#🍥|anime
1 messages · Page 215 of 1
made some awesome promo-art for my manga, thanks to stable diffusion
if you like what you see, please like and follow the instagram here. chapter 1 is free to read on the Instagram Bio
https://www.instagram.com/ilifa_inheritance/
civitai dropped a apr fools joke with text 2 odor....
any body knows how to set up controlnet reference only with diffusers??
no pc so im on colab with commandline diffusers but i cant seem to find the damn controlnet model
if anybody can reply with code thatd be gr8
about that image thing im uh getting somewhere
im on to something
loving the way eyes are drawn sofa
lemme see, hold my balls
Problems with faces usually come down to resolution and prompt. If you don't want :d face, then don't prompt :d and preferably do prompt what you want. If it's just not detailed enough then use hires fix and/or face detailer to give it more detail.
messin' around with the prompting styles from last night
I think the animagine v3 actually works a little better than 3.1 for this style
qt
I've noticed SD 1.5 models really struggle with modern military outfits, or anything tacticool in general.
most 1.5 models are pretty narrowly focused, but i'm sure there's one somewhere that's trained for that
a lora is probably the best option though
@supple raptor did you see the first episode of the new spice and wolf remake came out today?
horrendous
I haven't watched it yet, but yes I've been looking forward to it since it was announced
I watched it, I thought it was well done so far. It sticks to the source a bit more.
Very happy to get more anime
show an example of what you wanna make or just get better at prompting
couldn't have come at a better time either since frieren is done for the season
Frieren was amazing. Unfortunately for the rest of the anime this year they automatically lose best girl to Holo.
absolutely
combining my soft anime and line art style, interesting style
100 steps, 3cfg
"cute geisha chibi sakura" 🙂
messing around with that prompting from last night mixed with pony/autismix
seems to be just about as good as anything i can do my advice is just upscalse it or use impaint to do little touc up, also a vae should fix the color
"/dream"
Danger moth.
any body knows how to set up controlnet reference only with diffusers??
no pc so im on colab with commandline diffusers but i cant seem to find the damn controlnet model
if anybody can reply with code thatd be gr8
no idea
Bring the spray 
Afaik there is no actual model for it. Here's how the implementation in ComfyUI works if it's helpful though. https://github.com/comfyanonymous/ComfyUI_experiments/blob/master/reference_only.py
D::::: FUCK
Just found this. Looks like it should be all you need https://github.com/huggingface/diffusers/blob/main/examples/community/README.md#stable-diffusion-reference

More of my merge model.
you scared her
NAI? or some pony model?
some pony model
souka
check the algorithm overview. https://github.com/KohakuBlueleaf/Lycoris
@dull jackal
something tells me he forgot to set the bot XD
Mega cute!
@dull jackal Come back!!!
For a steam giftcard!?
Yeah already 150$ free today
Sasuke

You can have image of Yue instead
WHOAH!!! so much money!!!
we are all rich!
i just bough an house in gift cards
don't look me liek this <.<
😮 Was it a cardboard house!?
was the house of paper!
I hope it was some fancy origami!
Steampunk x Rias Gremory
Good day
Pls can someone tell me where to generate images
you first need to download the gui you prefer.
i suggest forge ui and than a model you like
they are all good
welp this sucks but this is the best my "poor man nai" can do
y is nai models so peak and y cant it be leaked boohoohoo
u can get something similar by merging some style loras into pony
ok maybe just maybe it is maybe just probably maybe not as hopeless as i thoght
ill dedicate tomorrow if it doesnt work then welp no more
did you see the images I shared? they were basically the same.
with animagine. I used v3.0, thought it did a little better than 3.1.
line quality was close but i dislike how the face looked
they blushing for no damn reason
but yes i did see them
they dont have to be blushing...
I mean some were alright but most of the time the faces bug me so much
i aim to fix that first
if you say so. 🙂
oh my gosh Ive been running prompts with weird results.. trying to figure out what's wrong for an hour.. and just now realized i had the wrong model selected
it does Cammy from Street Fighter pretty well...
oh, and han juri and chun li, too
Oh yea, I totally forgot I could be doing Juri stuff.
the best thing is the fat grinning cat in the background 🙂
just some randoms...
shameless plug... updated both of my image tools today:
https://github.com/GarlicCookie/PNG-SD-Info-Viewer
https://github.com/GarlicCookie/SD-Quick-View
oh wow this set came out so fun...
an extra Tifa for today...
Any lora creation guides out there for noobs like me? A lot of AI tutorial stuff is either out of date or just terrible.
I hand edited 25+ images for my OC and need to see if it is workable enough.
I never could find a decent guide for XL, they were blatantly wrong most of the time so I just ended up running the training and figured it out well enough to make it work
for 1.5 on the other hand, there should be plenty out there that is recent enough to work since it hasn't changed
I'll see what I can find. I was hoping there could be some new tools in the last 10-12 months that make things less of a hassle.
evenin'

