#๐ฅ๏ฝanime
1 messages ยท Page 209 of 1
lol sure
ha no one had it
You guys are going insane
it can't even caption images containing people 
i often add "adult mature" to offset the anime super-young look ๐
i will remember to try that out ๐
"close enough" (upper right corner! ha!)
mature female and mature male are the actual tags if you want to use boorunese
yeah, depends on what the model was trained on
had no clue till now, definitely something i'm gonna use a lot with xl models
not bad at first try,
anime, mature female, out in the garden. tshirt, shorts, bare legs, sneakers.
at least she doesn't look fresh out of womb
On pony "child" in the negative works surprisingly well
shrugs not feeling gW2
sigh I don't know what I want to play. just in a blah mood. oh well
but shrugs, watever. i'll figure it out or not sooner or later. shrugs
is that a new model? that looks pretty neat
lighting is good
oh it's a lightning model... low cfg...
so, does that make adherence to prompt really poopy?
returns to creating art, it's relaxing
well pony is already so... biased towards more "interesting" stuff that the age alone doesn't make too much of a difference ๐
same with realvisXL
just have rating safe and it should solve most of those issues
Idk about that on pony merges
what happened to the contrast?
was using some retro models vae and such
oh. was wondering cause the image above came out with sharp tone
was trying different things
uhh oh
stop it making me hungry
No grabby
hahaha

What SDXL model produces the most detailed anime stuff you think?
probably pony or autismmix
And the controlnet stuff works with it or nah?
some do work other ones are buggy like tile
Nice!
BluePencil XL 2.0 works for me so far with img2img, but a problem it has is making the eyes look towards the viewer. Oh well ๐
i like how you make your profile pic fit the avatar decoration
had a gen a new pic just for that ๐


Worst discord idea ever
Those cringe decorations
stooopid

at least we have em
someone's poor ๐
I'll go back to old discord so I don't see them

You click a profile and you get flashbanged for one second until you can do anything kek
tbh, I think they're mostly useless and way overpriced, but i couldn't resist the cat ears
would be better if they were just included with nitro, or at least significantly cheaper
It's only for the rich
They prey on us
not because we are rich but because this is how we spend our money, lmao
i'd be rich if i didnt buy stupid stuff like animated cat ears
right, lmao
If rich > spend money > not rich
Big brain

yes,everyone knows u get richer by wasting money ๐ฟ
Steven is the richest
also i haven't had a job for about a week now, getting less rich by the day
you'll find out when you grow up
I wonder when that will happen, im 33
๐คทโโ๏ธ i'm a bit older
pretty much been working full time since i was ~15 because i'm poor
job is a form of enslavement, and the only justification for such 9-5 labor is society doesn't yet have the technology to replace human labor but that will change with the development of AI and smart machines, with better safety protocols and precision, and that will usher in the golden age. not joking btw ๐
only question is when, not if
i spent the last couple of hours tinkering with model merges, sadly nothing was worth keeping

I haven't worked a 'normal' 9-5 job in years, most recent one was 2am-6am and only about ~2 hours of that was actual working, the rest was driving or waiting in the truck for my load times
fits my natural sleep schedule a lot better than a typical job
the cute girl got bored sitting around home, so she decided to take a bus ride to the beach, and was excited to be there, later she stopped by a cafe before calling it a night
Nah I think it just means more money for the rich and the poor get poorer
if you say so but the rich pushed the boundary way too much where the ecosystem is collapsing now
People are way poorer in other countries. We have a lot more to fall
this is a famdily frienmily discord server 
that's a very acute observation btw
imagine the toll on the outter gradience
I remember a picture of a lady with a 4 sale sign selling her kids during the great depression on her doorstep. We always say never again, but it always happens again. It will probably happen in our lifetimes
lets hope not, and not if society understands how corrosive the current economic systems are
society barely understands not to click on links offering gift cards, there's no hope for understanding economics, even on a small scale
you know why that is? everyone is caught up in the frenzy to survive in this economy, no time to sit and think
eventually the bubble is gonna pop and explode
btw i love this animepencil xl from blue pencil
I think people have a general understanding that the system is such that the rich can take advantage of the not rich. But to do something about it is like the story of the mice putting the bell on the cat- they know they must but nobody is willing to do it. It can't be reformed; revolution is needed. But nobody will fight for it
yeah that was the general idea, but that didnt work out well at least not for a long term solution when the same economic model was kept in place, today its not just rich vs poor, its entire civilization vs nature
I mean, the 2009 financial crisis was basically big banks making a massive gambler's bet that failed, and then the government paid their debts with taxpayer money. Not to mention the wars that take other countries resources using American people's debt to pay for it, then we buy stolen resources at the gas station. It's open and brazen theft on a spectacular scale.
I'll have to try anima pencil in koi2img soon
Guess I'd better sleep now ๐ด
near about every 10years or so that cycle continues where the world faces economic depression, its part of the economic model, increasing debt ceiling won't help, its all gonna come crashing down, there is like 30+ trillion dollar debt now
just copy/pasting something @nova remnant said like 3 hours ago
not sure if it's just some hacked account spamming random stuff, or if they think it's some kind of bot prompt
@lyric zealot come look at this and
with me
novel ai
@junior quest you anime more than me


