Last night I was tried to get it to produce an image of a Joro (I had already had one failed attempt as one of my first images.) I got this:
I just put in “Joro”, not “Joro spider”, so it must have went with one of the other meanings of the word…
Aaaargh! Kill it, kill it with fire!
The site generated a 400x400 image and the video for 5 credits. It is one of my favorite images from the image-creating AI yet. It looks like a mother watching over a child, a tender moment with eldrich horrors. I cropped the 9:16 segment because it would make nice phone wallpaper. I then ran that cropped image through the image-generating AI again, with the prompt of “yokai”. I generated two images, one using the “detailed” output system and one with the “artistic” output system. The results from each were radically different from each other and from the original. Here are the full square image, the 9:16 crop, the two evoltuions of the 9:16 crop, and an image compairing just how much the size and detail changed from the original image to the final one.
Getting a bit technical with the numbers, but this really impresses me. The original square image generated by the first AI was 400x400 pixels. But the site upscaling option doesn’t simply duplicate pixels but looks at an image and guesses at new details to insert in those new pixels. So I had it upscale the 400x400 image to 1,600x1,600. After I passed the cropped image back through the image-generating AI I had the site upsample those images, too, to around 8,000 pixels high (each is a slightly different size). Since my original cropped image was only 1,600 pixels high, I googled for and found free AI image upscailers on the web. I ran it through two of them. One did nothing but resize the image, no AI at all. But the second did a beautiful job, inventing new details that worked perfectly in the resulting 4,608x8,192 image.
And here’s the thing: that beautiful, detailed 9:16 area would have been only 224x400 pixels in the very first image. So the final image has 420 times as many pixels as the first one, all invented along the way by AIs.
Here are larger versions of the images. And a bonus image generated with the prompt “yokai in forest at fuji”.
Everybody tired of this already? I have a few topics/observations I want to touch on but don’t want to dominate things.
Here is one. I tend to keep copies of funny/interesting comic panels when I see them. There is one colorful panel from a Usagi Yojimbo comic spinoff that I decided to try using as a seed image for all sorts of image prompts that happened to pop into my head. (I roughly blotted out the speech bubbles because I knew neatness didn’t matter.) In every case, I used “oil painting”, the concise AI, and this image. The only thing different is the text prompt.
Here are a crap-ton of generated images.
And here is the seed image. I thought it would be fun if any of you wanted to try it, too.
(I want to thank @iiandyiiii for starting this thread—I haven’t had so much fun with an application in a long time.)
Thanks! I’m still having a lot of fun with it.
I started making some sort of sci-fi urban-ish landscape type pictures, at first just by chance but later on purpose, and I’ve been fancying that they’re all pictures of different parts of the same city or region, on an alien planet, that I’ve started to call “The Machine Kingdom”. Here are some examples:
I call the last one the “central business district”, another one the “old industrial park”, another one “city outskirts”, etc. As a sci-fi author this has been an immensely satisfying creative outlet, as well as sparking some ideas for new stories.
Personally, I’m more interested in figuring out how it works than I am in creating art, so I’m mostly firing it up when I think of some new experiment. And I haven’t been sharing all of those experiments, because I don’t know how interesting that is to others.
I would like to do so as well, I just have not have time to put my money where my mouth is, but there are a number of pre-trained models online, so you can see what effect changing the training set/training hyperparameters has. I think this would be more instructive than merely going through their web site.
I know how you feel. Some images you (the Royal “you”) make are just kind of funny for how weird and bad they are. But some are amazing. Case in point: last night I tried to see if I could recreate the first image I posted (I think post #78) in widescreen if I used the seed value from the original image. The result was wildly different (but in a way similar? both involved a character in the central-ish foreground with a mask-like face with large eyes—it is the most similar image that has been produced in that regard—maybe that is meaningful?) but the creature created was amazing. I evolved that one and te result was also amazing. (The third version was going a little down hill.) They seem like compelling, complex characters to me. At first I thought Generals of oposing armies, but then I decided that they were brothers who were soldiers in the same army and edited them together. I would love to see something with them as characters (even if it isn’t a Ken Burns style documentary: Dearest Grrzhlar, Today we remain under siege just outside of Ul’lug’tak. Once more we resorted to eating the rotting flesh of our fallen comrades for needed sustainance. I dream of the day we might once again feast on the fresh flesh of our fallen enemies…)
Here’s another one from last night that I genuinely love.
