Follow Me Down the Rabbit Hole (Patreon)

EDIT 11/26/23:  There was one more character I tried during this round of experiments, but didn't want to post the results until I got the OK from her creator.  Scroll to the end to see!

Hello, Legionnaires.  In a separate post I admitted to doing some experiments with an AI image generator this past week.  I thought I'd go a little more in depth with my experiences doing that by going over each image.  If you'd like to see the images first and see if you recognize the characters I tried to recreate, seek that post out here before proceeding.  Otherwise, read on!

Julie

Project QK's reigning MVP seemed like a reasonable place to start, but man, was she a mess to deal with!  Being a more public facing service, the Bing AI generator has content filters, and if your prompts generate an image that is flagged by one of those filters, it won't show the image.  While most of my entries would give me four options, my prompts for Julie would only give me one, or none at all.  I thought at first it had to do some how with racial concerns--I specified first Korean (most accurate) then when that didn't work, Asian.  It seemed odd to me though that it would prevent me from specifying a whole ethnic group.  I thought maybe the outfit was partly to blame-- a little too "schoolgirl fetishy" perhaps for the filter's taste.  So I tried switching to a different image:

Kinda nice having the background fleshed in like that, yes?  But none of the images still really said "Julie" to me.

I move on to other characters and at the very end circled back to Ms. Kim.  I tried changing "skirt with suspenders" to "pinafore dress" and "plaid" with "tartan."  Perhaps most helpful, however, was giving Julie something innocuous to do besides lounging on the couch, so I added the smart phone and briefcase.  That seemed to work and I was able to add the ethnicity back in.  One of those attempts is the image above.  Still, I never quite got all the elements to add up the way I wanted.  The sleeves would be the wrong length, the tie would be missing, the shoes open toed or the wrong color, she'd be sitting on the floor, or whatever.  I also started with a night sky and full moon in the background but realized that wasn't really accurate to the moment I was trying to capture.  I switched to sunset, which is above, and then dusk, which I liked better but none of the results for Julie herself were quite as good.  Our Ms. Kim remains elusive it seems.

As a bonus, I did take a swing at creating another shot of Julie as she would be a few minutes later:

Jenny

Recreating a panel from one of my classic works seemed like the next logical step.  But of course the content filter was not crazy about the idea of a distressed woman tied to a metal chair so I had to take out the ropes.  Still, with the pose and camera angle you get the idea.  Comic art filter this time.

On a side note, my first foray into AI was to see if I could illustrate a sequence using it.  Sadly I learned the hard way that unless you save the images they drop off, and the whole thing was lost.  Probably just as well, in the end.

Bobbie

I did this in a number of styles and they all came out pretty much the same.  Some were running more intensely or had a more extreme expression than the others, but the color palette and overal gist was the same.  I think this one is a "Pixar Style" one.  I chose it because the hair was the right length and the tie was included.  I did find the Army base background to be a nice enhancement.

Luna

Luna was fairly straightforward for the most part, but she's also a great example of how much trial and error is involved in this process.  This was the only one with the right length hair and the off the shoulder design of the dress.

Dana

Another rising star in the QK Pantheon. Dana came together pretty simply.  Some of the passes gave her too much of an MTV supermodel look, but overall it nailed her 1980's vibe, if not the details.  I could not get the AI to leave her shirt untucked, for example.  Also, fun fact:  Bing AI apparently has no idea what stirrup pants are.  Every full body shot it generated gave her loose fitting slacks that ended just below the calf.

Sandy

Sandy was another one with trial and error and one where getting all the different elements--pose, outfit details, expression, background--to happen at the same time was tough.  Also, the earliest passes had her disturbingly scrawny for some reason.  Not sure how I got out of that but I'm glad I did.  In the end there was a LOT of images that I like this round... it was hard to pick just one to feature.

Evie

This one was fun and gave me a lot of good results.  Got the basic gist of the character pretty much right out of the gate.  I had several to choose from, including a realistic one and a comic book art inspired one, but I went with this one--the Pixar filter again-- because Evie was actually inspired by a Pixar production--"Toy Story of Terror" which opens with a scene from a fictional vintage monster movie where a vampire is chasing a girl through a graveyard.  "Betsy" inspired the look for Evie quite a bit.

Jan

Could not end this experiment without the Canine Crusader herself!  I was a little nervous going into this after my experience with Julie.  Jan is actually biracial, so I tried first just putting in "youg woman with brown skin and reddish brown curly hair in a ponytail"  but it did not look right.  So I added in the African American prompt and of course the results were much better.  Getting the right hairstyle was a fight, though.  And OMG--while most of my shoe prompts went ignored (because the feet often did not appear in the image), the IA was determined to get Jan's purple running shoes in the shot even if she had to hold them or they were lying on the ground somewhere behind her.  Weird.

