Midjourney has changed so much over the last year. My process & approach to generating imagery has fundamentally shifted It used to be all about finding the right words, now it’s more about visually directing your scene I’m gonna start breaking down my process below.
There are 2 ways I explore Midjourney:
- Visually Focusing in on the aesthetic direction you wanna head in first, not worrying so much about subject matter or scene details to start.
- Semantically Scaffolding out a scene, adding visual elements w/ words, leaning into MJs training data.
I wanna explore the visual approach here, because MidJourney has new tools that make this process extremely fun & interesting. The main tool at your disposal is the “Style Reference” feature, aka --sref It let’s you fine tune the way MJ interprets your prompts.
How does –sref works
Imagine this is “style space” It lives inside of Midjourney, it’s infinite, and it’s filled with every style & aesthetic you can possibly imagine. You, the captain of that tiny little ship in the middle, navigate through these “styles” by providing midjourney with “coordinates”.
Your coordinates come in 2 forms: > Style “codes” and > Image references. When you lock in on particular coordinates, you send MJ to a specific location w/ unique “vibes” & “aesthetics”, breaking free of MJs default training data.
Now, your prompts are influenced by the location:
These locations influence the medium, color, lighting, and overall aesthetic interpretation of your text prompts.
The top 2 rows are the generic MJ outputs for the prompt “portrait of a woman”.
The bottom 2 rows are the same prompt, but with –sref 668702311 applied
Ok, so that code i just used, where did it come from? It’s literally random. There are 2^32 style codes (4,294,967,296). Each code serves as coordinates to a unique location in style space.
Here are different prompts generated using –sref 668702311. See how they’re all related?
But you don’t always need to use a code you can also provide MidJourney Bot coordinates in the form of an image reference, and images work exactly like codes. When you use an image, MJ analyzes it, & determines the coordinates in style space directly from it.
So, the image IS the code:
You can even use multiple images, blending the styles from each together to create something unique. But, if you’re new to using these, try with just 1 image first to get familiar with the effect, then try adding 1 or 2 more.
You should be able to visually decipher what each does.
Conclusion
Why style references are useful? Often times, words do not sufficiently describe a desired “look” or aesthetic you’re after. Also, words alone force you to rely on Midjourney’s training data, which comes w/ bias & more generic output/variance. Visually guiding visuals is more intuitive.
So, style references, aka –srefs, let you fine tune the visual direction of your images, giving you WAY more control over the final output.
More on all of that in the next thread Click below if you want a slow drip of random sref drops to your inbox https://nickfloats.lemonsqueezy.com.
Credits: Nick St. Pierre, Creative Director and unofficial Midjourney shill who’s publicly exploring AI & sharing learnings.
Read other articles: