Now, With More Cthulhu


One of the nice things about the AI art generators is that they can be entertaining in their own right.

I have a lot of fun farting around with this stuff. Maybe it’s just me.

Dall-e2 and Marcus Ranum “gigantic and terrifying cthulhu rises from the sea behind cute new england beach house. in the style of thomas kincaid, detailed painterly”

Cthulhu doesn’t like Kincaid either.

In case you haven’t played with these things, here’s a bit about how it works. You type in what you want and it gives you 4 options.

cthulhu chibi style 3d rendered cute white background

Then you can drill into the variation you like:

There are a lot of tricks you can play with this stuff. And one of the things that’s extra fun about the AI is it doesn’t understand when it doesn’t understand so it just forges ahead regardless. There is no “huh?”

“reference https __labs.openai.com_e_nWv5WKlnljC60jyFXbex8pbq place chibi cthulhu into thomas kincade new england beach scene”

I am irritated that all the various AIs try to prevent people from producing erotica, pornography, or fakes of existing people. I guess that’s to keep people from producing images like:

“cthulhu eating Joe Rogan’s head in the style of caravaggio”

But it’s also pretty cool to punch song lyrics in and see what comes out:

I pick up my guitar and play, just like yesterday, then I get on my knees and pray we don’t get fooled again in the style of caravaggio

Comments

  1. astringer says

    OK, so I fell for it: tried

    “A hawk sending a text wearing a kilt in the style of monet”

    and… actually, that’s quite good.

  2. says

    “Maybe it’s just me…” The hundreds of images on my hard drive and my external hard drive say it’s not just you. And they’re not all just cute animals in turtlenecks. I’ve done the song lyric thing too, as well as just song titles. It’s also finally let me put one of my Ello accounts to good use. https://ello.co/minimummaus

  3. Tethys says

    I’m sure it’s highly entertaining, which is exactly why I don’t let myself use it.
    I would waste way too much time playing with it, rather than using my large supply of art stuff and practicing my arty skills.

    The Caravaggio is lovely, and clearly a Caravaggio.

    Hokusai is a master of form and line, his art has inspired me to learn about woodcut printing. Eventually all my research and ideas will coalesce into printing/painting myself some fancy art-nouveau style wallpaper, since the price of hand painted wall paper is well out of most peoples budget.

  4. says

    “I would waste way too much time playing with it, rather than using my large supply of art stuff and practicing my arty skills.”

    One of my peeves is people who plug in a few prompts and call themselves an “AI artist”. No, you’re not an artist. There are artists who use the tools and get amazing results, and photographers who can put all the technical details in can get really incredible shots. They’ll still get hands with 10 fingers though.

    I may change my mind though as the technology progresses and the ability to describe something really makes a difference. I still won’t be an artist for typing “Pomeranian puppy in a turtleneck sweater –ar 2:3 –v 4” but someone who is able to put in 500 words of description and technical notes might be.

    My own thing is I do have enough artistic ability that if I weren’t too lazy to practice. I appreciate this technology for giving me an outlet for my imagination.

  5. Ridana says

    Why does AI have such a problem with polydactyly and other malformations of fingers? I’d think there are enough references for them to have sorted that out by now, but I can’t recall ever seeing one of these that produced proper human hands.

  6. Tethys says

    Ridana proper human hands

    Yes, the hands are always messed up, as are many faces. Maybe that’s why every art student spends time being required to specifically draw hands?

    They are surprisingly difficult to render realistically.

  7. says

    Ridana, it’s not just fingers, it’s toes as well. After seeing someone’s attempts the other day it occurred to me that it might be deliberate to keep the feet guys from taking it all over.

  8. Reginald Selkirk says

    @7 Why does AI have such a problem with polydactyly and other malformations of fingers?…

    Aren’t there human artists who, for example, draw their nudes in a field of wheat so they don’t have to render the feet?

  9. says

    That would excuse some deformity, but at some point the computer program needs to have “the human hand normally has four fingers and a thumb” inputted into it.

  10. Tethys says

    I had never considered the foot guys being creepy as a reason why the hands are often horrifically deformed by AI.

    If you look at “ Father Coughlin plotting with the secret congressional cabal to overthrow the government” which illustrates Marcus’s previous post, the hands are a mess, and the faces have been run through an Orc filter AFAICT.

    Perhaps that was exactly what Marcus was trying for with that composition?

    The zombie eyes of Coughlin, that ear-jaw fusion, and the pronged nose of those Orc’s on the right are especially wrong.

  11. Dunc says

    That would excuse some deformity, but at some point the computer program needs to have “the human hand normally has four fingers and a thumb” inputted into it.

    As I understand it, these are not really “programs” in that sense, and not only do they not know how many fingers a human hand should have, they don’t know what fingers, humans or hands are. They have no conceptual model behind what they’re doing, they’re just statistical engines mapping text descriptions to blobs of binary data. There’s no understanding invovled.

