CREATE YOUR AI CUM SLUT ON CANDY.AI TRY FOR FREE
x

VN Ren'Py Witches Pairing [v0.1.5p] [Kitty_SFM]

3.00 star(s) 1 Vote

Karnewarrior

Well-Known Member
Oct 28, 2017
1,635
1,913
Setting aside the ethical and environmental issues of AI art, it's biggest issue in the context of a visual novel is a lack of consistency between images. Like even with genuine expert prompting you just can't get the same character twice. Which means that as you move through a series of images you have what is supposed to be the same character changing. Facial details will be different, the outfit will have different details, colors will shift, hair in particular can get wild. It just visually throws the player off.
Actually, you can get pretty consistent imagery with img2img software, which isn't that much harder to get ahold of than regular AI art that isn't those shitty online ones that always make the same crap.

See, the trick is to sketch it out for the AI, have the AI do the work, and if you have the art skills touch up the result. Most people skip both the first and the last step, so their gens are weird and inconsistent. But then most people just type their prompt in in plain english like they're getting a commission, which inevitably results in horrible abominations.

With a combination of img2img and inpainting you can actually skip having skill in traditional artistry entirely and get something pretty decent, though if you want to avoid that uniquely generic AI artstyle you'll need some level of skill in prompt engineering.
 

Kitty_SFM(Developer)

Thanks for your support, friends !
Game Developer
May 10, 2024
110
116
i understand your point and its cool if thats your opinion on AI art and honey select. i wont say there arent those games that are guilty of your claims but there are also those diamonds in the rough that have really good art and dont suffer from the inability to create the same character image. the best example i can give is "Echoes: Cards of Destiny" i dont notice any inconsistencies that you mentioned in that game.
Thanks for everyone for giving me your point of vision on this topic !
As for the consistency - you must train your own networks, low rank adaptation networks(lora's), and then you will get the same character in any imaginable by the Base model(or any additional connected to the inferention loras) tokens(imagine it to be synapses in huge enormous brain of visions(textual invertions(functions)).
As for the hands and eyes - these details is way too smal for the SD1.5(wich has 64x64 latent space/512x512 gen-pixel size after vae decoding), As for the SdXl - it lacks the anatomy knowledge due to the censure. As for the PonyDiffusion - its the place where things get uncanny - It's very good at generating anatomy, but it staggles with realistic style eyes...so just stick with cartoony generations. This is why style loras is so popular at the Pony Diff internet space. Buuuut...PonyDiff is based on the SDXL...aaaand whants 8 Gb of vram to even run on full vram...if you got 6 gb you will end up with -lowram mode at the comfyui aaaand...it will generate image per 1-1.5 minutes with Hyper lora and 8 steps. While the 1.5 does 512x768 bare bones with hyper and 4 steps for 5 seconds on the 1060 6gb, and then upscale to 1152x768 for 16,6 secs...aaaand then inpanting comes, with 512x512 patches for 2,3 sec...for an hour you will inpaint it untill get some really good result. With Pony it is impossible(doable only through SD tiled upscale) because - it so hard trained (5millions images for 3 mounth) so it's latent space got very far from what base SDXL(on wich it's based) was...soooo the controllnets, segs and other stuff will ...if work then really not ideal.
Sooo...the only real up-to-date sollution to REALY work with ai images is SD 1.5(And only the 2D-2.5D , to avoid the uncanny feel of missfigured but so realisticly textured and rendered imaginery, wich portray "human entity".
 
