"So yeah guys thank you for watching, we did amazing and, we... we were just y'know fucking some dudes out there and, we're doing great!"Axel Syrios

AI Waifus and You 101

Clem the Gem

Unknown member
Early Adopter
Joined:  Sep 10, 2022
Thank you all very much, tons has changed, and all for the better! Early SD gens were absolute nightmares, so I'm glad that it's now comparable to professional products.
Looks like you're already putting this knowledge to good use, judging by your avatar.
 

Clem the Gem

Unknown member
Early Adopter
Joined:  Sep 10, 2022
Jannies (or anyone else) aren't watching, time to doublepost!

That loliposting over in the other thread earlier reminded me that I made an attempt to remake the dancing gif with another character.
I tried a couple of different methods, and as you can see, the results are.. bad. But it's a start I guess. Wondering if anyone else has tried anything similar?

 

Watamate

Previously known as Tatsunoko
Early Adopter
Joined:  Oct 8, 2022
Jannies (or anyone else) aren't watching, time to doublepost!

That loliposting over in the other thread earlier reminded me that I made an attempt to remake the dancing gif with another character.
I tried a couple of different methods, and as you can see, the results are.. bad. But it's a start I guess. Wondering if anyone else has tried anything similar?

Not really vtuber related, but I saw this clip someone made, and as far as AI-Generated (IMG2IMG) videos, it's probably one of the cleanest I've seen. Some frames definitely still have fucked up hands and it's still inconsistent, but it's crazy to see how far we've come in a few years.

View attachment 20109
The second one is an attempt I made, but that was made by manually changing the prompt for each frame. Not IMG2IMG which I'm guessing is what was used for the Loli Kami Requiem dance? I think it looks great though for a first attempt, my first instinct would be to try and add a ControlNet OpenPoses to each frame. But that would be VERY time-consuming.
 

Clem the Gem

Unknown member
Early Adopter
Joined:  Sep 10, 2022
The second one is an attempt I made, but that was made by manually changing the prompt for each frame. Not IMG2IMG which I'm guessing is what was used for the Loli Kami Requiem dance? I think it looks great though for a first attempt, my first instinct would be to try and add a ControlNet OpenPoses to each frame. But that would be VERY time-consuming.
Oh yeah I remember your Mio example (how could I forget), but that's something different. No prompting was used in my example except for the initial image, and everything was done with the Lineart model in Controlnet.
A batch script is run to look at every image in a folder (the extracted frames from the original dance gif) and apply the ControlNet and generate each image with the same prompt.

I think OpenPose would have improved things a lot, but it had trouble detecting the limbs on the source images. Maybe due to the small image size, maybe because it's an almost chibi style. I imagine the dancing examples posted before would have been made similarly.

If anyone wants to try their luck, I was mostly following this tutorial:
 

God's Strongest Dragoon

Well-known member
Joined:  Mar 20, 2023
Proctor posted a picture of his rooster.

Turned it into a foxgirl.
 

Thomas Talus

Εκ λόγου άλλος εκβαίνει λόγος
Early Adopter
Joined:  Sep 15, 2022
Here they come to fluff the rooster
You know she ain't gonna die
 

Alchemyth

Well-known member
Joined:  Sep 17, 2022
I have no idea if this is the right thread for it, but I've been really playing around with LLM's for AI-driven story generation. Most people just use it to coom and erp with their waifus, but it can be so much more.
Pic related on the bottom, I'm playing a scenario inspired by SCP-610. My team found some weird shit in the field, including evidence of some kind of infection that can do various things to humans and wildlife, and our engineer ended up with something inside of her. We returned to base and succesfully got some weird parasite out of her. A day of downtime later, said engineer fucking dissapeared without a trace from the medbay. We suspect the infection has reached the base, and is potentially controlling people or turning them into sleeper agents of sort.