On closer inspection the first one has stupid thing on her hair
I'm pretty sure it's an anime character
They all have something stupid like that

eh?
isnt dat nier automata
Yesh
yesh
Yoko Taro do be doing a lot of crossovers... https://nier.fandom.com/wiki/Crossovers
oh I was wondering too where it came from
2 people who don't know NieR: Automata?
thought I missed an important ending for a second
Oh good. Just the cat needs educated.
Only Naraka is the true game

I even went to the orchestral concert last month :p. No need to educate me
I'm so damn jealous!
The useless goddess
How dare you
You should be in #๐ค๏ฝtech-support helping people
I know you x3
Nyo. Go play Automata
Stoopid cat
it will only costs you 30 hours and a part of your humanity


Super huge
Wtf is that on background
nubby...khoya ss doesnt get caption files even tho I added them correctly. know anthing bout it?
WARNING No caption file found for 195 images. Training will continue without train_util.py:1459 captions for these images. If class token exists, it will be used.
@supple raptor
I can't tell if a gen is 1.5 or xl now lol
Both are pretty good rn lol
||SD3?
||
super huge btw 
sd3 released and then it's back to square one with anime finetunes lol

true
but i have faith
How long did it take for pony v6 to come out btw

6 months ish?
no idea, but the author was saying V7 will probably be on SD3
Oh wow
heard from @nova remnant btw
@zinc flax grabby
I have some grabby stuff that's definitely not safe for this server lol


IIRC there's an argument you have to use to specify the caption file type. Pretty much everyone uses Derrian`s trainer now which makes that simpler.
Im surprise you are not in the ||super secret server||

Those are some long fingers
I wonder if..
They can reach that place well
The Pringles can, the last ones are pretty hard to get

maybe she don't need to reach in horror
cursed and horror
loong loong neck
Created TensorFlow Lite XNNPACK delegate for CPU
behold
what... what's going on?
lol i often wonder why don't aryetis and cs10 interact here more
#๐ค๏ฝtech-support takes most of my "discord dedicated" time. I still have to somehow work a full time job, workout, go out, etc.... What a normie I am.
totally understandable. can't blame you ๐

I ll probably disappear back into tech support realm whenever sd3 get released
Oh you'll have lots to answer haha
i might need some answers too, any words on the approx release date,?
can you believe this is done on dalle3
i havent been able to get that much symmetrical perfection with sdxl
can you believe this is done on dalle3
yes i do, cause the one who created is popular in the txt2img community and he's done several other amazing works, including animation
can u believe a human bean drew this
you know, the way you send across a message w/o aggression is admirable cause i get what you are implying
steven 
he is an under cover agent maintaining a balance
Hello. Some can tell me wich can of discord i can use to generate some free picture ? With sdxl or any other model please ?
SD bots are temporarily down, might return in a few weeks or when SD3 nears launch
Oh ok thanks a lot for your answer . So wait and see . I hope they will come back i was so impatient when i see the video !
np. in the meantime you could also play around with either for free image gens 1) https://copilot.microsoft.com/ 2) https://gemini.google.com/app
Ok thanks again ! I will try . Have a good day and see you soon on SD ๐
good luck!
my apologies
I missed free money? Figures
free money? sounds like trouble.
Grabby ๐
dalle3 example
when you roll 1 for perception check
another interesting stuff https://ella-diffusion.github.io/
ELLA: Equip Diffusion Models with LLM for Enhanced Semantic Alignment
ELLA
My cute dragon girl can't be grabby 
I have others for that 