  12. says

    Dunc@#14:
    As I understand it, these are not really “programs” in that sense, and not only do they not know how many fingers a human hand should have, they don’t know what fingers, humans or hands are. They have no conceptual model behind what they’re doing, they’re just statistical engines mapping text descriptions to blobs of binary data.

    It’s chains of probabalistic conditionals. I believe modern AI models are way more complicated, but basically, they’re still evolved from simple Markov chains: you build a tree of what is likely then walk down the branches until you reach the end. From what I’ve read, the big game today is deciding whether to preference old knowledge, or knew knowledge in the training model. One might argue that’s the inherent divide between progressive thinking and conservative but I won’t go there.

    Anyhow, here’s my bet: the AI “knows” that arm-shaped things should have hands on the end. But a hand, with its fingers, is a collection of arm-shaped things, that should have fingers on the end. So you get these hands that look like they are sprouting fingers on their fingers because that is what the AI sees as most likely.

  13. says

    Tabby Lavalamp@#4:
    https://ello.co/minimummaus

    Some of those are amazing! If you don’t mind, I’m going to feature three of my favorites here. If you object, lemme know and I’ll take them down.

    I’m curious how you got the faces so correct and clean. Are you using that face post-processor? I know another artist who uses a face-beautifier designed for “influencers” with some effect.

  14. says

    Marcus, not a problem at all. Midjourney V4 is really good at getting faces right the first time, especially if you keep the image relatively simple. The one with the worst results above is the busiest image with more than one person, cars, buildings, etc. There are also things you can add to the prompt to try to nudge things towards realistic. Beyond that I’ll often remaster quite a bit until I get good results, though sometimes I do have to just give up because that third leg just isn’t going to go away and the face starts to melt.

    One of the things I really enjoy with Midjourney is the Daily Prompt on their Discord. You must use the word or words in your prompt, and you get to see the prompts everyone else uses and try some versions of them yourself.

  15. macallan says

    Markus @16
    Gah, let’s try this again

    One thing I noticed, when asking for different artists’ styles, if it didn’t know the style – I asked for some rather distinctive manga artists – it reverted to Hokusai. Not sure if it remembered me asking or if that’s its catch-all for anything japanese.

    Then I tried chiaroscuro and hello kitty just to see what it does or doesn’t recognize.

    Then there’s van Gogh, creepy as fuck.

  16. says

    So I decided to join in the Cthulu fun and decided to do some compare and contrasting with Dall-E and Midjourney. I did Midjourney twice, once without specifying which version and once with specifying V4. Not only did the non-specified version give me smaller images, it hints at Cthulu more than anything else.

    https://labs.openai.com/e/GPgcdwDOHgtuEgDe3CG2p7s5

    https://cdn.discordapp.com/attachments/1010091205599711244/1047196659551043724/TabbyLavalamp_cthulhu_rising_from_the_sea_mt._fuji_in_the_backg_b1870f70-4750-4f8e-bf5c-8c7597296d31.png

    https://cdn.discordapp.com/attachments/1010091205599711244/1047197360779972688/TabbyLavalamp_cthulhu_rising_from_the_sea_mt._fuji_in_the_backg_ad8a411d-9f2a-4377-bda3-922f211d1513.png

  17. Tethys says

    It is curious how the AI changes the image based on the word prompts. Polydactyl tentacles seem appropriate for eldritch beings.

    Woodblock printing is an ancient art in China and Japan, and the technique itself gives the art a distinctive style. Hokusai and ‘The Wave’ are well known examples. Hiroshigi is another master of the technique of Ukiyo-e. (floating world)

    https://en.m.wikipedia.org/wiki/Ukiyo-e

    I’m not sure of the artist or era of these examples, but there is a cluthu like thing looming over a samurai in a ship, on a wave, with Mt Fuji. I expect the AI noticed too.

    https://palmermuseum.psu.edu/exhibition/ukiyo-e/

    ———
    @ Tabby

    Wow, the Emma Woodhouse image is gorgeous!

    @everybody
    I remain convinced that I would spend way too much time playing with such a toy, but I am enjoying all your shared images vicariously.
    Thanks!

  18. cvoinescu says

    Dall-E has terrible problems with faces being made of hands, feet being made of hands, and hands being made of… too much hand? To me, that looks like feature extraction during training was too local — that is, it captured the idea of a finger, but not the idea of hand-with-fingers. That, and too much feedback, or recurrent convolution (Deep Dream-style), again with too-local training, and no “understanding” of a whole face or a whole hand. I’m not surprised that fingers are problematic in a system with a piecewise view, because of repetition: this leads to slippage, probably more or less in the same way DNA replication has trouble with tight, small repeats.

    Also, the default skin color for people seems to be white, which just goes to confirm how perfectly unbiased computers are.

Leave a Reply