#🍥|anime
1 messages · Page 21 of 1
XD
how nice of you to remove that big wall in the background
looks nice what did you use to get this image?
Nothing wrong with a good wall
of course, but now we can see the REAL important bits (the scenery)
Model or?
You can also use pnginfo on my gens
didn't work for me
That image is chill, but your modification to bypass clyde is against the rules.
True 
Works fine for me, should be fine since i have the stealh pnginfo
yeah like model or lora?
Model's cetus-mix coda2 with ligne claire anime lora
I almost always use those 2 together
why is it that when I download images here I can't png info
No problem, image removed. I didn't know about the update to the rule. But this is funny, good old Clyde always blocks innocent images and let lots of spicy stuff go through, it was blocking almost everything I was posting ^^, and the image in question did not have anything wrong
Clyde is pretty schizo
^
Clyde is just lonely and wants to chat with us
I wonder if they will put ChatGPT-Clyde around these parts
thanks! the hands were nightmare to generate
Eh, that section is a perfect mirror...
Meanwhile, this is supposed to be an anime girl.
She's just far away
Oh here she comes. She's just holding up the infrastructure apparently.
Doing her part of the community
Uhh
It's just systematically taking all of the waifus out of my gens--even on different models, lol. The heck is going on?
waifu said fuck you
She went to buy milk
She drank too much mtn dew
This is what mountain dew FIVE gum feels like.
Mtn dew collabs with 5gum
sidebooba lady
Cranking the close-up tag.
Good idea, let me do that
Not quite tall enough.
2🐝
2bussy
@rose tapir we got full nudity in here.
"no lineart, sketch, traditional media, (blurry), dirty" is all in my neg prompt. Maybe it's just a bad LoRA. CFG any higher gets me blurriness. Weight any higher gets me the same.
Sadge
raise your hands if you have questions!
Something normal once in a while 
Very cool
The last 2 made me think of this https://twitter.com/ueyamamichiro/status/1615714951284551680
That's... not the idea. You might wanna follow that link you posted to find out more about this server. There aren't any bots here to create with! You might also want to go to dreamstudio to get a trial: https://beta.dreamstudio.ai/generate
got it, thanks
There is no chat bot. This message will help you get started. #1072229020520947753 message
👍
ʟᴏᴏᴋ ᴀᴛ ᴛʜᴀᴛ ᴄᴀʀ ᴛʜᴏᴜɢʜ. ɪᴛ's ᴘ ʀ ɪ s ᴛ ɪ ɴ ᴇ
I haven't seen that one
Uhh
It came up in my Google feed a while ago. I never realized until I saw it that 2b and Violet Evergarden are both voiced by the same actress.
Oh yeah, good point
We need a washboard reaction, heh.
it just gave me 2b with no shoes, SD is into feet confirmed
I was trying for that earlier, but failed to get one safe enough to post 
or maybe you just have a feet model 
Here
Generating some sfw stuff for this
Last 2B for now
This is very hard to not get troll feet
Chinese foot binding gone really, really wrong.
I'm just showing the ones that look like feet
what is better than one 2b? 3 2bs
2b × 3 = 6b
Darnit, I was typing out that joke.

ricksha of the future?
👍
Nijijourney V5 ^^

Evil peach is best peach
Whoops, it was faint but there was a 🐫
We doing bowsette now 👀
Just Peach, but evil Peach
any good realistic upscaler?
I use 4x_fatal_Anime_500000_G for everything

Same; but I only test few other upscalers
So I don't have accurate knowledge on what's the best, maybe could come to usage

Hey there, it's the
police...
if you detect the NSFW, it's better not to post it usually, given the rule 4 #✍🏼|rules-and-tos
I removed 3 from you. Excessive nudity or intended sexualization in all cases.
End of
police, continue the fun 🙂
you said it yourself for a start, and then just what I was saying there. When the pose is complicated to interprete in any other way than sex, or that there is close to no clothes on a full body picture, those go over the different parts of the corresponding rule
maybe I should sell my 3070 ti for like 500 euro and buy a 4070 12gb for 650 euro
4gb upgrade 😛
or get another gpu altogether to go with this one and then run 1060, 3070 ti and 4070 together and different processes on each.
train lora on 3070ti, train model on 4070, run tests on 1060, support tests by running tiles on cpu
ASUS GeForce RTX 4070 12GB DUAL OC cheapest 4070 is 730 euro atm
it's the space DNE

