#🍥|anime
1 messages · Page 41 of 1
idk what to tag it as :/
I would give it photorealistic
ooo photorealistic?
thanks guys!

🫂
i am now working on my merge
sometimes it makes mak-up weird af
Should try your model out 
you should 
Whats the name of your model again? 
watched her struggle to get the hands right 
synergix
yeah haha i stayed up until 8 am
GIRL
sooo laaateeeeeeeearlyyy
i know but i got into the hyperfocus zone
mood
O, I just looked up your model, I think its mostly the model + fashion girl, I don't think the makeup one is doing anything XD
it did add some freckles xD
I did, a thing 

cough coughmommycough
Fair XD
for color scheming, does SD reconignze color hex codes?
nope
fawk
already tried that
i got different shades of maroon im trying to achieve in various places

downloading Synergix, seems like a lot of fun
I'm looking for good hands (:
whats that?
i've put 3 hours into generating the same image with inpaint to get one right so....
a model from @real nexus
link?
I'm downloading that one :o
i'll download the Megamix too i guess for "anime" style
im so happy ty ❤️
What do you guys think of my OC?
looks cool
https://civitai.com/models/54592?modelVersionId=59052 i also made an inpainting model for synergix
As usual you guys a model with perfect hands and details... im really picky like that! this one has OVER 50 MERGES !! please consider supporting me...
what's the difference beetwen normal and inpainting model?
its personally the best ive tried ... im not biased its the only one that allows me to actually make detailed eyes
well with inpainting u can use it for infnite zoom
and to change how stuff looks in a picture in the img2img inpainting tab
Next up, Synergix LoRA 
There is one already :o
Works pretty well too
Its a big download
But it just works™️
haha idk how to do that
cuz it took a lot of merging
damn it works well
I will do it 
hehe im glad you like it 🙂
heck ye model collab legooo
im excited to see this one with synergix
its really cool
Its a pretty nifty comic-like style :o
How much vram do i need to make loras anywho
i have 8, and thats enough
i need some help guys
?
my english is not the best, what do you call this kind of uniform?
Hmmmmmm. I need to think on that
Its not really something you see worn in the west from my experience. Having a hard time trying to find the right words for that ngl
I mean, it just seems like a school uniform
Or maybe kinda like, military school uniform
gakuran
Oh yeah thats its
Damn I didn't know it had a name :o
the one in that picture is more specifically like the one nobara wears in jujutsu kaisen
thats the original btw
like chikarin reaction 
yea, it was also made in koikatsu
yeah neither did i, learn something new everyday
🤨
what?
why she crying...
anyone know what padding in ultimate upscaler do?
because i stole her teddy bear
@real nexus You're model's realism is very stronk :o
the best model i´ve used yet
is that good?
what is it?
i hope its good
i like anime style more but looks awsome
I mean, its really good semi-realism :o
i didnt even know my model can do this
might be worth to train a teemo lora on it

That's horrible. Yeah, I've seen it before on blessed2, but rarely to that extent. I'll have to do more testing to see if I can find ways around the problem or to mitigate it. SD is a complicated web of strings, and each pulls on the other to a certain extent. 
A pic without realistic tags :o
stunning!! which model? 😮
What is going on with them? 
Yours xD
XDDDD
im dum


Makima in Paris
Well although the eyes are red instead of yellow lel
To put it simply, a repeating almost half-tone like pattern in fuzzy/indistinct elements of a gen, usually in the background. I've seen it before but rarely as bad as the example Nubby showed me. It's a speckled quality to what should otherwise be a smoother or more detailed presentation.
It's apparently common in NAID-based VAEs.
sorry for spam but hands
Yeah... hands...
😮
baguette powa
Oh I see what you mean, so kinda like blobs?
if u put no half vae inm commandline args it should be fixed
or maybe im wrong
I don't get that, but they do apparently
try that
It's a grid-based localized blurring or gathering of similar shades and colors.
u click edit on the user.bat file
So a blobbening
Hmm, I've had that for a while. Maybe that's why I really haven't had that issue in a bit. I was talking with Nubby who did have the issue.
@supple raptor Looks like this might fix things with the VAE-based pattern issue. #🍥|anime message
I had to do that a while back when I kept getting black gens from using VAEs with large models.
it can do both
why wha happened ?
I mean, the results don't seem that great atm 
A bit too uncanny, and a bit too bland on the lighting
hmm yeah the prompting on mine isnt like rev
so u dont need to go crazy and prompt engineer it too much
I'm using Holokuki prompting, works fine on Aurora 
u can just have your idea put it on the prompts
I like to engineer tho

