#🍥|anime
1 messages · Page 127 of 1
Who is this witch?
there is no face in the hands there is no face in the hands there is no face in the hands
for sure 
6x upscale failed 
i see so many examples of embedding on civitai where people just shuv and squeeze all kinds of TI they get their hands on
i do that
A luck thing there.
Most likely.
this is my new generic prompts
P: (medium shot), (highly detailed, masterpiece:1.2), (best quality)
N: easynegative, (bad-hands-5:1.1), (low quality, worst quality:1.2), (watermark)
medium shot can vary ofc if i wanted full body view or something elese
actually for negative i'm gonna use
FastNegativeV2, negative_hand-neg, (low quality, worst quality:1.2), (watermark)
lots of models out there have poor training on hands
but this mod calicomix is pretty sweet in rendering hands
anime models always suck at hands
😬 TI isn't helping a whole lot btw
i get better results with my regular prompts
this was the reason i moved away from using TI some months ago
everything looks neatly rendered there
the hands 
the spaghetti hands are generic to SD 1.5
wrong
these are from a series of armed anime girls
and the fingers issues hasn't been fully sorted out with SDXL either
i don't like xl, honestly
only good for landscapes and structures i find
needs more finetune
yeah hopefully in the next version, cause SDXL is too resource hungry
(medium shot), (highly detailed, masterpiece:1.2), (best quality), 1girl in a long elegant robe inside a castle, standing, bare legs, solo, (medieval)
bad anatomy, bad proportions, malformed, bad hands, (fused fingers, interlocked fingers:1.1), asymmetric, conjoined, distorted, cloned, (low quality, worst quality:1.2), out of focus, low contrast, oversaturated, monochrome, grainy, pixelated, border, (watermark)
i couldn't get away with making this perspective and the bare feet
what do you mean
i'd get called a foot fetishist
how about this artstyle
some oilpainting
this has bunch of styles to use inside A1111 https://github.com/ahgsql/StyleSelectorXL
works on both sdxl and sd 1.5
i tried it briefly but later removed it
i usually use more of this pastel style
i love oil painting style
some of the images from MJ looks amazing
SD hasn't been able to come close to MJ in oil painting
didnt want background to look somewhat blurred
deep focus should fix that
Wdym by fine-tuning?
More training time?
Cause then it would just be overtrained
Someone needs to immortalize this comment
Good
And also they suck their hands
How did you upscale? This doesn't look like good upscaling
Man you fr be spamming this emoji
1.5x hires ->4x tile upscale
tiles are too small
You have 503 messages with that emoji out of 17 thousand messages
That's like 3% of your messages
What upscaler
What denoising
How many steps
https://github.com/city96/SD-Latent-Upscaler uhh like this one starting at step 8 out of 20
or maybe 11 on that image
idk
"Think of this as an ESRGAN for latents, except severely undertrained."
Basically it's shit
it works quite well
And also
You should upscale pixels
Not latent
I mean depending on your goal. But if you want the image to be clearer
I mean
I would just use SD upscale script on a1111 img2img tab
I dont use a1111
GM eface
gm 🙂
I know im just tinkering with similar shit
Basically it upscales by tiling the image then upscaling each one and then combining them
afaik ulti upscale is just a tiled sampler with extra steps to remove the edges
it overlaps by (setting) pixels
im doing more or less the same thing but with slightly differant nodes
Ye
and then you can diffuse those overlapped areas again
Man comfy is overrated imo tho I didn't try it yet
nah, comfy not overrated, just different
i can't pull of my prompting shenanigans there 😦 so for me it won't work
a1111 is so bloated with addons if you want to do the same things you do in comfy
and weird workflows are easier to do on comfy
True
Idk man
Nodes
Fuck em
A1111 is superior UX
Ofc depends on the person using it
mmm 3rd arm yummy
You know what I need to do
I need to buy a starlink router thingy
And use it on a plane
Frfr
is this xl?
yeah
man. if only xl was easier to train 
i just merge
its easy if u have a cluster of H100s
hmm maybe I should tinker with extracting loras and merging them into xl
thats one way to follow a prompt I guess
I've been cursed by a spate of bad hands
changed nothing recently but every gen seems deformed
When is qlora coming out for sd
qlora?
"""
QLoRA is an even more memory efficient version of LoRA where the pretrained model is loaded to GPU memory as quantized 4-bit weights (compared to 8-bits in the case of LoRA), while preserving similar effectiveness to LoRA.
"""
dunno if it's compatible with sd tho
I think I maybe once heard someone talking about it
They said 4bit is too low for sd or something
Idk
Anyways I am on plane now
Very exicitiitng
seems to work
quite liking how this is turning out, even at lower than ideal res
@wispy canopy
what's up @nova remnant
It doesn't update
I was just watching mathil
And slept like always