i dont know if you guys saw the images, i made with niji
but to turn around and put them throw i2i with different models
is so crazy good
sometimes at least
she have nice ... eyes...
yes
its helltaker lora but for some reason i couldnt generate it till i just yolo now with random anime lora
mistoon with hell taker lora
is it ez to make loras?
lool why this guy have a sword
dam i didnt thought that random lora can work
go for 4070ti
@frozen oliveyou can make perspective by prompt locations also
i think with open pose positiont
The helltakers is it a lora?
@fading sageyeah
Nice, i really like the style of those pics
yeah bro they are looking high quality
I like the artstyle there
They look like some high quality digital art. But like, the kind where you had a proper hand sketch first
dam my gpu crash
Oh no
I'm guessing the artstyle is from the model or yet another lora
i am generating pose now
yeah the lora called helltaker the checkpoint mistoon
i took this pic as a pose
So you're using img 2 img?
I also have helltaker lora on my local
Its very fried but the model works really well with it
no just txt 2 img upscale with highres
omg
Thats kiara the vtuber
dam ai is innsane
@toxic dragonno just txt 2 img upscale with highres upscale called 4x animesharp 1.8 high res 10 denoise i do 50 to 70 till i find a good looking pic
50 to 70 gens?
no 1 gen
denoise i try between 50 to 70
i go up by 3 so like 50 53 56 60 if i dont find good picture
euler a
You mean steps? My denoising strength is usually around 0.4
yes
Usually a bit more refining but I honestly haven't been using stable diffusion for very long
40-45 steps.
guys,prompt ideas to make the background white, solid color
simple background
with 2 back ?
High res steps or what part?
Absoutely in love
thanks owo
All the helltakers without vae?
yeah :[
Sort of fits tho
forgot i dint enable
arf harr frrrroooo
The dulled colors works for it
i wonder how it is with
But try running one of them again, with seed and such, but with a vae
Reposting old gens cuz I'm not on my pc 
Try reusing the seed for a direct comparison with and without vae
ok
could be interesting
looks good with out vae i didnt expected it
Yea the dull colors fit that kind of style really well tbh
Oh wow, its good. Very different
You were running mistoon? Is it mistoon_anime?
mistoon_anime
IF YOU ARE NEW TO STABLE DIFFUSION OR YOU HAVE PROBLEMS WITH MY MODELS GO HERE: Guide ( inzaniak.github.io ) How to use this If you are new to the ...
masterpiece, best quality, helltaker lora:Helltaker:0.7
Nice
i use this prompt
wow
the colors
vae 84000
50 highres
insane how much vae enhance it
Is it the Hires steps, the 50?
Ah okay, how many sampling steps? I usually don't see any merit above 40-60 steps
When you hires-fix it, it kinda regenerates the image and makes the resolution larger. Heres my settings:
ok i woldnt pass 20 but i just test things
100 highres
rn its 28
steps
deno 0.4
Using the blessed2 vae, 40 samples, upscaler 1: R-ESRGAN 4x+, upscaler 2: 4x-UltraSharp. (Scales from 512x512 to 1024x1024 to 2048x2048)
I usually don't use hires, and just run larger batches of the 512x512 pics until I find one I really like. Then do that seed again but with highres.
That's cute. Got one that looks a lot like it actually
nice
nice eyes
can i try one too?
I use the Blessed2 Vae
hello everyone
Hello
Just run 40 highres. 150 is way over what you need I'd think
2:43
yeah i might go 40
looks good tho
i dont see diff
might stick with 10 less time
19sec
wow
im happy too :>
Identical to the one a little up in the chat
That's a wild model
this is a understatement
yeah, the dataset is a little wild, given it's the different themes of the PoW
but it got quite a lot of Mikus because of it :p
and cows
I was running just dynamic prompt on all the tokens I have trained
and it choose miku + cow x)
it's the latest PoW model :p
anime painting, cow miku mermaid cthulhu tentacles, detailed, masterpiece
clyde 😢
yeah no
none of this prompt will go through
works OK though
I am still making miku cows for some reason..
what have you gotten me on here
it kinda works though
XL?

how much resstep?
have fun with mikows, gotta cow for now :p
wow
better, the "cow" token goes away almost in mine
this is a good mix of the tokens !
less mermaid
these cthulu cow mikus are wild lol

