AI Image Generation

  • Guest, it's time once again for the hotly contested and exciting FoH Asshat Tournament!



    Go here and fill out your bracket!
    Who's been the biggest Asshat in the last year? Once again, only you can decide!

StJesuz

Graybeard Lurker
448
1,485
Experiment of having the NovelAI model generate an image drawn by different artists.

Fair warning: it's an anime girl:
Original.jpg



Some were clearly a fail, like Dr. Seuss, others pretty meh, but I thought it did a decent Norman Rockwell. A real missed opportunity by not including Bob Ross. Maybe the 50gig torrent will finish downloading while I sleep and I can try it myself tomorrow.
 
  • 2Like
Reactions: 1 users

rhinohelix

<Gold Donor>
2,860
4,662
The porn that is going to come out of this before too long is going to be amazing. Everything that is going to come out this tech before too long is going to be amazing but like every tech on the internet, porn will lead the way because that's people's true basest, innermost desires. When you are alone with the machine god, eventually civilization washes away and that's what people are left with. Afterwards we get things like e-commerce, chat rooms, and online gaming but before, unfettered sexual desire.

TLDR: AI Porn will be amazing.
 

Kharzette

Watcher of Overs
4,856
3,471
There are images aplenty but I haven't seen any video. When I first got started I tried some really vague prompts and the typical standard giant tiddied girls appeared. If you like big fat anime boobers you are in luck.

The F111 model is trained on real life porny stuff I think. I tried a mix of that and anime and it wasn't very good. I need to experiment more.

For doing heightmaps, I'd think you'd need a model trained for that.
 

Kharzette

Watcher of Overs
4,856
3,471
I've been experimenting with mixing models. This one is 1/3rd each of f111, gape60 (don't judge), and ordinary stable diffusion.

Seems to be less animeish, with more detailed backgrounds.

05288-680.png
05287-679.png
05269-661.png
 
  • 1Like
Reactions: 1 user

Kharzette

Watcher of Overs
4,856
3,471
I've been trying to do nighttime scenes, but the characters are always too bright. Any reference to darkness seems to work on the background but not on the character. If you () it and start boosting it with a number it starts to have an effect:

06984-2358.png

Yes it turns them into goth elves :emoji_laughing:

Sometimes they will get horns or batlike ears as well:
06854-2228.png
 
  • 1Like
Reactions: 1 user

Pasteton

Blackwing Lair Raider
2,579
1,703
Shits seriously incredible and im not sure yet whether this will decrease or increase the number of serial killers and/or pedophiles in the world. So how big of a leap is it to go from this stuff, to say 3D models that you can import into unity or even unreal 5? The degree to which this will make solo game dev’ing a reality is insane if this is doable. The question is can the ai reach a point where it can generate animations on request as well or is that still a ways off
 
  • 1Like
Reactions: 1 user

Kharzette

Watcher of Overs
4,856
3,471
Neats! I wanted to try generating T-pose orthographic trace helpers that can be put on frozen planes in the front and side views of 3d programs. I haven't gotten around to trying it yet though.

When drawn, a good one looks like this:
fd79a8b2effcea770425ac774209b495.jpg


Eventually you want them straightened into a T pose. I did this one by mirroring the arms and crudely pasting it on.
CopFront.png


What I was going to do is take a past model with proportions I like and just take a grey image, then image-to-image that with a text description and see if it would fill in the grey with character details, keeping the proportions.

Though there are alot of character profiles out there, they almost always are not in the right pose. Most look something like this:

yande.re 930454 bili_bili_douga bili_girl_22 character_design chinadress heels tagme thighhighs.jpg


So I doubt you could ever get it to give you a proper T pose without using image-to-image
yande.re 930478 bili_bili_douga bili_girl_22 character_design garter heels tagme thighhighs.jpg
 

Kharzette

Watcher of Overs
4,856
3,471
And here it is extremely hastily made in 3D. This is one of my ludum dare attempts so speed was the primary consideration and corners were cut. It is rigged and I even managed a few basic anims.
shitty.png
 
  • 1Like
Reactions: 1 user

Pasteton

Blackwing Lair Raider
2,579
1,703
Amazing looking forward to seeing the progress on this. Any chance you can make it less weeb/pedo-ish. Actually I wonder if the ai input of ‘same thing but less weeb’ would work
 
  • 1Worf
Reactions: 1 user

Kharzette

Watcher of Overs
4,856
3,471
Weeb factor seems to be how you mix together models. F111 is trained only on real images I think. I'm not really sure how stable diffusion itself is trained, but I know it helps with backgrounds and details. Waifu is of course all anime all the time.

My most recent stuff was done with 1/3rd F111 and Waifu and SD, and it seems much less weeby to me but I dunno!

Controlling age is really tricky because japanese artists really play around with our perceptions alot in that regard. Some artists go for super youthful faces and big heads but otherwise older proportions. It comes out in the training alot.

For instance when I was doing Eco art for fun, the Eco characters typically wear a backpack, but just adding that seemed to de-age all my results because backpacks in japanese art are almost always on school age characters. (Randoseru I think is the red backpacks).

Specifying large breasts helps alot with age results, but sometimes they will be freed from the confines of clothing. What I do is just have it crank out a few hundred results and throw away the bad ones.
 

Pasteton

Blackwing Lair Raider
2,579
1,703
I feel like a less narrow scope needs to be inputted esp on this board, hasn’t anyone tried to de hoof Kate with AI yet. Or conversely, made a hoof filter to hoof-ify anyone we’d like