I told my team to prepare to get the fuck out of dodge. Meanwhile I went to check the security foootage to see if there were any clues. Now my own fucking 2IC, that I shared my whole plan with literally 15 minutes ago, is either compromised or has a doppelganger walking around. This shit is turning into the fucking The Thing, and I'm probably gonna start doing the bloodtest when I get a chance.

horror.png
 

Clem the Gem

Unknown member
Early Adopter
Joined:  Sep 10, 2022
Back on the topic of animation, I've been having some fun with something that's been around for ages that I never knew about called AnimateDiff. Really quite pleased with the results on my little test clips:

With a basic prompt where you don't specify any kind of movement or actions, you'll end up with the subject moving in place slightly usually with the camera panning around a bit:


I found if your prompt is too long, you can end up with the clip being split into two different (but still nice!) scenes as seen here:



This is a test prompt without all the character description, trying to diagnose the above problem

You can also split the prompt to be different things at different frames. For example, start frame 0 looking to the side, and facing viewer half way and changing expression. Or anything you want:




But that's not all. You can upload a video and use ControlNet to guide the prompt. Of course I had to have another go at making that loli dance.
I clipped a few seconds of this video as a reference:



Using an OpenPose ControlNet, I get very nice movement while keeping the character I want, and very consistent too. However the background is left up to chance and chops and changes rapidly.
(I could have improved this scene by leaving "bench" out of the prompt, but the trees and stuff would still be random. I also did not describe the subjet at all, using only a LORA. I think I could have avoided the few frames where the clothes mess up by describing the clothes in the prompt)


Using Canny ControlNet, you get the opposite. A stable background, but the character's outline resembles the original too much and is too random.

Combining OpenPose with a weaker Canny made maybe a small improvement, (and also took a hell of a long time to process). What we need is some way to separate the subject from the background. More experimenting required.
 
Last edited:

Azehara

Well-known member
!!Foot Dox Confirmed!!
Early Adopter
Joined:  Sep 11, 2022
Novel AI did an update to their model (they had updated their 2.0 model about a month ago and they already updated it again to the 3.0.

Old model you would have to use a ton of prompts and dial their intensity down to get certain body types and they seem to have simplified things a bit. For example for a mature milf body I would usually go with:
mature woman, [[[[muscular female]]]], [[fat mons]], toned, linea alba, curvy, thick thighs, wide hips, large breasts

Now you can probably go for "milf" and it will do all of that to give you a more broad mature woman body. Additionally the hands seem to be better and you can more easily apply artist styles to change the art style.

I did a few Kirsche prompts using my standard Kirsche prompt (getting rid of the irrelevant tags and simplyfing the prompts up a bit).


{{{dynamic pose}}}, {masterpiece}, milf, smirk, {{white hair}}, aqua eyes, twint s-1315992526.png{masterpiece}, milf, smirk, {{white hair}}, aqua eyes, twintails, {very long hai s-2891176175.png{masterpiece}, milf, smirk, {{white hair}}, aqua eyes, twintails, {very long hai s-2076586457.png

And applying some of the artist styles to the same promp:
1girl, {{{{{urushihara_satoshi}}}}}, {masterpiece}, milf, smirk, {{white hair}}, s-2952601248.png
1girl, {{{{{shirou_masamune}}}}}, {masterpiece}, milf, smirk, {{white hair}}, aq s-1827695141.png
1girl, {{{{{{{sei_shoujo}}}}}}, {masterpiece}, milf, smirk, {{white hair}}, aqua s-2283419189.png
1girl, {{{{{{{otochichi}}}}}}, {masterpiece}, milf, smirk, {{white hair}}, aqua  s-2631767658.png
 

PleaseCheckYourReceipts

Well-known member
Joined:  May 6, 2023
Novel AI did an update to their model (they had updated their 2.0 model about a month ago and they already updated it again to the 3.0.

Old model you would have to use a ton of prompts and dial their intensity down to get certain body types and they seem to have simplified things a bit. For example for a mature milf body I would usually go with:


Now you can probably go for "milf" and it will do all of that to give you a more broad mature woman body. Additionally the hands seem to be better and you can more easily apply artist styles to change the art style.