quiet tonight

interesting.. this looks like the style that person from yesterday wanted
damn, im a day late 😔
it was @unkempt cape - maybe he will see later
on these i used loras for
- sanpaku eyes
- half closed eyes
& - jitome eyes
you can get it without loras but loras are great

the hand tho 😔
it do be like that sometimes with the 
Damn I really want to find a good natural chat LLM, that can remain in character on the profile it has
well one I can run anyway...
not have to pay for an expensive server
Fork over that apple
Two!? Well, guess there's one for me and you.
Then let us celebrate this glorious morning and nomf---!!!
yes
hhe
made a style and just playing around : " retro anime style, fine art, cel shaded, elegant, vivid pastel colors and shading, line art overlay, holographic triple shadows"

ik i said to not do this shit any more but man if its not close
shading off a bit, left chin?? lines lil too thick but everything else so close
just gotta find ways to fix some issues. guys whats the best way to fix artifacts or random errors
Did you use hires
i upscale
#🌠|show-and-tell goku in super sayen form
error no bots here only humans
dress name?
Rule 6 just saying 
then use hires
why delete?

hopefully u seen it dude im aint trying to get banned
I couldnt read it
DM
it blocked me once
I did it
We have those words on a blacklist for the bots, but they apply on all channels 
eyyo sorry bro didnt mean it )):
dm'ed u
arigatoo
my god if only xl has reference
xl has reference isnt it?
What model are you using? If you're still trying to replicate that NAI style then I recommend using a good SDXL model, this image makes me think you're using something outdated like a 1.5 model
There is no way to replicate without ref in sdxl
?
That's a good choice, it has a nice base style but doesn't know nearly as many artists as other Pony models
so its good to not to know many artists? or is it because we're focusing on only this style (NAI)
Which model are you using? Or are you using this ponynai3 model?
Counterfeit 3
yes its 1.5
That model was trained on a couple hundred images from what I saw, not to mention they're all AI images, not the best idea to use AI images to train a model. But to answer your question yes, a model that has artist tags is a lot better for style control
counterfeit is good for concept art or wallpapers I guess
but thats the only way i get ref. i also used lora (but no credit seaart to make sdxl lora lull)
basically my best bet rn is 1.5
Ah okay, that explains the messed up hands and the overall bad quality
and then canny or recolor xl
why cant use xl model?
Seaart? Why not just generate on local?
i hope to run that shit through canny or something sdxl
No pc
Haven't used that site in forever but does it not have SDXL models?
it has but sdxl has NO ref support
Ah okay, I understand now
has lora which im using but 500 credit i dont have
From what I can remember you can have up to 6 accounts on there so you could just make all of those and then you'll have some more credits to use
500 to train lora
Ah okay
my game plan is: sd 1.5 with trained lora for lineart => sdxl canny + img2img
if it works it works it doesnt well shit
Not sure honestly, never really tried using canny before so.. Hope it works tho
ye
Weird they don't have it. Reference only just works on XL models since it doesn't require a controlnet model.
stupid
astolfo
what about the yodayo one? their chat is pretty good.
actually, yodayo has image generation, too, that would probably also be very similar to the NAI stuff that @unkempt cape is trying to do. So, you might wanna check that. It isn't local, of course.
Yodayo just takes open source models (with or without the creator's permission) and charges people to use them. At $20/m for unlimited gens and chatting on Yodayo, you'd be better off paying $25/m for NAI. Yodayo doesn't list the context size their chat bots get, but I'd put money on it not being as high as the 8192 tokens that the $25 tier on NAI gets you.
anyone know the best way to implement 2 character or 2 loras in the same image
if it's 2 characters then regional prompter, if its two different lora, just include them both?
wish I could turn this off
just found out its not just one model, sorry guys im new at this lol ill do my research now
So these are the artists I combined the other day to get that style some folks were trying to mimic....
I tried various combinations of all these and it's been tricky to pick a favorite combo.
i really like the 2nd one here
I'm running a few large batches to see if I can narrow in on favorites
Hi, is there a way generate anime images with good hands and fingers with SD without having to inpaint?
use embedding?
can we downscale a image? in a1111?

