#🍥|anime
1 messages · Page 245 of 1
Nai
oh it makes sense lol
imagine asking "what model is this?"
of course it would always be nai 😭
Yup 😭
why "always"?
Because nai sweeps
Real
don't even know.
It's hard to explain but NAI just has such a good base style, and the fact that it's able to mix artists together puts it on a whole different level in terms of aesthetics
and if we abstract from the binding of artists.
If we don't use any artist tags it would look like any other pony merge lol
Unironically
Even then I find it to be better than Pony 😭, cause it can actually do anime
so... in the end it just turns out to be a Pony?
It's the best model atm
Pretty much
It depends, there are things Pony does better but for actual anime NAI is much better
I'll take your word for it. I'm very far from the standards of real Anime. But I like the style of manga and anime in general. and that's enough for me. Whether it will be done in NAI Sd1.5, SD3 SDXL does not matter to me in the slightest. I find the aesthetics of illustration in general.
Of course, at the end of the day we're very lucky to have both NAI and Pony. It doesn't matter what people use as long as they're having fun
Personally I'm more excited to see what the future holds for AI image gen. I think both Astralite and the NAI team know what they need to improve with their models and hopefully they will deliver
As long as they don't hash anymore stuff lol
they need to process the results
It’s not for nothing that people paid to use NAI
Nai is stupid

Really the good thing is vibe transfer

Yeah... About that...
It's so over
hash all the artists
Bring vibe transfer to pony
I always use vibe transfer plus like 10 different artist tags
Pretty sure vibe transfer is just a better IP Addapter
Or whatever its called
I never used controlnet stuff

But seriously
Nai prompting is extremely dumb
It really doesn't do much more than tags

not surprising if this is a set of artist names
How so?
You prompt with it the same way you would with Pony, at least that's how I've been doing it
I didn’t understand at all how to work with the Pony model
NAI emphasizes on the prompt order more though
Yup
@brittle rampart 
Pony kinda dumb too, but it could do more of my dumb ideas 
Because of the score tags the model doesn't work as well as it should. But you basically just use danbooru/e621 tags to prompt with it
e621 
Makes sense, it was trained on e621 tags which are better than danbooru, that's why Pony might know some concepts better than NAI
I think this is how any model works or am I wrong?
Well yeah but some people say you can also use natural language to prompt with it however I don't think it works that well
The score tags are the main issue, and the small anime dataset
Eating earrings
It’s because of them that I can’t understand “how it works”???!
strawberries again!?
It would be much easier if you explained what you don't understand about it 😭
Eat all the things
o woops
Didn't know why I replied
Booba 
Which Score_N tags should I write and why?
tldr:
score_n tags were supposed to replace quality tags but failed to do its job
Whose milk idk
That's the issue, because the score tags are messed up people don't really know what to use. However, from what I've tested, using just score 9 is better for anime, adding score 8 score 7 and those other ones will give your images more detail (it will make the characters more 2.5D/3D looking)
Again, it's not a huge difference but it helps
And the fact that Pony's anime dataset is small compared to the other images in there is also a limiting factor. The model always makes the images look more cartoony rather than giving the anime images that anime look
Hopefully the bigger dataset will fix this issue but we'll have to wait and see
Will the new model still be paywalled though
It was never paywalled?
Oh huh, the v6.9 one?
So, when using the Pony model, I can simply write the Score_9 tag at the beginning of the prompt and then compose the prompt according to the usual rules?
Yup, that's what I've been using for a while and it works quite well
Pretty sure Astralite already said it's going to be open source as always, it's just that they have a discord bot running for those who can't run Pony locally, that's maybe why you thought it would be paywalled
They claim it's not that, but wont give more info.
GOOOOOD evening everybody 
It's gonna be interesting to see what they're capable of at least
Hey lex
Ima do my usual scroll and admire the gens people make
wasn't meI swear I didn't eat the cookies
Yeah, they don't wanna reveal anything, which makes sense I guess
you are guilty! you lusted after a little Cookie!
yep, wd team is working on a model that actually looks very promising based on the pre-alpha version i have (which is more than i could really say for any of the previous models they've put out). the next version of animagine sounds promising, but i'll reserve my excitement until it's actually out because i can't stand using any of the previous versions. there's also another architecture that someone is working on that seems quite interesting.
animagine is super undertrained aaaaaa
yep, if they explained how it worked, we'd probably have a reverse engineered open source version within a week.