I did a few Kirsche prompts using my standard Kirsche prompt (getting rid of the irrelevant tags and simplyfing the prompts up a bit).


View attachment 56425View attachment 56426View attachment 56427

And applying some of the artist styles to the same promp:
The Shirow style is still lacking about 10x the glossy reflectiveness, but I'm really impressed how far Urushihara mimicry has gotten.
 

Azehara

Well-known member
!!Foot Dox Confirmed!!
Early Adopter
Joined:  Sep 11, 2022
The Shirow style is still lacking about 10x the glossy reflectiveness, but I'm really impressed how far Urushihara mimicry has gotten.
I feel as if you go too detailed on the description, if its something the artist has not really drawn much of then it fights the style too much. Maybe i fucked up the name in the prompt which is also a thing. NAI uses Gelbooru tags and they mispelled Masamune Shirow maybe I did not apply the style properly. Going to test it out a bit.

{{{{{{masamune shirow}}}}}}, 1girl, beach, {{{sundress}}}, {{{{see-through silhouette}}}}, see-through, mature female, no bra, no panties, dynamic lighting, straw hat, backlighting, {pubic hair peek}, very hairy, tanned, {tanlines}, glasses, wide hips, thick thighs, [puffy nipples], wind, blonde hair, blue eyes,
This is the prompt I used. I will be replacing the artist name. As you can see its not too detailed when it comes to describing the character:

For Proctor's sake I will be censoring the naughty bits.
1700401512170.png
1700401661614.png

Urushihara had a bit of an issue with the sundress top (then again most of his pieces the girls are topless so that might be the issue).



EDIT
So I ran my test again with that same Kirsche prompt. Seems I had mispelled the artist name and thats why it wasnt working. Seems that if you use the fucked up spelling used in Gelbooru's tags, it doesnt work. You have to spell the artist name prorperly.
1700401953121.png1700402001115.png
 
Last edited:

PleaseCheckYourReceipts

Well-known member
Joined:  May 6, 2023
I feel as if you go too detailed on the description, if its something the artist has not really drawn much of then it fights the style too much. Maybe i fucked up the name in the prompt which is also a thing. NAI uses Gelbooru tags and they mispelled Masamune Shirow maybe I did not apply the style properly. Going to test it out a bit.


This is the prompt I used. I will be replacing the artist name. As you can see its not too detailed when it comes to describing the character:

For Proctor's sake I will be censoring the naughty bits.

Urushihara had a bit of an issue with the sundress top (then again most of his pieces the girls are topless so that might be the issue).



EDIT
So I ran my test again with that same Kirsche prompt. Seems I had mispelled the artist name and thats why it wasnt working. Seems that if you use the fucked up spelling used in Gelbooru's tags, it doesnt work. You have to spell the artist name prorperly.
View attachment 56666View attachment 56667
Yeah, the second run is definitely his style. Glossy reflections everywhere, lol.
 

Clem the Gem

Unknown member
Early Adopter
Joined:  Sep 10, 2022
Since it looks like no one has made one yet, I had a go at making my own LORA for Mozumi Pichi. It's been a long time since I made a LORA, (and I only ever made a couple before) - this time I used Civitai's own model creation tool, manually tagging the source images myself but leaving the training settings at default. This cost me 500 Civitai bucks, which I believe I was just given when the currency system was first brought out, so if you're a new user you might not have any. I now have none left, so I'll have to go back to the old fashioned way I suppose.

Anyway, it's far from perfect, but I think it's a pretty good start:




I'm not happy enough with it yet to put it up on Civitai, but you can try out this current version here if you like:
[Edit] Didn't realise I picked a file host that deletes the file after it's been downloaded, and some mad lad actually downloaded it. New link above.
 
