Oh. and did you know? This is the first Kuon thread of /maho/. It can be considered a real board now. Pretty cool, huh?

This stuff goes way over my head. Unfortunately it seems like most of it is undocumented and just relies on experimentation. You can search around for the words but uhh.. most people just say stuff like "you'll have to test it out yourself and see if it works". I did download a preset I saw on 4chan's /h/ so I guess I'll see how it goes. I don't remember it being this complicated last time, although I guess people were more in the dark and just threw stuff out there.

Cool. I have no idea what the error is since it's listing it 10 times. I think it's the "save and resume" option that I ticked on? I guess I'll turn that off and see if that fixes it...

>>325
I was thinking I'd wait until I have something to show for it, but now I gotta say it... I've been working on a programming language for the past couple days and I'm calling it Kuon!!
Kuon Programming Language... has a nice ring to it, I think.

Anyway, yeah, sort of off topic for this thread, just thought I'd mention it. Kissu has a lot of tech people apparently, I thought it was just vermin and 1 or 2 other anons until this board opened

Hooray! It seems like it finished some training rounds while I was away! The triton error is some optimizer that only works on linux, but I don't know what the other stuff means. Well, if it works then it works!
Looks like it will be a couple hours until it's done with the last epoch thing? I think I'm "overbaking" this, but I'll find out once it's done and I can actually boot up SD to test it out. Although I'll be able to test out the 'style', I won't be able to specifically prompt Kuon. I could get the hair and eye color and stuff, but not her specific outfit and and face. She does look similar to other characters of the same style, of course, perhaps closest to Tama-nee, but still a bit different.

>>354
!! WHOA! THE KUON LANGUAGE?! Now you're speaking my language! It's going to be beautiful and elegant, right?
I'm not a tech guy, though, I just follow instructions and FAQs for this stuff to accomplish limited goals.

>>355
>It's going to be beautiful and elegant, right?
I'm more so going for something very practical but easy to write. So instead of going in a direction of pure elegance like lisp or haskell, I'm more so doing something like mainstream languages, merging a lot of different ideas together and trying to keep it simple and very light on syntax.
Now that I think about it, that sort of feels fitting for Kuon... her whole appeal isn't just beauty, after all, she's a very capable and independent person with lots of different qualities

Anyway, I'm still trying out a bunch of different ideas, but I've got a rough idea of what the language looks like in my head with some things written down

>>356
Cool! I have absolutely no idea how any of that works, but I love the name!

My hard drive is becoming slightly more Kuony.
I'm not sure why, but gelbooru has over 1,000 Kuon images whereas danbooru is at like 300. Gelbooru does have lower standards and has stuff like terrible nude edits and so on, but that's not going to be very many here. Very strange.
Funnily enough I was using Grabber before all this AI stuff happened and it turns out that it's also a fantastic tool for AI purposes since it bulk downloads and labels stuff with tags.

!!!!!!!!!!!!!!!!!!

Hmmmm... I did a bunch of testing and while it's somewhat good, I'm not happy with it. The style really isn't there. The general concepts are there such as animal ears being on the side and how the :D face should look, but general style needs work.
The total training of the 5 epochs took about 7 or so hours, but it turns out the last 3 epochs were actually too noisy, so future tests will be quite a bit shorter... I think.
Oh yeah, did I mention that PonyXL is apparently really bad at faces without using the adetailer extension? Because apparently that's a known thing.

And here is a set with Kuon-like characteristics.
Of these the first set is definitely the closest to the style so I'll probably look at these settings for adjusting them. But, on the previous image set the third one looked quite good.
Man, this stuff is annoying to figure out.

Yeah, adetailer does a fantastic job of getting the eyes right. This will be the baseline of my future testing!

But man, the hands...
This is something I ran into on my 1.5 LORAs, too, the hands just get messed up. I wonder if I can fix them with tags or something...

Another round of training complete with different settings. Seems to allow greater variety in outfits, which was a problem in what you see here >>364
But, I think style can be improved more.
I'm going to try a slower learning rate with more epochs. The unfortunate thing with these AI LLM things is so much of it is random. I can never actually reproduce the exact training because it branches out and it's uncontrollable. I guess 'exponential' might be a good word to use because if it veers into a direction it may continue following it.

I read about the 'Lion' optimizer and that will be the setting I use as it trains overnight with a greatly reduced learning rate running for more epochs.
There's this Network Dimension stuff which is uhh... I can't remember but the larger it is the larger the filesize of the LORA. It's an efficiency and diminishing returns thing, but I don't really care about efficiency since it's just one LORA.
You may notice how there's a LyCORIS preset thing there, and yes some guy did name it after the show which was very popular at the time. It's a subset of LORA and I can't remember how it differed.

Man, I really don't understand this. I guess the second from bottom is the best one currently, so I'll build off that? Bottom has some really good parts to it, but it also seems to be damaging the output. I bet if I had some saved steps between the last two the ideal would be in there somewhere. That's the one that takes 5 hours to get to.
I need to find a way to get the style without damaging the colors and lines.
Interestingly the third one has the same errors as my SD1.5 LORA in that it adds noses and lips. I wonder why.

After 12 hours of training with different options... it's awful. Well, there goes that day. Maybe this 'lion' thing isn't actually good for this purpose? Huh....

Another set done overnight. I noticed a queue system thing for different settings so I have a bunch to test out. I'll just test the 'final' version of each one at first and I sure hope I see something good from this...
The ones I saved as LionLycoris806 took a LOT longer.

But these filesizes are quite a bit lower than my first tests. If it works, great, but if it doesn't I don't know if it's a filesize/DIM issue...

Okay, these are super "fried", although the top one the least. I may need to remove some of the older training images since the colors are again washed out, but I saw them as valuable because they had different angles or clothing.

oh wait I just realized I had an older lora active in each of these so it's doubled up...
gotta test again

The training from the past 2 days hasn't been very satisfactory, but I guess it points me back to the 'standard' training setup. But, this looks so much more like Kuon than SD 1.5!
I checked and someone else made a style lora for this, but it's not very good. It completely lacks charm and beauty. They gave no care to identifying the differences between brown and yellow eyes (kind of hard with Amaduyu) and you can tell they didn't label the animal ears correctly and probably didn't fix any tags. It was obviously a bulk thing with no care put into it. I bet they didn't even crop and edit artbook and gacha stuff! (The Utawarerumono gacha is an insult, but the art is very good for AI training)
The top row is the version I saw online and the bottom row is mine which I'm still trying to improve.

This is how it looks without me even training it on Kuon specifically! I did have a weird issue with SD 1.5 when I combined the Aquaplus style with Kuon, so maybe I should make a second Lora that is uhh... hmm... just her clothing sets? Her hair? Ehhh....

OH MY GOD KUON IS SO CUTE! My GPU and VRAM has been occupied 99% of the time the past few days which is warming up my room in summer and I can't watch video without stuttering... BUT KUON! WOW!

Oh yeah apparently Pony is known for being bad at backgrounds which is exacerbated by the ones I had to use for this training, so that kind of sucks, but having better hands and detail and stuff is worth it. After I finish training this I can try to isolate the specific LORA channel things so I can get the ideal influence while limiting the damage to image integrity.

KUON!

>>489
I completed 2 training rounds overnight and a third one will be done in a couple hours at which point I'll make another comparison chart!
Also something cool I noticed with my tagging is that I can properly prompt "white cat tail" to mean Kuon's tail, whereas it prompts a white cat in the other guy's LORA since it was never trained with that in mind. One does pop up rarely, but for the most part it doesn't.
AND LOOK AT HER HAIR TIE! My 1.5 merge wasn't able to do the "low-tail long hair" thing, but there it is in the images without even training on Kuon specifically!

kitaaaaaaa

nope, this one isn't better. there goes another 8 hours of GPU usage

on topic sager

Making all these epochs took all night and then all day, and my python window thing does this annoying things sometimes where I need to open the window and hit enter for it to resume. It happens with SD and it happens with Silly Tavern and now it even happens with lora training. I just don't get it, but it's VERY annoying.

After more testing the thing I made yesterday does seem to have more prompt freedom than my previously-winning thing that I made a few days ago.
For my next training set I guess I'll make some minor changes to the learning rate... or something. No wait, I'll have it run longer. If it comes out better on later epochs then I'll uhh... hmm... reduce learning rate and have it go even longer?
Hands, though... I really wonder if there's something I can do because Pony seems far better at them when my trained lora thing isn't active...

blehhhhhhhhhhh

ashdnas9udhnasiuodnjasiudnas
GOD WHY IS IT SO HARD TO FIND SOME SORT OF PATTERN

The curve is hard to fit

>>1222
trippy kuon the destroyer

and uhhh because it's as blackboxy as they come iunno

>>1225
Yeah, blockbox for sure. At least it seems these days you can set a specific seed so MAYBE stuff can be reproduced? Probably not, though. The way I heard people talk about this stuff last year reproducing the same training session is completely impossible, so this is probably just attempting to steer it back in the same location.

Information I only now know because someone decided to ask a question in a thread and another guy was helpful enough to remember a post from 2 months ago...
Most of my source images have transparent backgrounds which I then changed by adding gradients or backgrounds from other Aquaplus games so maybe this is something I should do...
But Pony's backgrounds are already known for being bad so I wonder if I should bother.
These backgrounds here aren't great now that I specifically look at them, but they're not the focus of the image of course >>1221

I needed some motivation so I downloaded some other LORAs.
I think I'm on the right track and tonight's training should make a better LORA than any other... maybe.

I have a feeling I'm wasting all this time training it for TWENTY hours, but I want to be sure that this Locon thing is a dead end for this specific purpose. Not sure why the training time thing seems to grow over time (the IT/s thing).
Locons seems to train 4x slower, but supposedly they have more data available so in theory more style could be carried over, but my previous locon in testing wasn't very conclusive.

>>1229
>(the IT/s thing
err apparently it's s/IT
Seconds per iteration or something? dunno

Those 20 hours bore no fruit, so I guess I have my answer.
Also check out this Simpsons "do it for her" LORA hehehe.
Yes, yes I am!
DOOT HER

>>1233
Are you looking to achieve something in particular? A task as open ended as approximating a general style across all sorts of situations while dealing with the unreliable results of a ramshackle console nobody seems to truly understand would make me lose hair from stress, and it seems to me pics like >>1215 and >>489 are already pretty decent. What do you think it's missing?

>>1237
Well, apart from just trying to make it look as good as possible I want to minimize the damage to hands, colors and backgrounds and stuff. With some of these it's close to being there, but then there's something that sticks out as worse than other versions. I should test to see how it interacts with other LORAs more, too, but that's a low priority.
Some of it just hard to explain, but in this set of images I prefer the warmer, softer colors of the middle row, but it also has the most errors and some sort of prompt bleed thing where the 'white cat tail' prompt for Kuon's tail also summons a white cat.
In the bottom row, the floral-pattern kimono I prompted for Kuon looks very similar to Karulau's Uta 2 and 3 kimono in the training data, which I guess isn't a huge problem but it's less than ideal.
In all of the images you can see how the stained glass background in the middle is quite bad, but I don't know if there's anything I can do about that. I read about some masked thing and posted about it here >>1227 but I don't see any options in that UI thing and am really not going to research command prompt stuff. This AI stuff is so horrendously documented that it's a wonder that any information spreads at all.

>>1239
Aforementioned kimono. It's definitely leaning into it in that LORA but not any of the others, which is quite strange.

>>1239
and now that I look at it more the middle outfit is leaning into Irena's Sage outfit.
Hmm... these ARE cool outfits, but I think this may be a problem when it comes to prompting other outfits.... maybe?
arrgghhh

>>1239
and here's the same training process as that bottom row there, but with fewer steps. Each "step" is a processed image... I think. I have like 150 different images, but I have it set up so that higher quality ones are processed a greater number of times.

I think I'm happy with it, at least for now. Who knows when there will be a different model out, and if/when I get more VRAM I can make new versions much faster. I think I've minimized the damage to hands. They're still AI hands, but she doesn't have 12 fingers.
But, I still want to make a Kuon specific one. As it is I can't prompt her outfit and her hair accessory thingie is not accurate at all. I guess I'll get to work doing it sometime soon, but for now I'm taking a short break because all this tagging and sorting and stuff makes my eyes and brain hurt, even if I'm looking at Kuon.
I'm very happy that I can generate her tail now, though, as that was something I couldn't get to work in my older 1.5 SD Kuon lora. I probably could have done it, but I simply forgot how to make loras and kept procrastinating.

Did another round of training, this time training on 'autismmix' instead of base pony, but I don't have the results yet. I did try to prompt Nekone with my existing merges and it seems to work decently for having so little data.
I never defined her ears apart from generic 'animal ears' so maybe that's something I should have done, but her ears are so similar to Eruruu's and such that I figured more training data on 'animal ears' was better than splitting them up.

It's quite interesting to see what happens when you copy the tags for an image that was in the training data verbatam.
Source image on the left, copied tags on the right. It's very accurate because it's just directly copying it without introducing anything new, at least on purpose.

The training on Autismmix was a failure. I mean, it's there but it's easily worse. It was my assumption that this would be the case, but it's nice to know for sure.
ALso... working on getting SDXL and my LORA working in irc and /chat/!

Pony, or more specifically the SDXL base's higher resolution actually makes the regional prompt extension useful to me! Add in adetailer to fix the faces (although it's a crapshoot at which face is selected first with the correct prompt) and things come out quite well.
I'm not someone that will spend time inpainting and photoshop editing and stuff, so the image I generate is the image I keep. (Well, unless I'm making a thread or something but it's rare I use an AI image for that)
This isn't an ideal image since the style took a hit somehow, but wow it's pretty amazing to have Nekone and Kuon there, roughly.

On a lark I decided to search 'Kuon' on civitai and there was in fact a Kuon LORA for Pony.
But... it's not very good. The guy has 342 models on there, so it's obvious that he does the automated mass production things which creates models that are decent enough for most people for most purposes (masturbation), but definitely not ideal. I'm not sure why anyone would make that many models, it's certainly not fun and you're creating the material that other people use to try and make money without getting any of it yourself.

I saw another Kuon lora (civitAI search seems quite strange and didn't show this one directly) but it was disappointing to me again.
Alright, I'm refreshed enough to give Kuon the attention she deserves!
Now that I'm training a concept instead of a style the process is a little bit different. First, I'm going to separate Kuon by outfit so I can try to trigger different outfits of hers on command.
But, I'm noticing some angles are missing so I'm going to watch Kuon's charming and beautiful and amazing anime form again and carefully go frame-by-frame to collect as much visual information as I can. But, I actually do find this frame-by-farme stuff mentally exhausting so I guess this will take me a little while since there's 53 episodes to go through. They really should have made it at least 80 episodes, but I guess it was a miracle that season 3 happened at all.

Like this, this seems like a good frame to show what it looks like from behind. If you don't provide this information then the AI will make an assumption.