How to keep character style consistent across Midjourney images (the 2026 way)

Keeping the same character recognisable across 5, 10, 50 images used to mean fine-tuning or character sheets pasted in manually. With Midjourney's --sref (style reference) and --cref (character reference) parameters, the workflow is now a 5-minute setup. This guide covers both, plus the multi-character scene trick that is the actual hard part.

What you'll learn

• The difference between --sref (style) and --cref (character)
• A step-by-step workflow for locking a character across 5 scenes
• The multi-character scene technique that works in 2026
• The two gotchas that quietly wreck --cref results

The 60-second explanation of --sref and --cref

Midjourney ships two distinct reference parameters in 2026:

• --sref <image_url> = "match the style of this image." The model extracts style features (palette, brushwork, mood) and applies them to the new generation. The character is not preserved; the look is.
• --cref <image_url> = "match the character in this image." The model extracts the face and body proportions and tries to render the same person in the new scene. The style is not preserved unless you also use --sref.

You can use both together:

[character description], [scene description] --cref https://your-image-url.png --cw 100 --sref https://style-image-url.png --sw 50

--cw (character weight) and --sw (style weight) control how strongly each reference is applied. --cw 100 is "match the character as closely as possible." --cw 0 is "match the vibe but not the face." --sw follows the same scale.

Step-by-step: lock a character across 5 scenes

Step 1 — Generate a "hero" image of the character

Start with one good image. I usually generate 4 candidates with the same prompt, then pick the one whose face is clean (front-facing, no occlusion, even lighting).

A prompt that works well as a starting point:

A 30-year-old female character, three-quarter view, soft studio lighting,
neutral grey background, wearing a red wool coat --ar 1:1 --v 6

Step 2 — Upscale and copy the URL

Click U1/U2/U3/U4 to upscale your pick. Right-click the upscaled image and copy its image address. You will paste this URL into every subsequent --cref call.

Step 3 — Generate scene variations using --cref

Open Discord (or the Midjourney web alpha), and for each scene, run:

[scene description], wearing the same red wool coat, three-quarter view
--cref https://your-copied-image-url.png --cw 100 --ar 16:9 --v 6

Five scenes I usually generate to test the character consistency:

1. standing in a rainy Tokyo alley at night, neon signs reflecting on wet cobblestone
2. sitting in a Parisian cafe, late morning, soft window light
3. walking through a snowy mountain pass, dramatic clouds
4. in a futuristic space station corridor, blue ambient light
5. portrait close-up, eyes looking slightly off-camera, golden hour

Run each, pick the best of 4, and you now have 5 consistent images of the same character.

Step 4 — Evaluate consistency

Open all 5 in a grid. Look for: same face shape, same hair color and length, same body proportions. It is normal for outfits to change (the coat does not transfer — see gotcha below), but the face and body should be recognisable as the same person across all 5.

If 1-2 of the 5 are off, regenerate just those with --cw 100 (the default) and a slightly more detailed scene description. If more than 2 are off, your source image is probably not strong enough — go back to Step 1 and pick a cleaner hero image.

Multi-character scenes (the hard part)

The 5-scene workflow above is single-character. The hard part is putting two or three characters in the same image and keeping them consistent and distinct.

The technique that works in 2026:

1. Generate separate --cref hero images for each character (call them A and B).
2. In the multi-character prompt, list each character by description, then add both --cref URLs with weights:

Two characters standing in a library. Character A: tall man with
glasses and a green sweater. Character B: short woman with a red
wool coat. [describe the scene and interaction] --cref
https://char-A-url.png https://char-B-url.png --cw 100 80 --ar 16:9 --v 6

The --cw weights let you bias toward one character if the model is mixing them up. Higher weight for the character the model is "losing."

Multi-character gotchas

• The model occasionally swaps faces if both characters have similar hair color or skin tone. Pre-curate your hero images to maximize visual difference.
• Hand interactions (high-fives, holding objects) are unstable in multi-character scenes. Generate the characters separately and composite if you need a clean interaction.
• Background consistency is harder than character consistency. If you need the same background across multiple multi-character shots, generate the background alone with --sref and then add characters in editing.

Gotchas

--cref does not transfer outfits

This is the single biggest surprise for new users. If your hero image shows the character in a red coat, the new images will not show the red coat — they will show a face and body, in whatever outfit your new prompt specifies. To keep an outfit consistent, you must describe it in every prompt ("wearing the same red wool coat"). For a stricter outfit lock, use a real Photoshop composite, not Midjourney alone.

--cw 100 does not always mean "exact same face"

The parameter is a weight, not a hard lock. At --cw 100, the model tries to match the face as closely as it can, but lighting, angle, and prompt wording still introduce variation. If you need pixel-identical faces across many images, use a face-swap post-processing step (FaceFusion, ReActor in ComfyUI) on top of the Midjourney output.

Style references can fight character references

If your --sref image is in a very different style (say, a watercolor painting) and your --cref image is a photorealistic portrait, the model will struggle. Pick style and character references that are at least aesthetically compatible. Both photorealistic, or both anime, etc.

FAQ

Does --cref work with v6?

Yes. As of v6 (2024 release) and continued in 2026, both --cref and --sref are first-class parameters. They do not work in v5 or earlier.

Can I use a real person's photo as --cref?

Technically yes — Midjourney does not check. Legally and ethically, do not do this without consent. For commercial work, only use images of people who have signed a model release, or use fictional characters / your own face.

How many images do I need for a good --cref?

One strong hero image is enough. Multiple images do not help; the model uses the single URL you pass. (Multi-image --cref is supported but the practical difference is small.)

Can I use --cref and --sref together?

Yes, and it is the recommended setup for character-driven style work. Use --cw 80-100 and --sw 30-60. Start with --sw 50 and adjust.

What is the difference between --cref and the old /imagine with a pasted image?

The old way of pasting an image into the prompt box did influence the generation, but inconsistently. --cref makes the influence explicit, weighted, and reproducible.

Why does my character look like a different person in scene 3?

Usually the prompt is over-specifying. If you list too many scene details, the model has less "room" to maintain the character. Keep scene descriptions to one or two short clauses.