Midjourney Character Reference (--cref) Guide

Midjourney character and style references

Midjourney v6 added two reference parameters that tackle the hardest part of AI art workflows — consistency across images. --cref keeps a character recognisable from one generation to the next, and --sref keeps a style consistent. Both take image URLs and optional weight parameters. This guide explains how each works and builds the exact parameter string for your use case.

—cref: character reference in depth

--cref <url> tells Midjourney to extract the appearance of the person in your reference image and carry it into the new generation. The companion parameter --cw (character weight, 0–100) controls how much of the reference is applied:

• --cw 100 (default): copies face, hair colour, hairstyle, and clothing from the reference. Best when you want the character to look as close as possible to the original in a new scene.
• --cw 50: balances face fidelity against freedom to change secondary details. Useful when you want recognisable identity but do not want the exact outfit carried through.
• --cw 0: copies only the face, leaving hair and clothing entirely up to the prompt. This is the setting to use when you want to put a character in a different costume or period without the reference’s outfit dominating.

You can chain multiple URLs after a single --cref flag to blend characteristics from several reference images, which is useful when your character is not captured well in any one photo.

—sref: style reference in depth

--sref <url> conditions generation on the visual style, colour palette, and rendering approach of the reference image rather than on a specific person. The weight parameter is --sw (style weight, 0–1000; default 100). Higher values make the style override more of your prompt; lower values let the text prompt carry more influence.

Multiple URLs after --sref blend styles together — useful for combining a watercolour texture with a specific colour palette from a different source.

How it works

--cref <url> [additional urls…] --cw <0-100>
--sref <url> [additional urls…] --sw <0-1000>

References must be publicly reachable URLs. Upload your reference to Discord (drag into a channel, right-click, Copy Link) or to an image host, then paste the direct image link. Local files and platform-specific links (Google Photos, iCloud) do not work.

Practical scenarios

Tips for consistent results

• Keep your text prompt focused on the scene, not the face. Describe pose, setting, lighting, and action. Let --cref handle identity — repeating facial descriptors in the prompt competes with the reference rather than reinforcing it.
• Pin the model version. --cref and --sref require --v 6 or --niji 6 or later. Older model versions silently ignore these parameters and generate without any reference conditioning.
• Blend multiple character references by listing two or three URLs to average a character whose appearance is inconsistent across your reference photos.
• Iterate on —cw before generating at scale. Test at --cw 0, --cw 50, and --cw 100 on one prompt to find the right balance before running a larger set.