Last edited:
  • Hey there
Reactions: CptValor

CptValor

Member
Oct 8, 2017
355
487
Thanks for everyone for giving me your point of vision on this topic !
As for the consistency - you must train your own networks, low rank adaptation networks(lora's), and then you will get the same character in any imaginable by the Base model(or any additional connected to the inferention loras) tokens(imagine it to be synapses in huge enormous brain of visions(textual invertions(functions)).
As for the hands and eyes - these details is way too smal for the SD1.5(wich has 64x64 latent space/512x512 gen-pixel size after vae decoding), As for the SdXl - it lacks the anatomy knowledge due to the censure. As for the PonyDiffusion - its the place where things get uncanny - It's very good at generating anatomy, but it staggles with realistic style eyes...so just stick with cartoony generations. This is why style loras is so popular at the Pony Diff internet space. Buuuut...PonyDiff is based on the SDXL...aaaand whants 8 Gb of vram to even run on full vram...if you got 6 gb you will end up with -lowram mode at the comfyui aaaand...it will generate image per 1-1.5 minutes with Hyper lora and 8 steps. While the 1.5 does 512x768 bare bones with hyper and 4 steps for 5 seconds on the 1060 6gb, and then upscale to 1152x768 for 16,6 secs...aaaand then inpanting comes, with 512x512 patches for 2,3 sec...for an hour you will inpaint it untill get some really good result. With Pony it is impossible(doable only through SD tiled upscale) because - it so hard trained (5millions images for 3 mounth) so it's latent space got very far from what base SDXL(on wich it's based) was...soooo the controllnets, segs and other stuff will ...if work then really not ideal.
Sooo...the only real up-to-date sollution to REALY work with ai images is SD 1.5(And only the 2D-2.5D , to avoid the uncanny feel of missfigured but so realisticly textured and rendered imaginery, wich portray "human entity".
thats a lot of technical specs to keep track of! to be able to just know all that and give the breakdown here must mean you know a lot about AI. im impressed! i dont know if id be able to remember all that! it also sounds like a complicated process to make the images come out the way you want em. ive just messed around a little for fun but never realized how technical it all is.
 

Alanray64

Member
Jan 28, 2022
343
185
I may have missed it, however scroll back appears to be disabled? Is this intended?

Plus the games are rather unfair when a player has a muscular disability and is unable to click fast enough. Very disappointing. I gave up after spending ages on the first game.
 

Kitty_SFM(Developer)

Thanks for your support, friends !
Game Developer
May 10, 2024
110
116
Witches Pairing Update 0.1.4p:
-Long awaited Lina Longhood inspections and interactions;
-96 High quality renders;
-10 high quality animations;
-Great intro to the Donna MeanAss with 12 renders and fun interaction
Download:


 
Last edited:
  • Yay, update!
Reactions: ArhraCole

Takuji

New Member
Apr 4, 2021
6
12
even though that's not physically possible and is in fact literally impossible for a number of reasons.
Can you name any of thos reasons? That sounds just like pro-AI talk.
Whats true is that it depends how you use the AI , it can be a powerful and helpful tool or it can just create dogshit.
 

The Slithy Tove

New Member
Mar 17, 2018
9
14
I hate criticising creatives but I have never deleted a game so fast. What a total pile of click crap. This game will allways be free and it will always be ignored. Off to oblivion with you. Expeliarmus.
 
  • Like
Reactions: Shiningforcehero

Karnewarrior

Well-Known Member
Oct 28, 2017
1,635
1,913
Can you name any of thos reasons? That sounds just like pro-AI talk.
Whats true is that it depends how you use the AI , it can be a powerful and helpful tool or it can just create dogshit.
The most important one is that the AI starts it's work with random noise. Stable Diffusion works by taking a bunch of random pixels and then slowly tweaking values until an image comes out. It can't go in thinking "I'm gonna recreate X" (I mean, it can't go in thinking at all, it's a machine), it's given some mashed potatoes and is told to sculpt them into vague figures and then refine it until the figures ain't so vague.

The second is that the AI can't store it's training data. The training data isn't stored anywhere in the AI. You can see this if you ever look at the file sizes - even the big AIs are only like 50 GB. That's half the size of my images folder, much less the whole booru's worth of shit you need to train an AI worth half a crap. Everything is itemized and broken down into descriptions of lines and curves and then even further abstracted into a really convoluted math problem, which is the AI itself. But you can't get the training data back out of the AI because the process is inherently lossy and the information quite simply no longer exists. It's gone, reduced to variables.