It doesn't boot in safe mode too, I'm trying to see if restore files does something
Big sad
Blue like Hatsune Miku
doesn't boot in safe mode either? that would indicate something big is up
boot to recovery and then sfc /scannow (which always fails lol)
boba lora when?
No need
i always get multiple straws XD
you do git pull from the root install directory
if you have a fucked up venv/python, you delete the venv folder too
although reinstalling venv will take a while and downloads
Except for python which file i can delete so i can update sd
updating is just the git pull command from the command line
Ik that i just deleted venv
so i can like free up 8+gb for my storage
im gonna take a break from SD abit
thats why im asking again which file needs to be deleted so i can update it if i ever comeback
thx btw
save the outputs and models\stable-diffusion folder i suppose (move it away) and nuke the install folder from orbit?
basicaly, venv and the models take the most space
how much gb does that install folder use
in a clean case? not so much, in my case, a whole lot
if i delete install folder will it like break my user/bat
venv is the only safe folder you can remove without affecting the current install (as it reinstalls later)
or will my user/bat just automatically add it back
yes, it iwll
if you delete the whole folder, everything is gone, and the next time you install it will be default
well nvm im sticking with just deleting venv
yeah, that's the smartest thing i suppose
i had to delete all of my loras so i can have a break
and thx again for helping me
fr fr
np ^^ , see you in a while, when you get into sd generation again 🙂

see you within a few months
ill still be chattin though lookin at the pics
wtf is this tripophobia shit
if you remove venv and the dependencies are outdated you just bricked the entire installation
git pull, pip install -r requirements.txt, start
and if that doesn't work. well, good thing git clone exists too 🙂
very nice ❤️
If you wann get more space back then open up a Cmd and type
pip cache purge
That will delete Cached files for SD install
They get redownloaded again when running the webui-user.bat
CS1o spitting facts 😮 i'm just a noob 😄
money can buy happiness
Been here bro. I hope you had a save checkpoint
wow 1250 files purged with the pip cache purge. thanks
Jedi light sword?
Star Fleet Kimono
<---- trekkie, can ya tell?
ok went from star trek to NASA?
I did have for once in my life
But Nothing I tried worked, my OS is just installed on a smol SSD, only a full reinstall worked
But not a huge deal just gotta download the apps again like discord

wait what happened?
Look up to what he replied to
Windows decided to k.o itself
Tried some stuff and gave up 
rip so did it finished reboot yet?
Yes this is my first message on the new stupid os that microsoft keeps blasting you with useless stuff :D
ah
Yachiyo Nanami/Nanami Yachiyo of Magia Record fame, is that you?
define cheating
Manual input/controlnet stuff haha
then yes
X3
not controlnet tho
I find it really hard to get blindfold to cover the eyes
i'd rather spend 20 seconds drawing that regenerating for a good result
lmao
Rushzilla style 

Anotha one
Giga thicc
No worries, back to generating shimakaze 
Is this shimakaze ?
Those god tier surgeons got the cancer again for the third time. I LIVE
hips to the max
That is an adorable picture with her face on it.
And there is bigger 
Thats really cool :o
Tiny and adorable, this one is.