Oh no what
hahaha
this is not the right channel for this
in paint and alot of tweeking around xd
dam
thiss is so nice
i didnt used paint in mine
last time i did
i ll shw you
sec
no tatto blush
what do you mean
its abit hard to use it
u didnt want tattoo or blush?
i mean
inpaint is ass to use
i work alot with inpaint
like what can you do with inpaint i know you caan change mouth eyes
areas
change readd
u can literally change everything
xd
?
oh i thought its not upscale
its upscaled

this looks interesting
3d
i think i can create with hell taker
this looks like if there is no vae used tbh
wow man
Nice one, but you should use a vae for that model
Idk much about anything and idk what a vae is
And hey all
it looks weird if u ask me
Not if you use the right one
like the head looks like anime but the body i would say not much
Time for my example image:
well
this looks like anime
yeah
non vae right
this is niji
Let me try one I do have but not turned on
might be with anythingvae
For this image i would use the kl-f8-anime2 vae
let me cehck
Its semi realistic so it would fit
It's too late in the night for me to do any downloading so I'll just keep at what I have
absorb the latents directly, no vae
lol yea, cow miku mix still
xd
got a few I am trying over here
looks nice
xd
what negative prompt u using
Better but yea you should try kl-f8-anime2 or mse 84000
A lot, im playing around before bed so im just using anything I can
Not trying to get anything in particular
Also I forgot I had a "realistic" in there which was probably screwing up the 2d-ness of the image
here is with that removed
Much better
Tonight, getting chat gpt to write it rn
xd
the hair even looks great
im telling you soon artist will use chat gpt to create anime character
Adding "vampire" in there too, let's see what happens
not much happens apparently
The background got cut
Yeah
add vampire teeth
Trying
You can try add
Vampire, fangs:1.2, open mouth,
this isnt what you suggested, thats loading now, but this is cool anyway
Just need some braces
put more wieght
vampire teeth:0.8
if you dont see the teeth put a higher number
loooks insane
Have to reload, gave me not much
use paint
Too much work, maybe tomorrow or when I'm at my pc to draw (I'm using Google remote desktop on my phone)
2
Grabber best xD
Tool for Downloading anime pics from image boards
oh
what lora you use?
True xD
You're right lmao
this is crazy
Very crazy
what prompt?
Basically just a lot of "vampire teeth" and "messy white and red hair"
Im changing the prompt a little every time so its gone now and I didn't save it to png
oh ok
Sorry
angry
There was a really good one I just made but I won't share it because too much butt 
Yeah but the mods have decided they don't want their tool to make content like that (which I respect of course)
instead of butt, how about boots
same
i respect them
for making this great tool to make arts
did you put vampire teeth
or wwanna ?
model?
@grand geyserhttps://civitai.com/models/9291/sunshinemixsunlightmix
This model has been republished and its ownership transferred to Civitai with the full permissions of the model creator. They have asked that all i...
@nova remnant
@steel arrowwow this lkooks so anime like
I just asked for a sword for the one loading now, amd she is holding the blade.
Oop, now the sword is just two hilts
The trick I find is to add "sunny day" in there
If you know how to get longer teeth then sure
might be
Will do
Prompt for the second image?
Looks amazing, can I ask which lora/model is giving that style?
last one from me tonight, bye bye
amazing

wow
This is my model that I am now teaching this style to
almost done with it
it has not yet been published
Because I'm still finishing it up.
Not sure if this can be considered good
looks nice
i forgot this one was cooking, now goodnight
no worries, it's looking great so far
i think i put this in the wrong place initially
Looks nice makima?
same as before but with different prompt style
"((masterpiece, best quality)), a painting of the interior of a coffee shop, complex detailed style manga volume cover by sui ishida, sui ishida complex tokyo ghoul manga cover art, horror, flat shading, complex colourful pallete, complex shading, colourful"
what was the model again?
WD1.5-beta based model.(Stable Diffusion V2.1-768) Negative Embedding & more info: https://huggingface.co/gsdf/Replicant-V2.0 Licence: https://...
oh yeah
@finite herald we don't have permission to write in the thread
So, I've got a weird question that I'm pretty sure I know the answer to but I'm still gonna ask
should work now
Is a woman wearing only an oversized shirt with cleavage (but nothing else) showing considered to NSFW for this server?
ooo epic
I like making these anims with ai art
An app called motion leap
Last one I promise made it a tiny bit more badaaa
Evening chat
Cuteeee!!!
cant wait to try your model/lora
That looks magical !
Hello- sorry was gone for a while
Ayyy, no worries, hope it wasn't anything bad 
Yee, glad to have you back!

