#๐ฅ๏ฝanime
1 messages ยท Page 224 of 1
there strawberry fairy ghouls
kinda
falls over laughing
ok was not expec ting a close up
hmm she's part strawberry
ok last one and i'm to bed.. thud
ok i Lied I gotta post this then bed
strawberry hat
runs away
๐
is very very similar to Nefertim Osphe
Nya!
yep i see
New cute
true!
thats her lora but i got the ears by accident
ended up being cuter than ever ๐ฅฐ
Disney-style mermaid
Alas. while I'm busy with work. ๐ญ

the harsh truth of life
coot!
Working fox
nooo i canceled this
cool
๐
dont worry - be happy!
Track 6 from The Legend of the Lion King Soundtrack - Hakuna Matata
๐คฃ she look great!
nice
Do you take long to gen
yes
thats why
my cmd line args are --xformers --disable-nan-check --medvram --no-half
also 60 hires steps๐
what you use ?!
wdym ?
WebUI?
yes
A1111 or Forge?
what a gen time for you?
this is SDXL mode?
no
1.5?
why 60 hires steps?
most models dont really have an actual difference after 7-10
i'd test it at least, those are the slowest steps ๐
what Sempler?
DPM++ 2M Karras
some model/sampler combos (or loras) need the extra steps, so it might be worth it, but usually not
yes i use 6 loras
actually didnt no that
yeah, but do they actually benefit from the extra steps?
hmmmm barely
my standard is 30 steps, 7 hires (if i do hires at all)
oh ok
and i only hires images i want to keep, i don't have time to waste on an extra ~3seconds per gen!
i am think your gen time can be not a longer 3min. with HiRes
yes, without hires it`s like 5 min
i would do a x/y/z plot with hires step count with like 10,20,30,40,50,60 and then compare them and see where the best time vs quality sits
what a card you use?
1.5 with no hires is 5min? are you not using a gpu at all?
taking too much time for me doing that
or very slow card
mine is gtx 1660 super
on my 1660 6G 20-40 sec. with out HeRes!
???
yep
we have a same card i think
10 min in my case this is a SDXL gen
ok then give me your sittings ( i`m a noob) ๐ญ
not a 1.5
oh nice
ok
and have a build in ControlNet and some extention
and time?
u should teach me what to do
and what a Upscaler you use in Hires steps?
I could never do image gens with such a slow card, I'd lose my mind. This took 2 minutes for the x/y/z plot, 1024x768, hires x2
here all do it with pleasure
6 min
thank u so much โค๏ธ
anime 6b
yes.. gift me your card ....๐
i'll sell you my old 3080TI ๐
still try to install the Forge version
ok will try
it's much, much faster!
how much?
What new version lol
this is
nevermind, just checked amazon and prices are under $1k for 3080TI now.. I'm better off keeping it for llm stuff
fummmuuuuu.... ๐ญ greedy.... ๐
should have sold it a few months ago when they were going for more than i paid for it ๐
hey hey - this is not my
๐ช
this is gen @zinc flax
Super cute
and this is mine. for which I asked for forgiveness.
and I asked forgiveness to @zinc flax
unless of course I'm confusing anyone ๐ฅด ๐ค
I trained using an old 1.5 lora gen just to use that style with xl lol
she look like a little Kikki
๐
@vital raptor pirate demon girl sailing through lava