not in img to img
I mean from txt2img
and let a style bleed in for basically no improvement?
i tested all the popular ti's and loras for 1.x
result? basically nothing
well the only one that did do something was vbi, which barely made hands better at the cost of major style bleed in
adetailer/hires fix is better
oh yeah make sure you have a sane negative prompt
wait hires fix means with 0.7 den or smthing?
oh dun use latent, always ultra whatever the upscale
ultra and the other pixel based upscalers only work for post finalization
if u wanna use latent without deformities try with NNLatent for comfy
so that means latent is like a minimodel with upscale?
or latent upscaler
nnlatent is latent upscaler
latent upscale makes the latents (pre decode) go big
huh, will check out if they've updated that repo
prob not last i checked it was updated like 3months ago😔
Some models are just really bad with hands. If you want better hands, then try SDXL if you can run it
i was thinking of purging every 1.5 model from my pc and just genning with xl and pony. anyway, with turbo, the time is the same and quality is way better
Try the hand refiner
Works well with ADetailer
Tight
no grabby
grab what?
chibi shenhe
nope,the full size package one 
naurrrr
she needs to be expressionless, and tall

shes always mad cuz u always tryin to peek from below 
who? me?
naurrr

How would you specify tall
Without references
You can't say someone is tall from an image
Without any point of reference
That's my point
hmm like what?
u can if i put a guy gettin choked by her thighs 
You made that?
i made her,shes mine,she belongs to me now
ugly 
nandato...
I can run it
Thanks, I will
sdxl is really handy
can't wait to get my hands on better models
I'll post a handful of hand picked images from sd3
Is it possible to make an img2img anime 2 manga? Like, start with a colored image and convert it to black and white monochrome manga like this
The converters Ive seen typically use a ton of grayscale/screen tone, they don't have the blank white areas that is typical of manga
something like that?
if you are on SD1.5
there's a lora for it
name: animeLineartMangaLike_v30MangaLike
oh wait you meant img2img
Yeah that's where I got the Lora from. But I mean, not prompting for it. But getting an actual color anime still and turning it into a manga image. I tried so many loras and models and combinations
i guess there's controlnet for that
I tried controlnet too but the image doesn't look good
sus
Go away

No
Don't go
Post thick girls
👍
give her sm clothes
guys whats the best way to train a lora for animagine 3.1 online cloud gpu? i dont have pc
thats the only option you have right? if you dont have a pc
ye
i can wait a week for seaart but might as well not waste my time and do what i can rn
theres replicate but the model for training is sdxl 1.0 which is like cool i guess but the realism bleeds in 😦

