At The Night Cafe

For absolutely no other reason than I wanted to share them, I thought I'd throw together a very quick post using some of the AI images I generated when I was trying to get a masthead for my latest music post. All of them used the title of the post as a prompt, sometimes with an additional phrase or keyword. 

Usually it only takes me a couple of tries to get something I'm happy with, mostly because I tend to tag most of my prompts with something like "Pop Art" or "Anime" or "1970s Comics" or "Hanna Barbera" or "Poster Art", which tends to force the results in a direction I'm predisposed to find acceptable but for once, instead of accepting the first half-decent one I saw, I carried on for a while, using a whole range of algorithms out of curiosity to see what they'd get me. I even used some of the images as seeds for others and "evolved" those results through several iterations, something you're supposed to do but with which I rarely bother.

I thought some of them were pretty good, at least if you like wonky cartoons as much as I do. Others were interesting but not really what I wanted. A few were really boring, those being mostly photos. 

I really don't like the photographic results these things kick out. They're either incredibly bland and generic or unecessarily disturbing. I generally try to add a keyword to the prompt that makes sure the result is a line drawing or a painting not a photograph but a couple slipped through anyway.

The very first result was so extremely specific I find it hard to believe it isn't a mulched version of some pre-existing work. It also doesn't seem to have anything much to do with the prompt, which also makes me suspect there's some missing meta-tag responsible, possibly relating to the content or context of the putative original.

Prompt: "All The Things And Then Some More" AI: DreamShaper XL alpha2

Don't you wish you could read all those speech bubbles? 

Next out of the bag came something far more relevant and really rather good:

Prompt: "All The Things And Then Some More" AI: Juggernaut XL v5

Now that's definitely "all the things" although just what things they are is a little less clear.

My third attempt resulted in the image I actually used, which you can see at the top of yesterday's post. That came from the same prompt again, using SDXL 1.0. I think it's a lovely illustration. It looks a little like a ligne clair artist tasked with imitating Richard Scarry for a 1950s magazine ad. The whole thing is beautifully poised and balanced, both in the positioning of the items and the color palette. If I was the art director I'd be looking to add whoever drew it to the permanent staff.

It was because I was so impressed with the first three that I kept going. Sadly, my luck took an immediate turn for the worse. I won't waste your time with the results - some utterly tedious photos and some murky thing in a wood. 

Disgusted with those, I picked a semi-random style from the extensive list, one I hadn't used before: Splash Art. The result was very unexpected:

Prompt: "All The Things And Then Some More splash art" AI: Stable Diffusion 1.5

So that's splash art. Who knew? I did a couple more like that with other AIs but while they made for interesting pictures I couldn't quite see the relevance to the prompt. Any relevance.

I took the modifier out, changed the "Noise Weight" to 70% and got this from the same prompt and AI:

That looked promising so I evolved it a few times but it didn't really go anywhere. I tried taking the noise weight down but that didn't help either. I added some good old standbys like "Pop Art", "Poster Art" and "Newpaper Comic Strip". That got me a few half-decent images, the best of which was probably this one:

Prompt: "All The Things And Then Some More pop art newspaper comic strip" AI:Stable Diffusion 1.5

That's using a noise weight of 70%, tending more towards chaos than order. It looks remarkably similar to the kind of collages the less talented pupils in my art class used to come up with when I was about thirteen. I ought to know. I was one of them.

At this point I decided to vary the prompt a little. I'd been sticking doggedly to just the title until then but since the post opens with a triple treat of t.A.T.u covers, I thought I'd add the name of the band on the end. I was curious to see if the AIs would have any idea what I was talking about.

They definitely did because all of them immediately started churning out photographic images of female duos. None of them looked remotely like Lena or Julia but I could see where the AIs thought they were going.

As I said, I don't really like it when the AIs use photo reference. There's something a little creepy about it, somehow, even when it's just scenery. I tried throwing in a couple of modifiers to ameliorate the photorealism, which worked - up to a point. I liked this one:

Prompt: "All The Things And Then Some More t.A.T.u pop art Sunday Funnies" AI : Stable Diffusion 1.5

That doesn't look anything like the t.A.T.u two, other than the hair color, but I love the title. If I was twenty I'd be very tempted to call my band "The The Ani Tay". Hell, I'm tempted to use it for something now.

That wasn't bad but it still wasn't what I was after. There were a few more dead ends before I hit paydirt with this next one:

Prompt: "All The Things And Then Some More t.A.T.u comic strip" AI: Stable Diffusion 1.5

Now, that's almost perfect. I love the four images in a frame, the colors and especially the esprit of the whole thing. It's purely charming and it makes me feel cheerful just to look at. There are some very odd AI touches, like the third girl having only one eye, the really bad dye jobs and that loop of string floating out of the second girl's mouth but they just add to the charm.

Still, I thought it could be improved so I evolved it. The result was fascinating.

Is that better? I'm really not sure. It's smoother and arguably less glitchy, although girl #2 has no eyes at all and #s 3 and 4 look like they've been in a fight. Meanwhile, girl #1 seems to have wandered in froma different photo-shoot entirely.

I really like the addition of the blue beaded braids and the ice blue tees, though, and the hairstyles are ace, as people probably said last time they were in fashion. I was extremely tempted to use that one for the post header but I thought I'd go back to the AI I started whith and let it have one more try:

Prompt: ""All The Things And Then Some More t.A.T.u comic strip" AI: DreamShaper XL alpha2

 

Like that first full-page comic strip, this looks too good to be true. I get the feeling that an image search might be able to track down the precise source. I wonder if that's a feature of the specific data set that model was trained on? Also, notice how the entire thing is in greyscale except for their eyes? That's spooky.

Again, I love that picture but I have a nagging feeling that's mostly because it reminds me of Love and Rockets. Hmm. I wonder what would happen if I put "Love and Rockets" as a modifier to the same prompt. 

If you look up to the top of the post you'll find out.

In the end, I went with the third image I generated so you might say the rest of my time was wasted. I got this post out of it, though, so you'd obviously be wrong there. I also learned a bit more about how to get closer to what I'm looking for and possibly more importantly how to stay away from what I'm not.

Much though I really enjoy seeing might-have-been work by real artists, I think that probably is something best kept for private viewing. It's not so much about the assignation as it is the aesthetics. It just feels a bit too much like visual karaoke to be truly cool. 

The warped images that look like the actual work of cybernetic entities who don't see the world way we do, though? Those I find cool as can be. I'll keep on trying for those.