#π₯ο½anime
1 messages Β· Page 146 of 1
h8
the new adetailer update few hours ago must have changed something to make face inpainting much better
That rack is impressive if I do say so, myself.
oh yeah?
its updated quite frequently
Beautiful girl, blonde, long hair, freckles, green eyes, wearing a colorful, vibrant, detailed embroidered dress, outdoors, night
I liked it so much that I upgraded it
@dull jackal ^ 
Trying out KuroGishiki (its the same as Gishiki but it has noise offset
)
evenin'
hey maybe ya'll know
in auto1111, can the x/y/z plot be used to render across different negative prompts?
from what I know it can do one negative at a time
I was talking earlier
What you mean ?
you can do a list of prompts easily
it'll just iterate through them, using the same negative
I want to do that, but with negative prompts
hmmm
but isn't that what you want ?
You can tell it to change the negative , if you put the negative you're using inside the prompt s/r
prompt s/r would require me to not use commans in the prompt
which is a pain in the butt π
can you not do dots to separate them like commas
I believe that has the same ish effect
I'd have to change my negatives to use periods instead of commas. Can't use periods as the separator.
yeah
that's the annoying part
@keen hollow but what exactly do you want to test ?
prompt matrix can also do that but it will try each variation
Morticia and Gomez in love
prompt matrix is definitely not the way... haha
did a bit of googling and frankly, it doesn't seem possible
I just wanted to test a few negative prompts to see which one I'm liking most these days
the difference between . and , is very minimal from what I know
shouldn't be a problem to use s/r to test
if you have a big list of prompts just tell chatgpt to fix it for you :P
the thought of editing all my negatives is... daunting π
the period vs comma does change it...
oh well
gonna have to go manual
wait
... styles... you can plot styles
this might be it
ok for some reason (ironic I know) came out creepy
π
the hell is that
the guy that likes to bully π€
Bro, what did u use to lipsync your anime ?
ok .. what kind of pastry is that?
wholly cannoli that's a necro... haha... ummm
try this one... https://www.d-id.com/
D-ID, the leading AI Video Creation Platform, helps you make photorealistic videos using generative AI via D-IDβs API or Creative Realityβ’ studio.
yeah, lemme see what you come up with
X/Y plot with Prompt S/R. Syntax is:
"negative, prompt, one, here","negative, prompt, two, here","negative, prompt, three, here"
notice, no spaces between the quotation marks and the commas that separate them.
wouldn't that be three positive prompts?
I figured out how to do it with styles, so using those. π
Sure !
Nope, because the negative one would only be found in the negative prompt.
oh, duh.. was thinking, nevermind...
forgot about the quotes trick π
I didn't know about that one for quite a while. I don't think it was in the documentation in the beginning. Not even sure that it is now, now that I think about it.
interesting hands
genius
better
@nova remnant
ugh i can't get a prompt/model that is resonating tonight at all