I don't think these boats are doing well

guys
whats the reason im getting black images on forge ?
this image was with no hires and it worked but
images highres are black?
yes
1.5
any other extension active?
I was right... do not bing watch "the strain" and goto bed.. shivers
we just install Forge
not going back sleep now
worked but so bad image
copy one model
so what we doing?
when genning with forge base, hr gives black images
you can show screen shot your settings?
this feels unconfortable
i'll run it
this is your model
can you share the black image with metadata?
this is i am generate now on yer model + some Lora`s with hires
why the vae is set to none and not default? this way load the vae from the model?
hires works fine on sd
i was trying hires
but will copy them
copy a VAE file. and select it
ok
Strawberry!
strawberries mister?
looooooool
I hardly use hi-res
for first steps HiRes maybe unchecked
i can gen normally in high res
even better than ever. my vram is running in a temperature of 45ยฐ c
it's not even straining
yep
What 1.5 model gives the most detailed eyes? (I want to use it for inpainting)
maybe just use a tokens in prompt?
I use blender render of characters as I put for img2img but there is a little bit of clipping with the mesh. So I need to inpaint to fix them. So I thought I may as well use the best inpaint models possible
Because I use controlnet I can keep the composition, but it includes the clipped mesh (the iris cuts into the face and eye whites)
(the iris cuts into the face and eye whites)try a higher mask value?
? It clips when I don't use inpaint, just controlnet close up. If I use inpaint the surely it will be fixed
I don't quite understand.
In the blender program I can control which direction the eyes look. But the character's iris cuts into the eye whites so it doesn't look good. So that's why controlnet also looks bad when I use it with the image. So I need to inpaint without controlnet. So I may as well use a model with a lot of detail for eyes
See, the whites of the eye cover up the iris
๐ but maybe fix a crooked mesh in the blender?
this is not good mesh
maybe this is animation deformer
The white dots are just part of the cheek highlight texture. The characters are made in a character maker program, I just exported it and then imported it into blender.
Fine. then where exactly is the visible problem in the rendering? here?
Yes
yet. as I remember in blender when animating it was called a weight map
changing which corrected such an overlap
Maybe I need to move the iris more forward?
in any case, it is advisable to obtain a render that is correct in relation to the mesh
Yes but since the mesh is not good as is, it would be very limiting (if I can't make the character look in different directions.) I was thinking using 3d character maker and 3d objects with SD could make a colored manga very fast. But there are problems like these
such small artifacts, in my opinion, can cause problems in SD
is it just me or dall e 3 been getting worse at anime
it used to generate crazy good stuff
Yes SD is not good with small details. It can do main things, then manual adjustments need to be done
@terse ore try max corrent in Blender. btw. you can use a move mode with big radius of action on the mesh
you can also change the curve in this mode
Thanks I'll try when I get back from work
๐

@ivory pebble she look veryy bored
Colorful!
Playing around with pixart sigma
Sometimes its really good and sometimes I get utter trash 
๐ ๐
Basically every single model in a nutshell
Yeah but this is like the SD3 beta that ran different models and all 4 outputs were vastly different
Maybe my workflow sucks 
I know I'm just joking, to be honest I haven't even tried it so who knows
Also prompting for Zelda gives Link 
Did the SD3 channel disappear or am I going crazy?
Uhhh, I have no clue honestly
oh I couldnt find any images of Ouyang Xueli from Cupid's Chocolates anime series. I mean in danbooru
anyone got images of her?

Oh wow there's actually 0 images ๐ญ
Is she a new character?
ah noo
from 2016 anime
hmm woow 0 images (pixiv)
guess I have to rewatch and get some images

Yeah I can't seem to find anything either
Nice art man, keep it up! 
It has more grainy texture
You need to account for that
you need to seriously add tendrils in negative lol
I dont use any negatives 
well, you can still add without tendrils in positive ๐คฃ
cute
were you given access or is it public now?
it's been "public" thru api
im pretty afraid its going to be really slow on local
sounded like it runs fine on 4090 so should be fine
still not going past like consumer usage
30-40 seconds per generation even on a 4090, image the rest of us then, and thats only 1024x1024
optimizations will probably come, that's how like sd 1.2 started
was super slow, then greatly sped up later on thru optimizations
and i have a nseaking suspicion not everyone has 4090 here
seems like normal ai hardware to me, kekw.
I was on the closed beta server, but since the API released its not useable anymore
30-40 sec on a 4090 is on 8B model tho, im sure there are otehr ones
multiple sizes were confirmed afaik
and emad on reddit confirmed the biggest model works with 4090 at the very least. guessing 3090 can do it but slower
i only use the best; 8b for me even if smaller can produce comparable-ish results
but it doesn't look like it's going to be too good for anime off the bat, might stick with nai until waifu diffusion 2.1(3.0 based)
what makes u say that it wont be so good for anime?
as a base model

the anime outputs aren't that good off the bat for anime from what i've seen
Just fine tune it
where anime output good on the old base models tho?
No
i was able to make a lot of good things on sdxl when it first came out, better than sd 1.5 finetunes
but i don't see 3's base model being as impressive of a leap forward for specifically anime
if you put a lot of time into prompting & setting up a good workflow, it'd work.