OH also @shy totem new model by sis @quick hill dropped, its pretty dope if you wanna check it out 
https://civitai.com/models/40199/aurora
upscaleeee
time for u to go review 


OH, YOU RITE
Oh yeah haven't tried aurora
I forgor
Sure thing- I'll check it out ASAP
El gato
Looks... weird
The cat?
Ah no the pose of this one
Oh xD

I reviewed
@quick hill

https://www.reddit.com/r/StableDiffusion/comments/12pvhhm/animov01_highresolution_anime_finetune_of/

wtf
No Shutterstock logo!
O wao 
oo Cammelia is back on Civitai 😄
Lmao
I can't figure out what to call my model, if I should continue to call it Dark Gemini even if it has become more versatile
or make something new
Timer start until it's gone again
NEWW
Looks amazing! What model did you use for this?
new downside is that those following 2.1 won't be notified
Ye
KoreanDollLikenessV10 plus Lora model Ayaka.
It's made me wanna generate some Genshin characters now! Thanks
maybe use these as example images for my model, I think they do a good job showing off some of what it can do
I can't see it on there
what is it
Play
lol, i found this, way better than the original file browser https://github.com/zanllp/sd-webui-infinite-image-browsing/blob/main/README-en.md
I would need help to spread the model though cause not many people seem to see it, but maybe I just need 10 examples of hot girls
straight from SD for myself*
from SD
I don't touch up with photoshop but I do remove text or logos
with fill in with recognition or whatnot
it's much faster than prompting it away
Most of mine are straight from SD. Occasionally I do touchups through Krita and/or inpainting. Very rarely I'll use PS for color adjustments.
I think the time has come to release DG 3.0
hmmm whats the extension called that checks for scrap, clip deviations etc?
Be aware, fixing a broken CLIP model will almost always significantly change how your model responds to prompts. You'll need to redo all of your testing and preview generations if you find and fix a broken CLIP model. Ignoring a broken CLIP is also a bad idea as it can result in SD using the CLIP from a previously loaded model instead which makes it unpredictable. https://github.com/iiiytn1k/sd-webui-check-tensors
I know
sometimes broken clip can be beneficial though
for creating uniquely acquired prompts
Potentially, but again, it will make the model unreliable as SD may sometimes decide to just use the CLIP from a different model all together.
I will do whatever I can either way though and weigh the cons and pros
Dark_Gemini\DG3_Full_V1.safetensors New format hash: 021972d7ff Old format hash: 5b673d2a tensor([[ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76]]) Type: torch.int64 0: 0.00000 0.00000 1: 1.00000 0.00000 2: 2.00000 0.00000 3: 3.00000 0.00000 4: 4.00000 0.00000 5: 5.00000 0.00000 6: 6.00000 0.00000 7: 7.00000 0.00000 8: 8.00000 0.00000 9: 9.00000 0.00000 10: 10.00000 0.00000 11: 11.00000 0.00000 12: 12.00000 0.00000
does this mean it's fine?
Yep, you're in the clear.
nice
gonna make a pruned version too
`Model is 6.94 GB. Multiple model types identified: SD-v1, EMA-v1. Model type SD-v1 will be used. Model components are: VAE-v1-SD, UNET-v1-SD, CLIP-v1-SD.
Contains 1.83 GB of junk data! Wastes 1.42 GB on precision.
Model will be pruned to 3.69 GB.`
What does it mean "wastes on precision" ?
how can you waste precision?
This; Very important.
Ignoring a broken CLIP is also a bad idea as it can result in SD using the CLIP from a previously loaded model instead which makes it unpredictable.
I assume that is telling you that you have a full precision (float32 model). The additional precision of fp32 is unnecessary for inference because most WebUIs load models as half precision (float 16) anyways.
My experience with Aurora was hell until I fixed it, whole problem with consistency, versatility and weird brightness are all gone.
Not sure if I should include clipfix in my figma >.>
Hello JRW 
hi!!!
Yep, nothing worse than having the exact same configuration give different results simply because you changed models and then changed back. I think broken CLIP was the reason for people believing that WebUI was caching previously loaded models and needing a full restart in the past.
Aaaayyy! It's JRW!
Possibly, most models are broke.. and I believe no one understand much about this CLIP and Text Encoder >.>
hi nubby!
I have to do my own research for whole week and experiment, I don't like it at all lol
`Model is 6.94 GB. Multiple model types identified: SD-v1, EMA-v1. Model type SD-v1 will be used. Model components are: VAE-v1-SD, CLIP-v1-SD, UNET-v1-SD.
Contains 1.83 GB of junk data! Wastes 1.42 GB on precision.
Model will be pruned to 5.11 GB.`
Wtf why is it suddenly different?
wtf cute
@supple raptor so you suggesting I use fp16 on model and on pruned and no fp32 at all
oh now I see
fp32 is 5.11, fp16 os 3.69
trying to do something like one of those psychedelic blacklight posters but with anime girl lol
I haven't pruned in a while and don't have the extension installed. @quick hill would probably be better to ask for the specific settings you should select.
I'm using toolkit for pruning
and for converting to FP16 on the normal
maybe I should release FP32 and FP16 on both versions
then people can decide for themselves, but I'll tell them which one the images were made on
and that it may differ between versions
Yes, having both is a good idea.
Does adding BREAK in all caps in the prompt actually do anything
Anyone wanna make some cute/alluring kitsune girls with big fluffy tails for me 👀
on my way
Yes, i specifically use it for my lyricsprompting
Seems very minimal in detail
What does it help with
MUAHAHA MY PROMPT BUILDING MACHINE WORKS!
TLDR, it fills the current block of 75 tokens with blank tokens so that the following word will be the beginning of a new block. In essence it's another way to control the relative strength of different parts of the prompt.

