Hi substack readers, I’ve baked up another entry in my investigation of AI art tool platforms. I’ve found they are great at doing cool things but pretty hard to do complex specific things, or make small precise changes so far. Which is critical to their being able to really replace human artists completely.
In this I compare the results they get to P-hacking and probably not really the most elegantly try to explain at the start what that is? Here’s an explainer that nails it down a lot better.
That doesn't make it useless at all, but good to keep in mind when you’re looking at an Instagram post and things Damn, we’re screwed! It’s just so much easier to look good when it’s improve and you’re rendering skills are good.
True for humans too, but here it’s what the AI's excel at. Then ask them to dial in something more specific? Well, the developers are working on it but it’s pretty rough now. Easy to make variations but there’s not the greatest control on those variations, it’s a bit random.
Here’s some of the things I’ve been coming up with thought. These first were from the prompt “Exploded diagram of a robotic Kaiju in negative bluprints cgi” typo included. it didn’t really get ‘exploded diagram’ I think, but otherwise…
”Cat like Kaiju”
I really like these, think I’ll use them to inspire a hand drawn pin up for Mind Engine.
“H.R. Giger draws botanical illustration in pencil and watercolors”
”a diagram of a brain on fire with ideas done in pencil and watercolour”
Some interesting stuff, great starting points for work I could clarify and composite.
Another thing I attempted was to emulate the challenge taken on by Nikola M. Zivkovic for Rubik's Code, “I tried to recreate a comic book using DALL-E and Alan Moore’s script”.
I didn’t have the patience frankly to get it all done, these are my attempts to get the first panel down, first with MidJourney where I used the —h ### and —w ### codes to get the right aspect ratio for the comic page. But getting it to give me graphic slashes of rain or something very clear probably would take a lot more fiddling. I both tried using Alan’s scrip directly, editing it down, and rewording it. I’ve noticed big differences depending on phrasing so was one of my first areas of exploration.
First reworded: ”--w 256 --h 448 thin black border, close on rain hitting puddles, blue grey pallet, the moment a drop strikes the water, rain falls diagonal left to right through the pane, film Noir cinema”
After this first set I used the —no ““ constraint at the end of my prompt to get rid of the sky…
“--w 256 --h 448 down shot, dark asphalt, night, close up of rain hitting puddles, blue purple and grey pallet, the moment drops of rain strikes the water, rain slashes diagonal, film Noir cinema --no sky”
“--w 256 --h 448 tight close up, surface of a puddle, silvery white ripples spreading across a shadowy black liquid, night time, large droplets of rain fall through the foreground in diagonal slashes, we see one at the moment of impact with the puddle, so close to it, it’s mid november and bitterly cold--no sky”
--w 256 --h 448 tight close up, surface of a puddle, silvery white many ripples spreading across a shadowy black liquid, night time. large droplets of rain fall in the foreground in heavy diagonal slashes. we see one droplet at the moment of impact with the puddle, so close to it, it’s mid november and bitterly cold--no sky -
Some interesting stuff but, it’s not here yet is it? This is the panel from the comics I’m aiming for.
Ok, so more variations then!….
“--w 512 --h 896 close up, surface of a puddle roadside, large droplets of rainfall in the foreground in sheets, hundreds of ripples spreading across each other, gutter side of the road, night time, we see one at the moment of impact with the puddle, freeze framed, mid November, bitterly cold--no sky - “
“heavy rain striking puddles close up, night time, many overlaping ripples radiating out over eachother, dark liquid on streets --no sky - “
There’s interesting things here, but not getting what I wanted, so I try Dall-E 2, same exact prompt at first. I have not worked out how to get it to give me vertical rectangles, if that’s even possible…
hmmm, how about “close up, surface of a puddle roadside, large droplets of rainfall in the foreground in sheets, hundreds of ripples spreading across each other, gutter side of the road, night time, we see one at the moment of impact with the puddle, freeze framed, mid November, bitterly cold--no sky”
Ok, that’s a bit more there? I might go back and try it with line art or some other stylization cue?
Next I tried more directly quoting Alan’s script, which it truncated so it only read…
”IN THIS FIRST PANEL, WE HAVE A TIGHT CLOSE-UP OF THE SURFACE OF A PUDDLE. (SEE? AND THERE WERE YOU ALL WORRIED THAT I WOULDN’T GIVE YOU ANYTHING FASCINATING TO DRAW.) WE ARE SO CLOSE TO THE PUDDLE AS TO SEE IT ONLY AS AN ALMOST ABSTRACT IMAGE OF WIDENING RIPPLES SPREADING ACROSS A SHADOWY AND BLACK LIQUID SURFACE. IT IS NIGHT TIME, AND THE RIPPLES THAT WE SEE IN THE FOREGROUND ARE CAUSED BY LARGE “
That gave me…
I tried some variations on the previous one, and it was about now that I started to loose interest? It started to feel like, I could have just drawn this by that time. This partly because I could, and may not be the case for someone who can’t draw already? But I felt like the time it was going to take to iterate myself to something closer to what Bowland had done, was not going to be worth it compared to what It would take to draw it.
I’ve done some other experiments but the short version is it’s pretty tricky getting the prompts to give good results, definitely not mindless work a cheep admin could do? Course if they are ready to take whatever then they will. But for anything specific and bespoke? Hmmm, not so great so far.
I will be revisiting this over the coming week so might have different perspectives in time but that’s it for now!
Don’t forget, you can leave me messages over on Anchor! If you have a prompt you’d like me to try, I’m up for it? Take care and see you next week!