aurora with all due respect is a model that requires a looooot of work to get what u want
like a lot of prompt engineering
and nothing bad about it its great for a lot of different styles
Maybe, works great tho 
yeah its great for a lot of different stylesa
i make my models as easy as possible and usually gpu friendly as much as i can
Just like upload like, varied pics, like if you got uhhh
80+ pics or more, preferably 100+
Just throw em into a drive
Send it my way
And I'll make a LoRA out of it 

I would do extraction, but I don't know how extraction works, + probably takes longer 
oh ma gawd

mhhhh
ooo for the model?
need to find easynegative for testing
hehe my new model
I find that my limitations have been only knowing what proper language to use. At least I think that's why it hasn't felt super easy to get to do what I want. That's usually my issue with models anyway--they often have their own working language.
mhm thats a caviat with ai it seems
anime models prefer danboru tags for example
and realistic straight up descriptions
A LoRA for the style of the model, so the style can be used easily on any model 
albeit both work with my model cuz it also has my anime mega mix in it
okay ill get pics ready for u ^^
do u want nsfw in there too?
Yeee, nsfw is fine 
got it on it!
Also Makima, and Also In Paris.
yes
I am probably one of the first people here to ever notice that in nai vae
or at least associate it with it
@real nexus Tis what Derrian said, it might not be too bad to extract a LoRA if its not humongous in size like the ones I saw on Civit 
Well, which one was the first one you merged xD


This is the channel now. Kazoo showing us all the cool places that he's seen and we can't go.
XD
hehe
just get community archive role

stable society
that's the real flex
now I cry
XD


anyways, I wanted to root my phone to run oobabooga on docker but now I gotta wait a whole ass week
fuck xiomi

you are a musician?
You betcha. Composer, audio engineer, and musician of well over a decade.
Primarily cinematic, but I've done many, many others over the years.
wow
revanimated
K. Don't dox me though, lol. I have a YT channel. I'll PM.
how many dies irae references you done?
I doxx myself too much to doxx others
Sent
Oh dammet, I don't think I have rev
what if princess peach was in LOL
Nope I dont
its a good model for merges
Well, I'll wait for your pics then, too lazy to download rev xD
rev is really good
am i the only one who has problems with vladmandic img browser sending metadata to controlnet?
thats a very nice one 🙂 i wish we could see the moon on that one tho🌚

I've heard that tho
Oh she is a her
Fair tho
But I wan make style LoRA 
Well, LoCon, so anyone can use the cool style 
i dmd u pics
that sounds sus


