Testing the new Consistent Styles Tool (Patreon)
Content
Hello friends!
Prompt: ballerina dancing over the sea --sref https://s.mj.run/uQF0uPWNirw --ar 2:1 --v 6
So I believe that most of you know that we had two HUGE updates this week: Niji 6 and the Consistent Style Tool. There is so much going on and we barely know MJ6 alpha! Niji 6 is impressive and, as usual, we have a big Niji 6 guide on works.
So I want to focus on the new Style Tool for today. I am still testing but very satisfied with most of the results. As someone who tried to train style models on Stable Diffusion, I think MJ's new tool is impressively faster and easier to use than its cousin. Of course, it is not perfect but it's good enough to give you control over your creations and include in your options styles that Midjourney never had.
Prompt: black young woman on bussiness suit, resting on her couch, golden hour --sref https://s.mj.run/oLfu1R-Ya6Y https://s.mj.run/fkUPPn55OXA https://s.mj.run/DPqoO0xlqy4 --v 6
How it Works
MJ's announcement is straight to the point:
Do your regular prompt, but after that, include --sref along with a URL to an image (see our first image)
MJ will look at the image URL as a 'style references' and try to 'copy' its style or aesthetics. It works only with v6 and niji 6.
You can add multiple urls to multiple images like this --sref urlA urlB urlC (see second image above)
Set relative weights of styles like this --sref urlA::2 urlB::3 urlC::5
Set the total strength of the stylization via --sw 100 (100 is default, 0 is off, 1000 is maximum)
Regular image prompts must go before --sref like: imagine cat ninja ImagePrompt1 ImagePrompt2 --sref urlA urlA
Let's try it
Ok, you have no idea of how happy I was when Midjourney announced this new tool. There are artists that I have been testing since v.3 and they never triggered the right results (or got something closer to its original). One of them is Shinkiro the legendary Japapense artist who has worked in the videogame industry for decades. Another artist is comic book artist Greg Smallwood, who recently won the Eisner award. Using a few images as reference and same seed gave me the following result:
The prompt was wonder woman in a pub --sref url1 url2 url3 --v 6
by Shinkiro
By Greg Smallwoold
Niji 6 responds really well to styles too. For example, in the case of Shinkiro, it got a retro gaming feeling, like an old videogram cover from the 1990s
By the way, I prefer using more than one image of the same as reference. I feel that a big dataset makes clear what you are looking for and Midjourney will have to find an average result of your images. Unless you really want to copy that specific image.
Changing style weight
Style weight helps to control Midjourney's influence on your new, but If you put --sw 0, you get your regular Midjourney result. Values around 1000 are more and more influenced by your style of choice.
For example, the next image was based on the art of a famous Brazilan artist, Ziraldo. Notice how --sw 1000 seems a bit more bold.
-- sw 0
--sw 500
--sw 1000
It's not perfect
Although I am very happy about including neglected artists in my options of style. But the truth is that this tool is not perfect. For example, Ziraldo has a distinctive style but if you use his work as a reference, you will get a generic cartoon. (see above)
Sometimes, Midjourney even decides to force its opinion. I am a big fan of comic book artist Doc Shaner. Shaner's art is clean and very beautiful, but if you do not pay attention, you can think it's just one more generic/traditional comic book style. (And I think that Midjourney does that). Style weight helps to correct that and avoids MJ imposing its opinion.
I tried to generate Lois Lane using my Doctor Shane style, and MJ was always generating Superman shirts for her. For example:
prompt: Lois Lane, centered picture, full body, streets of Metropolis --sref URL
When I changed the style weight to a higher value, my style was able to remove the S-sign from my images:
--sw 300
The generated style is even closer to its inspiration!
--
There is still a lot to explore! But which feature was your favorite?
Share with us on our comments! See ya next time!
Jos