Acidtester wrote:I’m not a huge fan of AI “art”, but some of these come out okay.
I have a question tho, can AI not figure out how to make a scared face? All of these girls look bored, like drowning to death in quicksand is on their to-do list, but it ain’t their whole day.
Part of the cause is that my reference images - my "studies" - use neutral model expressions. Consequently, the AI adapts those expressions, which for the purpose of what I was experimenting for, was sufficient. It's possible to input preferred tags, such as "scared" or "concerned", and the AI draws upon its database of artistic expressions which match them. Often they don't really match what we expect, and part of it is because attach a lot of subjectivity to the, well, subject matter. A "scared" face might not match as well as a "shocked" or "surprised" or "disgusted" look in the same context. You can go even further to specify exact objective expressions, such as a slight grin, or squinted eyes, or blushing cheeks. The challenge in generating good "art" is that you're not using the pen. You're understanding what to plug into the machine to get the desired result.
Plus, having excellent reference images crucial. The closer the image is to the desired output, the better and more consistent the result. That means not having to change colours as much (such as changing clothing colours or big hair colour changes, like red to blonde), the right facial expressions, the right quicksand texture, etc. What I'm doing now is saving a lot of my "failed" experiments because they contain a lot of useful references, such as clothing, texture and angles, to adapt into other pieces.
As an aside, I tried using your Ruby Roundhouse artwork to see what the machine could spit out. Might not be your thing, but here's what I got.
ruby1.png
ruby2.png
ruby3.png
ruby4.png
Believe me, it took dozens of attempts too. I can't recreate the exact costume without being uber-specific on what she is wearing, and the more you throw in, the more it confuses the AI because it means you have to be even more specific, and the AI will still be liberal in interpreting the possibilities, leading to more branches. I was never going to attempt to include the pack strap because it will never get it exactly in the same place and alignment, whereas as the artist you just paint the thing on. The gloves were interesting too. I started with "black gloves", and a bunch of images threw up bulky military combat gloves. "Black leather" worked better, but then I remembered that Ruby has FINGERLESS gloves, so that got put in too to generate these. That's really the battle (or the "skill") with AI "art".
MadMax359 wrote:it was worth asking... and i think Quicky in the mine worked great--- in the end, the supporting details are secondary to the damsel in distress--- but i know you have standards!
Yeah, well, what it did generate was pretty decent. The art style, alone, is very nice. It's that it took about 50 attempts and a lot of fine tuning because the AI doesn't recognise what the scene is. For us, visually, we know it's a character strung up by a pipe with her thighs sinking into quicksand with her skirt slipping, her blouse ripped open, and her breasts popping out of her bra. But convey that to a machine and they like "So, like, her arms are stuck in mud in the ceiling and she has three legs, right?"
It also didn't help that my artwork wasn't tailored to get a good read for the AI. It was too dark, with too many details that we can piece together, but it's all noise for the AI. It was almost impossible to get something resembling mud or quicksand on the floor because, frankly, it ISN'T mud/quicksand. Your brain is telling you that it is because of context. But it isn't, and the AI struggles to piece that together clearly. Tag it enough and you can nudge it close enough in the right direction to get SOMETHING that resembles it. Ultimately, what I generated was a "weak" version - that is, a high "strength" generation allows the AI to recreate the concept more freely, while a "weak" generation stays more close to the original. While this is good for keeping the construction of the piece, it can fail artistically because it's not what you want from the AI.