Hi, how use the XL checkpoints and models?
the colors & lines on this are cool... new model?
I trained a new LoRA :o
Kobeni pic in that style
clean white background
Well, I just trained a style LoRA
Rushzilla is the artist
Wait
@elfin geode Yo Jae, are leotards fine?
You'd need my LoRA XD
Though my LoRAs most of the time don't have trigger words
bah I do not. I can create masterpieces!
No, I haven't released it yet :o
Sure, just avoid lewd stuff lol
nice dress
I like running plots of a prompt across multiple models, but hate the loading time between models. ughhh
Trying to recreate your character, @magic wave ...
Fair, so this is fine? :o
If you just do (shimakaze_/(kancolle/), with a anime model you should be able to do it
Anime models seem to recognize her
Well, at least Aurora does
True 
she's promoting unhealthy body standards tho XD
I mean true
Her internal organs are probably suffocating XD
A while back? 
Not very sfw lol but doesn't break the rules. It's easy to cross the line with similar content so just be mindful of that!
anime being unrealistic? no kidding...
Fair fair, as long as no 🐫, or too much skin close to any bits right? 
True 
mind = blown
you know i didn't do it on purpose tho 😄 (as all my prprompts are not anything specific)
oh I didn't realize it was an actual character
It is yeah xD
tough to get thicker lines on some of these models
True
and the darker shadows
https://civitai.com/models/81313/boldline I recommend this if you want bold lines
interesting
Question, what models do people usually use when training anime character Lora’s?
I have some models, but they’re catered to… not that.
i got textgen ui working uwu
eface helped a ton and hours of googling and trying random conda commands and stuff and guesswork about settings and i still am not totally sure im doing is right
Nice but guys stop the thickness, it is too much to handle.
Oh wow this is amazing
Do it at 3840x1600 for mah screen
"a view of the end of the world"
nice
coool
XD
quiet tonight
NAI
Though some people use AnyLora
But I preffer NAI
What you mean infer?
Like when you make the inages
Inference
Like semi realistic models is not the best for this
Though for characters I think pretty much any model works
Uh.. which would you recommend then?
Most anime models are NAI based
how you gettin' those colors? 🙂
Though then the problem becomes finding a model that is not too overtrained/overfitted on specific concepts/characters and would therefore have a harder time replicating the character even with the lora
Aurora

that uses a Hyakka Ryouran lora i made a long time ago image should be encoded
Can I get a link to it?
hit that letter emoji
Please tell me it’ll be obvious which one is the right one.
right on
You should get used to both huggingface and civitai preferebly
Civitai is a website where people host models and you can download them
Civitai is a platform for Stable Diffusion AI Art models. Browse a collection of thousands of models from a growing number of creators. Join an engaged community in reviewing models and sharing images with prompts to get you started.
https://civitai.com/models/40199/aurora @valid bridge
Thats Aurora
Was searching for it but my internet too slow
Gotcha, thank you!
just be careful on civit... it's like landmines out there
My browser just knows 
What's Aurora? ... how far behind AM I? 😐
nice
just another model
If I searched that it would probably bring me to the AUR website
Arch user repository
I use arch, btw
I mean, Aurora came out like...
think it's that one, another sidegrade models vs something like based66
Ye already outdated frfr
AuroraONE?
auroraTWO
i'm still over here using lyriel and ghostmix and aniverse and revanimated and.... and other random stuff
Now tell me why the samples look like THIS
Nice cool
Um idk
looks fine
How did you train then how did you use the lora
dont worry about it theres no best model u either download a lot of them and try them out or just stick to the one u like the most
i think i downloaded nai when i was using sygil - before realizing it was incompatible, but i think i kept the model but im not sure if its legit or what
lol I know man, I'm just being flippant 🙂
What were your training settings and what did you use to train it
most 'models' are just 50 shades of nai anyways
Literally

That is how you train a LoRA
Or technically night
Those settings are atrocious
Get Derrian's training stuff
And use these settings @valid bridge
You can load them up in the derrian GUI
Remember to change ouput directory, model used, saving name, and dataset folder tho
and yes it did work
Or am I confused
I think that might be the buckets
Still confused
I just... put that because I didn't know what Anything V3 was trained on...
Resolution I mean
Not sure
Most are 512x512
You can use 768x768 if you feel like waiting longer, usually gives more detail (i think)
i tend to do 768 x 768 now helps a tiny bit

I think 1024x1024 for 1.5 models is too much tho
For SDXL is the norm I think, but for 1.5 models I think its not good
I think thats fine
How can gpu handle tho?
I mean on 12gb VRAM I get it
I can do 1024x1024 at 1 batch, and I got 12 gb vram
Inference?
their really isn't a 100% right way to do this, shit is still in very wild west days and it's going to be that way for years
Training
Well, then yeah
There is definitely a wrong way tho
XD
time for some burgs
Unless you go to projectai
Does this lady have a name?
They do theses on this stuff fr
no..
Project AI people are scawy
Throw a name
True and real
They too smart for me 
Not only smart but they are way too skilled
What was that one girls name
Pog
You know right?
Wait... locon? I'd be training a locon?
OH right, you can change it to LoRA if you want
LoCon is better for style
😐 Ok.
Oh thanks god
I was told that by Project AI person with japanese letters in their name, it is the truth