I tried to train a pony lora, it gave me 1day time estimate,
too much
canceled
4070 isnt enough I guess

dude at least its possible for u
yesh at LEAST
i cant even use colab it straights up doesnt have compute power to do it
yeah...
If pony is on the SDXL architecture, one day is definitely too much on a 4070. I have trained many Loras on my 4070 for sdxl.
I see
so you trained with 1.5SD?
I don't understand the question. With a regular 12gb Rtx 4070 it's possible to train 1.5, SDXL and cascade Loras. Depending on your dataset it shouldn't take longer than 1 hour.
skull issue
"Oh goddess of the thighs send me some thick goth ones"
👍
@cerulean bane 's favorite
10/10
guys whats the best way to copy say an artist anime style if u have the images
like as close as possible
(asking for my impossible quest of nai images lel)
(not literally)
u train a lora
theres no 2nd best option
yea ip adapter doesnt work for anime styles if it did ppl wouldnt train loras

why not use a sdxl model?
- ref only just doesnt work for some reason- other controlnets dont copy as well as that one.
- without ref only i cant
- no pc for lora
ouhg right
- no $ for epic cloud nasa
NAI style but chibihttps://civitai.com/models/277571/nai3-chibi45
ok not the style i was looking for but this very good stuff
so what style are you looking for?
oh fuck is there a not chibi verison
the same style he's been looking for for like 2 weeks
yeah I couldnt find
the problem is replicating a particular twitter users style with no tools, no money, and no actual effort
I dunno what style?

doesnt hurt to dream yk
at some point you either have to give up on a perfect copy, or start paying for something
no money no waifus 
1st option i guess ):
i mean, the artist uses NAI, so you HAVE to use NAI to get 100% the same
what style?????
if you want something as close as possible, then you're going to need a PC (or online service) that can run the tools needed to do that, NAI is cheaper
yeh i think i should just gave up
would have been faster to search this channel
kay
intercourse style

...why is it sd1.5
n not sdxl
I guess that guy trained a model from NAI images
why do you think it needs to be sdxl?

its sdxl
Ground view of the central pyramidal spire, reaching up through the ice like an obsidian obelisk
I feel like a few of us replicated the style he wanted already
Not sure what could possibly still be missing.
What even was the style
If he cant get it perfect surely close enough should be good enough
I think you posted it the other day... let me see what I can find.
If he can't find a single model/lora that suits his tastes he could just add more loras to fix any issues he sees with the ones he's found
get creative ong
Original he was asking about. #🍥|anime message
Oh nice
Yeah I've actually made some pretty similar art recently
I used a combination of loras to get it like that
He should be able to do the same 
If he's wanting the style I think I have a few loras that might help him but I'm not home rn 😔
my attempts to do similar: #🍥|anime message
#🍥|anime message
#🍥|anime message
#🍥|anime message
#🍥|anime message
#🍥|anime message
#🍥|anime message
Dawg those are so close its crazy
The ones you ( @nova remnant ) posted that I thought had similar elements. #🍥|anime message
Maybe the eyes??
And I also suggested trying yodayo, which I'm pretty sure that original prompter uses sometimes.
I thought yodayo got yeeted tbh
And lastly, I gave him the actual metadata from one of the images created by that prompter.
Yuh
yeeted == gone?
Yes
it's not gone.. seems to still be around...
maybe it was gone before? i dunno, I can't keep up with it all. Is it bad?
I'm pretty sure it's actually really good but like it got poofed for some reason last I heard??
I don't remember why
So like, what exactly is the issue he's having 
If he has all that
I don't really know
maybe nothing, maybe I misinterpreted it 😛

I thought you got pretty close
Maybe he was just wanting a white bg
There's so much crazy good ai art here 
Yalls hands are pefect
cool, what model?
just gotta add 'white background' to the prompt 😄
I know 😩
NAI then local upscale + Hires fix & face detailer with Kohaku-XL-delta.
I'm weird, I prefer the white bgs
thanks
What is NAI 
If you have a NAI sub, the full comfy workflow is embedded in the image.
NovelAI. Closed source online subscription service.
NOOOOO
I remember that
Pretty easily the best SDXL model around.
I have to buy it?
their original model is the main reason that our SD1.5 models were good. It was leaked and then everything after was trained with that as the base
yep, $25/m will get you unlimited normal res gens and several hundred hires fixed through them. My workflow gens at normal res through NAI and does the upscaling locally so it's unlimited gens.
Yikes
yeh, it's pretty expensive. I got a free month from someone who works there.
That's crazy
We need another leak
no we don't
come on
Lots of cool stuff in the open source world. Let's not leak stuff. 😦
wuddup nubs
Another leak is very doubtful but I have reason to believe that WaifuDiffusion may have a competitive opensource model before too long.
hi omnia