ugh i hate all my anime models right now
surprise..cat
definitely different.... not sure which is best though. I think I need some new negative prompts...
oh there's a bug
either 0, 5 or 8
it doesn't unload the TI between negatives if the next in the series doesn't have one
interesting
i should probably post to the github
then again, 4000 issues are posted and ignored, so, maybe not π
ah ha more anime
reordered to get around the bug
this is more accurate
tough to beat 8, but it's the most different
you could try closertodeath's negative works really well on some models. It's basically required on her models. I don't remember where to get the etone negative embed though, I can send it if you need that one. easynegative, (etone:0.8), (worst quality, low quality, normal quality:1.5), (simple background, white background, abstract, abstract background, uneven, double exposure, odd, disconnected:1.5), holding, teeth, tongue, nsfw, midriff, ((multiple girls, 2girls)), ((watermark)), ((no humans))
I'm finding the embeds kick it a lot more change overall and are a bit more unpredictable
like easynegative in that one
but but double exposure is fun
I need a prompt to test this some more. Someone gimme a fave go-to. π
you can always remove that one when it's what you actually want. if you aren't going for it then it isn't good.
kirisame marisa, blonde hair, yellow eyes, side braid, superb, witch hat, large hat, looking up, eyes closed, relaxed, sitting on a rock beside a stream, (barefoot:1.1), shoes removed, nature, outdoors
tests hands, feet, general anatomy, and character recognition all at the same time.
reworks his prompt
ill give it a whirl
waits
rendering...
ponders trying multiple layer prompt script
should give something like this
only doing these at 640x768 for the test...
yep
i'm using monstercoffeebang or whatever
looks like she's in pain or constepated
not liking my models right now π
based on that one, I would eliminate the last 4. 7 definitely looks the best here.
neg3, might look better with hires fix, so that one might be worth keeping.
i'm gonna run another example
just gotta find a decent base π
here we go
adding one more neg too
this one is another one of my go to test prompts
masterpiece, best quality, BREAK 1girl, long hair light purple hair, curly hair, cute, eyelashes, sitting, white dress, pink ribbon, belt, pink flats, white thighhighs, outdoors, nature, tree, bush, flower, rustic
cool ill try that too
maybe ill switch over to aniverse or something
Once in a blue moon, RNG blesses me with coherent fingers.
that one is from the new version of Kakigori which I'm still working on.
one of my goto for anime stylish is : fine art, soft colors, black outline style, elegant, very detailed Cel-Shaded With shaded detail's, soft shadows
latest addition to the merge was based67 which looks like this on that same prompt and seed.
last three are rough again
ok ill try the new prompt with hires
will take a little longer with the hires fix... running now
you don't know long tillyou've had to wait an hr over dialup for 1meg file
I know aging myself a little there..sigh
been there
I had dialup until 2009.
my earlist modem was.. hmm umm.... 1200 baud? with a phone reciever hookup
on a rotery dial
that was a pain
believe it or not there are, or where, bbs's online
but anyhow
just think how long it'd take to download a single waifu back then
there we go nubby... after hires
a trend i'm seeing is that NO negative is honestly fine a lot of the time...
I use Negative prompt: Scribbles, Low quality, Low rated, Mediocre, 3D rendering, Screenshot, Software, UI, watermark, signature, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, normal quality, jpeg artifacts, username, blurry, artist name, one tail, no extra legs kkw-extreme-neg
8 decided it needed a teapot π
looks like it has an embedding
ill throw it in there, but it wont pick up the embedding
yeah the negitive
kkw-extreme-neg
got tired of adding and got lazy
it's about done
the last 3 are the obvious ones to drop...
one more... then igotta get some sleep
Uploaded this one to hf
https://huggingface.co/sleepotimer/SweetMoe
Kakigori definitely likes very minimal negative prompting.
most negative embeds also make it worse in my experience.
taking photos my one demise
Me too
Not narcissistic and confident enough to take selfies
Only thing I have a picture of myself on is tinder lmao
Which I don't use, been years 
You didn't miss out much for not using Tinder.
Idk, maybe it's different if it's not in NA
Now too lazy to do anything, sleep instead 
I've never used tinder or any dating apps.

thankfully I met my wife in elementary school and dated her in high school so I never had to experience modern dating.
I'd be single if I had to try online dating.
I will never use dating apps. Let's see how many years will it take before I break my own vow.
@nova remnant make more chopsticks

But they come out bad
And very slowly
I need smol
What's elementary school
How old are students in it
America naming on things too hard to understand
English is hard anyways
time to learn Chinese
You get similarly spelt words with very different pronunciations
Drought
Thought
Thorough
Through
Threw
5-12 years old. we met when we were 5.
I'm trying to talk to people to not lose my social skills, people fucking bully me for being awkward.
Fucking bullies
read and read 