I don't think wd has ever made a good model
dataset is the biggest problem with it. since animagine 2.0 each new version has just been a finetune of the last and 2.0 had a relatively small dataset. usually foundational training would be the biggest dataset and you would reduce the dataset for finetunes, but they've been growing their finetune datasets instead. the result is that some concepts are extremely overcooked while others are under trained.
1.3 probably would have turned out pretty good if NAI hadn't leaked during their last bit of training which made everyone just ignore it.
Nubby go train a good model
you know all stuff
The thing is, they have a whole H100 cluster this time, meaning that they can run hundreds of test runs before actually making a good model. Being able to actually experiment and see what works best is always a problem for finetuners cause they just don't have the compute to be able to do that. This time it's different
oh good
the braindead pre-alpha version of wd-v that i have can do broom riding more consistently than NAI can.
I've also heard that the person who made that model is against having NSFW images in their dataset. That could also be why it's not that good but I have no clue honestly
i just have a 3080 12gb. i have to make compromises even to train a lora for sdxl
Are you allowed to post sample images here? I would love to see some 
biggest problem is that it started with a dataset that was basically just 5 games worth of gacha characters and instead of starting over, with the bigger dataset they have now and the (probably) better hyperparameters, they just keep continuing training on from the previous version.
They got big booba waifus so I'm not complaining 
If you do multiple img2img passes without increasing the resolution of the image, it comes out garbled and ruined. Why is that?
I'll never understand why all the local finetuners limit themselves like that, and it's always the people who actually have the ability and the compute to finetune
keep in mind, this is from an early test run, not an early version of the model intended to be released.
1 more from WD-V pre-pre-alpha
second one is a cute style
yeah, that style is probably my favorite that i've gotten from it so far. i haven't even really used it very much though. most of my gens were just nai prompts that i dropped into it.
yo that first pic is clean af
4 5 6 and 10 already look like something you could get out of NAI, less detailed but still really good
has some potential
And like I said, now that they have a whole cluster they can run as many test runs as they can until they figure out what works the best
that one had no artstyle tags either. someone just asked if it could make mahiro so i threw her description tags in.
there's no way it didn't just memorize a training image for that one
this is most likely using "vibe transfer"
ah yeah, 4lex mentioned he used it lots
but composition wise it feels like an official anime art
It was vibe transfer using some dark souls or elden ring image, I only use it to change the art style
"official art" is an actual prompt on NAI, which is what I used, along with "anime screencap"
after all, the screenshot was the basis
What screenshot?
That just gives it the style from the anime as far as I know, I use it to inpaint the eyes
I'm just thinking that you won't get that composition with any other anime
Doesn't have anything to do with the actual image composition
big bro standing behind lil bro style
like if you were to swap that out with megumin and aqua
or shinji and eva-01
I feel like NAI learned a fanart or official art
It was a pretty lucky gen, I generated like 3 or 4 of those but I was to lazy to inpaint them
Yes, that is what AI models do 😭, they learn from actual images
There is plenty of images of edward and alphonse on danbooru, you just have to use the tags "2boys, brothers, siblings" and you'll get some good results
I think you could, you just gotta roll the dice a lot 😭
And a lot of Inpainting, that's very fun to use
Now make her eat the microphone
(saliva:1.4)
Hands look really good for alot of these pictures how do you guys get them so consistently?
Depends on the model, most of the SDXL models can do pretty good hands, and if not then it's pretty easy to inpaint them
Shall we force her to eat the microphone too?
SDXL is mostly good, if it's not good I re-roll
NO! 😭
lmao
weird dog
because it's Chanterelle!
Chanterelle!?! damn, what is this!?
🤣 Google Translate!
This is top tier. I would've killed for a wallpaper like this back in the day. This is going in my hall of fame folder, well done.
toooo looong!

