That's the funny thing about it. It's no specific artist's signature (and is obvious gibberish). Rather, it's the idea of a signature.
When Deep Learning models like these are trained, the images they're based on are completely broken down into nodes with weights that decide the strength of their connections to other nodes. They don't actually do collage, because they can't. None of the original images exist any longer in the model. They're 100% atomized and vaporized into oblivion. The AI generates the entire image de novo on the spot, based on the associations between the prompt words and the collections of shapes that the model is statistically approximating.
Through the magic of latent space and manifolds, this is how you can generate things like... a very drunk and disorderly Stormtrooper eating pizza with his helmet still on:
If you mix and match bizarre, polar opposite concepts, what would the AI create?
For example, "Borg Assimilated My Little Pony" or "Medieval Mammoth Tank"? This tool has caught my interest.
I used my free image trial on Replicate, which uses Disco Diffusion. You have to pay to render more images, but each minute taken costs fractions of a cent, so I'll probably throw some spare change at it in the future.
Behold, an assimilated Twilight Sparkle and oh my god what the fuck is this--
Not that bad, but the things on his back next to the pauldrons look wierd, like they are growing out of the non-armored soft material at the joint sections.
Very weird pauldron again and oversized legs/underarmored ass and in the first pic the knee guard looks out of place/misalligned and asymmetrical, otherwise it looks quite awesome.
I can tell you exactly what would happen with Midjourney v4. The model isn't very well-trained on MLP images (I used NovelAI's Stable Diffusion for that last Twilight one, rather than MJ), but is trained on a very large corpus of images that includes a lot of IRL horses rather than cartoon characters, so when you do things like prompt it to do a pony from MLP without any further elaboration, what actually comes out is something out of a nightmare.
NovelAI's Stable Diffusion-based "Furry" model isn't well-trained on the Borg from Trek, though. Rather, it's trained on a plethora of furry images, where cybernetics tend to be drawn with a more generic appearance.
MJ does, however, know what Borg implants look like, and can put them on a dwarf horse.
For example, this is a result of "a pony from my little pony: friendship is magic assimilated by the borg from star trek --v 4"
That's kind of disappointing though, since we want it to be MLP-styled, so let's try "twilight sparkle assimilated by the borg from star trek, in the style of lauren faust, cartoon, 2d, flash --v 4"
Way closer, but still a little... odd.
Now, if you ask Midjourney v4 for a Mammoth tank, you have to consider that the associations between the text and images are built from analyzing a machine learning database of text-image pairs (like LAION).
LAION, Large-scale Artificial Intelligence Open Network, is a non-profit organization making machine learning resources available to the general public.
laion.ai
If Midjourney were to do an accurate GDI Mammoth tank, it would need, ideally, hundreds and hundreds of images of the subject in question to be in the training data set. Otherwise, if you put in "medieval mammoth tank", what I bet you'll get will be something more like "medieval tank with the traits of the animal referred to as the mammoth", like a turret with a trunk instead of a gun or something.
I just tried it, and yep, that's what you get. Not a Mammoth Tank, but a literal mammoth-tank.
This is some real Metal Max shit.
That's not quite what we're after, so let's try a more generic prompt. MJ struggles to do even a main battle tank with a dual-gun turret of any type, because there are so few photos or illustrations to use as examples for that, to train the model with. Instead, let's try regular "medieval" tanks, but in the styles of various artists.
Da Vinci tank:
Rembrandt tank:
Bayeux Tapestry-ish tanks:
And here's what you get if you do the above, but insist on caterpillar tracks and some type of cannon in the prompt:
This really showcases just how these types of AIs work. If you prompt them with multiple subjects, they blend and hybridize them like a goddamn telepod from Cronenberg's 1986 version of The Fly.
For instance, if you ask for Colonel Sanders and Keanu Reeves separately, you will wind up with fairly close facsimiles of how they actually look, but if you ask for them together, then you'll get both of them as hybrids of each other.
Prompts have to be structured a very specific way to get the desired results. You can think of each word in a prompt as a certain "influence". A prompt for an AI image generator is not an exact, word-for-word semantic description of elements of a scene. Rather, every word affects everything in the scene. You have to picture that you're adding stuff to a Vitamix blender, because it all gets blended in, tugging the entire scene this way and that. This could be good or bad, depending on the effect you're going for. Sometimes, simpler is better.
If you put "kowloon walled city --v 4" into MJ, this is what you get:
Fairly basic. Let's try, "kowloon walled city with rain and flaming 55 gallon drums --v 4".
It doesn't know where to put the fire, and instead of making 55 gallon steel drums, it produced "Kowloon Walled City style" casks. The drums took on the attributes of the other features of the prompt.