am i the only one who has problems with vladmandic img browser sending metadata to controlnet?
@untold glacier
None as far as I know. 
My influences are usually more contemporary.
Johann Johannson, Olafur Arnalds, Yoko Kanno, Michiru Oshima, HZ obviously, Jeremy Soule, etc.
EDM: Porter Robinson, Madeon, Wolfgang Gartner, Uppermost, Mord Fustang...
ᵗᶦⁿʸ ᵖᵉᵒᵖˡᵉ
anyways I'mma just upload here a 55mb wav file
Lol
not related to anything
am i the only one who has problems with vladmandic img browser sending metadata to controlnet?
idk I use a1111
there is an extension for better image browser
I stopped using Vlad after my first attempt to upscale anything, sadly.
that would be?
wait a sec imma find
goreshit enjoyers arise
Name had me worried before I hit play.
hehe
after seeing his sad video on his life idk what to think anymore
it was like way too sad
Drums be like 
thats literally what i have
There is two actually
rly?
didnt watch that video yet now im scared
yes
that one does not have sent controlnet to txt2img, and thats what i need
chinese be like
I'm all for DnB, but not a fan of anything with -core in the name. It's a bit too harsh/fast for my taste. 😅
This would fall into that category--it was great until the frantic beats, lol.
breakcore
then what about jungle liquid dnb intelligent dnb, something something?
First movement (I. Adagio - Moderato) of the Cello Concerto in E minor (Op.85) composed in 1919 by Sir Edward Elgar (1857-1934). Performed by Jacqueline du Pré and the London Symphony Orchestra, conducted by Sir John Barbirolli.
Image: Lane near a Small Town by Albert Sisley (1839-1899).
just cello
i feel like breakcore is something you at the beginning neglect you show interest in it and then it slowly creeps up on you until you listen to it every now and then and suddenly you remember every album name from goreshit he ever produced
very nice
still debating on releasing this lora
as an autistic + adhd person, breakcore actually makes me feel the deepest parts of my soul
MRBEEEAST
venetian snares more than goreshit though
wha
that meme
i totally feel you, i have adhd as well 🙂 also i could suggest you XI-world fragments album its really good
same to all
never heard of that thanks
This is nice. I'll have to see if it's on Spotify.
I knew you'd like
simple, but cool 🙂
It gives you nearly all known genres of music and lays them out on the page according to how organic, mechanical, dense, or dynamic it is. Very fascinating.
2 in one day
a lora and a model
This is a LoRA for a style i call animepunk ^^
if yall want the style of the images i just posted
there are some wild genre names in there haha
any keywords to that lora?
nope ^^
I need someone who wants to test a new neg ti. I dont want to give any infos on how it was trained to prevent bias while testing :)
wut
someone down for it? :O
well its not a LoRA and well yes, just testing, but its okay I guess. Will do it myself with some variety.
yes
Need to step up my game again since im back
send it
I would like to try it too!

Nope. I've tried that. It mostly solves the frequent NaN errors that those VAEs give when run at half precision but doesn't solve the texture. It also makes the VAE take more room in VRAM which noticeably affects your max generation resolution.
Yeah, I figured you probably tried it.
One more reason to step up my game. Although idk how good easyneg2 is
where
It does say "easynegative" not "goodnegative" :3
easyNeg v2 ?
I only used no half vae for like the first 2 months or something
You uh, mind closing the embeds?
There's an easyneg v2 now?