I approve
scary to open..... 🫣
she'll bite you be careful
I suggest you sign a contract! I'm using a pen! you are blood! 😉
Thank you 
she'll bite my knee if I don't ?
woooow
I just tried what I recommended @ivory pebble
what`s the prompt ?
``` (in style of anime Fullmetal Alchemist:1.6), 2boys, brothers elric, edward elric, alphonse elric, best quality, amazing quality, very aesthetic, absurdres, dynamic illustration, (official art, anime screencap:1.2)`` `
can`t get it right 
This was my 8th or 9th attempt.
no more cookies left, I ate all 
big green slime
why did she tie a rope to her leg!? or did you tie it to her?
@nova remnant did
hey hey! I did not eat!
if the first jar contains milk.. then the second jar is very similar to one of the emojis!
😁
Ouch
Sometimes it'll lessen and only take like
10 minutes
But the images also come out looking DOG
@nova remnant I read the correspondence. and wanted to ask. I understand that Forge failed to launch?
what video card?
Video card???
I think it was an internet issu??
I went to Walmart for wifi
But my computer was fucking up
your video card?
you can look in the task manager
this??
Yes. how much video memory does this card have?
i heard forge was less intensive on computers so ive been trying to download it but the install keeps fucking up
6Gb
yes yes
From my own experience I can say that Forge works much faster
6gb is trash
ik
but yeah i heard forge is faster
and easier on computers with even 6gb of ram
i wish it didnt need internet to install
its crazy that it does
it should all just be there with the initial download 😭
well not true. I have NVIDIA GeForce GTX 1660 SUPER 6GB
i get mixed reviews on the 6gb thing
and even now I do generation using SDXL models
some people say its enough to run even sdxl
some people tell me i should have 16 gb just to run them
some people say lmao thats trash, others say thats god enough to run

on forge??
i need forge
a1111 is the same and all but its not, optimized ig??
thats the least amount of time it takes for me to gen on a1111
but sometimes
even without switching the model
1 to 3 hours
its a literal gamble
and then the images look shit
might just be my prompts but, idk, i think even simple prompts should result in good images like that
Well that prompt is quite simple
[ningen_mame], [ciloranko], [sho_(sho_lwlw)], [[tianliang_duohe_fangdongye, as109]], happy, blush, girl, white hair, long hair, blunt bangs, long sidelocks, blue eyes, pointy ears, {{{vampire fangs}}}
Well NAI is just too good
how bad they look sometimes
ive heard of nai
i have some credits on there
ill gen a few while waiting on a1111
but NAI just doesnt have what im looking for tbh
its gotten really good but its lacking tools i like and extensions plus its not free 😔
💀
they look like this 8 times outta 10
sometimes its like this though which is fine
but then its like this again
At least it's better than this 
its a mixed bag
NAH THAT LOOKS BETTER THAN MINE
can I see your settings in A1111?

which ones
those that you use when generating
Tbh this is old Nai leaked model
this is my base settings
sometimes ill turn on hires but
it takes so looonnggg
is this a Pony model?
is this SDXL?
score_9,score_8_up, 1girl, (long hair, white hair, blunt bangs, long sidelocks, blue eyes:1.2), pointy ears, long eyelashes, choker
Negative prompt: score_4,score_3,score_2,score_1,ugly,bad hand,bad feet, censored, abdominal muscle, artistic glow
LMAO
gotta ask whats withthe score stuff?
It's quality tags for Pony based models
ahh that explains it. I dont use pony
well , at least not the pony model 😄
i just know to put them there
animate that yourself??
Luma
never heard of it
i animate things by hand 
with after effects
when i do it if im animating just like
an image i found
i have to cut it all up where itll move
and then draw underneath so theres no gaps
move it all to ae
then animate
its... actually easy??
its just two things tbh, but you just have to know how to use them
ooooo thats pretty
trying this fr
gah the faces
This is feminine "charm"😁

Yui is beyond competition
Model
it looks cute whats it call
NaiV3
i need a cute model 
link 
i cant remember how to prompt correctly on it though
my images end up a little weird sometimes
is not it so?!
ok, let's try to launch Forge and generate something in T-ponynai3
I hope I will soon
is it bad?
but this is generation without HiRes.fix?
ye
then what is the quality?!