Normally you submit your prompt in sets of 75 tokens, with each tokenset being weighted average (so if you have 125 tokens, that's 75+50 tokens)
how 😮
but at some points, you want to break the first token set early, because some stuff needs to be together, for example, girl, in armor, (just example)
So basically after u put break ur just starting in a new prompt box/space?
so, you add BREAK so it breaks the first tokenset of 75, and as @supple raptor explained, it pads it with _

so girl BREAK wings would consume 76 tokens -> 75 for girl, a lot of padding, and then wings in the next set of 75
which. would be a bad idea in that case
but that's how it works
uuuh pruned and regular FP16 turned out the same size
7b vicuna + homemade LLM prompt
I…think i get it
@final river here you go ^^
Thank you! Do you want me to make you anything in exchange 
@compact warren here's an example of how i use it with lyrics prompting (requires dynamic prompts)
Truth be told It's a lonely road Chasing your dreams And all that we are And could ever be Is within our reach (Woah) Ignite the fire (Woah) Pure hearts can never expire
| May the bridges that we burn Light the way From the ash we rise Like phoenix taking flights Blaze an immortal flame
| Shine like the sun Piercing the sky Pure hearts can never die Rise from the fire Begin again
| From the ash we rise Like phoenix taking flights Blaze an immortal flame, oh Woah-ah-ah-oh Waited all our lives Just to burn this bright Blaze an immortal flame
| Woah-ah-oh Woah-ah-oh Woah-ah-oh Blaze an immortal flame Blaze an immortal flame
}```
basically, 2-3 verses get selected at random, but each verse will be made to weigh as much as the others this way
Sort of, yes. The only reason that we can use more than 75 tokens on most models is because WebUI already breaks your prompt apart into 75 token blocks which is all that SD can handle. It then averages the weights from each of those blocks of 75 tokens together to create a single 75 token prompt so that it can be processed by SD. BREAK allows you to control specifically in the prompt where the prompt is separated.
maybe at a later date but ty ❤️
this prevents verses from breaking in strange places, and each verse begins at the fresh start of 75 tokens
So when SD processes the tokens in a prompt, by itself, it usually breaks it down randomly, but using Break allows control over where it does so?
Exactly
Aha
the use cases are limited for i suppose, but it can really help
i had a prompt that had a break on the word blue
that just makes the image blue. period. blue everywhere
Any idea how pruned FP16 and regular FP16 turn out the same file size?
And it's important because tokens inside of a break are more closely associated with each other and prompts towards the beginning of a break are intrinsically higher weight and reduce as they get closer to the 75th token.
until you break 75th
As Eface said, the use cases where it is really important are limited, but it can be a very useful tool at times.
Was gonna say, it sounds like it helps with keeping important tokens together, and if so i really need that
yeah, if you have some words and they get split up between tokensets they just lose any and all coherence
careful, don't badmouth catgirls
there is no
about it
but if we are talking about catgirls I can mod abuse
Fox girl>Cat girl
BANNED

TanukiGirl
Hmm... fluffy tail is appealing.
numeric cat girls
TA-NU-KI
I need more catgirls in my dataset
there isn't enough catgirl in the PoWs
uhh... my LLM associates catgirl with horny.
Hmm... guess I need to do catgirl world for POW this week 
I'll open a "best off SD art" collection thread soon by the way. I'm looking to complete the "megaPoW" dataset with more subjects I don't have enough of.
Miku, we have, but not catgirls
she has all the clothes once SD is done generating though
Have you combed #1055215351731851354 for Mikus? There has to be hundreds in there and mostly good.