this was pure sdxl
ok
did you test it yourself
idk, ive seen some anime from sd3 and it looks quite good, but idk
but its a little worrisome that staff only show the same fox girl in here with sd3 lol
All I'm posting in the chat is sd3
These days
That's just pretty much all comfy ever gens and he's the one that posts in here most.
yes haha, makes me wonder why tho, but its okey
he just likes fennec girls. even before SD3 she was pretty much the only thing he ever posted.
Lmao
Comfy will make sure every model makes fennec girls
A spy working for the fennec girls community inside sai

@cerulean cosmos
๐
still the same further distance eye problem and bad handz. it'll still be good when we get some finetunes off of it, hopefully sooner rather than later
someone had asked about my artist run...
Hires would be good
Yummy
ew
I don't see their names can you turn it up a little bit
Like turn up the upper uppingtons of the upperitons
I like this one better.
Lmao
Yung one, my glasses don't work as well as they did back in my days
Can you zoom in more in that picture
Cx
yo guys
how to fix this
i think this is why im getting black images
is there a line to ignore it or fix it ?
watchu looking at
from what i can tell it has something to do with a broken lora, or upscaler model but it's hard to tell
Maybe different upscaler can work idk
?
I think i said something about the set all looking samey
her face a very very cute
hello funyaa

shes litening to space man
for sure
๐
what model?
autism confetti
i see today you using a new model?
It's SD3

I mostly use that and pony cause they are the 2 I can use while I'm on my phone
And I'm always too lazy to be on PC gaming stuff cx
@nova remnant SD3? not interesting... pressed the button that said โmake it beautifulโ and thatโs it ๐ฅด
kondo nanda?

Cheating 
I don't know what that is
just forget bout dat
runs a batch
runs a batch of this
ooh
grinz evilly... let's get crazy
ok it got better.. whta?
guess it's just me this morning
was cooking up some sort of thing. lol. I think i have figured out how dynamic cfg works ๐ฎ
nice
not sure why i'm in a ninj girl mood
but.. shrugs
hey, everyone has moods at every time
when i'm using playground v2.5, i love the creepy porcelain dolls
aight, lets see what my upscale workflow will do with this image ๐ฎ
it'll just take 2 minutes for all 28 steps ๐ฎ
oh, the name in the upperleft is the artist name that pony responds to, not sure if it's an actual trained artist in pony -> but i know pony responds to it
I think batman's in trouble
ngl, this one looks quite well ๐ฎ
perhaps lower the initial sigma on the upscale ๐ค
vivid colors, medium view, dynamic action, Full body portrait, illustrated air brush, Kinuchi-Ninja Quinn joker Bat-Girl, wearing stockings, ninja dress, headband, leaf background, realistic, photo, movie, soft edges, multi-shaded colors, anime overtones, line art style, Midjourney style, hypernet:mjv4Hypernetwork_v1:0.8octane render, finely Detailed, elegant, sharp, focused, Gradient color, soft shaded, UHD, Hires, HQ, elegantly detailed, 4k, black thin outline art style
that's on a sd1.5 model i suppose?
else i'd have said -> pony doesn't support quite some of those tags
well a mix
I've saved a bunch of styles I like to mix and match ๐ , saves time
well, who hasn't made their own mix/model at this time? ๐
i took the liberty to change your prompt up a bit ^^ think it will be a cool initial result
interesting way of holding lol
might not be your taste, but i like it nonetheless
unhinged canadian ninja?
hey, you specified leaves ๐ฎ
changes it up
there we go: Naruto Quinn
ooh a female Naruto Raised by Harley quinn after she left the joker
or the joker died, would make a interesting fanfic. ya know I bet it exists
checks
there's always the clusterfuck called injustice
shame about the right hand tho ๐ฎ
dmn pc locked up and I lost the one I was making. sigh
checks only fanfiction not archiveofourown.. that way leads to danger
hmm ok I guess it doesn't exist
very little quinn left in this one tho
oh boy Naruto: My Harem Academia--- Chapters: 125 - Words: 151,775 -
@bitter latch here we go!
"hiya bats! I been training!..(hand appears out of thin air, smack on back of head) I mean cough, Greetings Honored Bat.. umm.. man? "
that shadow!
heh.
tho the best naruto fanfic crossover was where naruto found an ancient armor with blades on it and a school with the sign of "The Foot" on it, and new era begins again.
๐
hmmm thats an idea Naruto Shredder
I really should go back to bed
๐คฃ
@brittle rampart how's your day going fuuunya , here's a meow
uhh oh