The WD team is still around? that's neat
It's largely different people working on it now, but yeah they're still doing stuff
They've been quiet because they lost access to the compute from coreweave that they were using before (because it went to a paying customer instead of free access for an employee), but one of the members now works at 11labs and has access to some free compute there.
nah, it's just my hobby
😔
Understandable
is there any way to make NAI based model gens look good?
You mean NAID1 (the one that was leaked) based models or the current NAID3 that's only available through subscription.?
Leaked one
I conjure starboard
I mean, pick any SD1.5 based model that you like the look of. It's almost certainly trained on top of NAI, or is a merge of multiple models that were.
She looks very HAPEV
those stars are so pretty 😭
I just got another one
:D
Just happened to see that after checking what ping I had here
Isnt wd absolutely trash at this point??
like just out performed by every model known to man
we need asf
this would be an entirely new model based on a much, much better dataset.
Can i see where u heard this from?
sorta doubting stuff here yk last time it was NOT that great lel
Touhou Project AI server, the people who run the server are the WD team.
loooool
i think i found their discord
whats WD now
WaifuDiffusion. It was the og anime model back before we even had SD 1.5.
ooo, tyty
Some of its output was seriously cool tho
just that no matter what i do the output doesnt look 1 bit like the official pics
Yeah, I still have a collection of neat images from WD1.2. Back in those days we also normally generated batches of like 50+ images at a time and felt lucky when 1 was kind of good though.
Then SD1.5 came out and they started work on WD1.3 which was going pretty well, but NAI leaked during their training and everyone switched to using that. And since then we've only gotten a couple of not very noteworthy models from them.
well we can only hope i guess
we are skipping SD3 and going straight to SD4
./s
just incase people take me for real
stands for you 

I like S

S and TT?
Anime Snoop Dogg the cute japanese girl finds a free job posting at a maid cafe serving up delicious snacks and a double cheeseburger with dr dre
what you thought i was just going to do the short parody video i did a couple of months ago? :^)
Parody of the intro for the anime Blend S.
Lot of hidden work in this one, a ton of rotoscoping, ther...
s░s░e░x░y░ ░a░n░i░m░e░ ░g░i░r░l░s░
this one is better 4 u
One piece
for no reason i shall present you all this 🥰 awful color combination
also can i ask what model that is lol
NovelAi
oh god f-
Sry for asking but do u know what Strength and Noise mean in img2img nai? Idk why there are 2 parameters there

