Sogna Digital Museum Forum
Sogna/VIPER Series => VIPER Galleries => Topic started by: 黒い灯影 on February 01, 2023, 04:34:25 am
-
So, i got into the whole AI art craze going on right now,
been doing some Stable Diffusion, which is crazy powerful.
Trying to create a model that would recreate Cala from RSR,
having only moderate success so far.. lol
The images SD creates do have some of Cala's elements, but it also captured something in her hair like horns.
so I gotta redo my training images and I want to try and capture Katsura's drawing style
as i'm writing this, i think i messed up my training images D:
well, gonna retry tomorrow.
-
Some of it looks like half-Cala, half-Carrera.
Callera :P
-
I wondered when someone was going to do this. Always knew Cala's hair was gonna be a problem.
It'd be far easier to train an AI to replicate Miki or Asuka.
-
lol,
this was my earlier attempts of trying the image2image function,
trying to make a realistic Cala, and one of the backgrounds from RSR
-
Having a bit more success this time,
just need to refine the training images a bit more,
and training methods.
Unfortunately the container I was training this model on started to run out of disk space.. :cry
well, at least i know to rent more than 35GBs of diskspace now
-
Been playing around more with Stable Diffusion and Dreambooth settings.
getting better results
-
more results
-
The ears and hair look alot truer to form.
-
Can also recognize more of the specific pics that would have been used to train it. So it looks more authentic, though also less like anything new
-
Can also recognize more of the specific pics that would have been used to train it. So it looks more authentic, though also less like anything new
im curious, can you give some examples?
-
Trying to work on my Miki model, the results are mixed,
they are from a few different models as i was training the Miki models with different parameters
-
These are the ones that mainly came to mind seeing the first couple in your previous post. If they weren't used for the training data, then I guess it's an impressive convergence.
Can also recognize more of the specific pics that would have been used to train it. So it looks more authentic, though also less like anything new
im curious, can you give some examples?
-
These are the ones that mainly came to mind seeing the first couple in your previous post. If they weren't used for the training data, then I guess it's an impressive convergence.
Can also recognize more of the specific pics that would have been used to train it. So it looks more authentic, though also less like anything new
im curious, can you give some examples?
All of them are really good tho, I dont like the one with long hair. Not a fan of long hair much and most when a character I love has short hair, I love my girls with short hair.
-
These are the ones that mainly came to mind seeing the first couple in your previous post. If they weren't used for the training data, then I guess it's an impressive convergence.
Can also recognize more of the specific pics that would have been used to train it. So it looks more authentic, though also less like anything new
im curious, can you give some examples?
im still not really getting it..
i can see maybe the breast grabbing ones, and the ones where she's looking down?
-
All of them are really good tho, I dont like the one with long hair. Not a fan of long hair much and most when a character I love has short hair, I love my girls with short hair.
Oh, i was just experimenting with that, see if i could modify certain qualities, like short hair -> long hair... and throwing in glasses.
trying to see if the model is rigid or flexible, i'll probably be throwing these models out, and re-training new ones
-
Started using another method to make the models,
I think its better.
-
tried making a realistic version of Miki
-
Started work on a Cala LoRA model, but i dont like the way it handles, so I'm gonna make another one
-
Those AI models are getting scarily good, especially with being able to nail Cala's hair. Any images from her backside?
I've only messed with AI from an editing standpoint, but i'm really impressed with the results with the logo removal in the attached image. Did some manual editing but nowhere near as much i had to in the past. Using Photoshop's Generative Fill btw.
Here's the original -- https://c.bookwalker.jp/coverImage_2503930.jpg (https://c.bookwalker.jp/coverImage_2503930.jpg)
-
The hair and the line quality are pretty spotless in that last batch. The "bruises" are a weird new flaw on top of things like the hands, but things like that could be edited out manually pretty easily. It's definitely getting to the point where even if it's not 100& perfect, it does enough of the work that a human just needs to do clean-up.
-
I dont really like this model, but I will re-do this model to a better one... because of the way I handled all the tags
the issue with this one is if you forget a tag like "pointy ears", she ends up with normal ears, so I need to associate such features to a tag like "Cala"
-
Any images from her backside?