IVE SEEN 1.5 MODELS MAKE LESS BLURRY EYES 😭
better ?
looks nice
and yet tPonynai3 is an SDXL model or not?!
dawg i told you it was a pony model 💀

pony model - what is it?!
here
It’s a SDXL pony based model
still SDXL!
Unless you are talking about the SD1.5 version of pony tho
thats a thing?

Yup, but it’s not that good from what I heard
i mean, i figured
1.5 isnt the best
i love it
but thats because im used to it
There’s some good 1.5 models tho
My old time fav was holokukiv
well it's not bad
and the time is only 2 minutes
i used like 30 different 1.5 models at any time
i used them all
igotta wait 30 minutes rn
Then maybe you will like using refiner
on my video card
never used it tbh
and this even without HireFix!
yeah I dont' use hi-res fix either
must be nice 😭

I always do hire fix 1.5 time for generation of 768x1024 in about 50 seconds

it takes like a minte for 1.5 models with hires fix
idk why pony and sdxl just jumped that HIGH
true in my case it took 2 minutes. but without HiRes.fix
Did you try using the multi diffusion extension?
i forgot about that tbh
send the link again?
That should make it so that you dont use your ram for generation
I just woke up and am on bed 
well could be worse you could be under it
how come it takes like 10 minites just to interupt and image
and what should the settings be?
I think put first 2 at 64 should be fine?
Might be because your image generation is on vae stage, and that usually doesn’t cancel properly
wdym
Spam your interrupt button and/or use your preview to the most least resource one, forgotten the name
spamming doesnt work ive tried but but wdym by the second thing
Wait, might be wrong, I can’t remember what’s the best setting, try 128 128 64 4
So there’s that live preview that you can see when generating your image right?
ye

the cute girls
how dare you

uhhh which one is it
Using that, is a lot less resource usages, and still decent preview
Under settings
thats better kinda
so long
Under preview or something
Well damn ! I only have 6GB!
Because I’m not on pc, so can’t check for you
one sec

changes some of the propmt and tries the multi diffuser thingie
I'll just generate more and more goth ones :<
Might be VAE
Okay, it’s basically that, but don’t change those
We wanna change it for only live preview image

??
heres all the stuff with "live" in it
Live preview method
wtf nite horror bs is this?
TAESD is pretty good quality for being fast
Perfect

is this fine
i got the multidiffuser

I think the settings are off, I can’t remember the best settings for multidiffusion

If I remember correctly, you want your first two values to be as high as possible without going into your ram
Might be wrong tho 


uhhhhh
(I can't figure out how to do a heart plushie) 
i still dont know which numbers to put
strangely, when using MultiDiffusion, everything slowed down a lot for me!
What vibe transfer are you using now stoopid
Then try increasing your first two values
I guess it’s a trial and error
noooooooooooooooooo 😭

................
it was faster
but the multidiffusion isnt working all that gooooood
just gonna keep making the numbers bigger
DynamicThresholding (CFG-Fix) Integrated
Seems like some tiles upscale with 1.0 denoise
why are there 4 of them?
Increase your first two values by like 2 times
I’m gonna create a new command called “hall_of_shame” just to publicly view your vibe transfer data
used multi-diffusion , same H and W , only 1 batch size
but look like this
Maybe tile overlap needs to be higher as well? Maybe 128
Or even higher
I wish I remember my settings
Does multidiffusion work in t2i?
It should
Or at least I think so
Do yall have the tiled vae thingy from that extension?
ok she's missing a hand? I think
Doesn't forge have that enabled by default?
ooh k then
and....?
And what?
exactly
that's not in Forge's multidiffusion

Well, should help if yall intend to get it, no idea tho, since it’s been long since I used that

To my knowledge Forge does that automatically already
grinz evilly
x2
I have doubts
no winnings!
