How to keep the same face in Midjourney?
To keep the same face across Midjourney images, give it a face reference. On V7, use Omni-Reference by adding --oref plus your image URL and tune strength with --ow. On older V6, use --cref with --cw 0 to copy just the face. Reusing the same --seed and consistent descriptions also helps hold the face steady.
Why — the first-principles explanation
By default, each Midjourney prompt starts from fresh random noise, so it invents a new face every time. To lock a face, you have to give the model a reference and tell it how tightly to follow it.
On the current V7 model, Omni-Reference (--oref plus an image URL) carries a face and appearance into new scenes, and the --ow (Omni Weight) parameter controls how strongly. On the older V6 model, Character Reference (--cref) does this, and its --cw value is the key trick: --cw 100 copies the face, hair, and clothing, while --cw 0 copies only the face, letting outfits and scenes change while the face stays fixed. Use a clean, front-facing reference image for best facial accuracy.
Two extra stabilizers help. Fixing the seed (--seed) keeps the random starting point identical, so results vary less. And describing the face consistently, same age, eye color, and features, narrows the range of faces the model produces. Combine a good reference, the right weight, and consistent wording for the steadiest face.
An example that makes it click
Imagine a makeup artist working from a headshot you keep handing them. If you say 'match this face exactly, everything else can change,' they keep the face identical while swapping outfits and backgrounds. That headshot is your reference image, and the 'just the face' instruction is like setting character weight to zero, only the face carries over, so your character stays recognizable scene to scene.
How to do it
- Pick a clear, front-facing image of the face and copy its URL.
- On V7, add --oref plus the image URL to your prompt, then adjust --ow for strength.
- On V6, add --cref plus the image URL, and set --cw 0 to copy only the face.
- Keep facial descriptions (age, eye color, features) consistent across prompts.
- Reuse the same --seed value to reduce variation between generations.
- Regenerate until the face matches, then upscale your favorite.
Key facts
- V7 uses Omni-Reference (--oref) with --ow to keep a face consistent.
- On V6, --cref keeps a face, and --cw 0 copies only the face (not hair or clothes).
- --cw ranges 0-100: 100 copies face, hair, and clothing; 0 copies face only.
- A clean, front-facing reference image improves facial accuracy.
- Fixing --seed and using consistent descriptions further stabilizes the face.
A leading text-to-image generator prized for artistic quality.
Affiliate link — we may earn a commission at no cost to you.▶ The 60-second explainer (script)
How do you keep the same face in Midjourney? By default you can't, because every prompt starts from random noise and makes a new person. The fix is a face reference. On the newest model, version 7, use Omni-Reference: add --oref and your image link, then use --ow to control how strongly it copies. On the older version 6, use --cref with your image, and here's the pro tip, set --cw to zero, which copies just the face while letting the outfit and background change. Use a clean, front-facing photo for the best match. Two bonus stabilizers: reuse the same seed number, and describe the face the same way every time, same age, same eyes, same features. Combine a good reference, the right weight, and consistent wording, and your character's face stays the same across every image.
What authoritative sources say
People also ask
What does --cw 0 do?
On V6, --cw 0 tells Midjourney to copy only the face from your reference, leaving hair, clothing, and scene free to change.
Why does the face still shift a little?
AI references guide rather than perfectly clone. Use a sharp front-facing reference, raise the weight, keep descriptions consistent, and reuse the seed.
Can I use a real person's photo?
Technically yes, but avoid recreating identifiable real people for public or commercial use to reduce legal and policy risk.
Does this work on the newest model?
Yes. On V7 use Omni-Reference (--oref) with --ow instead of --cref, which isn't compatible with V7.