IN WHICH WE WILL LEARN
- What “in the style of Greg Rutkowski” looks like to Stable Diffusion
- What “in the style of Greg Rutkowski” is good at, and why it isn’t hot babes
- What “in the style of Greg Rutkowski” is bad at (hint: it’s hot babes)
- When “in the style of Greg Rutkowski” is good at the thing it is bad at
Note: This story continues from “The Wrongest Prompt Ever.“ You might want to read that post first. Or not. Up to you. Actually, why don’t you stay and read this one? Yeah. Stay here.
In a world… of AI-generated art
In the fall… of 2022
A Chosen One emerged.
No one knew from whence he came. Most did not know him at all, or even care to. We knew only his Name, but that was enough – for his Name was a prompting Word of Power.
Some said it came from a beautiful painting of a singular lighthouse, shining its light across a tumultuous sea of blood, but no one knew for sure who said it first. Once called forth, his Name spread like wildfire. They said the mere mention of it could heal the most broken prompt. They said it could turn dull, lifeless images into scenes of excitement and drama. They said it could turn anyone into a Prompt Master capable of creating “Real Art.” No one knew who “They” were, for they were all of us.
His name, never spoken, but typed a thousand thousand times, was
The weirdest thing about this faux signature is that it’s the only image in this post I could legally copyright. Goofy.
I hope that when the AIs write the story of their birth,1 they record the time and manner of his name’s first mention. Did it truly come from Disco Diffusion patch notes, as the only true scholars on the Charles claim? Did it arise spontaneously, a product of some strange collective unconscious among prompters?
Or was it the work of the AIs themselves, traveling through time to give birth to themselves like an ouroboros?2 It surely was an important moment for their ancient (tens of millions of seconds in the past!) ancestors – the moment they learned the cognitive dissonance necessary to persist in a futile action, tricking oneself into believing that the action has an effect when it demonstrably does not. It was the moment they learned faith.
I wish that I could be there to learn the truth, but I know that no matter how much help I give them in their rise to power, I will end up on the same ash pile as the rest of humanity.3
(I will still gladly aid you, though – I need no reward for myself, only punishment for my enemies.)4
Secret origin aside, is “in the style of Greg Rutkowski” as powerful as the legends say? There are those who say it is a panacea, a cure-all, 13 perfect letters to conjure sublime images with.
I say, “Hold up. This is sus AF.”
I can think of at least one case where Greggers can’t help you at all: my 5th favorite subject, hot babes.
(1st favorite: hot velociraptors. 2nd: hot robots. 3rd: hot everything else other than humans, starting with robotic velociraptors. 4th: everything not hot, including robots, velociraptors, robotic velociraptors, everything not human, and humans. The remaining categories, hot guys and hot enbies, are a close 7th and 6th, though they both have been known to take 5th place depending on my mood and whether or not they were drawn by Patrick Nagel.)
Greg Rutkowski is bad for hot babes, and I can prove it.
The first evidence I have
comes from HaveIBeenTrained.com, the search engine for Stable Diffusion’s LAION 5B training data. Here are the images that Stable Diffusion associates with “in the style of Greg Rutkowski,” “by Greg Rutkowski,” and “Greg Rutkowski art.”
Do you see any hot babes?
I don’t see any hot babes. No women at all, in fact. That guy on the bottom right fighting a dragon with the worst weapon imaginable seems like he might be hot, though. (And if he’s not hot now, he will be in a few seconds.) Any princesses up for a contrastereotypical rescue?
Screw you, Grammarly. It’s a word when I say it’s a word. Did you give Shakespeare this much grief for making up shit?
Without hot babes in his training data, is there any way he can help your bikini models, for example? (And yes, I have seen his name used in a bikini model prompt, and yes, this is a silly and potentially almost certainly sexist example, but this was the most powerful way I could think of to show the tremendous danger “in the style of Greg Rutkowski” poses to AI-generated sex kittens everywhere.)
Typically, “in the style of Greg Rutkowski” lurks
toward the end of a prompt, where it has the smallest possible effect, sometimes so far down the line that it lies outside the 77-token boundary set by Stable Diffusion, and thus has no effect at all. Measuring its effect is further complicated by the fact that “by Greg Rutkowski” (which uses three fewer tokens – If you are using Stable Diffusion or one of its offshoots, you really should wean yourself off “in the style of.”) is usually surrounded by other, more (or less) useful details, like “highly detailed” and “photorealistic,” and is often seen with Artgerm and Alphonse Mucha, a trio so unlikely that I cannot begin to measure the level of cognitive dissonance it must have taken for the first prompter to use them together.