Drew is not the same!
huh?
but he didn’t give up trying to steal the Moon
"Despicable Me!"
umm o..k
is running low on wakeful ness
hmm where everyone go?
not long till bed for me
and I’m already falling asleep... it’s already very early here
changes to a basic anime model instead of the merged ones
same seed: 3544849180 and prompt
ooooh she grew up
yeah, that's it. just trying to use the model
oh boy.. ummm (backs away)
hmm sorta crispy
last one and must sleep
goes to sleep THUD
i used these settings
it took a sec
but
not bad
If it’s working better than before, then it’s a win
its like 3% better but ill take it fr
Better than nothing 
hey 3% of the 3 hours is some 5 minutes
not bad
👍
runs away
booba 
LMAOOOOOO
this is very charming
love it
back to the 80s?!
you did it!
how long does generation take?
like 30 minutes
how is that!?
i place my hopes on neta 2
@wooden coral
hello


anybody from India
i know some from india, there
one lamb + second lamb....
charming
the mouse doesn't come out beautiful
cute
she looks like a pokemon
she is, eevee
Well damn... why don't I have enough memory... there are still some shortcomings!
the fact of the matter is that this is without Upscale
the res is weird
I don't have enough memory for this
This is not 1.5
oh
yeep
i'm trying to work around using img2img
hand drawing rough forms to guide the end result
What do you mean?
Wow! I'm lucky today! I have started the Upscale process!
🤣
on krita i draaw a rough idea of what i want, and i use region to prompt just a precise zone, in the hope to gen what i want, where i want
a bit better!
but when i draw a window in a angle and a library and a desk on the center, i usually get a window as big as the wall and the library as big as the window
i don't know if it's because region is not precise, or because it's not very compatible with an image genned at 1260x1260
i'm still testing
maybe i should add regions where i want a the wall, to limit the dimension of stuff
and things like this
but i never attempted till now to force a model so much
i'm not even sure it's possible
shrug it’s hard for me to even imagine what exactly you’re doing
eheheh
i like to push stuff to the limit
i was just wwondering if i can gen 512x512 abusing the comprension of xl and use the reference to upscale it and img2img it on 1.5 lol
another experiment for later
but ithink i will face the same problem
so many ideas and so little time >.<
damn, i like this
Hi
as usual, don't fall for it
Grabby horns 
nuoooo lol
Yes 
Fluf
No goth
lick


What resolution are you using misery
no clue
512x928

That's a weird one
limited by vram
Where booba
my prompt has so many cute in it, its just too cute
your prompt contains just two words. "(cute, kawaii:8)
leaves this here
anime, piano, guitar, pop, female vocals song. Listen and make your own with Suno.
breaks over back to work
cute
This is unfair!


There's an orange there's on the left
Trying to blend in

Imposter
@brittle rampart is this big enough
👍

Oh! She's not athletic at all. who runs around in such a fluffy dress?!
maybe avoiding nsfw 😂
yeah I sorta cheated and used bing create
but now i'm off work and can do it right
ok making a cute kitten is harder than I thought.... sigh
gives up and plays a game instead
supposed to be ayaka but this is ok i think
Ah been awhie since I poked my head in here
go sleep
Can you share the prompt?
its sd3... so just "anime art, ultra comfy bed, aesthetic anime art, window" lmao


tries it
wonders why his gen is taking so long wtf
hmmms
oh wait
doh
i'm talking to myself.. stop that
Cat 1Girl, 90's stylized , aesthetic anime art , Cute Pin up art
that is some long legs. WOW
that is.. umm.. o.. k
haha
tried that prompt in anima pencil
nice
tries scramble prompts and tweak weights extensions... let's see how this goes
Cat 1Girl, 90's stylized , aesthetic anime Cute Pin up art-poster, art by Garry Trudeau
goes to bed.. thud.. ow
niteee
ninomae ina'nis is such a cutieeee
all the holo designs are nice. can't wait for new designs to get implemented
Here smol
instructions unclear, floot stuck in mouf.

Wonder what would happen if I tried to bring dark gemini into the newer model climate and make it up to date
and hows the SD scene these days? What model is most prominent and how does it fare on GPUs, how much memory is needed?
