this post was submitted on 22 Sep 2023
92 points (100.0% liked)
Technology
37719 readers
112 users here now
A nice place to discuss rumors, happenings, innovations, and challenges in the technology sphere. We also welcome discussions on the intersections of technology and society. If it’s technological news or discussion of technology, it probably belongs here.
Remember the overriding ethos on Beehaw: Be(e) Nice. Each user you encounter here is a person, and should be treated with kindness (even if they’re wrong, or use a Linux distro you don’t like). Personal attacks will not be tolerated.
Subcommunities on Beehaw:
This community's icon was made by Aaron Schneider, under the CC-BY-NC-SA 4.0 license.
founded 2 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
This is the first actually decent article I've read on AI art. It absolutely covers my concerns and fits with my own experiences using the technology. A lot of this is stuff I've been saying, and it's nice to see that I'm not alone here.
AI art is not at all straight-forward and absolutely requires creative input in order to get something usable. Prompt engineering, to me, is itself a type of art. You may at times find that there's something you want AI to generate that it's actually quite good at, like old women playing cards around a table, but usually you're going to be looking for something it struggles with. This is when you need to be creative and inventive with prompts, thinking about things from the perspective of what's probably out there in abundance.
Recently I needed to make some rings for my upcoming Planescape-themed Conan Exiles server. I started out by asking NightCafe for rings and it output a bunch of useless junk. Using my brain, I realized it's probably much more likely that it'll have a reference point for 'wedding band', and then I'll have my ring shape and can work from there. This worked quite nicely and it started producing rings, but it was often shoving them half out of the picture as it tends to do. There are some negative prompts that sometimes help with this, but I have my own technique that I think works better.
My method for framing objects in AI art generators is to surround them with something. If you add 'surrounded by' and some other object that's slightly smaller than the object you need a clean shot of, usually you'll get your image of the whole thing. Which object you pick to surround your target object with makes all the difference.
In this particular case, I first tried to put the ring on a table surrounded by small dogs, but it got caught up in the dogs and forgot to render the ring. Eventually I landed on 'surrounded by berries', and that was the jackpot. I'm guessing this one was a good choice because of Christmas themed ads, because suddenly my pictures were all full of mistletoe blasted with the kind of bokeh you only see in jewelry ads and wedding photos. And within each shot, a nicely centered ring in half-decent focus.
Now this is where I got really lucky. During my 'surrounded by berries' iterations, the generator decided to do something weird. It put a bunch of tiny little purple berries along the outer surface of a ring, standing on its side. It was perfect. I took this iteration and used it as a base, feeding it back into NightCafe and decreasing the noise ratio down to like 20-40% while turning up the prompt weight a little and changing my prompt. Now instead of berries and wedding bands, I go back to my initial search for magic rings encrusted with glowing gems. And in one step, my berries are a gorgeous array of gemstones. A few iterations later, I have a couple of decent rings to bring into GIMP and do some work with.
After pathing out the rings, separating the gems, adjusting the colors, and creating a few iterations with different colored bands and gem stones for my different finalized options, I had my results. A little blurry, okay, but totally fine for an inventory item thumbnail. Looks gorgeous.
Personally, I've dipped my toes into all sorts of art. I write, I take pictures, I sing, I play around with painting, I've made animations, games, mods, videos, weird unpalatable noise music; if it's a method of creative expression I may not be particularly good at it but I've almost certainly tried some version of it. And to me, this feels like creative expression. It feels like art.
Working with AI art feels like trying to collaborate on a project with an alien robot. You're trying to take these presumptions that we have and figure out how to get a workable result from something that fundamentally doesn't understand any of it. It's this really fascinating exercise in exploring this almost dream-like logic that's heavily rooted in media consumption, and weirdly enough your own understanding of media culture can be a way of teasing out what you want.
I don't just go and say 'please give me one rabbit' and get a rabbit. It doesn't feel like handing the project off to another artist with some notes and coming back to see what they made, it feels like an actively creative process. It doesn't feel more creative when I'm editing those images than when I'm digging through this strange robot-logic dreamscape looking for them.