Cursed
Bat and bat
Smol sus is online and coming home soon, now you can gen more skinny legs later.
no i was not, nubby is talking about SDXL model with SD1.5 model as refiner
hopefully he isn't busy making tensor engines
smol is here
Gonna bonk me

then what were you doing

i've started learning japanese over the last few months (slowly) and as a native english speaker i've been surprised how much more sense most of it makes so far.
I started learning English very early
RuneScape days
When things on the internet weren't available for my language pretty much
:p
Runescape made me not forget English when I was in China.
yeah, i feel lucky for english to have been my first language
I was reading English stuffs too in the net cafes in China, I wonder what would the people nearby who saw my screen think, lmao.
i would not want to try to learn this confusing ass language as a second language.
Hmm
Russian is one I want to learn more
Wait
What is Russian like for you guys
It feels similar to English
Russian makes absolutely 0 sense to me.
thank you 
You mean like the order of words?
Yeah
Cyka bylat
i was literally scolling from where i last read, and replying to messages 
starts hearing gunshots and invisible mutants
Chiki briki iv damke
to be fair, I have never tried to learn any amount of russian, but looking at the cyrillic alphabet is even more confusing to me than trying to use 2 different alphabets in japanese (i haven't started kanji yet, but that scares me).
Chiki briki panchiy vriki (I have no idea how the scavs pronounce the 2 last words there)

how can you read those alien runes
Is how I feel
Portuguese and Russian... They just work
I think cyrillic confuses me because a lot of the symbols are similar enough to the alphabet I know that it seems like I should be able to guess the sound, but it usually isn't even close.
kinda how everyone feel about all other languages probably

Why can't I just learn them in a day

guess what time is it now

it's making more engine time
Go to engine jail
@lyric zealot been using that blazing drive model stuff
looks super nice
really, maybe the negative prompt that i have doesnt work well with it?
I dunno
Also
I have to try more
I don't remember what I did for positive and negative
I was liking euler a 30 steps
@lyric zealot later you try euler a 10 steps
how bloated with extensions is your webui ? 
think extensions run on my gpu ?
you meant with the blazing model? 
yeah
I was trying each 10 steps they all looked nice and completely different lmao
maybe thats what you mean by the noise at the start
Pretty much the only ones that have a chance of looking decent at 10 steps are Restart and the various DPM++ samplers.
same thing but 30 steps



i got silly tavern working
text gen thingy llm
Oh
its like an alternative ui from the standard one but u can make bots talk to eachother in group chat and stuff
do omao gens 
I can't really get the lora to work 
because its one of my generic quality prompts lmao
Niice
liquid rubber clothes
Cute round eyes
shes ready 
Can't get shocked if you're wearing those
smart
@lyric zealot

trying prompt matrix with the quality tags to see the combinations
@lyric zealot funny, did the combination of quality tags, and did a little blind test to see which image I liked the most and it was the one with all of the prompts x3
fastest amd gpu 
real
Fake but realistic disney movie poster with detailed credits and text "Hogwarts Strikes back"(hogwarts star wars hybrid film:1), fine art, soft colors, black outline style, elegant, very detailed Cel-Shaded With shaded detail's, soft shadows, sharp focus, elegant detail, HDR, detailed shadows, cinamatic lighting, unreal render, realistic
used a xl anime model
Sign saying "I don't know"
SDXL isn't extremely good with text

Its fun to get cats with little signs
Me praying for more thick thighs
proper girl clothing according to openai
KEK
terrorist?
nun now let us pray π πββοΈ
try restart
the sampler

Are you sure about that
let me find some rtx 1070ti for sale
oh my god the mini people there haha
Or is she the giant one
Either way I'm in
shes massive 
whatever makes it hotter
1 girl blonde hair, white shirt with printed words "I'm racist",anime style
@native halo come lets pray 
the omao images i posted earlier were from SD
Local stable diff is best
This
((Antonio J. Manzanedo, Anne Bachelier)) toss these names at the end see if something good happens
Generating, it will take a few minutes 
π oh nyoh
dalle

