How to Use Midjourney to Create Consistent Characters

March 6, 2026 · Editorial Team · 5 min read · midjourney ai-image character-design

Getting the same face to show up twice in Midjourney used to be a lottery. You could write the most precise prompt in the world and still end up with a completely different-looking person in every generation. That changed when Midjourney introduced --cref (character reference), and for people who do sequential illustration, storyboarding, or any kind of character-driven work, it was a genuine turning point.

This guide walks through the practical workflow: building a source character, using --cref effectively, understanding the --cw control weight, and pairing those with seeds so your character stays recognizable from image to image.

What --cref Actually Does

--cref lets you point Midjourney at an existing image and say "use this as the character reference." You paste the image URL (or drag it into the prompt box in the web editor) and append --cref <url>. Midjourney then extracts the face, hair, and general body proportions from that source and tries to carry them through.

It does not lock down clothing or background. It works on the person's appearance, not their full outfit. That distinction matters a lot in practice: if you want outfit consistency too, you need to handle that in your prompt text.

The syntax looks like this:

/imagine prompt: a woman standing in a rainy alley, cinematic lighting --cref https://yourimage.url --cw 80

That --cw 80 is the character weight. It runs from 0 to 100. At 100, Midjourney clings tightly to the reference and tends to copy the original pose or framing more than you want. At 0, it barely influences anything. In my testing, 70 to 85 is the sweet spot for faces: you get consistent features without the model just cloning the source image.

Building a Good Source Image

The quality of your reference image determines everything downstream. A blurry, small, or heavily stylized source gives Midjourney almost nothing to work with.

The best character references share a few properties:

Face is clearly visible, not obscured by shadow or extreme angles
Resolution is at least 1024px on the short side
The character is facing roughly forward or at a slight 3/4 angle
The style matches what you want to produce (if you want anime output, your reference should be anime)

I usually generate a dedicated "character sheet" image first. The prompt for that looks something like:

/imagine prompt: character design sheet, young woman with short auburn hair and green eyes, 
neutral expression, front view, white studio background, flat lighting, 
illustration style, multiple angles --ar 3:2

Run a few variations, pick the one where the face reads cleanly, then use that image as your --cref source going forward. Don't skip this step. Using a scene image (character in action, dynamic lighting) as your reference introduces too much noise.

Using Seeds for Reproducibility

Seeds are the other consistency lever. When you find a Midjourney output you like, click the envelope reaction on the message to get the job ID, then copy the seed from the response. Add --seed <number> to a prompt and Midjourney will start from that same noise pattern.

Seeds work differently than --cref. A seed keeps the general composition stable but does not lock the character's identity. --cref locks the character but lets the composition vary. Use them together for the most control:

/imagine prompt: a woman reading at a cafe table, warm afternoon light, 
cinematic, detailed --cref https://your-character-source.url --cw 75 --seed 4821903

A small table of what each parameter controls:

Parameter	What it locks	What it doesn't lock
`--cref`	Face, hair, body proportions	Outfit, background, pose
`--seed`	Composition starting point	Character identity
Prompt text	Outfit, setting, action	Face features

The practical read: describe the outfit and environment in the prompt, reference the face with --cref, and use --seed when you need to reproduce a specific framing.

Maintaining Outfit Consistency

Since --cref won't carry clothing through, you have to be explicit in every prompt. The trick is to build a short "character description block" you paste into every generation rather than rewriting it each time.

Mine for a recurring character looks like this:

young woman, short auburn hair, green eyes, freckles, wearing a dark green 
linen jacket and black turtleneck, slim fit

I paste that text into every prompt, then vary only the action and setting. Combined with --cref and a consistent --cw of 75, the character stays on-model well enough for storyboard and comic-style sequences.

Style Reference vs. Character Reference

Midjourney has two separate reference flags that people sometimes confuse: --cref for character and --sref for style. They are not the same.

--sref takes a style reference image and tries to match the aesthetic. --cref takes a character reference and tries to match the person.

You can use both in the same prompt:

/imagine prompt: a woman sitting on a park bench in autumn 
--cref https://character.url 
--sref https://style-reference.url 
--cw 80 --sw 60

--sw is style weight, same 0-100 scale. This is useful when you have a specific illustration style from another artist or your own previous work that you want to carry through the series.

Handling Multiple Characters

Midjourney's --cref technically accepts a single URL, so multi-character scenes get tricky. One workaround is to use an existing scene image that already contains both characters as your style or composition reference, then describe both characters explicitly in the prompt text.

The honest limitation is that --cref is best suited to single-character consistency. For scenes with two or more recurring characters, you end up leaning more heavily on detailed prompt descriptions and --seed for scene-level stability. It works, but it takes more iteration.

Practical Workflow for a Character Series

Here's the sequence I use when starting any project that needs consistent characters:

Write a neutral character description prompt and generate 20 to 30 variations at --ar 1:1
Pick the cleanest result (most legible face, best expression) as the canonical reference
Save that image locally and upload it to a persistent host (Imgur, your own CDN)
Build a short character description text block you'll paste into every prompt
Test the reference at --cw 70, --cw 80, and --cw 100 on three different scene prompts to calibrate how tight you want to go
Lock in your working --cw value and start production

The calibration step in step 5 saves a lot of frustration. What works as a --cw value varies depending on how stylized your character reference is versus how stylized your output prompts are.

Midjourney's character consistency tools are not perfect. You'll still see drift across a long series, especially with extreme lighting or camera angles. But --cref paired with careful prompt descriptions gets you close enough for most storyboard and illustration work without having to train a custom model. The key is investing time upfront in a clean source image.