Little bit of weirdness: I ran an image and it decided to just draw in maybe a third of the space beyond a few flourishes elsewhere. Decided to evolve it a second time and it just said “Nope, we’re sticking with what we got” and stayed within its self-decided borders.
Similar thing happened on the way to Landscape #10. Had to add another prompt to get it to fill in the mostly blank space.
I actually did several variations of prompts for normal “Girl With a Pearl Earring”, some of which were monstrosities but two of which I really liked. Interestingly, it rendered several patterns/shapes/textures for the head cloth, all of which were plausible. My favorite face came from “girl with a pearl earring hyperdetailed pencil drawing junji ito manga”. Except that she had two upper lips. I pasted in the head cloth cut from a second image and from a third image took a ribbon that ended up being a cool faux- Chinese/Japanese-style script. I then edited the face in a few tries, finally settling on a favorite. I ran that through an upscaling AI, but did not like what it did to the face. So I did a simple b-spline upsize, and cut out the ribbon from the AI version (which was improved) and pasted it in. Finally I cropped it to 3:4.
Liked image in stages:
Final version. (A click because of large size.)
The other one I liked, plus one of the Abominations Unto God and Man.
Another art term the AI seems to understand is tronie.
I give you troll tronie triptych:
Because it isn’t concerned with discrete objects. The way this (and similar AIs) works is that it starts with random noise, and it looks for anything in that random noise that at least vaguely kind of resembles the thing or things in its prompt. Then it accentuates those parts of the randomness, and repeats. If the bit that sort of resembled a fox was next to the bit that sort of resembled a car, then the finished fox would be next to the finished car, but if they overlap, then so will the final images. This is especially seen with things like “terrier fighting a dragon”, or the like, because on some level, a terrier is similar to a dragon, and so the AI ends up taking the same elements of the picture for both, and making it both more dragon-ish and more terrier-ish, and so you end up with a dracodog.
The random noise is part of the image generation, but to answer my question we probably need to see what the classifier, by itself, thinks matches “car fox” versus “a car and a fox”, “a car next to a fox”, “a car that looks like a fox”, “fox car”, and so on. This can be tested by feeding it clip art. I may or may not be able to get around to doing this myself. But I have been wondering about it.
I have a photo of a mushroom with two white spots on the cap. I wanted to see if I could get it to make those into eyes. I was going to use a prompt with “mushroom monster”, then went with my classic modifier and asked for “mushroom yokai”. More specifically, “Mushroom yokai full-length portrait tronie oil painting by James Gurney horror Gustave Doré Greg Rutkowski hyperdetailed charcoal drawing”. This is the photo, iteration 50, and iteration 200.
Later tries with different seeds produced one with a single eye in the middle of the cap and a face under the cap, but no use of the white spots.
Woah, that’s cool though.
What have you done to me? Friggin’ addictive. I’ve run up only a few credits so far, but I’ve created a few cool ones already.
I started with The Church of Batty’s Left Nut, but then morphed it into my old band’s name and created this album cover:
And then in tribute to my recent trip to a Cleveland Guardians game, I created “The Guardians of Baseball” using some of the added modes and filters and shit. I hope I can keep this up without buying credits (my Twitter feed may become irritating at best).
More salvaged failures:
It seems to associate the original Greek name for the element mercury – hydrargyrum – with greenery. The amount of greenery being in direct proportion to the weight of the prompt.
Those are remarkable.
I’ve been goofing around with evolving my images through various iterations of the filters and the artistic/coherent thingie. I’ve also discovered the deep levels of filters (or whatever you call them) in the Modifiers menu. Here are my latest two.
Superman in the Hospital with a Broken Leg - I ran this one through several iterations of the Marvel Comic / Pop Art / Photo-realistic modifiers and I couldn’t really get what I was looking for, but this one is interesting:
https://creator.nightcafe.studio/creation/6LrigUs9QG9BaoJUAjuL
And then there’s Man’s Inhumanity to Man, first run with the Candy and Heavenly filters, artistic. Then with just the keywords run through the Coherent filter and then a gate to hell opened behind the refrigerator when this image appeared:
This was the original image:
An artistic rendering of the art on my arm … my new tattoo, with keywords American Traditional Dagger and Rose, run through the Artistic filter, with Matte weighted at 1.0 and Steam Punk weighted at 0.5:
Landscape #12 started as an attempt to get something resembling a jumpsuited Elvis. Here’s the image process:
It boggles my mind how it can generate such detailed images from noise (or very low-detail downsampled start images) in just 200 tries or less. I would have imagined something more like 20,000.
And the details that it invents—some of my images have added objects that look like little bronze or silver spheres, with plausible lighting for a shiny sphere and even hints of environmental reflection in them. Or like the middle image below, attempting to produce reflections in a wet surface.
(First image is the start, second has the reflection, third added mainly to make it wider than tall, but also to demonstrate what can happen when you simply plug the word “origami” into the prompt, as I did with both of these.)
Here it is with “tattoo art” added and with Coherent. It is…somewhat different.
https://creator.nightcafe.studio/creation/40M4T5QoG5FNmzNjXpGw
But again, look at the invented details—it produced an object sitting on top of a surface casting shadows.
Huh, I guess it thinks a guardian is an owl?
And I did some more experimenting with its understanding of concepts like “and” and “next to”. For my tests, I went with a dunkleosteus and a koi, as two objects that are similar enough that they could plausibly be merged, but which probably weren’t found together in its labeled training library. First, a prompt of just “dunkleosteus”, to make sure it knows what one is (all of these are oil painting, coherent):
OK, swimming thing with big jaws and teeth, check.
“dunkleosteus koi”:
As expected, it merged them: Bright orange fishhead with jaws that are way too big.
“dunkleosteus next to koi”
…Maybe? The bluish thing in the foreground is probably a dunkleosteus, and the grayish thing above it might be the head of another, larger, dunk, but the only trace of “koi” is the orange thing in the top right, which looks more like a ladybug than any sort of fish.
“dunkleosteus | koi”
Another merger. Weird, but definitely a merger.
“dunkleosteus and koi”
Not sure what happened here… The only fish in the image looks more like a bass than a dunk or a colorful carp, and then there’s some reeds, or maybe a heron, behind it? And maybe a roaring T. rex in the top left?
Later, I’ll try the same experiment with cats and dogs. There, I expect the “next to” and the like to work much better, because there probably are a lot of images in the library that start off with a cat next to a dog, and which may well be labeled “cat next to dog”.
“dunkleosteus, koi”
(dunkleosteus) (koi)
dunkleosteus
Koi
(Separate prompts, both at 1.0)
Not to resort to begging, but if I get one more ‘like’ on this one I’ll get some free credits…
As the Dall-E review pointed out, multiple subjects happens to be something this approach is bad at, at this stage. Maybe there are some tricks, but there is only so much you can do by fiddling with the prompt. (The masking approach may be the way to do it right now; first just ask for a dunkleosteus, then freeze that part of the canvas and have it put a koi somewhere else. If Nightcafe gets around to implementing that feature.)
Done!