The third and most minor is simply that reproducing the training data is actively avoided by AI developers since, well, that's counter to what AI's supposed to do. People aren't looking for an overengineered copy-paste program after all, they want the AI for their weird niche edge cases, like a bunch of portraits of specific anime characters for a porn game played by at most 1000 people worldwide, or their OC donut steel D&D character. The term for an AI reproducing it's training data is called "Overfitting" and basically ALL of AI development is focused around preventing that. Most AI these days have special little... Algorithmic "hooks"? Sorta? They got things in 'em that prevent overfitting. The most important of which is the lossy process of learning - the AI crunches the information to such an extent that it can no longer overfit because the original image is lost to it.


All of which is comparable to human learning, by the way. AI's aren't people, but we've gotten as far as we have by closer and closer emulating human beings. Humans also engage in lossy learning processes that take aggregate examples and boil them down into patterns (Stereotyping). Humans also have in-built resistance to overfitting (boredom).

The only real argument for AI taking artist jobs is the suggestion that people are turning to AI instead of artists which... Well, come on bro, if you're getting outpreformed by this:

1gm79jt86_00075-2356251980-Masterpiece_best_quality_Mami_Tomoe_pullups_baruma_lifting_self_pon...png

You probably weren't gonna make a living anyway.
 

toccata

Active Member
Nov 24, 2020
628
415
Today I found this game and played through it... My version has a bug. The "roll back" doesn't work. Neither with the mouse wheel, nor with the mouse button, nor with the back menu. Is that normal for this game, or is it just me?
 

Kitty_SFM(Developer)

Thanks for your support, friends !
Game Developer
May 10, 2024
110
116
Today I found this game and played through it... My version has a bug. The "roll back" doesn't work. Neither with the mouse wheel, nor with the mouse button, nor with the back menu. Is that normal for this game, or is it just me?
Hello)
Yes, this is normal for this game.
I would advise you to play Witches Trainer as well.
How's your favourite mouse going on?)
 
Last edited:
  • Like
Reactions: toccata

toccata

Active Member
Nov 24, 2020
628
415
Hello)
Yes, this is normal for this game.
I would advise you to play Witches Trainer as well.
How's your favourite mouse going on?)
I've now looked at "Witches Trainer".
The type of programming, the pictures and the way the passwords are handled. The pictures are very nice. I particularly like the landscapes, which are very imaginatively designed.

In my opinion, the passwords are lying around far too obviously. If you're interested, I'll tell you a easy way, how you can make it more hidden, and "noeon" can find them. (Ask me via PM)

Unfortunately, the rollback function is also blocked in this game. I don't like that at all and I really miss the option of rolling back and forth in your game. The game is too annoying and unplayable for me without this roll back.
Since RenPy has that as standard, you must have blocked it on purpose. That's a shame. Because I need that when looking at the pictures.
 

Karnewarrior

Well-Known Member
Oct 28, 2017
1,635
1,913
...Why would he want no one to find the hidden passwords? That defeats the point of gating things behind passwords.
 

toccata

Active Member
Nov 24, 2020
628
415
...Why would he want no one to find the hidden passwords?
So that you pay on Patreon before you have them !!
And – if you don’t want to pay – you won’t find them.
Not even if you look for them in the RenPy code. :)
,
 

Kitty_SFM(Developer)

Thanks for your support, friends !
Game Developer
May 10, 2024
110
116
Witches Pairing Update 0.1.5p:
New animations :

-Jessica Wallters gain 4 new H animations;
-Prachariah Prabhti gain 11 new H animations;
-Genna Whirlpool gain 25 new H animations;

DOWNLOAD:


 

Vreejack

Member
Oct 7, 2018
230
205
Just utterly boring. And the silly puzzle was annoying enough before he made us do it perfectly three times in a row. At that point, even with a cheat, it was more satisfying to simply delete the game.
 

DetectiveGirl

Newbie
Mar 3, 2024
54
73
The first pictures looks interesting...urghh VN....double urghh running since 2021.
Maybe I'll check it out if it's halfeway finished....so lets see again in 10 years.
 
  • Like
Reactions: zaqrwe
3.00 star(s) 1 Vote