I left my computer afk generating images.
why not, horns make for good handles ๐
soushiki?
wat
What model are you using?
autism mix but if you want to get the style you need the granblue fantasy lora
Model?
1.5 or xl?
2.1
don't have that 

NAI isn't a style
ok, but it's not
it's a particular artist
pony can do that
but also they said it's animagine
that one looks like animagine because of the big ass light on the hair
pretty sure this conversation has happened multiple times in the last few days, pony (and other models) can have more than one 'style'
pony/autism specifically knows thousands of artists

everyone walks around with a light beam on their heads

just use the artist tags ppl use on nai to get the same images 
whos the artist?
quasarcake?
imprrressive

running the same prompt/seed on the entire pony artist list eface made
interesting seeing all the differences
wdym?
prompt is in the image,
Post some gens 
Im traning a lora
22 try

23
both would be my guess ๐
yup
the whole point is figuring it out though
I dunno all loras coming so flat
1st one is my 4th version, after that I did smthing to setting now I cant figure out what did I change
finally, 654 images done
now all are coming so baddd
what's wrong with the 4th version?
its not flexible
What's wrong with the 4th
Oh

I'm special
deym after 4th ver all loras are cursed
Is the data set flexible
Does it have multiple poses and stuff
same dataset with some regular images
kinda not much
eh?
With good images you can use very few images to train a lora well
Generate images that look good to use as training later cx
ouh
That should help make it more flexible
some interesting styles

ugly
ofc I agree
looks overbaked. lower LRs and/or fewer training epochs/repeats.
LR- 0.0001
epoch - 1
repeats - 20
here
Ugly

help me carry these fruits
i don't think she really needs help, she has a whole dumptruck
are there ais that can gen in a kaneko art style
Did you try searching for a lora of that
Artist?
Whatever it is 
It's more likely to exist a lora that does the style than a model
want some honey-dipped gold coins?
thicker?
randomizing expressions at the same time as randomly mixing artists and sometimes the random expressions have me rolling
White and grabby
hgh
What style is that?
i will give you the pnginfo
1girl, white hair, sidelocks, forehead, forehead, red eyes, big eyes, stud earrings, ring, dragon horns, lingerie,
source_anime, mizushina minato, (bee \(deadflow\):0.7),
Negative prompt: (worst quality, low quality:1.2), 3d, censored, abdominal muscle, artistic glow
Steps: 24, Sampler: Euler a, CFG scale: 7, Seed: 62954268, Size: 768x1024, Model hash: ac006fdd7e, Model: autismmixSDXL_autismmixConfetti, VAE hash: 66f78ad136, VAE: sdxl_vae-fp16fix-blessed.safetensors, Denoising strength: 0.4, Clip skip: 2, Hires upscale: 1.5, Hires steps: 10, Hires upscaler: 4x_fatal_Anime_500000_G, Version: f0.0.17v1.8.0rc-latest-269-gef35383b```
First dozen or so invites gone out testing, will gradually scale it up as we iron out kinks.
Imagine we will see a lot more #SD3 in the TL soon.
Also working on the API side for preview so can pipe it into ComfyUI as we build the various add ons etc.
Thanks all for patience.
Any XL model recommendations for a sort of 2.5D half-anime half-realistic look? I really like Breakdomain Realistic's style but would rather work with SDXL stuff if possible
There's a break domain xl
I'd try autism anyway
Get a lora for the style you want
1`
Depends on how you count, ~3 month to train but data collection, experiment and other tech took much longer (but also split between other versions)
I hope so, but I have no idea right nowโฆ
having to go through artists for pony models kinda kills huge time and patience
Donโt use artists I guess?
well im trying to find a style that i can use, default is crap
I like it (thatโs why itโs default)
score_9
i use that but im talking about artist name
do you use something like ...
score_9, score_8_up, score_7_up, score_6_up, rating_safe, source_anime,
lol
im trying to generate female characters
sigh.. pony/autism mix aren't cutting it for me
If there is one thing pony is good that would be female characters
not my queen ๐ฟ
(latex:3.5)