"noise_schedule": "native" - This one I don't know
"controlnet_strength": 1.0 - This is the only "Strength" I can see and it's related to the controlnet model
But its img2img
isnt controlnet and img2img different things
Nope, it's complimentary
uh i dont get it
For example, if you use a controlnet "recolor", you can change the colors of the image you're trying to modify
Or you can use Openpose to get the exact same pose as the base
so "add a base img" is controlnet not img2img is that what u mean
You can also use it if you "inpaint" something in the image
Not sure I understand the question
It depends on what you think img2img is, or does
there is an option "add a base img" and "vibe transfer"
the second i think its controlnet ref
so i think the first 1 is img2img
which is weird that it has 2 options
Oh, you're talking about NovelAI which I never used and know nothing about.
Sorry, All I said was probably trash.
mornin'
ive given up on this thing but just wanna point out, this prompt is completely off
dunno where u took it from but the actual artists had these names {artist_hxxg,{pigeon666},[ciloranko,wlop],[artist...
I took it from the metadata from one of the prompters images.
it is literally pulled from the image
ok idk tbh but this was straight up the name of one of the images
specifically this image.
discord ffs load the damn image
...its obviously not the same style... Anyways how did u get the metadata? I downloaded it but all metadata gone
sorry that was just a copy/paste.... metadata included in this one
no like u pull it from twitter right?
just any twitter images at all
no matter what image from twitter all metadata wiped for me so i wanna ask how did u get it
I did not pull it from twitter.
... ): oh
I did a reverse image search on twitter images to find that this prompter has a collection of images on yodayo.
Some of those images are retouched and the metadata is gone, but this was the most recent image and they got lazy. 🙂
if you have a different artist list from some of the other images, then go with those, of course
here's a fun set
splendid
you can find a description in the docs: https://docs.novelai.net/image/strengthnoise.html
Strength is equivalent to denoising. Noise is adding additional noise to the image before beginning to denoise.
cute
man i was training a lora and the out had the time like (something like)
epoch 1/10
6/780 [0:55<1:45:40] and i was like "nahhhhhhhh 10 hours lora training??? bro i aint doing it"
turns out its the entire process phew scared me there ngl
evenin'
what's cookin'?
making darkness now 
ive discovered wildcards
its so fun, never knew i needed this in life
whatcha mean?'
im saying i wish i knew about wildcards sooner 
theyre really fun for random prompt stuff
whatch amakin'?
I see
hows the search for the specific art style
yeah?

given up
well @nova remnant - let's see!
bet
this one is my favorite of the two
tragic man 😔
Try Coffee Donut for free and many other models at AIEasyPic.Checkpoint baked with ClearVae 2.3 = https://civitai.com/models/22354?modelVersionId=88156A checkpoint merged consists 3 of :DonutHoleMix = https://civitai.com/models/78433?modelVersionId=83231line and light - v1.0 = https://civitai.com/models/42930?modelVersionId=47602AingDiffusion v5...
this is the second one

i just downloaded two random ones and im liking them so far
its so hard to find any sites outside of civit 😔
No i mean I wanna see the images. 😄
pretty neat
try a white background one
i'm actually not generating cartoons for once, and am trying more photorealistic 😛
i dunno.. time to go back to anime models 😛
bet
bet?
ah ok
that one is the first model
didn't they release something that works with xl models that gives consistent character creation?
What you mean weird
i wish i could run xl models 😭
Just the usual chimera angels

tbh thats totally normal for me
What GPU do you have
2060??
i think??
rtx 2060 or smth

@nova remnant are you using forge
a1111???
try forge
what is forge 
The better a1111 for optimization

Just Google forge
And do the thing
haha well there ya go
since i'm on github, shameless plug for my image viewers....
https://github.com/GarlicCookie/PNG-SD-Info-Viewer
https://github.com/GarlicCookie/SD-Quick-View
i use infinite image browsing 

mine are a bit lighter than that, but if you use those features, I get it
mine are also offline local programs, not web driven, fwiw
hold up, got smth that might explain everything


@lyric zealot
Look at this gaming machine
W H Y
who are you summoning 😭
listen
it runs
idk how but it does
A stupid one

it runs photoshop, after effects, a1111, anything
but it screams at me
if i play spotify and gen an image my laptop starts speaking in tongues

Who's Kaily?
hopefully you can get some stuff tweaked and run the xl models
idk how im gonna do that tbh
if you are running forge
a1111 wont even run unless i have one of those storage devices plugged up
a specfic one
then you might already be able to run sdxl based model, but just a little slow
im so horrible at this stuff
im broke 
let us know what you buy tomorrow
i get like
yay haul!
nah
id have to become an egirl then 😔
what exactly is an egirl?
is SSSniperWolf an egirl?


i feel like no one can really define it
female who uses her looks for views/gifts specifically online
Tupid
not even looks tho
just being female

not as stoopid as you 
I feel like that isn't the full definition 😛
how slow 
i have a 8gb vram 3070, so i can't tell you how you will fare
😔
try it 
12GB here
the flex 😭