it was so sad when I realized it was practically useless and just took vram and made it slower
always the pinkys that missing
I constantly get NaN errors with VAE and nothing has fixed it 
I'm looking on the gsdf huggingface page, I only see the original easynegative
sometimes i can work around by swapping them
woa
There's likely no clear superiority or inferiority between this and the previous embedding, so feel free to choose according to your preference. Note that I'm not specifically recommending the use of this embedding.
oh
No, you're pro
There should be an X to close embed right here, to not take up as much space. This is what I was referring to when talking about closing embeds, 😅
NaN errors occur because [redacted] based VAEs aren't able to be run at half precision, which is what Auto1111 does by default. The two fixes are either switching to Anime2 or MSE VAE as those do not have a problem at half precision or add the commandline arg --no-half-vae, which will mostly solve the problem but also increase VRAM usage and slow down generations.
I use anime2 😭
I just switched to vlad, but full args I used on A1111 were --autolaunch --precision full --no-half --no-half-vae --use-cpu interrogate --opt-sub-quad-attention --opt-split-attention-v1 Keep in mind, I have an AMD card and use the DirectML branch
I've experimented with different settings but was never able to figure out a way around it that didn't result in my speeds utterly tanking
Hmmm. There may be another problem that causes it on AMD then. I've never heard of anyone getting the NaN error while running anime2
my thought is its something related to either the sub-quad or split attention, but im not 100%. I was able to get it to stop more often then not in A1111 somehow, but in Vlad its back with a vengeance
Do you have to use those to run it on AMD?
at reasonable speeds, yes
otherwise I couldn't generate any images over 512x512 and it took forever
basically only slightly better the nCPU
man I wanna watch redline but it's too late😭
I wonder if SDP or xformers are possible to use
xformers, no unfortunately. I think current I am using SDP becuase its on by default in Vlad
Any tips for training style composition lora?
Also, yall who aren't watching Tengoku Daimakyou, go watch it 
We need a SD server moving emote that's just a hand wiggling 
If youre not too reliant on extensions etc you could run shark. I dont like the lack of features it has with auto but eventually shark should be better down the line, heck if its true that we getting rocm support for windows soon™️ auto1111 will deff be stronk
I did give shark a try, and now that i know I don't have to use Dynamic Thresholding for everything to get it to follow my prompts I'm more likely to
use it more i mean
I never use dyanmic thresholding tbh
Part of me wants to dust off my development chops and see what would need to be done to integrate the vulkan code from Shark into A1111
but that's an area of code I dod not know enough about 😅
Now that we have SAG, there's a lot less reason to.
oh right, didn't Vlad integrated SAG?
I dont use vlad either
whats sag
I just switched because Ishqqy, blessings upon him, put in a pull request to give Vlad native directml support
So vlads repo works on amd?
i dont think the pull request was accepted yet, but I pulled the branch for it
um
give me a sec
You'd pull this https://github.com/vladmandic/automatic/tree/directml-stable
Ahem. Sorry, lost myself there for a second.
SAG stands for Self Attention Guidance: https://ku-cvlab.github.io/Self-Attention-Guidance/
I am confused and abit dumb could you clear what im looking at?
ic
Oh its like... a git thing. It's a separate branch of code that cna keep changes isolated. Ishqqytiger, the guy who does all the DirectML stuff, made a separate development branch of Vlad specifically to work on integrating directml into it
and that's the link to the branch
Ah cause whenever i looked at it just brought me to vlads repo
Without all that extra stuff you mentioned.
Cos i tried running this and i was bound to my cpu
UhOhOhuh testv1
why ride a spider chair, when you can ride a spider+table set
really unexpected result
cant stop laughing at this lol
this is honestly such a good invention, imagine in a restaurant you dont have to search for a table, the table searches for you
lol
goes into the back to get sprayed down
maybe one day
could do it with wheels first
gahhh so pretty!! many moth!

I wish, its Solution Epsilon from Overlord, shes kinda like a slime girl if you touch her boobs she pulls you in and eats you alive
I don't think thats her specific power
ahhh, Epsilon
I spot the similarities 
I mean, it is her, I'm using a LoRA 

Can't make her do the crazy smile 
Waifu hair change? 
Looks good to me :o
what model is this?
Thats adorable :O
OOOOOMMMMGGGGGG FINALYYYYY
took me fucking 7 hours straight
finally rooted my xiaomi phone
bruh
if I were to do it the normal way it would have taken 7 days
but I prefer to ruin my sleep
why tf they made it so fucking difficult
at one point I thought it was bricked
grz
I don't know why but it has a manga style
Can someone remind me (if they know) how to set the probability for the ai to select a certain token out of two choices?
Only way I know is doing ( token1 | token2 ) and it'll blend them 50/50
I think you can set specific time frames the prompts will be applied for before switching, but im not quite sure about the syntax
it'd be on the A1111 page though
Yeah ik but if u use dynamic prompt theres a way to set the chance number
Sunflower Princess
(story) India and Japan independently discovered ritual magic in the early 13th Century, which allowed the user to channel energy given from spirits to accomplish greater and far more complex forms of spellcasting. This became a dominant method of spellcasting for much of the known world, but frequently clashed with established religions that were less tolerant of working with outside spirits (Christianity, Islam etc.)
beautiful 
POGGIES
ghostmix?
finally, lillia is coming tomorrow OwO
yea, I just tried out the fractal stuff, was thinking about using it for AI video projects
nice should be cool, deforum doesnt work for me sadly
I haven't used deforum, I've been using controlnet with img2img batch processing
using reference video to transform
no idea what any of that is lol
cept controlnet
ig i mean
idk what that workflow is
basically turn every frame of a video into an image, then process every image in img2img with stable diffusion with the help of controlnet, then turning the output images back into video
neat, should check out deforum, can do some cool stuff with that
supports contrlnet
I've seen what people can do with it but I went through like 3 different SD extensions before I got sick of the headache of AI video extensions. I'll try it next but that stuff just drives me crazy cuz the render times are so long so testing phases are excruciating
Yeah nothing really cool is instant lol timr and patience
I just need a better graphics card, for a 20 second video it can take 4 hours to render all the frames, cuz its img2img on 400ish frames
Yeah im using rx 6600 lol could be worse
I didn't know AMD even works with this stuff lmao
Thats pretty nice :O
only 4 fingers but dont worry about that
im going to worry aboutit now