I never really tried to render her from the back, some weird artifacts, but still good me thinks
-
The "bruises" are a weird new flaw on top of things like the hands, but things like that could be edited out manually pretty easily. It's definitely getting to the point where even if it's not 100& perfect, it does enough of the work that a human just needs to do clean-up.
Yeah, i usually don't both with those smaller details, because i know those can easily be fixed up in post or inpainting
-
might as well throw in some prepped images for Akira, and see how it would perform, also had a bit of Saki and Karin in the training data.
Saki's training data was not that that strong in this case
-
I have no idea where Akira is actually sitting in that next to last M.C. Escher car (00117). Akira is the front seat based on driver's wheel and by virtue of being in front of Karin. But Karin is in the front seat based on the door and side mirror, and by virtue of rear seats behind her. And why are the seat backs so low or the girls so tall? What's going on!?! :huh
It's hurting my brain.
-
The spatial geometry of those cars really distracts from the question of why they're in a car wearing lingerie in the first place.
-
hahaha, thought that was funny, being so high
-
That's awesome.
How do we do this? Whenever use stuff like stable diffuse it immed blocks explicate conent.
-
That's awesome.
How do we do this? Whenever use stuff like stable diffuse it immed blocks explicate conent.
Don't use ones that are designed for public use, they usually have a NSFW content filter installed.
You'll either have to do a local installation or use a GPU renting service like RunPod.io
the easiest way to set one up locally is cloning the repo https://github.com/AUTOMATIC1111/stable-diffusion-webui
Runpod (which I like to use) has templates that installs that for you
-
... why they're in a car wearing lingerie in the first place.
You say there's some problem with this?
-
Those backside images are great!
-
miki
-
Miki is probably my 2nd fav, next to Cala ;D
-
i swear i produced more images of Carrera,
im not feeling confident about her model at all, might have to run it longer, i am missing her promo materials in my dataset though.
-
I've been experimenting with some Deep Learning models lately,
and got inspired by DeepCreamPy's efforts but my focus is more on mosaic censored images.
Some Love Fetish series results
(https://jettatura.media/DeepLearn/ProjectMosaic/preview_7_e0006.png)
(https://jettatura.media/DeepLearn/ProjectMosaic/preview_8_e0006.png)
(https://jettatura.media/DeepLearn/ProjectMosaic/preview_31_e0006.png)
some viper results
(https://jettatura.media/DeepLearn/ProjectMosaic/preview_39_e0008.png)
(https://jettatura.media/DeepLearn/ProjectMosaic/preview_40_e0006.png)
(https://jettatura.media/DeepLearn/ProjectMosaic/preview_49_e0008.png)
(https://jettatura.media/DeepLearn/ProjectMosaic/preview_59_e0005.png)
this is from a model that im still trying to fine tune,
i know i won't get all the detail restored,
but i want to get it close enough so that I can train another GAN based model to try and get it closer to the original image.
I did this by chopping up a lot of screenshots from Viper RSR and CTR into 256x256 chunks, rotating them and flipping them until i had about 20k images
copied that to a 2nd folder, then i applied a 16x16 pixelization to them.
So my model tries to work out the differences between the two sets,
and learn to rebuild as much as it can, but a lot of information is lost from the pixelization.
but the goal of this one is to try and recover as much information as possible from this one.
This is my Unet model that does pix2pix.
The 2nd stage which I haven't started on yet, is to use a GAN model,
meaning it will generate on its own to work out the details, like to clean up all the mess the first model created
and to try to bring it back to a clean image, along with the line art.
i wont get started on this stage, until im satisfied i've created a great model on the first stage.
then of course i'll create a masking component, that will mask the areas that only need to be processed.
This project is costing me money though,
as I have to rent a GPU and a system that has like 250GB of memory and 48GB of VRAM to fit my dataset of 40k images.. lol
-
Huh. I wasn't sure anyone used DeepCreampye anymore, after the program was dropped off the web. (shortly before all this AI art stuff REALLY took off).
I know that because I still have a copy saved. I tried using it to decensor various pictures, but they weren't good enough for me, so I instead started using it to decensor pictures that I would THEN use as references for commissions, lol.
