- Joined
- Mar 26, 2020
Someone made a short comic from their render and an anon decided to translate it: https://twitter.com/h071019/status/1585963841858551808
Stable Diffusion, NovelAI, Machine Learning Art - AI art generation discussion and image dump
- Thread starter Near
- Start date
-
🔧 At about Midnight EST I am going to completely fuck up the site trying to fix something.
- Joined
- Mar 26, 2022
In case anyone is wondering what kind of cursed text gets good results in speech sampling, it is abominations like these entries:
Some text replacement (accent)
i -> ae
ts -> xs
on -> ong
ow -> o
do -> tu
Emotive and Filler text blocks (to get more natural patterns)
t HA t Ha t Ha
tt tut nghh buh goj cpiw xmzp
dowj jooe joob
phjj iushf
ahhhhhhhhh (often helps with producing natural intonation, sighing sound, and spacing, results depend on neighboring phonemes)
Examples (before replacement):
ahhh They have tut to ghhh give us a chance aaaaaaahhhhhhh
what why why how ahhhh what is your question ahhhhhhh
(repeating words that effect the pitch of the entire sentence creates emphasis)
tata HA No matter what you saaay ahhhhhhhhh I believe that eeeee it will be possible DOWJ aaaaahhhhhhhhh
T HA T HA T HA
- Joined
- Jul 25, 2019
- Joined
- Mar 26, 2022
Yeah will have to be later this week my audio rig is in another castle.
- Joined
- Apr 3, 2021
- Joined
- Apr 27, 2022
With Deforum We're going to be making anime cutting out several middle men
Video anon made:
powerpsi
kiwifarms.net
- Joined
- Aug 9, 2022
- Joined
- Mar 24, 2016
Huh. I don't recall it being said that it requires anything high end. Just that it takes time.
CwMkCW
kiwifarms.net
- Joined
- Dec 16, 2020
doing it on a 2080 TI, but I guess thats kinda high end... not compared to a 3080 TI tho
- Joined
- Apr 1, 2022
Yes, there are largely two parts (in layman's terms): tagging and flagging inputs and building the decision tree. Tagging and flagging always ends up having a lot of manual work that cannot be automated away. Building the final checkpoint takes a farm to finish in a reasonable time period. And it's a back and forth between those two until a subjective final product is reached.
- Joined
- Jul 25, 2019
Can someone explain what exactly a hypernetwork does? How to is on the guides, but why to do one is not.
CwMkCW
kiwifarms.net
- Joined
- Dec 16, 2020
Null linked this on the front page https://rentry.org/hypernetwork4dumdums
- Joined
- Dec 28, 2014
Has science gone too far?
- Joined
- Aug 28, 2019
Just wait until the RTX 6090 drops.
- Joined
- Dec 28, 2014
And burns up people's computers when it catches on fire because of Nvidia's shitty cables?
- Joined
- Apr 3, 2021
I've spent about 2 hours trying to recreate Sandy Cheeks cockvore and generate furry foot crush but apparently the only "nsfw" shit these ai models generate are stupid stock image porno, you know the calendar playboy types, the way these models completely ignore the reality of the overabundance of autistic degeneracy that clogs your search engine whenever you search a cartoon character is baffling and also answers your question, no A.I still has standards, unlike us humans.
- Joined
- Aug 5, 2022
I think vram is the main requirement
I think they're just an additional layer as in a model that sits on top of another model, which you can train rather than retraining the whole base model. A neural network is a graph of parallel activator nodes, and each parallel 'layer' operates on the output of the previous layer. Each layer can have any number of little shits. So your system can do whatever you like before or after the network and tacking on another network to the end is the same thing as having one big network, so when you define your hypernet for training you set up the internal structure of its layer composition like normal. Specifically how this should be composed is mysterious magic and nobody fucking really knows what to do, it's all based on rules of thumb for different applications and comparing performance basically.
What they do is basically whatever you want but people are using them as a final style process, like by training them on a particular artist so all your girlies will consistently look like theirs. Here's a big compilation by wonderful korean masturbateurs for example: https://arca.live/b/hypernetworks/60940948
An alternative is aesthetic gradients which is now also integrated in WebUI: https://github.com/vicgalle/stable-diffusion-aesthetic-gradients
I haven't messed with it yet and have no idea how it works but I think it's more for general aesthetic styles (like making any image look like it was painted by a particular artist's brush, rather than making the girls look like the ones in Ivan Aivazovsky's doujins) but apparently it's a lot faster to train and works on tiny datasets.
Disclaimer: I study this shit but I don't do image processing stuff and big pervert data terminology is different from what I'm used to anyway
- Joined
- Jul 25, 2019
There are special checkpoints for that shit. Saw them on the list of models. Just a sec.
Edit:
A list of models there are ponies, specific anime girls, particular sex acts, whatever. Enjoy.
Edit 2: Where the fuck is the extensions directory? Nothing just outright tells you and there's "extensions" subdirectories for several directories. The documentation for this needs some work.
Last edited:
- Joined
- Aug 5, 2022
made some zingy wallpaps
but uploads are fucked so here's an imgur link
reprocessed some of em too
being able to do high-res gens in under a week is fun
if you don't do a git pull
but there's an integrated extensions manager now so you can get the common 'official' ones inside the web interface
though if you're trying to use an unlisted one you might want to do a git install for updating reasons, there's an example on the extensions page of the webUI wiki (which I assume you didn't see or you'd have the directory already)
but uploads are fucked so here's an imgur link
reprocessed some of em too
being able to do high-res gens in under a week is fun
you should have an "extensions" folder under the webUI root with a txt file that says put extensions here
if you don't do a git pull
but there's an integrated extensions manager now so you can get the common 'official' ones inside the web interface
though if you're trying to use an unlisted one you might want to do a git install for updating reasons, there's an example on the extensions page of the webUI wiki (which I assume you didn't see or you'd have the directory already)
powerpsi
kiwifarms.net
- Joined
- Aug 9, 2022
This level of autism is truly a thing to marvel at. Whoever the anonymous person is that compiled this list, God bless them.