Astra is the best
Always making sure the pony bot is working x3
And everyone can enjoy some pony
shes better
Wait
Is that the Yu-Gi-Oh girl

Anime character 
would
have a morning coffee...
mmm, nothing like a nice warm cat in the morning
Well now I'm obligated to
๐
fail?
of course - "anime artwork a friendly barrista winking at you . anime style, key visual, vibrant, studio anime, highly detailed" with realvisXL
"a nice warm cat in the morning"
Didn't really work on pony but here's something similar 
Key visual kept giving her a key necklace cx
@native halo 

The images you generate are very creative
That mdastaro style really makes it lewd lol
I feel much better now
very cute
i'm dm-ing you
๐

It should've been me


I love popsicles
Very sexy
i feel sorry for cute girls, they cant even eat ice cream w/o getting censored
i dont support smoking btw, but she is cute
cause smoking is bad of health and lung
yes absolutely, not a good habit
ashes are bad
Yoo what is that artstyle? It looks nice
all the info is in the image, but heres the prompt:
score_9, score_6_up, score_5_up, score_4_up, source_anime,
1girl, petite, fox ears, frustrated, from side, looking at viewer, dress, goth, V,
extremely long green hair, twintails, [wild hair|messy hair],`
multicolor eyes,```
it's mainly the 2 artist tags at the begining making that style on autismmix
thank you, what is the [ | ] for?
u use it when merging 2 things together , for example [ girl | snake | tiger ] it would give u a mix of a tiger snake girl , that`s what i know
it alternates the prompt every step, first one on step 1, second on step 2, first on step 3, etc
for artists/styles it helps blend them a bit
oh thats cool
some artists/styles will overpower the others enough that using artist1, artist2 doesn't blend them as well as [artist1|artist2] does, but it all depends
Any of you know if there's a way to do a certain prompt on 20% to 30% steps and remove it after
on comfy is easy
model loader, conditioner, kesampler set with only 30% of the steps, 2 more conditioners without the tokens, a ksampler with the noise from the precedent ksampler and that starts at 31%
for example, you have 30 steps, you set the first ksampler with a total of 30 steps and starts from 0 to 10 and the second from 10 to 30 steps
easy peesy
charming
i'm going to sleep, so now i don't have time to load comfy
see ya

Syntax is [from:to:when]
[word1:word2:20]will change "word1" to "word2" at step 20[word::20]will remove "word" from the prompt at step 20[word:20]will add "word" to the prompt at step 20
You can use a float value instead of an integer to specify a percentage of the way through generation to do the thing like:
4. [word1:word2:0.5] which will change "word1" to "word2" halfway through.
The step count does NOT reset for hires fix so if you have 20 regular steps and 20 hires fix then all of the above examples would have the prompt editing occur at the end of the low res generation.
It's far easier on A1111/Forge.
I want to add a word to the prompt and remove it midway through the generation
It doesn't look like it