This isn't based on any specific image of Jan.  I took a typical outfit of hers and put her in an autumn setting as I've done in the past.  I did try to recreate another specific image, which incorporated her alter ego a bit, but man did that give me a mixed bag of results.  The original image:

One of the notable results is the image at the top of this post.  Here's another:

It's worth noting that at no time did I ever use the word "werewolf" as a prompt when creating these two.  But between the pointy ears, and teeth, weird eyes, grey skin and the fact that I was trying to get the wolf head logo on her costume with mixed results, it didn't take much for the AI to put two and two together, apparently.

And finally:

Battle Duck

Swung for the fences with this one.  I did try a couple of other costumed characters but quickly decided against pursuing that.  One of the controversies with AI is the way it learns by analyzing and incorporating other people's art.  If I generate an image of, say, Galaxy Girl and the AI incorporates some new design element to her costume that I really like, I'd have to ask myself if I'm stealing someone else's design work to make my character look cooler.

But Ducky here is so out there I had to see what the AI would come up with.  I mean she's got a helmet shaped like a duck's head, for crying out loud.  I was a little concerned after the "She-Wolf in burning office" debacle what kind of monstrosity it would come up with, but I'm happy to say that almost all the passes captured her spirit nicely.  Leading off the prompts with "big smile" probably helped.  A lot of them actually skipped the helmet altogether.  I chose this image because it did make an attempt and I like the solution it came up with.

Ericka (Added 11/26)

So any of these images were the result of a good deal of trial and error, but not Ericka Samuels, the alter ego of the Savage Girl, created by Boggs Chaucer.  Nearly all of her trials looked good.  Having something concrete and easily scraped off the internet like the 70's era NYPD uniform shirt surely helped.  Not so much the Savage Girl herself, who I tried add to one of the Ericka images in a sort of Incredible Hulk vol 1 number one or Savage She-Hulk #1 homage,  but somehow "tall muscular woman with wild long dark brown hair, tan skin, white eyes and leopard skin bikini" gave me a gorilla in leopard print briefs, I kid you not.  It was something I tried quickly at the very end so I'm sure some more trials with just the Savage Girl could yield better results.  It's interesting I'm having similar problems with CroMagna, although as that character is more brutish in appearance that's understandable.

So after all this, what have I learned?  That I can see both sides of the AI argument.  All the trial and error, trying to get the right combo of prompts can leave someone with a sense that they created something from their head, even if the heavy technical lifting was done by the software, based on the works of other artists.  I can understand why they'd want to share their results, just as I did.  I'm glad dA has made an attempt to filter out the AI generated art, because while a lot of it is very cool, I hate having to second guess whether it's a cool image someone programmed or if it was created from scratch by a visual artist.    Allowing people to post by requiring them to identify the method of creation so it can be filtered out is as good a compromise as any.

I see the validity of the pro AI arguments.  It has useful applications as a source of inspiration or reference.  It does allow individuals who don't have the skill, time, and resources to learn art a chance to bring their ideas to life.  And it's just kind of fun to do.

But the downsides are also real.  AI can undermine the artists who have taken the time to learn to create something from scratch.  I especially feel for the digital painters because a lot of the AI work I see most closely resembles theirs.  And of course we're already hearing stories of concept artists and graphic designers who are losing work because a studio or business can just hop on MidJourney and ask the AI to generate a poster or a logo..  Maybe a human has to fine tune it, but that's still less paying work for the regular artists.  And it's addictive!  I'm always wanting to try one more variant.

And, at least with the AI generator I was using, a lot of the glitches are still evident.  You may notice Ducky has six fingers, for example.  Some of the images where I added perspiration, the sweat would appear on the skin and the clothes.  Random extra limbs pop up, or Julie's sitting halfway down the couch, or Sandy appears to be phasing through the stone railing.  Buttons pop up in random places, and so on.

I think for me personally,  I can use AI as a reference tool and that's about the extent of it.  Rather than google endlessly for the right interior or so on, I can generate a concept image, then refer to that in my drawing.  I will need to resist the temptation to generate a background then drop it behind one of my figures.  The biggest issue I have with AI is that it can only do what it does by assimilating artwork created by the very artists it threatens to replace.  So can I (and will I) use it to generate images that I then can refer to in my own drawings, perhaps to add some more variety and detail to my backgrounds, poses, and hairstyles?  Probably.  But I can't see it taking the place of my actual art.

Which finally brings me back to how I feel I fit into all of this.  As much as I wish my art was as detailed and polished as some of these images, the characters as they appear in these just don't feel the same.  For better or worse, I bring something to the table that this generator can't touch.  It's evident by the fact that all of you are here, and that I still have people who are ok with paying me to bring their idea to life.  I think I'm a little lucky that I'm not a higher profile or more professional artist.  I've carved out a small little niche for myself, and it's one that AI can't quite seem to fill. At least not yet.

If you've stuck with me this long, thanks for reading.  What are your thoughts on all of this?  Let me know in the comments, as the influencers would say.

More "real" art to come soon!

(Ericka Samuels appears courtesy of her creator Boggs Chaucer.)