I'm kinda looking toward the new APU by AMD that gives nearly 128gb of memory. It's not pure GPU but I wonder if it works...Since it's about as fast as a 4070
ohi
@green temple trying to get particle effects to render just via prompt is pretty hard but I'm playing and trying stuff out 😁
oh nice !
well, editinjg the image and processing it as img2img should work
usually that's how i fix stuff
do I have to go back to a1111 to do that? I'm new to comfy and have no idea if and how one would do that there?
i use krita
adding some stuff you can simply open comfy and use it as ai for krita and work on it
very cool stuff
i had problem to install inface, so for now i delete comfy and only using krita
but in some time i have the intention to reinstall it
without the setting of comfy, genning with pony is too slow
ah ok... so I mostly do video or at least am breaking into video animation which I find the control of comfy to be a necessity... I see it uses comfy as a backend... hmmm I'll have to give this thing a look it looks really cool.
the funniest thing is that i can gen 1600x1600 fast enough without upscale
there is a video tool
okay now I really will check it out
i don't even know if my hardware can gen video lol
how much vram you rockin?
but for what i saw till now, the coder did a very good job in coding this stuff
4gvram
I'm on 16 on a 4080
eheheh
but i'm a wizard, i always find ways to make stuff work lol
that's why i like comfy. it gives me more freeway
I think 8 is probably the bare min and 16 allows me to actually do shit
agreed. mhmm mhmm mhmm.
i need to find how it works
no problem I'll test it out for sure when I get back... thanks for the tip about krita. I'm excited squee
nice
I'm currently trying to see if and how it chooses to animate, lol.
oo fancy
which animation technique?
svd img2vid
the first attempt was very static 🤔 time to mess with some settings and pray the next one is better
bzzztttt ... sparkly
still no go on the character motion at 140 I guess I crank it up to 160 🤷♂️ I'll try one more time but if this doesn't work I'll have to build out a cntrl net workflow for this kind of thing.
Good luck 🙂
Still pretty cool, tho
😭 ran out of VRam at the frame interpolation stage of my workflow and it crashed... it still was a static character video from the earlier preview I got... 💢 but I'm gonna try again and go do something else while I wait.
rip
This is the time where I'm thankful for my ADHD--my tasks often have me rollin' from something else to the next project that it's like a surprise coming back to something I generated. "Oh, yeah, I forgot about you."
So if it crashes, sometimes, I can just go, "Right. Well, I suppose IT DIDN'T MATTER ANYWAY." Rl or somethin'
unfortunately my life is so tied to this pc and I don't have a second machine...
Sooo cute! Who's got a cewt wittle face. you do, yes you do.
looks like old anime style
sgeu
ummmm....forgot the name of this YouTuber!
by the way. does he only use a combination with Comfi?
y
I mean the module you use in Krita
I'm a little familiar with this editor
Well, she's not small"

she is an artist
krita diffusion can be set with an online workflow set by them, a local default workflow, and a local manually set workflow
the default workflow works only with krita, the manual set, once loaded, can be use on krita and when krita is not working, can run a workflow run by you
so, what you can't do with krita, you can still do with the loaded comfy workflow without opening and closing different gui
for me, this is the best configuration, but i had problems installing inface on my confy setup, so i deleted everything and now i'm only using krita
but in a nutshell. This is a process of working in conjunction with krita + Comfie*
y
ye, replicating old style is easy-ish.
krita diffusion use comfy as a backbone to run models and krita as a gui
but he is beautiful in his own way! I really like him"
because I haven’t heard about the other connection
Krita - very nice editor
and he is free
if I don't confuse it, like Blender, in its idea it remains Free Software
i know a lotta ppl who use krita & really like it (who don't use AI stuff). does it have good stable diffusion support now or something?
but even if we abstract from AI. This is a very good graphic editor
and it's completely free
yep
bepore krita
i was a corel painter user
it was pretty easy too change
except for some difference, krita was easy enough to figure
and have even some more functions
I worked with him like as a graphic editor
but guys, tell me Pony - is it just a separate model or is it like a new standard?
What do you mean?
It's a fine tune of sdxl if that's what you're asking
i don't really think it's a new standard either... more impressed with nai diffusion 3 for paid or for local anima pencil XL right now... but i think one of the major driving forces behind anima pencil XL being so good is animagine XL 3.0/3.1 as base.
what they claim they wanna do with pony diffusion xl v7 though sounds good, but been waiting for that for agesss or the new waifu diffusion
no