meow back
Fluffy white fur ball!
fluffy in prompt can't go wrong
my first 4k !!!!! @brittle rampart thanks for help
I'm eepy
!!!!!
2.5 min per image
damn! it must be a cat! ๐ฅด
noice !
noice?! hmmm ok raddit ears - is noice
yes nice !
help
hmm, the restart sampler is half the speed (5-6it/s vs 10-12it/s) but produces just as good, if not better, results at 10 steps instead of 30 ๐ค
of course, what is 'good' or 'better' is highly subjective, but it seems like a reasonable trade off for a bit more speed
Looks the same honestly, is it better than euler a tho?
I remember people were talking about that sampler when Pony got released and apparently it's better than euler a, however I've never really tried it
tried wariza pose, seems like its not work on pony

is this already out? wut model is this? (but also only hundreds of characters? pfffff.)
4th tail
ahh, doesn't seem as ambitious as animagine xl based on wut it says in desc
o there's also 5th tail which is a finetune off of 3.1 though
The hell are tails
wat
What does this mean
a model
It's better than NAI at tails
That's pretty questionable
Both on why tails and why it needed to be said lmao
nai is only model that gives this obscure af char her tail. other models might produce like a partial result on the char but never her tail xD
granted it messes up her hair color 80% of the time (shuld be green)
Just make a lora
can't use loras on this model
I'm saying for any other models
i dun like loras
i prefer all-in-one models
xD
animagine xl 3.1 for local & novelai for private
but just genning her without that she always gens with like
blue/pink/brown hair
nu i luv her
and she isn't randumb
she is the sister of wynn
and connected to charmers, eventually becomes shaddoll-ized
then reborn
she has tons of lore
i use her reborned form the most tbh, super useful in random decks
That's extremely random to me

It's a Yu-Gi-Oh card
wat isn't a yu-gi-oh card
there's even jotaro jostar as a yugioh card
and the entirety of dante's comedies
magical girls, vtubers, etc.
No clue, got that randomly lol

it's either V for victory, or I"ll take 2
Yeah usually you prompt "v" on pony and it does that
But later I noticed it was 3 fingers up lmao
or peace. i know it as peace sign or victory sign (and for nai you have to use "peace sign" instead of v)
"hi "
lol
One question, what is the seed option for in the stable?
not smol
๐ญ
there were artists in the prompt
now coot

@green temple more smols for you

mottoo !!! ( more ! ) @nova remnant
euler a tends to have more variation the more steps you have, but ultimately most generations are basically the same across all samplers, it's the smaller details that change and it's all subjective on which is better for your own uses
What model?
wtf
๐
Just prompt I dunno

it was "A cute small girl with pink dress and cat ears with bright eyes, pretty skin, and a cute smile, adorned with a crown of flowers and a tangle of vines down the trees, her hair is a light shade of blonde, in the magical garden of the castle, whimsical style of japanese anime style"
don't go overboard. if something of dubious taste or offend you, just ask politely the author to delete it, or ping the staff to take action. don't sour the mood of the channel plz
sam price presence offends me pls delete him
We need to do something about it
And it's gonna be thick
one of the characters show her underwear. it's not a little too much explicit?
dw im gonna talk to her right now
Wtf

not from below ? 