Yesh
@gloomy lodge oh god those did a lot on the images x3
never good quality but still @gloomy lodge
cross eyed one kek
Touhou or no?
I do not see frills, I declare not touhou
I'm sending an email
I'll ask them to remove that character or add frills
π
tohou or not???
holy bjesus
I tricked you. You forgot the other rule I told you. If it's not Miku and I gen it, then it's probably Touhou.
idk what happened here tbh
hard to do it because dalle only does square images π
π
And if you ask for an anime girl who's sitting then you're likely to get bonked for the inevitable upskirts.
I don't think dalle likes π¦Ά
i cant walk so feet are offensive π
Asked for a ghost. SD just shot the Cirno to make her into the ghost I guess.
shes witnessing the third impact and feeling the effects already
Ghost of a hentai monster or something. IDK
@supple raptor
I put (touhou:1.3) on the church girl prompt
idk what is this
looks like this cat whatever it is
MORE Prayer
praying is offensive π
one out of 50 tries
asukachads can't stop winning
"Please make steven not have access to dalle anymore"
done
@dusk seal help 
most normal looking girl in attack on titan
how does my open pose controlnet input give my character a beard 
the white dots in face look like a beard 
OH LORD WHAT IS IT LICKING
Its normal
do what I do. block'em
my bad we will only generate girls with burqa 4 u π
I'm so close to saving the children
Wildcards just threw a hammock prompt at the new merge I'm testing and it got it almost right in a single try 
yep that looks human
rip
why 
very cute 
I swear to god i dont get it, how can an open pose skeleton affect if the character has a beard
like wth
try with another model
I tried like 5 already
all that helps is putting more strength onto 1girl and using beard and 1boy as neg with extra strength as well

or try with a girls lora like a character
still makes no sense to me 
tensorrt such a gamechanger tho
real anime
Turtle Power! Get ready for a FULL EPISODE MARATHON featuring the Teenage Mutant Ninja Turtles, straight out of the β80s! Leonardo, Donatello, Raphael, and Michelangelo are the worldβs most fearsome fighting team, and alongside Master Splinter, theyβre here to stop the evil Shredder and protect the world from evil.
Binge the COMPLETE FIRST SEAS...
Keeps making weird hands in the bottom right, but this is really solid 

playin with low brightness latent
no more dark loras or prompts 
cd tuner extension 
woah thats cool ty
heres a darker one 
oh god the settings theres so many
this one is so dark it got full of artifacts 

SDXL is very fun to use
(when you can generate 3k images in 10 seconds)

Idk
3000x3000
Whatever that is
darker 
Too dark
Can't see the tight suit well

its ok ill take my suit off π
this is cool i think i was been loking for an extension like this
yea it also has add detail option havent tried yet tho π€
Cat found the pastel paint buckets
i always play with the color correction stuff when i post my stuff on instagram and ive been wondering if thered be a way to pre-process color correction
i wanna figure out the color map thingy
yeah that seems to be true. haven't used them in a while, and sees I won't be going back
ew
i dont know cmyk very good
people were up late last night, woo. ya'll are night owls or European π
mike 
i was gonna use an image program like gimp to get cmy values of the color i want but looks like it doesnt accept color values like 255.255.255 just strenght values from 1 to 20 π
we aint talking about that its about cd tuner 
... what?
I've seen that cat before >.>
this one

least bloated a1111 
needs more extensions π€
@native halo 
not sure what "Detail" is, but the C/M/Y color values are there
nope it only goes from 1 to 20 not 1 - 255 so its a strenght value
least bloated a1111 install
gtx users smh my head

@native halo Your bloated thing needs healing
i need my daily π

yeah gonna need to figure something out might ask dev for a color picker or something
yea its a new extension so maybe he will add more stuff later 
need to see it from behind to verify it properly