When I tried to get an updated version, that's when I found out the DCPye had vanished and been deleted by it's creator. :cry
-
Huh. I wasn't sure anyone used DeepCreampye anymore, after the program was dropped off the web. (shortly before all this AI art stuff REALLY took off).
I know that because I still have a copy saved. I tried using it to decensor various pictures, but they weren't good enough for me, so I instead started using it to decensor pictures that I would THEN use as references for commissions, lol.
When I tried to get an updated version, that's when I found out the DCPye had vanished and been deleted by it's creator. :cry
Yeah, shame he vanished after his laptop got stolen, but i still find it crazy he would ditch the project because of that, i donno why.
yeah, I used to use DeepCreamPy a while back, i think i still have a copy of it on my harddrive somewhere,
im doing this from scratch, building my own model to try and achieve results like DeepCreamPy, and hopefully improve upon it.
-
the results from the latest model, the Cala one is manually masked though,
wanted to see what it would look like when mixed in the original
(https://jettatura.media/DeepLearn/ProjectMosaic/0000e0009_latestmodel.gif)
(https://jettatura.media/DeepLearn/ProjectMosaic/Julietta_latestModel.gif)
-
Just thought these were funny... Its AI art of Cala from yahoo auctions, but they are on acrylic which is cool, but the design of the clothing is just off...
-
Here's some realistic attempts at Cala
-
Just thought these were funny... Its AI art of Cala from yahoo auctions, but they are on acrylic which is cool, but the design of the clothing is just off...
Well at least someone is trying to introduce some new VIPER material. There's maybe one doujinshi a year now, two if lucky.
-
ok
-
Some images got from prompts. Got abit addicted having fun with it and made hundreds, maybe pack them.
https://perchance.org/ai-anime-generator
Seems very hard to get characters to do stuff "together", but for something free, it's really good. Could also use these are templates for drawing upon and editing.
-
Diff styles as well. Fairly random at times.
-
Prompt -
Dark-pink-haired elf, short hair, huge large breasts, stiff nipples, green eyes, gold arm and leg gauntlet, red sandals, blue arm band on right arm, nude, naked, prison collar on neck metal, in dark blue brick prison cell, on knees, covered in globs of semen fluid all over breasts and on floor, miserable, eyes looking down in submission, POV of camera looking down at her at dramatic angle focusing on he face, 6- degree side angle of character, white globs on floor indicting sex happened, characters legs spread wide to show off crotch area at camera, pink pubic hair, cum dripping mouth
Vintage/Neon seem to give closest resutls
-
Neon vintage one to use imo.
-
Not bad.
I messed a wee bit with AI prompts. I'm pretty far from getting the desired results.
-
Here's a prompt I use for Miki with edits. Jump between default, Vintage and Neon Vintage are try out a few others. Sometimes it likes to do stuff like make her socks not red, and the hair isn't quite right, but it's mostly on point.
Dark-short-blue-haired woman, (extremely short pixie cut style hair on woman), (woman wearing thick bright red lickstick on lips), large breasts, stiff nipples, blue eyes, (woman is wearing thigh high red socks), (woman is wearing white sneekers), blue pubic hair on woman, red pearl earings, blushing aroused cheeks, nude, naked, in white restroom, white tile floor, blushing, unwanted intimacy, sex about to happen, anime masterpiece, intricate, highly detailed, 4k, VIPER CTR art style, character looks like Miki from Viper CTR, vagina fluid dripping from groin, POV woman is being held up frontal at camera to show her vagina area at it, two ugly men force her legs wide open, one man on either side of woman, high quality, pixie cut hair, tomyboy hair cut, womans arms bound from behind, gang raped, (Ken'ichirou Katsura art-style)
Negative promp -
(worst quality, low quality, blurry:1.3), , low-quality, deformed, text, poorly drawn
If you set if to 12 pictures as well you get more results to get closer. The aspect ratio seems to have a huge difference as well. Adding "POV bla bla" as if you're a camera director seems to have a good effect.
At worst it's a really good base to get a picture to edit. Which may have try, haven't drawn in years.
-
:)