“Olan Mills double exposure”…is not a mod.
It went sorta in the right direction with the tiny head growing out of an eye socket, though?
So close! Maybe actual photographs aren’t included in the dataset the learning algorithm pulls from.
Or that trait just isn’t labeled. I think they use something like this:
New feature added—symmetry. What it does is at some point in the rendering process mirrors half of the image. In horizontal symmetry it mirrors the left half, in vertical the top half, and in both you get four top right left quarters. The percentage 0 through 100 determines the point in the render it does the mirroring, the later the more identical. It doesn’t help the face problem unless it is situated directly on the mid-line, though.
It shouldn’t be too hard for them to put the axis of symmetry through the center of mass of the primary feature on the image. That… probably still wouldn’t solve the face problem, because a face that’s tilted either in yaw or roll won’t be horizontally symmetric. But it might help.
It is just a bandaid on the fact that the AI(s) can’t do good human faces. But we know it is possible. I’m sure within a few years somebody will train something that can do good artistic and realistic faces (and bodies, and hands…)
That seems a lot better than the last time I tried it. They seem to have largely fixed the problem with eyeglasses.
The AI generated art looks similar to some of the entries in this thread. The article says Carson Grubaugh entered text from C.S. Lewis’s book, and I assume he used other prompts to control the style of art generated. It looks like he had the same problems we did of keeping human figures steady enough to not look totally disfigured, but was able to properly use the context to justify the slight aberrations.
Lately I have been trying to come up with things considerably different than I have before. A seamonster attacking a Viking longboat popped into mind, but, as has been noted before, the AI doesn’t seem to think much in terms of non-humanoids.
I then tried a satyr
And a satyr parade.
Minotaurs are pretty similar to satyrs
But nymphs are all fucked up.
It tries for some snakeyness for Medusa, but it does it badly. (Ironically that first one looks more like the krasue I was trying to get weeks ago than I ever got with krasue prompts.)
So I went back to trying putting yurei in places. In a chapel. In a mosque. In a mountain lake. In space. It does a pretty fair job of understanding all of those requests.
I put some in an onsen. It understood the need for nakedness even if it doesn’t draw it well. But it does an especially good job, imo, of drawing splashing water.
Doesn’t do too badly on Chris Foss:
Not quite as geometric as the real Foss, but it got the idea. The second one could almost pass for a real book cover.
The 3 likes for this sad clown has ironically pushed me to a thousand likes which is worth 80 credits. Woohoo!
Sad clown my ass. That’s one of those evil forest clowns. And I should know because the epidemic started only about 10 or 20 miles from me.
Ha ha. In that case, you’ll love my realistic Krusty the Clown.
Playing around with a setting that can generate something like a wall covered in posters.
Here is a Theodor von Holst | Rembrandt | Margaret Keane
Here is a Robert Crumb | Kei Toume | Margaret Keane
And here is a Tim Burton | Wes Anderson | Margaret Keane
These two work as another demonstration of differences between Imglarger and the Russian supersampler. The bottom one is upscaled in the Russian AI and if you zoom in on the guy right of center you will see that it invented for him an eye with iris, pupil, and even a bit of reflected light. In the upper one from Imglarger, the eye is a featureless black hole.
(I also trimmed the margins on the bottom one. The AI, of course, didn’t.)
Did more experiments with the settings above, which are
Yurei parade | full-length wanted poster | tronie | hyperdetailed oil painting | (names of artists or other mods that you want to try)
And you can replace the word “yurei” with whatever you want to try. Zombies, Gibson girls, androids, whatever.
But one of the images produced with a vertical aspect ratio was really good (and nothing like the walls of posters) so I decided to use it as a start image in my main account. Same prompt, random seed, low noise and high image weight. Original was in 1 credit resolution but I went for 3 in the evolution. The result is one of my favorite images that I’ve produced, which I upsized to the max.
If anyone wants to recreate something created anonymously in their main account I’ve had good results with simply:
- copying the address to my browser with my main account
- evolving (not duplicating)
- dialing weighting original image to 100%
- dialing down weighting noise to 0%
- leaving everything else alone
Yeah, those new image weight and noise sliders are very useful. Would have been great to have back when I was searching for superresolution sites.
Tonight this simple prompt
Summary
Wanted poster | tronie | full-length portrait | hyperdetailed colored pencil drawing | kei toume | washi paper
Produced Moldman
I’m always wanting some images to look realistic. Not like a pencil drawing or an oil painting, but a photograph.
I decided to go very simple.
Field of wildflowers | photograph
Gives reasonably decent results.