XD
Ok sorry
Gomenasai
I think all my controlnet stuff is way out of date
Same
But best for style
Sometimes with locon there is too much detail. Like if you have not that big of a dataset in general I prefer lora
isn't there some sort of controlnet tile thing for large renders?
Fair, I rarely use LoRA so I dunno
There is
Idk overrated frfr tho
Ong
No cap
hmm
teach me 🙂
Thicc doesn't cut it.
now make her pregante
Am i pegnate??
preganante
peegat
true and real
Just because it must be posted when mentioned for anyone who hasn't seen this absolute classic:
https://youtu.be/EShUeudtaFg?feature=shared
A glimpse into the wonderful world of Yahoo! Answers. Song is Curley Shirley by Otto Sieben.
Twitch channel is still underway, just been having some technical difficulties. Gonna try and get it off the ground within the week. When it's ready I'll post a quick announcement video with a scheduled time.
I have, but I'll watch it again
Hell yeah you will. 😄
my repo got a new star 🙂
gremlin miku
Nice man
Which repo?
You should link your github to your discord
my image viewer - https://github.com/GarlicCookie/PNG-SD-Info-Viewer
oh, didn't know you could do that. 😄
I wanted a local lightweight program to quickly browser my A1111 images, so I made this. Just been refining it for a while.
man... the last couple % of an sdxl generation take foorrrevvverrrr
Now make one for linux
No even better make it chromium based
Like all "good" apps should be
no i didn't want a web one 🙂 local native executable 🙂
Write it in javascript
I could probably do linux with .net core
javascript is too loose for me, never got the hang of it
I am just trolling a bit
But like for combatibility reasons it might be a fine idea to use something along those lines
File "A:\LoRA_Easy_Training_Scripts\sd_scripts\train_network.py", line 873, in <module>
train(args)
File "A:\LoRA_Easy_Training_Scripts\sd_scripts\train_network.py", line 203, in train
train_dataset_group.cache_latents(vae, args.vae_batch_size, args.cache_latents_to_disk, accelerator.is_main_process)
File "A:\LoRA_Easy_Training_Scripts\sd_scripts\library\train_util.py", line 1422, in cache_latents
dataset.cache_latents(vae, vae_batch_size, cache_to_disk, is_main_process)
File "A:\LoRA_Easy_Training_Scripts\sd_scripts\library\train_util.py", line 770, in cache_latents
cached_latents = np.load(info.latents_npz)["arr_0"]
File "A:\LoRA_Easy_Training_Scripts\sd_scripts\venv\lib\site-packages\numpy\lib\npyio.py", line 260, in __getitem__
raise KeyError("%s is not a file in the archive" % key)
KeyError: 'arr_0 is not a file in the archive'```
What did I do wrong?
Like if you wanna use the same app on both windows Linux and androdi
I did a little bit of android programming, enough to know I didn't care for it. 🙂
unless it's in unity or something and it's all the same
Unity
but with all the unity backlash right now, I think I gotta move on from that 😐
I'm confused about it... what happened to cause this?
no idea, not a trainer
I justa makea the images...
these 3840x1600 runs take forever 😦
Turn off 8bit Adam thingy
Idk this is what first google result said
What do I replace it with?
ya'll tried stable audio yet?
But there is also like a checkbox
8bit Adam something
"""
Under training parameters and then advanced is the checkbox you are looking for.
"""
What UI is that
Oh that one
I only heard about it
Never actually saw it
Anyways
Try just using adamw instead of adamw8bit
If that was what you were using
Might just fix it?
If not then I completely have no idea
nothing to do with training I don't think. has to do with files and software
idk that's what I assume
I've never had that happen to me
lion8bit is the optimizer I use if you can't get the others to work
Why can't anyone just directly tell me what's wrong? It's always something that goes wrong, and yet I'm the only person in the entire server with this issue? U G H
I'm sorry... I just... I just want to fix this.
reinstall numpy? make sure venv is activated
There we go, 3840x1600.
it's basically just saying "I can't find this file in the archive"
I don't understand, everything installed fine when I started it...
Hold on, I'll send the config I have.
the thing with tech problems is they can have like a lot of causes, hard to know unless you reset everything and narrow it down
reset just to ensure it's not any of your files
well yeah, but it's not gonna use or recognize the error until it tries to execute it
because most of us here dont train
Someone said that it would be better to use this.
hm, I'll try and see if I get the same error
Who? I wanna know why it is better
I heard so many people talking about it
@magic wave
@magic wave please eggsplain
more specific settings. allows pausing of text encoder, use data subsets with better ui, and lora weight block training
hmm that sounds interesting, would be cool to compare messing with blocks on 128dim vs 64dim etc
also a queue for stuff which is...whats making me switch LOL. I'm at school for 4 hours might as well get something done
mostly rocket scientist stuff i think but would be useful for specific loras, like the weird ones that only apply to the eyes or a layer of an image
@valid bridge i ran into no errors, its clientside. maybe python installation, idk. if that doesnt work after a reset and reinstall try kohya
or could be triton. i did not install it when running install bat
because i dont have it on kohya and didnt see the point
Man I need some better self control
Every time I tell myself I am gonna go to slep
And then 2 hours go by
And only then I slep
Anywayd
GN
yeah I know how that goes
I was gonna fire up Starfield like 3 hours ago
and get stuck on youtube and crap
Just starting to mess around with animations with animatediff... interesting
length limit is a shame
infinite length is possible with the right repo https://github.com/neggles/animatediff-cli
oh, nice, ill scope it out
most of what I was reading said comfy supported but not auto yet
ah this is the ol' cli implementation
This is nuts ... https://www.reddit.com/r/StableDiffusion/comments/16t9xrb/chibi_girl_dance_animatediff/
I dunno how this person did it, but, wow.
Cubey over in the Touhou Project AI server is doing some amazing things with AnimateDiff
this looks like vid2vid or gif2gif
They have some much longer ones, but most of them are just a little spicier than I would post here.
not familiar with that server
It's the Waifu-Diffusion server.
nah, the quality is not already at this level sadly, but maybe in some months
i wanna learn animatediff
i figured out deforum even though i thought that was the more complicated one
after using text gen images are gonna seem fast
I want this lora
For research purposes
How does Tohsaka keep on being able to 'Rin' a lot of her fights?
https://www.youtube.com/shorts/2oa5WCUpwD8 hopefully a full version in Neuro style 
#shorts #animation #hatsunemiku
Anamanaguchi - Miku ft. Hatsune Miku
https://www.youtube.com/watch?v=NocXEwsJGOQ
Happy birthday!
Pretty sure that's their plan.
Some re-trained styles, faces are a bit wonky tho
foooound yooou
For yall that might be panicking over discord being wonky
clearly result of nuclear war