Last edited:

Clem the Gem

Unknown member
Early Adopter
Joined:  Sep 10, 2022
Improvements have been made! But so have some sacrifices..



Ran the LORA training locally this time using the same sample images, but with different settings and modified captioning. As you should see, I'm getting a more accurate result, but the downside is it is harder to control. You can see on the last image I asked for "green hoodie" and it is trying too hard to give it the same style as the LORA.

Still more experimenting is required, but once I've got this formula down, I'll be able to apply it to any character. Also, I guess this is where you all can make requests.
 

GOD'S STRONGEST BUILDERBEAR

"Shut up, Dazzle. I will clip your balls" -SB
Early Adopter
Joined:  Sep 12, 2022

Clem the Gem

Unknown member
Early Adopter
Joined:  Sep 10, 2022
Was that my LORA, your own, or purely from prompting?

I've spent the last while struggling to get the different outfits working consistently in a single LORA while being flexible enough to allow completely different clothes. Would be kinda bummed if someone beat me to it
 

GOD'S STRONGEST BUILDERBEAR

"Shut up, Dazzle. I will clip your balls" -SB
Early Adopter
Joined:  Sep 12, 2022
Was that my LORA, your own, or purely from prompting?

I've spent the last while struggling to get the different outfits working consistently in a single LORA while being flexible enough to allow completely different clothes. Would be kinda bummed if someone beat me to it
by 3 days, I think. Its baked on live2d to all shit tho. I'll try yours later

that said, i did get few good pics
00015-3273299965.png 00004-1802048668.png
 
Last edited:

Clem the Gem

Unknown member
Early Adopter
Joined:  Sep 10, 2022
by 3 days, I think. Its baked on live2d to all shit tho. I'll try yours later

that said, i did get few good pics
View attachment 58229
Alright, alright.. I think my failed attempts are already better than this one. I will continue to improve it and release when it's done.


You can't fool me. I refuse to believe!

Prompt:​

(((masterpiece))), high res, (best quality), simple background, 1girl, waist up, portrait, <lora:mozzu:.7>, mozzu, animal ears, wolf ears, short hair, multicolored hair, orange hair, aqua hair, aqua eyes, tail, animal tail, wolf tail, orange tail, white tail tip, wide open eyes, happy, holding a glass of bright yellow liquid,
 

GOD'S STRONGEST BUILDERBEAR

"Shut up, Dazzle. I will clip your balls" -SB
Early Adopter
Joined:  Sep 12, 2022
Alright, alright.. I think my failed attempts are already better than this one. I will continue to improve it and release when it's done.



You can't fool me. I refuse to believe!
describe piss, COWARD

Your model is way better tho, yes, heres some stuff
Your tail fucking sucks tho, her tail is thin not fluffy go redo
00107-1890648614.png
00092-775463672.png 00070-515743876.png


00096-3981618099.png
00079-3325872939.png
 
Last edited:

Clem the Gem

Unknown member
Early Adopter
Joined:  Sep 10, 2022
describe piss, COWARD

Your model is way better tho, yes, heres some stuff
Your tail fucking sucks tho, her tail is thin not fluffy go redo
View attachment 58282
View attachment 58283 View attachment 58285


View attachment 58286 View attachment 58288
Will work on it some more when I get home tonight (like I have been consantly for a week now..) and maybe release what I have.
I'm trying to get a model where you can call for outfit1 or outfit2 (coat / no coat) while still being able to use any other clothes you want without them being affected by the style, and also get the character without describing her at all (you can already do this with the test version). But those outfits just aren't working properly. I'm starting to think I've reached the point where it's as good as it can be with the limited dataset available.

Using a Hololive character for example, you've got thousands of fanarts and official pictures to use in your dataset. For mine, I've had to make do with screenshots from streams where it's mostly them standing in the same pose with a couple of different backgrounds, plus a couple of full body artworks. I'm sure it doesn't help that the two outfits are so similar so the AI is probably having trouble telling them apart.
 
Top Bottom