Let's say we have 10 steps, I want word1 to appear on steps 3, 4 and 5 only
just nest them [[word1:3]::5]
though, if those numbers are what you're actually planning to do and not just what came to mind at the time, it won't have much effect on the image unless you're using super low step counts like with LCM/Turbo/Lightning/whatever.
What black magic is that
Let me try cx
dont 
it's simple [word1:3] adds it at step 3, so we just use that as the "from" portion of the other operation [from::5] which removes it at step 5
morning
basically using tiled vae in SD-forge, when upscaling img2img at 1.25 it has some weird resolution object type non error
once its over a certain size it lets me use 1.25, but before that only 1.5 works
might be VAE error
think switching vae would help?
what vae are you using?
depends on the cup size, pun intended
coffee with lots of milk
Yeah, there's a bunch of similar/same prompts, various art styles, and I haven't gotten rid of the bad ones though so you'll have to go through them. I can zip them up in the morning and upload them.
arigatoo
how much vram do i need if i want to train an style lora for sdxl?
16+GB I guess, (you can train with 12gb but probably it'll take like 5hours or so)
(you can even do it on less but it takes more.... knowledge and dedication)
I got 12gb card, it takes 1hours to train a 512*512 lora
1 epoch
5 batch size
20,000 steps or smthing like that (200images)
that means you training at 100 steps per image ?
yups I tried diffrent steps
that feels like too way too much but I don t have my training master degree yet
but I guess 20 steps is better for me
probably yes, that way you can even train for 5 epochs and get more intermediary lora to check for under/over train
yups, after all Im just a guy who just interest in these stuff so I dun know bout best settings or optimized settings
Im doing these for ENTERTAINMENT

np :p
can t help you much on that myself. Like I said, I don t do lora training on my mere 8gb gpu. I just started looking at it properly a few days ago.
I just trained a (checkpoint) model with 340 images (512x512, 40steps, batch size 2, epoch 1). got 1.3hours
When I use booru tags for models that support them, should I write them with the underscore (long_hair), without (long hair) or does it not make a difference?
one is a tag and the other is not.
yes use the same syntax
use https://github.com/DominikDoom/a1111-sd-webui-tagcomplete for booru tag completion
There's this also
https://huggingface.co/nubby/blessed-sdxl-vae-fp16-fix/tree/main
And for 1.5 models
https://huggingface.co/nubby/kl-f8-anime2-blessed/tree/main
,
.
!
'
oh
#๐ฅ๏ฝanime man, yellow short hair, white short-sleeved shirt, Miyazaki style, pondering on a big bed, surreal atmosphere, shopping mall setting, girl, pink long hair, blue eyes, earrings, necklace, black collar,
guys , what`s the recommended dionising strength for chromaNeoFT checkpoint ?
if at 0.7 which is the regular , it changes too much , if it`s less it becomes undetailed
in general you want around 0.4-0.5 if using an esrgan upscaler or 0.55-0.65 for latent upscaling.
oh , yeah i use eresgan anime6b
thx

i'd say train on whatever model you like to use
I used 1.5 based model. it makes so many errors, like hand and legs
what if I use smthing like anything or sm other model? will it decrease those errors?
if that model has those errors less to begin with, yeah
but also, only on that model
you're not going to make a lora on a good model and then have the same results on the base model
You most likely want to use anime_full_final_pruned.ckpt (nai leak). That will allow it to be compatible with most anime models. If you train on a specific anime finetune then it will most likely not work as well on other anime models.
don't hesitate to increase the steps higher than usual when working on img2img. The real amount of steps done when upscaling with img2img is : denoiser strength*steps+1
that can fix your """undetailed""" issue
depending of what settings you had
dunno how to use img2img.... but thanks , will try ๐ ๐
nice
thanks ! it worked , got the idea


meet miku, miku, miku and ...
Meet @dull jackal
@dull jackal @finite herald
they keep coming
I wrote hatsuni miku in my prompt

fixed
๐ offering for the mods
SO LONG, FAIR WELL, I HATE TO SAY AUDIEU
Are you trying to tell me... the only reason i dont have a goth girlfriend...... is because i never visit bread shops??? 

fingers kinda fucked but...
naaa, we have had it much worse haha
the 1.5 hands... 
lovely.
too much simp for gothies 
@cerulean bane I tried to generate a cute goth dragon girl but it ended up a bit nsfw so I censored with inpaint so I could send here
๐ !
Shes just sitting at the park 
approved.
very. very. approved.
coot
Does 50 steps break autism ?

It went nuts lmao
started doing random images when I put 50 steps
seems ok to me, this is 50 steps
I'll dm you what it did
bored based on the prompt
that expression W
naurrr
now im mad
I want these 2 mads on top of me

#bot-x
He's ugly 
Like Steven
burn that evil fox 

she doesnt have shiny skin
/ban
she will take that personally ๐
like that ๐
Save the fox from the rain

not anime 
I like her