@native halo (Something) girl will get you, she angy

needs more dark 
ha ha ha it works π
problem is the watermark...
This is an amazing image.
haha right on, glad it worked for ya
all girls are inherently angels, all packed together in sugar and spice, but then there are harmful social conditioning that can affect a healthy mindset (basically the outcome of a preying economic model). also aging is unaccepptable which is actually a shortcoming in medical science
wildcards π§‘
when you're running a 90 minute generation set, and you realize about 70 minutes in that you messed up one of the parameters...
i changed up the prompt a tiny bit and got dis
pic looks much better than anything i could make with my own prompts
Civitai going full corpo now
bond yield curves have been negative for quite some time now. will be exciting when they go back up again.
Welcome to the server!
If your message wasn't meant as a joke response, then it's likely more suited to #πΆο½off-topic or #1008652627204132904
age of easy funding is ending.
π€ ?
Solve this:
looks good, idea with prompting is to use certain keywords that reflect on the core idea, try not to make prompts too big because that will get stretched too thin
im trying to prompt some stuff on my own do u wanna see?
sure π
i wanted it to be bright and yellowey
but ik theres probably alot i could do to make it look better
that;s pretty nice, i like it, simple yet very focused i also like the idea with color tone
it coot
but theres stuff with brackets u can do that im not doing, and i thought they make the pic look better
you mean the emphasis with weights? i don't use them too often, i only apply them on generic defaults and keep the rest of the regular prompts in normal weight
let me show you an exampe...
so this is my default generic that i apply too all anime
(highres, highly detailed:1.2), (aesthetic, masterpiece)
but for regular prompt i'd use the words w/o weights in most cases
so sometimes u dont want to use weights?
yeah i avoid them in most situation with my regular prompts
because when you apply weights that can affect other words too
so unless its very necessary to bring out a certain aspect in the image i avoid applying them too much
oooooooh
but then i look at other peoples pics and they look better, so then i feel like i should use weights to make my image look better too
well the pic you just showed above is uniquely creative so don't underestimate your own ideas, having too much noise and chaos in image isn't always great
what model are you using btw for that image above?
uhmmm, the one nubby made
In the end, experimentation and finding what works best with the model and network extensions you're using is a process. In my opinion, if you want to know how other people are making such great images, look them up on something like civitai: https://civitai.com/images
And click an image you like to see the prompt (if they've included it) and various settings. This will help you learn common uses, tags, methods of prompting, etc.
seriously? what you just did with it is pretty amazing
i really like this image that was just made
your prompts are right on
u can see the prompts?
its as creative to keep your images simple and focused as it can be to make them complex
Yep. If you pull it into your PNG Info tab, it'll show up if the metadata is included with the image.
i added the last bit myself with all the environment details and clothing
i yoinked the first bit from someone elses prompt
That's a great way to see how others' prompts affect the gen, especially across models.
i never know what type of language to use when doing prompts, like i mean on image boards the tags are very short, like 1-2 words so before i tried to do it that way, but doing the opposite also gets better results, using lots of words for a single detail
As for the amount of attention per word, it's best to experiment a little at a time to better understand what exactly is going on with the changes from prompt to prompt. Also best to use a standard seed, with multiple batches, e.g. 4, so you're not thinking that something is changing when it's not. If you have, "1girl, standing, peace sign, frown" and you're wanting to see how attention works, then try it one word at a time for the more difficult to gen ideas: "1girl, standing, peace sign, (frown:1.3)", etc.
you can take two approach with prompts, using keywords and using descriptive sentence, usually with sd 1.5 model keyword appraoch work pretty good, but with SDXL latter is a better approach due to its context algorithm
i always used keywords and not sentences, but now i see sentences also work well too
This is because some models were made with booru tags in mind. Most of the anime models are like that, although there are some exceptions (DarkAlfa being one). It's a difference in prompting language. More realistic models usually are more lenient with language and wording, and generally work better with natural-language prompting as opposed to booru-language prompting.
SDXL vs 1.x is quite different when it comes to prompt-language. SDXL uses a much more natural language system, and will not work well with the old booru-style tags.
sdxl works very very slowly on my pc
Yeah, and it'll likely be that way for a while--if not forever. If all you're wanting to do is gen anime, then it's best to stick with 1.x for now.
Also, welcome to the server! Glad to see new faces getting into SD.
I-
its hard to understand things
your explaining is very very concise but im also an idiot
It's a lot of information coming at you all at once, yeah. I get it. One of the best ways to progress is to try out new things. Try other people's prompts. Test out changes. Always do so with a standard seed (setting the seed to something and not changing it) and with multiple at once (i.e. seed 1000, 1001, 1002, 1003) so you can see what's happening accurately and not come to the wrong conclusions.