Though I think the other hand is more concerning xD
oh dear lord
woahh
macro anime xd
Damn that looks dope :o
?
Not sure, perhaps though :o
They have an artstation page with their stuff if thats what youre looking for
No I'm looking for physical books. Looks like there's a couple.
Oh unsure
gonna have to look into this
Fair :o
Wouldn't mind a Nixeu one either.
I made LoRA to get those eyes 😄
so how ya'll doin
yeah I gotta hit sleep soon too
quick debate: deliberate or dreamshaper?
At least in terms of variety of styles including anime and realism.
doubt those two can do anime
Deliberate can do Anime apparently.
Dreamshaper, yea, but to an extent
Not an option, but could try

Gavendra, Goddess of Love, Art and Beauty, as well of some of her followers, during the Basant festival in Lahore in 16th Century CE
Can I have a photo
very large, giant girl
about stories
Yeah, I was experimenting with using multidiffusion for text 2 image. I haven't been able to get the background to be truly coherent with the subject size.
XD
Yeah, it can make some cool images. I just wasn't going for giantess.
why what happen?
Basically attacking me for making another semi realism model without even looking at the civitai page
My whole thing is that I don't release models unless the hands are good 95% of the time
I literally can't my brain won't let me
And without even looking at the photos they curse me out lmao but all I wanted to do is share it in case other ppl wanna use it eh whatever it happens
what did they say?
I don't think the style is unique I was simply asked to make one in this style cuz of the way I do hands
Is all
eh who gives a fuck, some reddit idiot, just ignore lol
If these are bad hands then 😭 what are good hands ?
He then said the hands are bad and image quality is shit
Bro what
Eh reddit sometimes is funny
or maybe a peace sign lol
I mean the hands are good
Aw maybe
🤔
If it dosent always give good hands
Or even 80% of the time
Then rng
Definitely rng
got this on a couple of tries
waving, hand on hip, and peace sign usually gives perfect hands lol
Interesting :0 I wonder what that would look like with my anime punk model but I'm in bed ><
And honestly might just pull an all nighter at this point
Yeee
What is the quality difference between fp16 and fp32 models?
Not much iirc
What if a fp32 is pruned into fp16?
nth much

Thanks
Is there a space between those colons?
Stupid question maybe
Oh ok
Cuz thats what was throwing me off mainly
I just realized how well SD could be used to make an ARG
how is this pose called
cool.
what did reddit do now
What model is this?
base model?
Yeah
dark gemini 3
Yes that one
I am impressed
when i get outa hospital im gonna make a lora with a bunch of different eye types, same with hair styles and merge it into DG3, test it and maybe release DG 3.5
thank you, been growing the base since like january I think
started with V1 then
do it with mbw though so it only keeps the eyes or hair styles and none of the other data
okay
there's a guide over at projectai
I'll try to remember it when I get back home
Is the earth flat
Stupid question achieved