@nova remnant early morning coffee?
She has a bouquet in her sleeves
Super secret gift

she had to free her hands to hold the cup! ๐
oki-doki, then!
What about cup with hands

Scarryyyyy! how drink coffee from this is ?
chin massage while you drink
๐
the main thing is just the cheeks and nothing else with such and such hands!
https://github.com/Akegarasu/lora-scripts
anyone use this?
dunno how to install dis
i think guy`s from tech support chanal help you with this
againe Sd3?
โค๏ธ
bubbol

You okay
pony?
No, sd3
A not so safe to touch bubble
I added Lego movie to the exploding bubble I'm going off rails here 
oh wooow
wait is he flying?
floating
Super Saiyan fox!
Hmmmmm
Maybe it's just the first power level thing
weak fox
Hopefully when people inevitably finetune sd3 for anime, they will have artists in tags
Don't finetune with images of a specific set of anime styles (that sounds weird as hell), because result will be garbage
anything else?
skibidi toilet
damn bro u got famous ๐๐๐
*you're
Our*
are you using the api key or some other options?
I rlly wanna try using this for myself
no credit )):
Look dm
Man first time using kaggle
And like did alot of things and wondered why nothing works
APPARENTLY INTERNET IS NOT ON BY DEFAULT๐ข๐ข๐ข
Bro wtf is this design choice
I'm not sure honestly, even NAI removed artist tags from their new model, and artist tags were pretty much the only reason why people even used it/payed for it
It's also pretty much confirmed that PonyV7 won't have artist tags, so yeah...
Unless someone who has enough money to make a model as powerful as NAI/Pony with artist tags then it's not looking good ๐ญ
because artists loras are better than merging artists onto the model unless you have NAI money
but its kinda pain to use loras
Innate artists will always be better than artist loras, every single artist lora that I have used hasn't been as flexible as an Innate artist.
But then again it's just like you said, you'll need a lot of money to make a model that can actually replicate artist styles without losing that flexibility, which is why I mentioned Pony and NAI
thats just bad training
yes if your card can fit thousands of artist images(or if you can find them) to add to your checkpoint its better to do it that way
models are moving away from using artists tags because they're (possibly) infringing on copyright by taking a bunch of an artists works and training their model on them and then leaving the name in the data so people can replicate their style, most of the larger providers have been dealing with lawsuits over it
removing the tags means they can still 'steal' the art but not have direct proof of it
and also, you won't be able to directly reference it to replicate the style as easily
what i'm wondering, is how will they reference the styles though, because a model like pony knows hundreds of styles, but without tagging them directly it just spits out the same ugly default style
To be honest the main reason why I (and many other people I assume) use these models is to replicate and merge different artists together. But I understand, I guess there isn't much we can do about it
maybe they will get a better version of ipadapter and u will only need to upload a pic of your fav artist and it will copy the style
that would be pretty cool actually
Astralite talked about it in the PonyXL discord, seems interesting but I have no idea how it's gonna work
for realistic images it's a bit easier, since there's not really a 'style' in the same way as illustrated art, so it's easy enough to train a lora on a particular face/person and then merge it with a model and have something custom no matter where the images came from
for anime stuff it's a whole different thing
They've been talking about these "super artists" for a while but I'm not exactly sure what they are
as long as there is some kind of master list of all the trained 'styles' that's fine
most likely all it is, is changing the names before training
It will be fun to mess around with, that's for sure
You mean like, changing an artists name to something else?
instead of tagging an image with 'nemu' you tag it with 'notnemu' or 'superartist34' or whatever
Ahh okay, makes sense
it's still using their art and style, but putting it under a different token/tag
And lets just hope we can get a list for all these new tags, cause if not then it's gonna be such a pain to prompt for all of them one by one ๐ญ
? wut since when
Their new Furry v3 model has no artist tags
Which makes me think if they're gonna do the same for their next anime model
NAI don't you love fucking money??
My guess is that they probably will end up removing them, they're a big ass company and they donโt wanna risk dealing with lawsuits
of course they like money, that's why they're removing tags so they don't have to pay the artists they stole the work from ๐
tbh, i don't think anyone should have to pay for training a model with peoples art as long as it's derivative enough, which pretty much means not being able to replicate their exact style/character/whatever
but, i have no interest in copying anyone either, and i don't generate existing characters, so i'm a little biased
Now if they somehow decide to leak the NAIv3 model then that's a game changer
Highly unlikely but I got hopes
hmm hmm I like dat
but whos gonna do dat