Don't worry about downloading tons of models and LoRAs and trying to understand how they work. Start with one, like Nubby's you're using and see how it works. Then move onto a good popular model like Aurora (for booru-based tagging) or DarkAlfa (for more natural language based prompting.
the CLIP system really just sees "Tokens". Tokens are just numbers that relate a word (or any string of characters like a phrase) to a specific node.
Also, ensure your clip skip is always at 2.
ooooh, right i get that, because i always click the dice little icon and it changes the seed to some ungodly high number, but if u put the seed at a set number and change prompts u can see what is actually being changed and stuff
Exactly.
downloading more models is tough fo me tho
u needs the right vae too for the pic to even end out good and i can never find those
Use this one and don't change it, it doesn't matter as much as some others think when using anime models: https://huggingface.co/hakurei/waifu-diffusion-v1-4/blob/main/vae/kl-f8-anime2.ckpt
you would usually just need 2 vae for all your renders kl-f8 yumeshiro just linked and the other https://huggingface.co/stabilityai/sd-vae-ft-mse-original/tree/main first one is good for anime the 2nd works better with realism
no changing?
I'd argue that it doesn't work "better" with realism, but just differently. I rarely change these days. π€·π»ββοΈ
If you're using an anime model, then I wouldn't.
i will be always using anime models
okay maybe not always but a majority of the time
its just my personal opionion but sometimes the realistic ai images look very odd and creep me out
at least when i see bad anime pics its not that bad
There's definitely an uncanny valley there that doesn't happen with anime, as it was never real. 
Though some could say that people are just used to bad anime art. π
I've looked through my old anime wallpapers folder and found the average quality a lot worse than I remembered.
100% agreed. Even realistic images that are pretty good tend to be a bit creepy.
theres something i dont understand, whats a sampling method?
whenever i change it stuff looks weird
Here, I just woke up and don't feel like typing an explanation: https://stable-diffusion-art.com/samplers
Haha, here I was expecting a standard Nubby explanation.
Good morning to ya.
Good morning
morning nubby
nubby and yumeshiro are the best teachers
where would we all be without them
Aww, I'm here to help out once in a blue moon. Nubby is far more in that category. I appreciate it though.
shudders in #π€ο½tech-support 
Lol
i used tech support in other servers and tbh it was better just to ask in a main channel
not entiely sure how it is here, but tech support in other places sucked
I think it's better here than it used to be. When I used to hang out in there from time to time, it was usually just the people asking for help trying to help each other. Hardly anyone who knew what they were doing was in there. It's been a long time for me though.
it depends, people try to participate
but this is a free product
nobody gets paid for support
so good attitude and waiting power is usually key to good help
The ones that I recommend you use though are DPM++ SDE at 18-25 steps for solid gens with good cohesion, DPM++ 2M Karras at 30-50 steps for gens where you want more detail, or the original great sampler: Euler A at 28-40 steps for the simplest gens like anime-screencaps and such.
Wait, you mean it isn't good practice to ask the same question every 30 seconds and get mad when nobody replies by the 3rd message?
That's an actual thing that happened in here a while back XD
it happens a lot, I participated a lot in tech support over the years
a lot too much to be honest
so reminding that nobody gets paid and everybody that answers you is just doing it for your pleasure is usually good :p
I'm so tired right now, I've been dying a lot in Lies of P x=) you tried that game ? quite the great time
I haven't, I'll have to take a look. BG3 has somehow managed to keep my attention for 500 hours so far.
damn, quite the good run
cause I posted at the wrong time :p
Holy cow
I need to get back into it. I've been in 2077 lately.
I need to get back into BG3, I wanted to earlier but I've been working on my streaming channel, and BG3 is quite heavy on good streamer, hard to take any place in there
thus Lies of P x)
You stream? Wanna link me?
it's in french but sure. still a new channel, and I learned yesterday that I could record my lives... so the channel is quite empty still
not the fingieeees
Followed. 
thanks, it does help x)
Question with this. What has your experience been with Karras samplers as opposed to the base DPM samplers?
I have defaulted to DPM++ 2M Karras for a while, but I don't really have a robust reason for that over base DPM.
Yeah, I made an error and corrected my comment there. It's 2M Karras.
you juste described exactly my problem too, and my situation xD
I've done some looking into things, here's a recent comparison of the two on Aurora and Breakdomain:
2M Karras
2M
General sampler comparison on Aurora
I wish I had a better idea of the effect each one has. They produce relatively similar results, so it might be nice to be able to look at a gen and say "I think I need to use SDE for this prompt."
Be aware, the images are quite large. π
Generally speaking, what I said here holds true in my experience: #π₯ο½anime message
If I want a gen with the best cohesion, I usually go with DPM++ SDE. If I want more detail, I'll go with 2M Karras. This is all for anime anyway.
For more realistic stuff, I often with with 3M SDE--unless it starts glitching up my gens with the LoRAs and network extensions I'm using.
I'm still a huge fan of EulerA
Euler A is great. It's just a bit too indistinct on the finer elements compared to SDE for me.
Hi Guizmus, I hope you've been well
this one came out nice
Hey there panda π it's been going mostly ok, thank you π you're all right too ?
moooooooose
Restart Sampler