I edited the original slightly to put the moon behind the buildings. I thought about rounding it out, but I decided that since that moon obviously isn’t Earth’s moon, that’s just the shape that it is.
“A lively futuristic city at night.” Drawn by VQGAN. Only changed the model.
Wikiart version:
S-Flckr version:
So no enormous difference there. Then I would have to do each one a bunch of times, compare different kinds of prompts (e.g. the “faces” model probably does not know about architecture), etc.
BTW, the AI upscaler there is amazing. No other upscaler that I’ve found on line gets such good results. I wish it could be used on any image.
Whatever it’s using obviously can be used on any image. Question is, which model is it; I bet you can find that exact one running on a different site.
I was going to post some links, but there are many, many, many, e.g
https://iterative-refinement.github.io/
https://upscale.wiki/wiki/Main_Page
More cities (LAION-400M, latent diffusion):
And I obviously meant “I wish I had access to that engine for any of my images”, not if it is technically possible. And I have already tried at least a half dozen sites that are findable by Google and free to use. I’m not an idiot. (The output of all those sites is garbage compared to the Nightcafe AI.)
Not only do I not think that, I am operating under the assumption you know a lot more about it than I (in particular, how to find various sites and links). What I meant was 1. Somehow figure out which algorithm he is using, by going through his blog posts like this one I clicked a little bit on Reddit but haven’t identified it yet. (2) there are sites, including but not limited to Google Colab, where you can run literally any algorithm you want.
I decided to experiment some more. I prompted an image that I hoped would have the kind of fine details that I’m concerned about (“Straw scarecrow in a field of dry winter corn”.) I downloaded the original 640x384 image, then upscaled it 4x to 2560x1536 on Nightcafe and downloaded that. I then ran the original through a number of on-line “AI enhancement” sites from the top Google hits.
Some of the sites I couldn’t use because I had already used my free allotment in earlier experiments. Some I didn’t use because they demanded that you sign up with the site even for offered free samples. Some were buggy and I couldn’t get them to work at all. And one site apparently was trying to do the processing on my phone. The speaker would produce static as it was “processing”, the processing onky advanced with the window in the foreground, and at the rate that it reached 5% it would have tanen around 30 minutes to an hour to finish the single image.
When I had a batch of processed images I zoomed in on one detail of all of them. Imglarger is very good, and seems to use the same AI as Nightcafe. (I had produced one excellent upscale before, but could never find a site to reproduce that quality again. I think Imglarger.com is that site.) Every other site is just plain garbage.
(I know that the AI invents new details and not recover lost details in a forensic CSI “zoom and enhance!” sense, but I’m good with that.)
Recently saw The Andy Griffith Show mentioned in a thread, so it gave me mind to plug that in.
That is probably the most realistic human face I’ve seen yet. (Andy, not Goober. Is that Goober?)
Most recognizable, maybe, but @Knowed_Out 's Katherine of Aragorn painting was, I would say, more realistic. The AI seems to have a tendency to over-emphasize facial muscles.
(and I think the other Mayburrian in your picture is Barney-- He seems to be wearing some sort of official hat. Though he might be an amalgamation of multiple characters)
“M*A*S*H”, however, didn’t work out so well…
I think the figure in the middle might be wearing Col. Blake’s hat? And the… thing… on the right has Hot-Lips’ hair?
Whereas this one, I think, is at least good enough that you can probably guess the prompt, but it’s still definitely very impressionistic:
He’s got some big eflin ears.