is it not?
I'm a flat earther, just so you know 
Right? Imagine believing that you live in a fucking ball of dirt/water.
Probably starting to watch this, although I wonder if I can survive without crying.
Somewhat already depressed before watching
Also this
what do you think
@merry tendon @queen void just for fun, this is what i saw u guys interact last time 
.
Ai 😭
I feel like theres a rly sad thing i missed
Tho I actually also wonder if there will be any lora for aqua and ruby
There's already one for Ruby
actually nvm the whole gang's already being made
https://civitai.com/models/40182/oshinoko-characters-lohaloconfullckpt-will-update-along-the-season-hoshino-ai-hoshino-aquamarine-hoshino-ruby-arima-kana-saito-miyako-saito-ichigo-gotanda-taishi
The model is trained on ACertainty Please check the version notes for further details including the clip skip to use Normally you should be able to...
ah nice
damn thats pretty good
what happened
i like it
nerfgod what loras have u done recently
only one
Pina Artstyle LoRA Use Cases The LoRA is in itself very compatible with the most diverse model. However, it is most effective when used with Aurora...
this is insanely beautiful 
a quick test of my model's capabilites of non characters with regional prompting
i like it
Someone mentioned trying out koikatsu as a character maker, so I used the default character, put it in blender, made a render, and fed it to SD. Initial blender image:
Turned into this, and stuff like it
It looks like it got shadow confused with clothing designs
I think for consistency, better to simplfy things, ie avoid multi-color and details.
I was hoping it would make the character's shadow on the ground. Maybe I have to change lighting in Blender a bit more
wat?
This is really cool! I've thought about tryign to do comics with SD but I was always worried i'd constantly be in a battle with the program to get it to do what I want. Do you just use controlnet, or what?
you stole my idea
jk
do you use koikatsu sunshine?
You gave me the idea lol, thanks for that. I use sunshine, yes, it has wider capability.
made this yesterday
nice, do you use better repack?
yep, everythng I use is the "free" version
got my region smixed up but i like how it turned out lol
me too, but im still on version R7
you are prob. on version R9
can you give me the controlnet parameters?
I think I'm on version 8, I can't remember. All my infrequently used programs are on a trashy old laptop, more common stuff on desktop.
Didn't use controlnet
Let me send you the screenshot of settigns
dm please
yeah i used controlnet, mostly openpose
what?
ah lol
Do you end up having features mix when you do that, or are you using something like Latent Couple to avoid it?
I have issues even with two, where the features described on one show up on both or switch between the two at random.
using regional?
do you use region prompt?
Yes i used regional prompter
such a good feature
that sounds like the other extension, i dont use that cos im too dumb lol
ohhh your tiled diffusion looks cool
i use kitchen theme
oh i have the same one i'm just dumb
I wonder what type of workflow is good for character interaction. For example if one character is grabbing another one by the collar, it should show its hands being curled up, clothes twisted where its being grabbed, etc. I assume 2 openpose modes would be needed, but maybe not regional prompter because the 2 openpose models automatuically means 2 characters.
bahahaha
does kitchen theme work with vladmandic?
You'll still want some sort of region control if you're trying to specify how the characters look.
does kitchen theme work with vladmandic?
still can't figure out how to have different eyes, but i think its the model i'm using rather than region prompt
Would it work if you use img2img and just color in the eyes very basically and re-feed that same image back in as well?
yup that could work, but i want it to be text2img only
creating full scenes with different characters// interacting with each other
@lyric zealot sooo going back to the weight prompt thing, is it supposed to add the weight number to the result it picks? Like if i put “0.3::shoes|0.3::heels”, and it picks the former, it pops up in the recorded prompt for that image as “0.3::shoes” instead of simply “shoes”
isn't that just for the chances of popping up?
yes
I mean yeah but for some reason it shows in the prompt as well, beside the option that its still picking seemingly 80-100% of the time
I also have UmiAi installed but im being honest when i say that almost all functions regarding dynamic prompt that i use have no reprecussions or issues with it also being enabled
just wanna make sure, you did use { and } right?
Always
my example
1girl, {0.1::green hair | 0.9::pink hair },```
what does that do?
using dynamic prompt, you can choose the percentage of how often it comes out. In my example above, green hair has a 10% chance whereas pink hair has a 90% chance
Yep
Thats pretty much what I had
How much do spaces matter
almost doesnt matter, tho im not too sure when it's inside dynamic prompt bracket aka { }
Maybe it’s because i trying to do skin tone. I havent tried any other tags
But even still it shouldnt appear in the prompt like that
Using a Blender render with the character from afar, its great at getting perspective right. But...
how do you place the char there?
Output character looks like artifacts
How do you mean? I moved it with the move tool (accidentally messed up the alignment so its not right on the character but w/e)
how tf did you export the model from koikatsu?????
SD isn't able to make good quality stuff at low res which is why hands and eyes can be better at high res. So I think for far off shots, it might be better to do it seperately and downsize
https://www.youtube.com/watch?v=JSdggnGtcmU Enjoy! Super powerful
Series playlist: https://www.youtube.com/watch?v=JSdggnGtcmU&list=PLhiuav2SCuvc-wbexi2vwSnVHnZFwkYNP
A video series that shows how to export a model from the game Koikatsu to Blender. Once the model is setup in Blender, it can be exported to an fbx file for use in other programs like Unity and Unreal Engine.
KKBP plugin pack: https://github.c...
There are tools for it, super easy (as long as the right versions are used)
regional prompter can be used with img2img too? I haven't tried because I get errors using Vlad for some reason, I'll try in a couple of weeks once things are updated a bit more
This is pristine. I remembered reading the convo this is based on as it happened, lol. 
if it was not for the fingers... but i loved that expression
I just wanna be able to make normal 2d art with them but training koikatsu OCs is difficult
I swear you are literally the only person i see actually use consistently different styles of art
dont thrain then
just use controlnet
works pretty good
But
The ai cant generate them if they arent trained…
as i said, use controlnet
Perhaps generate many img2img outputs based on the koikatsu character, and then use the outputs to train instead?
I tried that, unreliable
nah, img2img is not raally good for koikatsu
then use controlnet
But what is this supposed to solve tho
i dm´d you
I think it's unreliable because the character has too many details? Maybe a simplified outfit would help
I think its the 3d artstyle getting in the way
bro, i sent you a dm
thats how to solve that
well since I'm training n artstyle loras xD
but thx
model?
Yeah im but theyre always significantly different
I have to wonder why a lot of people dont actully try using other style loras and things
aurora + 2 loras
What 2 LoRAs tho? 