Let's try "kowloon walled city with fire raining from the sky --v 4".
It just makes a version of the city that's on fire, while it's raining, because the AI has no idea what it should look like for it to "rain fire", because there are no examples of this phenomenon in the training data.
This is how a lot of prompts are structured, unsurprisingly. Basically, just say "super detailed masterpiece rendered in a 3D program" and it will try and make something that looks similar.
Oh, sweet summer child. That is literally the least disturbing thing you can have NovelAI generate.
You can straight-up toss "Twilight Sparkle ahegao" in there and get, well, exactly that.
Oh god, reroll, reroll!
There we go.
Horror, pure horror.
This really showcases just how these types of AIs work. If you prompt them with multiple subjects, they blend and hybridize them like a goddamn telepod from Cronenberg's 1986 version of The Fly.
For instance, if you ask for Colonel Sanders and Keanu Reeves separately, you will wind up with fairly close facsimiles of how they actually look, but if you ask for them together, then you'll get both of them as hybrids of each other.
Prompts have to be structured a very specific way to get the desired results. You can think of each word in a prompt as a certain "influence". A prompt for an AI image generator is not an exact, word-for-word semantic description of elements of a scene. Rather, every word affects everything in the scene. You have to picture that you're adding stuff to a Vitamix blender, because it all gets blended in, tugging the entire scene this way and that. This could be good or bad, depending on the effect you're going for. Sometimes, simpler is better.
If you put "kowloon walled city --v 4" into MJ, this is what you get:
Fairly basic. Let's try, "kowloon walled city with rain and flaming 55 gallon drums --v 4".
It doesn't know where to put the fire, and instead of making 55 gallon steel drums, it produced "Kowloon Walled City style" casks. The drums took on the attributes of the other features of the prompt.
Let's try "kowloon walled city with fire raining from the sky --v 4".
It just makes a version of the city that's on fire, while it's raining, because the AI has no idea what it should look like for it to "rain fire", because there are no examples of this phenomenon in the training data.
This is how a lot of prompts are structured, unsurprisingly. Basically, just say "super detailed masterpiece rendered in a 3D program" and it will try and make something that looks similar.
Yep, that's because most examples of obesity in the model are human-ish (or at least biped anthros).
Fat horses don't actually look like that, though. They look like this:
It all goes in the center of the barrel, lol. Everywhere else looks basically the same.
Fat Twilight Sparkle, on the other hand, is a perfect example of that blending effect I mentioned. She ends up more upright, more anthro-looking, because that's what the most common examples of illustrated obesity in the model look like. You have to roll over and over again before you end up with something a bit more anatomically correct.
Yeah, that fat distribution will IMHO not be permitted by a horse skeleton.
Also, I don't think those horses are fat, remember, they are grass eaters, and digesting thet shit is a bitch, also, they'd need to store up fat for winter.
Decided to throw a few bucks at Nightmare AI: prompt was simply "cyberman".
Thoughts:
AI could be used in a traditional concept art design flow as a compliment to the inkblot silhouette method
Basically, random inkblot shapes are looked at for features that could be fleshed out on/look interesting, such as texture, design for insectoid legs, mechanized walker profiles, tank silhouettes, et cetera
The cyberman above had some interesting features/ideas that could be adapted by an artist when sketching out a new cyberman design
The "uneven backpack"
The glassy eye recessed into the eye socket
The borg-like implant on the right side of the head
The silvery, suit-like skin used by earlier models of cyberman (seen in OldWho), mixed with more solid metal limbs (seen in newer cybermen in nuWho)
The "crest" seen on the cyberman's forehead
The more traditional cyberman eye (the round hole) and cheek could be used in another design for a more traditional, nuWho cyberman composite
The idea of an uneven cyberman face, like how military goggles of various types have three lenses on one side and a few less on the other side, each having different functions
I'll consider doing "zerg" and "Starcraft Space Marine" next, just to see what pops out.
I'm only throwing spare change at it though (it costs like a buck and a half per image), so I dunno when I'll get around to them.
Decided to throw a few bucks at Nightmare AI: prompt was simply "cyberman".
Thoughts:
AI could be used in a traditional concept art design flow as a compliment to the inkblot silhouette method
Basically, random inkblot shapes are looked at for features that could be fleshed out on/look interesting, such as texture, design for insectoid legs, mechanized walker profiles, tank silhouettes, et cetera
The cyberman above had some interesting features/ideas that could be adapted by an artist when sketching out a new cyberman design
The "uneven backpack"
The glassy eye recessed into the eye socket
The borg-like implant on the right side of the head
The silvery, suit-like skin used by earlier models of cyberman (seen in OldWho), mixed with more solid metal limbs (seen in newer cybermen in nuWho)
The "crest" seen on the cyberman's forehead
The more traditional cyberman eye (the round hole) and cheek could be used in another design for a more traditional, nuWho cyberman composite
The idea of an uneven cyberman face, like how military goggles of various types have three lenses on one side and a few less on the other side, each having different functions
I'll consider doing "zerg" and "Starcraft Space Marine" next, just to see what pops out.
