Midjourney: How I use Cref to create a big happy family like a mad scientist
An in-depth review of the Cref parameter and new discoveries
The character References (Cref) feature was released in March 2024, allowing Midjourney fans to create consistent characters using a single reference image. It is a significant improvement over previous versions, which required users to employ a series of workaround steps to ensure character consistency in photographic style.
Almost all of the Cref use cases that I’m aware focused on creating a consistent character who appears in various scenarios - one time in the city shopping, another time enjoying hi-tea with friends, and so on.
Cref in V6.1 can only be applied to one humanoid character at a time. Adding more characters to a scene requires some clever workarounds (for example, inpainting and [Pan]), which may be effective but still have some limitations.
But … wait a minute; this article is about the fun of being a mad scientist who creates artificial humans and families in Midjourney, right?
While I want to get right to the fun stuff, I need to make sure everyone is on the same page with Cref, so I'll summarize the main points.
What do we generally know about Character References (Cref)?
You can create a consistent character by referencing an image of that character.
To use the Cref parameter, add it to the end of a prompt. For example: a very cute cat --cref https://thisisacutecat.png --v 6.1
You can use the --cw (character weight) parameter to specify how strong the Cref should be. It should be used in combination with Cref in a prompt, like this: --cref URL --cw 50.
The --cw value ranges from 0 to 100. The default value is 100, which means that if you only use --cref without specifying --cw, the bot will automatically set the --cw to 100 for you.
The default or higher --cw value is effective for capturing a character's face, hair, and clothing.
If you set --cw 0, the bot will only focus on keeping the character's face consistent.
You can use one or more images of the same character as references. To use multiple reference images, add them in this format: --cref URL1 URL2.
The Cref works best with images created in Midjourney. Cref should not be used to refer to actual people.
Cref parameter is similar to regular image prompts, but it focuses on character traits. If the prompt has other regular image prompts, those URLs must be put before the --cref parameter.
Cref is unable to precisely and reliably replicate the referenced character's details, such as fashion accessories, freckles, t-shirt logos, tattoos, and so on.
Cref can be combined with other parameters like Sref to keep the style consistent.
Cref is supported by Midjourney (V6 and V6.1) and Niji V6.
Cref can be used in both the External Image Editor and Retexturing mode.
Lesser-known aspects of the Cref parameter
Unlike Style References (Sref), which support multiprompting, Cref does not. For example: --cref dogeating.png::2 dogsleeping.png::3 will return an error.
The same image cannot be used as a reference for Cref more than once. For example, --cref dog.png dog.png dog.png will only reference one image --cref dog.png. The bot will simply ignore any duplicated image references.
In this article, I will share with you how to work around the above two issues, as well as the new discoveries about Cref.
The parents
OK everyone, let's meet the stars of the day: Jack and Emily, the father and mother of our genetic experiment.
These are the prompts for creating Jack and Emily.
a front view portrait photograph of Jack in his 40's in the style of Wes Anderson --ar 3:4 --v 6.1
a front view portrait photograph of Emily in her 40's in the style of Wes Anderson --ar 3:4 --v 6.1
To maintain a consistent aesthetic style, the style-related keyword "Wes Anderson" was used instead of the --sref parameter.
Let's use this prompt template to create three more images of Jack and Emily in various poses for the genetic experiment.
{front view, head tilt pose, 45 degrees turn pose} portrait photograph of [Jack or Emily] in his 40's in the style of Wes Anderson --cref [URL] --ar 3:4 --v 6.1
Here are links to the generated images. These links will be used to create new babies (aka the new characters):
Jack's genetic assets
https://cdn.midjourney.com/47e1d2c4-f058-4fbb-ba5d-084d991acfca/0_2.png
https://cdn.midjourney.com/130c0250-c4b2-4560-a3d8-7b8bcd6098f3/0_1.png
https://cdn.midjourney.com/17fe436a-c527-4ac7-b07d-c7405816b9ee/0_1.png
https://cdn.midjourney.com/23c8f3f4-c827-45e0-8ab9-860cac873547/0_1.png
Emily's genetic assets
https://cdn.midjourney.com/b6fc573c-e3d4-4379-8dc7-a18d9046410d/0_3.png
https://cdn.midjourney.com/2cf7b53b-43a8-4187-a7e8-eac20980852c/0_0.png
https://cdn.midjourney.com/ecc8e427-934c-4635-9143-bd9081c62829/0_1.png
https://cdn.midjourney.com/faa7bd13-291c-4ad9-9cb8-5173675367ea/0_0.png
Everything is set. Now, let's combine their links (or "genes") and create artificial babies…
The family
First, we will meet with the family's eldest daughter and sons.
These characters were created using Cref and mixed URLs ("genetic assets") from their parents.
Sarah: She inherited three genetic assets from her mother and two from her father. Additional keywords (feminine, soft) were added to lessen the effects of the father's genes. Otherwise, she will also inherit her father's mustache.
Brad: His genetics are heavily influenced by his mother's.
Brad (another) inherited two copies of his father's genes and three from his mother. The keyword "fat" did not significantly change his appearance.
Brad (another): A 1:1 blending of father and mother's genes.
By combining the URLs of the characters, we can adjust the new characters' similarity to their parents.
The aesthetic style can be maintained by specifying a style-related keyword ("Wes Anderson") in place of the --sref parameter.
Blending the Cref URLs allows the creation of new characters who inherit traits from their parents.
The creation of multiple URLs for a character and then blending them with other characters' URLs in a ratio is comparable to using weights in multiprompting. This is a workaround for Cref's limitation of not supporting multiprompting.
The younger siblings
Sarah: We can also change the new character's gender by using keywords like "girl," "feminine," and "soft."
Brad: Although this young man inherited 75% (3/4) of his genes from his mother, the keywords "boy" make this character a man.
Sarah (a teenage girl): We can also use the keywords to change the age of the character.
Baby Brad: We can create a toddler or baby who shares similarities with the parents but with some limitations (see the later section).
Comments regarding the creation of feminine characters
Jack's masculine genes are extremely strong, and they must be toned down by keywords and more URL references from Emily when creating new feminine characters.
If we mix the parent genes in a 1:1 ratio with no adjustments, the daughters will inherit the father's mustache and masculine traits.
Extramarital affair and interracial relationship
Families in the artificial world are more complex than those in the real world.
Everything is possible. We can mix genes in any combination, including male-female, male-male, male-female-male, and so on.
Let's make this family more exciting by adding another gene donor: Ken, the couple's lover.
An interesting mix, no? (Haha)
Emily + Ken
Jack + Emily + Ken
Jack + Emily + Ken (2 genes/URL references)
Mum (3 genes), Ken (2 genes), Jack (1 gene)
Their mixed-blood children all look fantastic, inheriting the best characteristics from their parents.
Using the keyword "Korean" helped to make new characters look more East Asian.
This technique is useful for developing characters who are not stereotyped by race, country, or region. You can create a spectrum of new characters who look more Caucasian, East Asian, or from other races/countries.
Extraplanetary relatives
We can go one step further by including genes from the outer space.
This produces a wide range of humanoid aliens that could be used in sci-fi stories or videos.
We can control the intensity of the alien traits by varying the number of genes/URLs in the mix.
Dear paid subscribers,
I'll share with you the following paid content in the next sections:
Comparing --cw 0 versus --cw 1
The impact on Cref when we adjust its character weight (--cw)
Summary of discoveries
Additional tips and tricks for using the Cref parameter
Keep reading with a 7-day free trial
Subscribe to Geeky Curiosity to keep reading this post and get 7 days of free access to the full post archives.