I would say okay too. Just keep doing your best, that is enough.
I'm already restarting the sampling process with hires fix, why would I want to have its attention split 4 ways as opposed to 2? π
Here waifu
Sponsored by Restart sampler
you got to keep on being yourself for sure. And doing your best, not always the simplest
Working on that channel does help me feel good those days, a thing to do each and every day at least :p
Yeah, it's good to have something to keep you busy. Even better if it is something you enjoy.
Nun mommy
Also sponsored by Restart samplerβ’οΈ
What's that model you're using?
cute
KuroGishikiv1.1
- My Sciamano240 style LoRA, and hadanuri1
Lol. Time to reboot SD
O, I never get that xD
I hoped NaN exception came with kebab
they don't
Once I finally got it working, it... Restart doesn't seem to like me very much.
It looks like a bad fanart, lol.
the left hand looks normal to me though
You got hires fix? And I mean
I haven't sent my LoRA :O
And I also got Adetailer on
I had most of the LoRAs, and adetailer is on. I didn't see any other LoRAs in the PNG info. 
Sciamano240, hair slider and faceage :o
Yep
Well, its pretty important for the style :o
Indeed. That's why I got them, heh.
Wel, faceage and hairlength are just for the hair length and face age :o
They don't do too much for style :O
? Then what were you referring to when you said "its" pretty important for the style?
This is how it looks without my Sciamano240 LoRA
My Sciamano240 LoRA :O
I trained it, so you can't have it
xD
You may have some other sciamano240 LoRA
But I trained a better one 
Yeah I think theres up like 3 on Civit
But mine is the best 
Ah, well that's the difference for sure.
Like to hold on to your LoRAs then huh.
Though, model still looks good even without my LoRA
I have like
30 unpublished LoRAs XD
Cause I'm lazy to post em and no one asks for em xD
been experimenting with negative prompts, are starting to wonder if just using NO negatives often yields my favorite results...
careful owo

