AI-Generated

Yeah sure, I'll give it a whirl. Not sure what generation you're doing to cause the file size to be that big unless you're upscaling them.

Last edited by suomynonA on Mon Oct 10, 2022 8:24 am; edited 4 times in total

More coming in a bit, promise.

edit, I'm generating whole libraries worth of "stock" pose base images to use to make everything and anything. Here's one, that came out, during a run of very very low detailed stuff. Not sure why it looks so nice.

some

suomynonA:
hxxps://www.youtube.com/watch?v=obVpFRTgQ8A Gives a pretty decent explanation on things. A big factor is using the right model. Basic Standard Diffusion 1.4 has a very hard time getting the anime asthetic. Thankfully Waifu 1.3 is a very good one and freely available. hxxps://rentry.org/sdmodels Documentation link above will give you a good idea of what your prompts should look like.

There's a lot of knobs and dials to turn, though not everything will have (Consistent) relevancy. Don't be afraid to post some of your images.

Been having fun with making 'Bowsettes' and Jack'o.

Thank you So Much! i'm looking foreward to giving this a try myself. I often find there are never enough Neko's in the world

Now I'm wondering, if one were to collate their own reserve of base material, how would one create a ckpt file to make their own base form? Does anyone know? I've been googling without much luck.

You can currently do that in two ways. One is with the official SD 'Dreamstudio', I'm not terribly sure of the details but essentially you can train a sort of 'mini-model' with a selection of art or photographs to better capture specific subjects or art styles. People have already been making these to better make certain characters or... *Ahem* Anatomical parts. Then piggy back it off of another more complete model.

You can do the same locally if you have a beefy enough system (Or you can 'rent' video cards on the relative cheap from other people) with a method called textual inversion. as of this writing, you need a card with at least around 12 gigs of Vram.

I haven't the foggiest on either method since I haven't looked into it at all. Training your own model, however, I know is going to cost you a decent penny. Waifu 1.3 cost something in the lines of 6.3k USD.

Also, don't forget if you have even a modicum of ability to draw something, you can greatly influence things to get something closer to what you want with img2img.

suomynonA:
You can do the same locally if you have a beefy enough system (Or you can 'rent' video cards on the relative cheap from other people) with a method called textual inversion. as of this writing, you need a card with at least around 12 gigs of Vram.

That's what I need to find a guide for. I've tried Img2Img, and it's all right, but eh. I still get a lot of distorted bodies and I want to thin that out.

Naturally body distortions are things you're always going to have to deal with to some degree with this. Some models are better than others at this, but on the whole you are going to likely generate pictures with at least some oddities. There are ways to course correct ones that are close to what you want by using in-painting to regenerate certain sections of a picture (I've done that a few times with what I've uploaded.) or even drawing basic colors or drawings on top of generated art to actually get it to something that you want (I.E. The 'Feather-chan' uploads I did, I basically colored in the correct color for the wings and did img2img.).

Either way, a simple Youtube search got me these:

Textual Inversion: hxxps://www.youtube.com/watch?v=7Lxdk89W2K0

Dreambooth: hxxps://www.youtube.com/watch?v=tgRiZzwSdXg

A fun thing you can try is prompt alternating in A1111. Using [prompt 1|prompt 2|Or more] and it will kind of blend things together, so you can get a sort of middle ground in things.

Lolibop:
suomynonA:
hxxps://www.youtube.com/watch?v=obVpFRTgQ8A Gives a pretty decent explanation on things. A big factor is using the right model. Basic Standard Diffusion 1.4 has a very hard time getting the anime asthetic. Thankfully Waifu 1.3 is a very good one and freely available. hxxps://rentry.org/sdmodels Documentation link above will give you a good idea of what your prompts should look like.

There's a lot of knobs and dials to turn, though not everything will have (Consistent) relevancy. Don't be afraid to post some of your images.

Been having fun with making 'Bowsettes' and Jack'o.

Thank you So Much! i'm looking foreward to giving this a try myself. I often find there are never enough Neko's in the world

ok.. i think i have it installed... here's hoping i have something to show for it

some

Very nice! I've been tinkering, still. I don't know how to set up a lot of the stuff out there, like Automatic1111 or whatever. I have put some work in making characters.

This was Brandon. Now it's Brandy. Hey, if a wealthy relative you don't know well offers to pay for transition, you better believe they'll want something in return.

some

more

some