As a Midjourney beginner trying to generate pictures with a certain “look” one of the most useful stylistic tricks is the –sref function. The reference pic can’t be any old image though. The more distinct and stylized it is, the better. The simpler and clearer it is, the better. This is where real art education has value for the user, and a very broad schooling in graphic design.
In my post on creating imaginary Velazquez paintings I touched on this, as well as in my beginning post of the series, Fiddler on the Ref. In that latter I used one image to riff on a series of six different subjects. In this post I’m going to hone in on choosing and altering a suitable style reference image.
As an example I’ll use this one.
It was most likely a throwaway illustration from the 1950s or early 1960s, created and printed in haste. The artist hadn’t bothered to refine it. But that’s what gives it is charm, IMO. It’s cheap and looks cheap. It’s also of a medium that is not being used today — the quick two-color printing press job, one of the two colors being that sickly Pthalo green. Using that image, I will generate a series of pics that looks like they came from the interior pages of one of those cheap, pulpy hardback books that used to be sold at mass market stores in 1960. The ones that are yellowing today because of all the wood pulp they used.
But my pic has that pesky lettering in it, so I’ll just Photoshop it away.
(Not really Photoshop, just a free knock-off program that does the same thing.)
The lettering excision is important. AI doesn’t know typefaces from hoot. It can’t think “this is typography, therefore not part of the central image of two spacemen and a drunk alien.” To AI, it’s part of the scenery, so it’s better to erase it.
OTOH, in this particular case, the lettered version got results that were just fine, so YMMV. I guess it depends on the pic. Or Midjourney’s mood that day.
Here’s the results I got from a variety of simple subjects.
“A giant spider in a black web.” In all the generations, I didn’t get one black web. But I’m very pleased with this simple, direct, scary image which might have appeared in a Nancy Drew or Hardy Boys book.
“A happy Medieval prospector holding a pick and a bag of rocks.” Rather gnomelike if you ask me.
“Watercolor wisteria wyvern from the Lord of the Rings.” A wyvern only has two legs, but it’s a nice dragonish creature anyway.
“Taylor Swift.” To me she looks a tough lesbian on the cover of a trashy 1960s novel. In this pic, the medium is canvas (you can see the hatching) as if she’s been printed right on the grain of the cover.
“Don’t Bite the Sun.” This is the name of a SF book by Tanith Lee. I wanted to see what Midjourney did with a more abstract prompt. The execution is rougher, perhaps, because of it, like a quickly done woodcut or linocut. Note the introduction in these last two pics of a vivid red-orange ink, which was of its period but also a color Midjourney likes to inject on its own.
Even though the medium of the reference pic is obvious, it’s also wise to include it in the prompt. For example, “Timothy Chalamet” using the same –sref as the above pics, got me the pic on the left, which is clearly a photo with some 1970s-style rubylith cut overlays (don’t ask) in the background. It might have appeared on the box of a cheaply printed 1970s board game.
However, specifying the prompt to “a crude woodblock print of Timothy Chalamet” got the image on the right, which is more in the spirit of the original — a two-color illustration from a pulpy book or magazine. Yes, I know I didn’t spell his name right.