i'm coming back from a break, what are the current best practices?
is kuki still king? did we get better vaes?
whats kuki?
There's Aurora, new model by our very own Luna 
model
i'll check it out
hf or civit?
it's definitely very versatile

Well im exaggerating ofc
idk if people realize this is purely vanity lmfao
Also, on a separate note, we mustn't forget where we came from:
Wdym
I did about 3 seconds of testing a month ago and found that it's better than 0. :3
Its nai’s normal seed number, and nai is the foundation of almost every anime model
im training lora for the first time, can anyone tell me i should i be worried? i have no idea what to expect lol
So i feel like it makes sense to use that
Expect coherence, not chaos
and then you realize it's for the "elite / ELEET / 31337" meme
meaning i should be seeing something coherent to my dataset?
Yes
so i should deff worry then if im expercting stuff like
I've tried other numbers. Maybe it's all in my head, but 31337 produces the best results for me.
model?
Holokukiv2
that could be an issue xD
100 epochs? 
too much? i need guidance pls 🙏
one new and one private :/ (the private has legal reason why I dont want to release it)
an artist?
yes

Oh
for the most part i left settings default but i did change 1 thing, maybe i didnt do it right?

but everyone uploads artists
i haven't done any artists...

im completely innocent
I know but I personally support the artist :)

cutesexyrobutts
I know but none are "good" :)
No kidding, I just deleted an old embedding of that and was going to look for a LoRA.
if theres a good video tut on dreambooth lora id appreciate a link
Wait wot 

What are you doing? I thought you were training a LoRA 
this shit makes no sense 
I'm not familiar with all the webuis
But uhhh
I think the console should say the steps 
i dont like dreambooth and only heard bad stuff about it, but thats personal preference i guess xD
Also that seems like just the settings for the samples, and not the settings for the training 
I cant kohya
Oh rip
Well, I don't think I can help then, I've never used dreambooth training xD
Hecc
so you dont use Koyha ss for lora training?
any loras? still holokuki? prompt?
no lora, just prompt, and yes still holokuki
just doing very simple prompt to get interesting result 1girl as a ghost, horror \(theme\),
removed best quality, masterpiece and stuff
i think this is what u meant
with allot morre going down obv
You got uhhhh