No negative does look kinda cool 
It's just been that way across multiple prompts
running another one
I was checking about 10 different negative prompts, narrowed it down to these three as typically the best
I mean... it's subjective obviously, but... no negative wins again
I mean yeah I think no negative seems to be winning :o
The Karras part is the scheduler. It controls how noise is added to the generation between timesteps. enough information here to probably make your head explode
I find that Karras scheduling usually looks better on the DPM++ 2M sampler than without, on other samplers, I've found it to mostly look different, but not necessarily better.
No Negatives VS With Negatives
which do you prefer?
I think in this case, I preffer with the negatives because it looks closer to the style I want :o
what's for dinner tonight everyone?
wow.. the negatives altered this one like crazy...
idk yet. my wife will probably pick something up from the store since she's going shopping after work.
whats a lycoris?
lycoris is an umbrella term for several different types of LoRAs.
We got Mommy Miku Milkus β’οΈ
uhm, idk how to get one to work on a1111
theres no tab thingy for it
If your A1111 install is updated to latest, you don't need a new tab
It works like normal LoRAs
do i put the stuff in the lora folder?
Yeah
ooooooh
yeah, since I think webui v1.5, lycoris is natively supported in A1111 so you just use them like a normal lora.
prompts with steakpunk and wire elements that are more complex yield a lot of changes
simple portraits are a little less dramatic..
still kinda like no negative...
Yeah, can't really tell much on these without more seeds. I usually recommend at least 4 for comparisons like this.
just found out the wildcard list I'm using has jack-o challenge in it
she is now XD
it's better than 1.5, but it's really not good at it.
if you are using a finetuned model I would guess those are even worse at it than the base model is.
i'm so happy with SDXL, no special anime models for me anymore (except for... concepts that SDXL is not trained on π )
Dall-E 3 is the best I've seen at text and even it gets it wrong more often than not unless it's super common text like "happy birthday" or something like that.
yes - this specifically is with realisticfreedom XL, but base SDXL has all the styles you ever want in one model, just prompt for what you want, it's crazy (there's also SDXL style selector for automatic1111 that makes it even easier)
and 1024x1024 native means no more hires fix or anything for normal images
of course
i have wildcard extension thingy downloaded but idk how to use it lol
aalso
cool or nah
mods may say "that's not dress that's body paint" and question SFWness, though... but it's a bit random IMHO
what on earth have you just sasid
you hearted the image
therefore ill take it as you liked it
which one did you download and do you have the auto-complete extension installed as well?
i have autocomplete but i think its for danbooru tags only idk tho
ill find the extension
some text sample π
Fabulous.
i dont actually know what the extenssion is @supple raptor i just have smth at the top of my sscreen that says WIldcard Manager
ah that... ((((((sign saying "sample some text")))))) a cute anime woman with a big halloween pumpkin, evil smile, cats Negative prompt: loli,underage,teen,child Steps: 30, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 208406145, Size: 896x1152, Model hash: dce7eb8449, Model: realisticFreedomSFW_alpha, Style Selector Enabled: True, Style Selector Randomize: False, Style Selector Style: base, Version: v1.5.2
i like it, but mods may not. i am not a mod, no harm done
now it's getting surreal π
ah, I don't really like that one. The extension is really bloated and most of the wildcards that it comes with are really bad. I like this one better https://github.com/Tsukreya/Umi-AI-Wildcards
Using the wildcards is the same for either one though, you just have to enclose the name of the wildcard in double underscores. The easiest way is to use this extension https://github.com/DominikDoom/a1111-sd-webui-tagcomplete.git which will autocomplete your wildcard names so you just start typing __NameOfYourWildcard and select the one you want.
automatic1111, not quite recent release
yes it's a mess - but AFAIK the other infrastructures do SDXL now as well
auto1111 was the only option way back when i started and i stuck with it
i have no clue about comfyUI but i bet help is just a quick google search away for you π
etc etc
wats a wildcard list?
it's a text file or yaml that contains multiple words or partial/full prompts, that can be randomly added to your prompt. So for example, instead of having to pick the clothing you want your character to be wearing, you could have a whole list of clothes options in a file named clothing and just type __clothing__ in the prompt to have it randomly pick one of them for you.
wait a second, when u download the file off of like civit does it come in a zip folder?
or is it something u dont download?
maybe you open a thread to discuss this? not much to do with anime π
for illustration π
-_-
From civ, it most likely would yeah. You could also type your own. Most of the ones I have just came with this extension https://github.com/Tsukreya/Umi-AI-Wildcards
and you can see the wildcard files that it came with here: https://github.com/Tsukreya/Umi-AI-Wildcards/tree/main/wildcards
upping the image-to-text ratio
how am i able to recreate this quality for myself?
well, anime artwork color sound smell . anime style, key visual, vibrant, studio anime, highly detailed Negative prompt: photo, deformed, black and white, realism, disfigured, low contrast, loli,underage,teen,child Steps: 30, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 3701037164, Size: 896x1152, Model hash: dce7eb8449, Model: realisticFreedomSFW_alpha, Style Selector Enabled: True, Style Selector Randomize: False, Style Selector Style: Anime, Version: v1.5.2 with automatic1111 and a good SDXL model, running on my PC locally
there are good SD 1.5 based anime models, but with SDXL i can simply go to photorealistic style with the same seed and model, just a prompt change
good for lazy folks like me who just load one model and do all with it
yes it is, just look a her π
hi Lazzy ;P
SDXL is better with spelling but not good yet
need to get comfy or fooocus to get working before i can try sdxl first :/
but nice style, and just with the prompt