And like, given that I could literally just take a picture one of my own rings and edit that, I don't see how it's less mine via creative output. I bought the ring, I didn't spend an hour digging it out of a morass of nonsense.
Frankly, I care little about law and less about money. I'll make the things I'm inspired to make with the methods I'm inspired to make them with and we'll see what happens. Maybe I'll make something people like, maybe I'll cut my own ear off and die penniless, but the opinions of the copyright office on what legal claims I can make around my work aren't really a primary factor in that or in my decision making when it comes to art.
I do hope they'll read articles like these and talk to folks with similar perspectives and find a better take, though.
Thank you for describing your process in such detail. I'm sorry if this is going to come off as overtly contrary; that's such an impersonal convoluted way to make an image. There are good illustrators out there that can sketch roughs and make a beautiful finished painting in a night, all right out of their head. Frank Frazetta would be laughing in his grave at AI art.
I'm in it to make stuff, not to impress you or any dead person. What are you making?
Right now, a game, an album, and um.. oil painted MTG Myr Tokens?
Cool!
I am not a good illustrator though, but I'm good at working with AI and Photoshop. Using generated images as a base gives me the best quality in a reasonable time.
I'm pretty sure there are musicians who would roll in their graves at the thought of generating music in synthesizers, yet without we would have electronic music.
I don't think that is a fair comparison. Electronic musicians don't outsource song construction to an algorithm that copies all the other songs on the Internet. Even though they can use midi instruments, sequencers, and samples (which do carry a known risk of copyright violation) they're still composing or performing.
The entire point of generative AI is to generate things not present in the training set by teaching it to abstract the concept.
It's a very fair comparison because in both cases you take the physical skill requirement that takes years to learn and even longer to master out of producing art. To make a good electronic song you need to compose, but you don't need to know how to physically play the instruments. To make a good image you need to know how to compose it, but not how to physically draw it.
That's a good argument. I get this. The problem that I see is that you aren't very present in the art. The AI is 100% leading you with what it knows. AI is essentially helping you create a collage of all the styles and bits of image content on the Internet. How are we going to develope new styles? A human can use their imagination and skill to create something groundbreaking and pioneering (artists had to break ground and fill the world with this art for AI to be even able to do this). AI is just going to continue to remix remixes of remixes. It's sad to me. That's not really what art is about. I'm not saying AI art isn't useful. It's a remix machine.
Now that depends on how much agency you give yourself, doesn't it? If you just give Midjourney a prompt and call it a day then yes. But the result won't be very good, will it? Similar how you could just input random notes to a synthesizer and get shitty music in return.
The problem is that the majority of people does exactly that and then shares the resulting images online, making it appear that is all there is to it. You can however express yourself artistically by using prompt engineering to get something good and than working with that to further approach what you imagine by editing the result. There are many people out there who could not artistically express themselves as they lacked the ability to translate their vision to a canvas. With the help of image generating AI they can finally express themselves. I think this is something beautiful.
While I do agree with you that our current AI image generators won't be very innovative, this is by design and not necessity.
This is what you would have gotten in let's say 2017 when asking an AI what it thinks a dog looks like (s. DeepDream).
And a couple years later you can achieve this with Midjourney.
Things are developing very fast and in the end of the day, even if we would never get an AI that can innovate art there is nothing stopping humans from just doing it themselves as we have always done over millennia. You can already greatly increase the creativity of existing image generators by tweaking the randomness factors and those algorithms don't just remix existing images, they are actually creating their own. You need the training set of existing labeled images to train the AI as it doesn't know what a frog is, nor a tree or anything really.
This is indeed a concern. If you feed too much AI generated images into the training of an image generator it causes a sort of degenerative disease in the AI that results in inferior results. Some sort of AI incest so to speak. The prevalence of AI art on the internet and the inability to reliably differentiate it from human art is proving to be a challenge for making new training sets.
I'm really good at searching Google. I'm a "prompt engineer" too