I'm only throwing spare change at it though (it costs like a buck and a half per image), so I dunno when I'll get around to them.
That's a terrible value for what you're getting. If you don't have a video card that can run Stable Diffusion locally, try getting a Midjourney subscription for $30 a month. It might sound steep, but it gets you 900 GPU-minutes of fast time a month, which is basically something like approximately 900 generations. Pretty much 3.3 cents a generation, if you use it all up (no, it doesn't roll over into the next month, unfortunately).
The results you get for a prompt like that are something closer to this:
That's a terrible value for what you're getting. If you don't have a video card that can run Stable Diffusion locally, try getting a Midjourney subscription for $30 a month. It might sound steep, but it gets you 900 GPU-minutes of fast time a month, which is basically something like approximately 900 generations. Pretty much 3.3 cents a generation, if you use it all up (no, it doesn't roll over into the next month, unfortunately).
The results you get for a prompt like that are something closer to this:
Decided to dip my toe into the AI Pool and start learning it. I'd rather figure it out early on as even if I'm an extreme grognard that still uses old-school ink pens, I want to keep on top of future art technology.
I'm taking a bit of a different path though, rather than going with trying to make portraits I have the notion of having the AI generate backgrounds in my own drawing style so I can put them behind my hand-drawn characters, since I have a tendency to laze out and leave a blank background.
In general, the AI tends to add a lot more lines that I'd like, even when I try to specify simple line drawings.
Windows in medieval buildings tend to turn into bookcases, as do doors. This took a huge number of iterations to gradually simplify and remove random extra lines and it's still only so-so.
The AI has no idea what a straw hut is, I haven't yet figured out what keywords to give it to produce any kind of primitive structures. This is the closest I've managed after dozens of iterations, and an extreme majority of attempts at producing huts give me an interior view of a corridor, and also it often just generates a bunch of squiggly concentric circles when asked to draw a hut.
Been trying to generate a castle background for a while and getting only complete failure. It invariably goes for pointy Disney Castle-style structures and can't seem to understand what I mean when I tell it I want a rock/masonry/stone wall with crenellations or any other variants of that prompt. It definitely has no idea what a machicolation is and more exact castle jargon isn't helping as it doesn't seem to know what the technical terms are in the first place.
I'm doing stable diffusion on my own system as I don't really want to operate using somebody else's server. I've run a couple different CKPT files, mostly either NAI or HD18.
In general, these AI are heavily focused on characters so getting landscapes and architecture out of them is difficult in the first place. I may decide to train it myself if nobody produces a CKPT that does that in a reasonable amount of time.
Decided to dip my toe into the AI Pool and start learning it. I'd rather figure it out early on as even if I'm an extreme grognard that still uses old-school ink pens, I want to keep on top of future art technology.
I'm taking a bit of a different path though, rather than going with trying to make portraits I have the notion of having the AI generate backgrounds in my own drawing style so I can put them behind my hand-drawn characters, since I have a tendency to laze out and leave a blank background.
In general, the AI tends to add a lot more lines that I'd like, even when I try to specify simple line drawings.
Windows in medieval buildings tend to turn into bookcases, as do doors. This took a huge number of iterations to gradually simplify and remove random extra lines and it's still only so-so.
The AI has no idea what a straw hut is, I haven't yet figured out what keywords to give it to produce any kind of primitive structures. This is the closest I've managed after dozens of iterations, and an extreme majority of attempts at producing huts give me an interior view of a corridor, and also it often just generates a bunch of squiggly concentric circles when asked to draw a hut.
Been trying to generate a castle background for a while and getting only complete failure. It invariably goes for pointy Disney Castle-style structures and can't seem to understand what I mean when I tell it I want a rock/masonry/stone wall with crenellations or any other variants of that prompt. It definitely has no idea what a machicolation is and more exact castle jargon isn't helping as it doesn't seem to know what the technical terms are in the first place.
Try Midjourney v4. It puts out absolutely amazing ink drawings.
If you have Midjourney, just put --v 4 at the end of your prompt to use the Version 4 model. Don't forget context, it really helps. That first one there has the prompt "ink line drawing of a primitive straw hut, shelter, survival --v 4"
If you don't include "shelter, survival", you get something more like a house:
