8 replacement?

Not really. Omni Reference is documented as compatible with V7, and attaching an omni image "will automatically run the prompt in V7"^[3]. On V8 the chart lists it as "Uses V7"^[1], so you trade the V8 render to get it.

What is the closest thing to a character reference on V8.2?

Personalization. V8.2's release notes describe it as focused on aesthetics, image quality and Personalization, with profiles drawing on a larger image pool than before^[1]. Combine a profile with a style reference and a fixed seed.

Can I still use Niji on V8?

No. The same chart marks Niji Version as unsupported on V8.1 and V8.2, while V6 and V7 have --niji 6 and --niji 7 respectively^[1].

Does the version change what a generation costs?

On Midjourney's side, HD carries a documented GPU cost difference: 1.3 minutes versus 0.8 minutes for SD^[1]. On seedance2.so, price moves with speed tier only, never with version.

How many images does one Midjourney run return?

Up to four, billed as one request. Content review can trim the grid, so fewer than four sometimes come back for the same cost.

What aspect ratios does V8 accept?

The chart lists a maximum of 14:1 for V8.1 and V8.2, dropping to 4:1 for HD images^[1]. Aspect ratio is a free-form width-to-height value rather than a fixed dropdown.

Why did my V8 images change style after I attached a reference image?

Most likely the version switched under you. An omni reference forces the job to V7^[3], and V7 renders differently from V8.2.

Where that leaves a character-driven project

The honest summary is that Midjourney character consistency is no longer a parameter you set. It is a habit: one Personalization profile, one style reference, a seed you write down, and a written description you refuse to reword between prompts. Teams shipping character work on V8.2 are pairing it with an instruction-following image model for anything that needs exact layout, then animating from there.

That is more moving parts than --cref was. It is also less fragile than a flag that can be retired between versions. If you want to try the pattern end to end, the Midjourney step starts at seedance2.so/midjourney.

References

1. Midjourney. Version — feature compatibility and comparison chart. Retrieved August 2026 from docs.midjourney.com/hc/en-us/articles/32199405667853-Version
2. Midjourney. Character Reference. Retrieved August 2026 from docs.midjourney.com/hc/en-us/articles/32162917505293-Character-Reference
3. Midjourney. Omni Reference. Retrieved August 2026 from docs.midjourney.com/hc/en-us/articles/36285124473997-Omni-Reference
4. Kōda [@aimikoda]. Midjourney v8.2 + GPT Image 2 + Seedance 2.5 workflow. Posted August 8, 2026. Retrieved August 2026 from x.com/aimikoda/status/2085803816553685091