WIP 190 - AI give up (Patreon)

It's been a while since I've had a meaningful WIP to share, so I hope you're interested in this one.

As someone who works a lot with different comic artists, I've always eyed AI art very sceptical, because of course I do not want AI art to replace real artists. However, working with AI, I soon realized that this will not be the case for a very long time anyway.

The AI system parses requests very differently than a human would. Most of the time, if I ask an AI for a specific image, I get something back that no human would even dare to deliver to me, simply because it does not fit my prompt the way I want to. Human beings, unlike AI, also ask to clarify if they have an issue. AI just ... does.

So the standard practice in AI is not to do one picture, but instead let it do twenty and maybe there's a passable one in there. Do two hundred, maybe there's even a good one in there.

As someone who simply isn't good at drawing, I still fiddle with AI. and try to find appropiate usecases I know I can't make comics or anything with them, but I have been using AI images as background images to the character sketches. Because I can take the best snippet out of a picture, it makes it a bit easier and it "livens up" the image.

I still need to generate quite a few pictures in the "quantity makes quality" rule, because a prompt that says "hallway of a 17th century house in France" might come back as this:

Needless to say, this picture is hopelessly overloaded. In addition, the placement of some of these chandeliers is plain laughable. So you adjust the prompt and make a couple more.

The prompt "not too decorated empty corridor in blue tones of a 17th century house in France" DID give me the background of Babette's panel:

But it ALSO gave me pictures like this ...

So much for empty!

The AI sometimes just doesn't care if you tell it to NOT put something in. The word is in the prompt, so it's getting delivered. "remote lush greek island with caves and no houses" may still produce something like:

 if you tell a human artist to draw a picture of an island with NO houses and they come back with this picture, you would (rightfully so) ask if there is a money back guarantee for their services.

So sometimes AI just doesn't do what you want it to do or it forgets part of what you want it to do. For fun, I wanted to make an AI image of Syphilis and came up with a rather basic prompt of "a short and spunky tavern maid, headband, holding a tray with beers in old mugs, standing in a tavern".

 The tavern is there, all right. And the barmaid, too. But she isn't doing what she should. There's no tray and the glasses are too modern.

Now if you try to correct this by giving more information into the prompt, these instructions will simply confuse the AI and make the picture worse instead of better.

I thought maybe the commas were too confusing and I tried a couple more images with "a short and spunky tavern maid holding a tray with beers in old-fashioned mugs, standing in a tavern" in anime style. Only ONE image came out where the tavern maid is holding a tray. Well ... Kinda.

Why the hell it was decided to fill the tavern with more maids,but no customers? Well, because it wasn't specified any other way and the AI simply doesn't know better (yet). 

Also let's not go into the way the hand looks ... ugh. But that's an old AI problem - one that seemingly won't be fixed anytime soon.

And the point where you give up is when you enter prompts that clash with the presets of the AI. In that case it simply can not fulfill your request.

I tried an Espada like image and went with "Warrior woman, muscular, very short red hair with curls, freckles, leather vest, leather pants, two-handed sword" and got several variations of this particular nightmare:

 Putting the fact aside that AI can't make persons holding swords properly, all images of "very short red hair" were at LEAST shoulder length. Less was not doable.

I also wanted "A cobra that has a body looking like a rat's tail" for some reason (at the time it made a lot of sense, I swear - I think it had something to do with the term "cobrat" that I found funny).

A human artist would parse this easiily, but for AI, it seemed rather impossible.

I tried with various prompts and preset, but in the program I was using, it wasn't possible.

Maybe it's doable with patience and more experienced AI programs like DALL-E, though I doubt it after trying to have it make a four-leaved clover and repeatedly getting clovers with three leaves.

And to come to the final point of "AI give up". Sometimes you just CAN'T get what you want.

 I thought that maybe I would make a panel for Felicite as well. She's a recurring character and readers new to the comic can look at the panel and know the names. For Felicite, I do not have a "clean version", but I do have a sketch that Carmine made for me that would suffice.

After mulling it over in my head I thought that the best background image for Felicite would be a scaffold with a guillotine as she is a French noble who fled a revolution. But, alas, it was not to be, because up to now it seems the AI system has no clue what a guillotine is. Or it's not allowed to show one for some weird reason.

So, at the end of this (rather lengthy) WIP post I will show you the "most creative" images I got for "French scaffold with a guillotine in the 17th century" from the AI. Enjoy!

Creative scaffold?

Wait. Where are the things I specified in my prompt?

Okay, I didn't specify no people. I guess. Let's add "no people" to the prompt!

Ah yes. And ... uh ... interesting building. Is that my scaffold and guillotine?

So, even though the AI system is limited in its use in general, in some cases it becomes ... unusable.