Midjourney v6: Character Consistency Guide

For digital artists and AI enthusiasts, maintaining a consistent character face across different scenes has traditionally been the hardest challenge in generative AI. You create a perfect character in one image, but in the next, they look like a completely different person. With the introduction of the –cref (Character Reference) parameter in Midjourney v6, this problem has largely been solved. This feature allows you to “pin” a specific face and body type to your prompts, ensuring your protagonist looks the same whether they are in a spaceship or a coffee shop.

Understanding the --cref Parameter

Released during the Midjourney v6 alpha phase, the --cref tag stands for “Character Reference.” Before this update, users had to rely on complex workarounds involving seed numbers, image prompts, or training custom models (LoRAs) in Stable Diffusion. Midjourney has simplified this into a single text command.

The tool works by analyzing the pixel data of a reference image you provide. It identifies key facial features such as eye shape, nose structure, and hair color. It then attempts to map those features onto the new image you are generating. It is important to note that this feature is specifically optimized for Midjourney v6 and Niji 6 models.

Step-by-Step: How to Use Character Reference

To use this feature effectively, you need to follow a specific workflow within Discord.

1. Choose Your Reference Image

First, you need an image of the character you want to replicate.

  • Midjourney Images: These work best. Generate a character you like (e.g., “A portrait of a cyberpunk hacker with blue hair”). Upscale the image you want to keep.
  • External Images: You can use a photo of a real person or a drawing, but the accuracy may vary compared to images natively generated by Midjourney.

2. Get the Image URL

If the image is already in Discord, click on it to expand it. Right-click and select “Copy Link” (or “Copy Media Link”). If the image is on your computer, paste it into the Discord chat first to upload it, then copy the link.

3. Write Your Prompt with the Parameter

Type your standard /imagine command and write the new scene you want to see. At the end of the prompt, add the tag --cref followed by a space and the URL you just copied.

Syntax Example: /imagine prompt: sitting in a 1950s diner eating a burger --cref https://image-url-here.jpg --v 6.0

When you hit enter, Midjourney will generate the “1950s diner” scene, but the person eating the burger will share the facial features of the URL you provided.

Controlling Accuracy with Character Weight (--cw)

The --cref tag keeps the face, but sometimes it keeps too much information. By default, Midjourney tries to copy the face, hair, and clothing of the reference image. If you want to change the character’s outfit or hairstyle, you must use the Character Weight parameter: –cw.

This parameter accepts a value from 0 to 100.

Strong Consistency (–cw 100)

This is the default setting. If you do not type a --cw value, Midjourney assumes --cw 100.

  • What it does: It pulls the face, hairstyle, and clothing from the reference.
  • When to use it: Use this when you want the character to look exactly the same in a different pose or lighting environment, but wearing the same clothes.

Face Only (–cw 0)

This is the “face swap” setting.

  • What it does: It focuses strictly on the facial features (eyes, nose, jawline) and ignores the clothing and hair from the reference image.
  • When to use it: Use this when you want to put your character in a completely different outfit or change their hair color.

Example Scenario: You have a reference image of a wizard in blue robes.

  • Prompt A: A wizard casting a spell --cref [URL] (Result: Same face, same blue robes).
  • Prompt B: A wizard wearing a modern business suit --cref [URL] --cw 0 (Result: Same face, but now wearing a suit).

Advanced Tips for Better Results

To get professional consistency, consider these nuanced techniques.

Blending Multiple Characters

You are not limited to a single reference image. You can merge features from two different images by pasting multiple URLs after the --cref tag. Syntax: --cref [URL1] [URL2] This is useful if you want to create a “child” of two characters or if you have two different angles of the same character and want Midjourney to understand the facial structure in 3D better.

Using Niji Model for Anime

The character reference tool is highly effective with the Niji 6 model. If you are creating a manga or comic book, this is essential for maintaining character identity across panels. Syntax: --cref [URL] --niji 6

The “Uncanny Valley” Warning

While --cref is powerful, using real photos of yourself or celebrities can sometimes result in an “uncanny valley” effect. Midjourney v6 tends to “stylize” real photos to fit its artistic bias. If you are using a real photo, try lowering the stylize parameter (--s 50 or --s 0) to keep it looking more realistic, though it will rarely look like a perfect photocopy.

Limitations to Keep in Mind

While this is a massive leap forward for AI continuity, it is not flawless.

  1. Accessories: Specific details like jewelry, glasses, or tattoos might not transfer perfectly every time unless you use a high --cw value.
  2. Non-Human Characters: While it works well on humans and human-like figures (elves, aliens), it struggles more with complex non-humanoid creatures or abstract objects.
  3. Profile Views: If your reference image is a straight-on portrait, Midjourney might struggle to invent the “side profile” of that character accurately. It is often helpful to have a reference image that is a 3⁄4 view.

Frequently Asked Questions

Does –cref work with Midjourney v5.2? No. The character reference parameter is exclusive to V6 and Niji 6 models. If you try to use it with older versions, the bot will likely return an error or ignore the command.

Can I use a URL from Instagram or Google Images? Yes, as long as the URL ends in a standard image extension like .jpg, .png, or .webp. However, privacy settings on some social media sites might block the bot from reading the image. It is usually safer to upload the image to Discord directly or use a hosting site like Imgur.

How is this different from Image Prompts? Standard image prompts (putting a URL at the start of the prompt) influence the composition, style, and colors of the entire image. --cref specifically isolates and targets character features, ignoring the composition and style of the reference image.

Does this cost extra fast hours? No, using the --cref parameter costs the same amount of GPU time as a standard text prompt.

Why is my character wearing the wrong clothes even with –cw 0? If your prompt is vague, Midjourney might still pull influence from the original image. Ensure your text prompt is very descriptive about the new outfit. For example, instead of just “walking in a park,” try “walking in a park wearing a red t-shirt and blue jeans.”