Why are you putting a choker on a child?
Sexualized poses, or borderline underage sexualized characters in particular cannot be shared in this server. Heavier actions such as time-outs or bans will be considered on repeated offences.
@vital raptor
Wooooow !
which model
NovelAIv3
link?
Just type in Novelai, literally the first thing that pops up on google
how is it different from midjourney?
Not sure how midjourney works but Novelai is uncensored and it has artist tags + it's really good at mixing those artists together. Also knows pretty much every anime character that has over 100/200 images on danbooru
NovelAI is a closed source finetuned SDXL model and also has a few unique features available through their service. Vibe Transfer which is similar in use to ipadapter, but they claim that isn't what it actually is and SMEA samplers which don't have an opensource equivalent. SMEA samplers somehow handle larger resolutions better (I have no idea how or why it works, but they really do reduce the typical problems that you would get when generating above the model's trained resolution. They also provide API access which can be convenient for allowing you to do things like doing a hires-fix pass (upscale+img2img) using local models.
Midjourney... we don't really know what it is because they're more secretive about that. Allegedly it's not based on Stable Diffusion at all.
How big of resolution can those samples do

stupid people with money
~1536x1536 equivalent seems to be no problem for them which was about as big as they let you go on NAI. I mainly used them when upscaling images since they tend to make styles a bit softer than I usually want at normal 1024x1024 equivalent resolutions which is the max size you can do unlimited gens on the Opus Tier sub (Opus Tier is the only one that's actually worth while). Depending on how often you gen and especially how often you want to upscale your images, most people could probably get away with only subbing to Opus Tier once every 3-4 months and using the 10k credits that it gives you to gen normally in between.
yesh
A few of us have been bugging one of the guys at NAI to make SMEA open source since they created it.
It's not likely to happen anytime soon

Not SD3. SD3 doesn't use samplers as we think of them.
It's basically an LLM.
That just makes images instead of text.
Maybe if/when NAI makes an SD3 model they'll grace us with opensource SMEA.

fingers cross

hmm forgot to change the model
was making a Picture of two mules side by side in full tac on farm land... ok then
gues sit's from an anime?
oh I put framing instead of farming.. hahahaha
well tried a different and it did it again... hmmm?
hmm kinda reminds me of that game, death stranding

so fun to prompt with sd3
"she has a pink themed outfit and is prepared to the fanciest party in the castle"
lmao
cool
cute adorable girl "she has a pink themed outfit and is prepared to the fanciest party in the castle"
BING
๐
non-bing better

that first one, at night, definitely turns into a creepy doll inside the big mansion
Im outta here
want a laugh, i'm tryint to do two mules side by side, well it did it but....
lmao
yeah i got their sub to try out midjourney , their pics are all over the place lol , but it feels like its highly tuned somehow

!!!!!!!!
This gives me an idea
@green temple gave present to smol
๐
And one day I'll know how to not generate snow inside

๐
Box

She found a portal
Inside a box
cx

Just like the emote!
@brittle rampart look at him

they again so cute โค๏ธ https://www.youtube.com/shorts/ccE3geiSV6Q
#viral #xuhuong #trucxinh #vanphong #dilamcogivui #nghile30_41_5 #genz #gogovnz
the model for sd3 better release soon
๐ @brittle rampart
Is that a threat? 
what's with all these randoms tagging me lately
Randoms do random things
had one earlier pretending to be a mod and threatening me with punishment for my image, talk about needing a hobby..
It's just a burden of popularity! ๐คฃ
i'm popular? ๐ฎ
why not?
Try harder 
deym
tbh my loras werent any better 

Im currently trying to make PixArt work, but stand alone pixart is kinda bad quality wise
Gotta try anime with it 
Found this on 4chan, it's quite a long read but I think it might help you
Don't mind them, they're just an alt acc lol
I got the same treatment
Here we got something more unique 
hmm read, eh but I dunno cant find a option
Reference_Only on PonyXL
Maaaybe ๐
ofc !
ahh its not going
is webui allowed on kaggle guys?
catodeejay!
yeeeeiiiii!
you must be in form!