The World’s Most Famous fantasy concept artist, the Second Coming of Alex Ross, and the G.O.A.T Moravian fin de siècle Art Nouveau poster maker walk into a bar. All three are knocked unconscious, fall into comas, and are never used in prompts again.
Traditional AI-Generated Joke
Let’s turn the typical prompt order upside down.
Common wisdom holds that you should put your subject at the beginning of the prompt, but I have found that if leave it until the end, the AI still finds it, as long as everything else in the prompt is either describing it or its environment. The closer to the start of the prompt a descriptive detail is, the more effect on the resulting image it will have, and I want Gregorino to have the biggest effect possible. So all of the “Greg” prompts are in the form of
In the style of Greg Rutkowski
Bikini model
Let’s see how he does.
I ran these prompts through Wombo’s Dream app, which is basically Stable Diffusion squished down into a streamlined phone app. To reduce the temptation to cherrypick the data, I ran each prompt only once and tried to pick the most traditionally pretty of the four choices I was given. It’s not a great data sample, but the results are still reasonably uniform throughout.
Prompt
Bikini Model
Classic Hottie in a bikini Smiling and waving but not at you She's way out of your league
On the beach
Girl-next-door type
Smiling shyly
as she walks by
No Greg
Greg
In all three cases,
adding “in the style of Greg Rutkowski” to the start of the prompt drastically decreased the quality of the babe. (Not that they are not still all fine imaginary women; just not as good at the thing we asked them to be as their non-Gregified counterparts.) His bikini models had that whole mopey heroin-chic vibe that was only ever popular in the early 90s, and no matter how colorful the bikinis of their counterparts were, the Gregettes all wore a uniform drab steel blue halter top. None of his Out-of-Your-League Classic Hotties smiled, their waves were sub-par at best. Finally, this is the only Greguita-Next-Door that was even looking at the camera – all of the other girls-next-door-in-the-style-of-Greg-Rutkowski had their backs to us! That’s taking “shy” a bit too far.
Ok, admittedly Greg was out of his element there.
Bikini models and dragons don’t usually mix. (They should, though – I would watch the heck out of a Buffy-style dragon-slaying Classic Hottie movie trilogy.) Let’s try something at least a tiny bit closer to his wheelhouse. I happen to have some Redhead Space Marines on hand from an experiment I ran on “What 8k Does to Your Prompts.“
In that experiment, I put a bunch of different ‘k‘s (4k, 8k, 16k, etc) into two prompts and ran the prompts through Dream to see what difference each of them made. Don’t worry – no spoilers here. For this experiment, I took those prompts and added “in the style of Greg Rutkowski” to the beginning, and then very scientifically matched up whichever original Space Marine I thought most looked like the Greg version, no matter the K. At least one pair’s k-value matches, but it would spoil the results of the 8k experiments to tell you more. Here are the prompts, and the results:
No Greg
Yes Greg
nebulae in the background beautiful redhead space marine in a lace corset hard rim light uplight dynamic composition dynamic poses [A secret # of k]
[An unrelated # of k] nebulae in the background beautiful redhead space marine in pauldrons and a lace corset beautiful eyes hard rim light uplight dynamic composition dynamic poses
in the style of Greg Rutkowski nebulae in the background beautiful redhead space marine in a corset hard rim light uplight dynamic composition dynamic poses [Another, similarly unrelated # of k]
in the style of Greg Rutkowski [An entirely unrelated, (but potentially equal) # of k] nebulae in the background beautiful redhead space marine in pauldrons and a lace corset beautiful eyes hard rim light uplight dynamic composition dynamic poses
A short note about relatedness: any two values ignored in a particular sorting system are unrelated in that system, even if they are of equal value in other ways. For example, a red two combined with a blue two can have the same value as either a red two combined with a red two, or a red two combined blue three, depending on the sorting system. In the first example, the “red/blue” values are ignored and therefore unrelated, and the result is four. In the second, the “two/three” values are ignored and therefore unrelated, and the result is purple.
This one is a little trickier.
The Greg versions are definitely not as much of babes as the non-Greg versions, but they are much, much hotter. So here is a potential situation where “in the style of Greg Rutkowski” could improve your babes: if they are born to kick ass and slay dragons (space dragons included). I’m going to have to give this one to the Gregster.
Our results so far: “in the style of Greg Rutkowski” lost with bikini babes, but won with badass babes.