I love that xD
True and real
ow
Oh no discord banned all of my vpn IPs
Cryo why no use adetailer?
in hospital so can only use my creds on dreamstudio
I mean they invented engines
They just didn't think it was useful
So they stopped
i don't even spend time on rendering anymore, playing so much games
playing starfield and hogwarts then in no particular order I. queued tales of arise. lies of P. FF7 Remake. Persona 5, Persona 5 tactics, baldurs gate 3
Testing 123
etc
Bg3 (insert gigachad emoji)
Time to try
But first I need to reinstall an OS into my computer
What distro should I choose?
I am gonna be using kde plasma as a desktop env
I have the same, right now I left home, hope they can fix it soon for when I get bavk
Security error when you start?
That's what I got anyway, and it's not ip-bound, tried different public ip's
Said something like my ip was blocked
Unless my 4 are blocked, I'm pretty sure it's not your ip
I think some enginee done goofed, or some software went haywire
It was a ddos attack, or at least that’s what I think
So basically it glitch the system enough for them to “block” all ip in
Or something along those lines
So, I wonder how's her day.
pwease
My dream today be like
She is a hard worker.
For some unknown reason this seems like a biased suggestion
Also I am not talking about user name or computer name
Why does she look kinda sad?
Feeling smol. Might change later
banner? 
Still seeing if I get better before I start gaming
I got bored so I made a grid with my published and unpublished LoRAs
Its a 60 MB image btw, so they're all in full res if you zoom in or open it in browser xD
what model was this again?
metadata included
Lil soosi 

Hows ya 
doing great, but getting my ass beat in by Velkhana in mhw 
O, fair, I haven't fought Velkhana :O
the AT version is such an ass 
@magic wave @lyric zealot
hewo
my sd is working again
Where are the damn thick loras at 
@nova remnant Welcome back, there.
rip clothes
Howwos 
@nova remnant Here the thiccness
@lyric zealot are you using these two by any chance?
not anymore. If i wanna get similar effect, i usually regional prompter
can you share the link for regional prompter please, i'm having technical issues with those two i posted
thanks!!
cool thanks again, i might likely have few questions when installing it ... in a bit tho
sure thing sure thing
it might be awhile for each reply tho, since i'm hunting monster in mhw
lol ok, yeah i will go through the readme as well

Send me THE best one's like the REALLY 2D with minimum rendering
Color rendering
I think thats probably SwamiTsunami and MinLona, one of those is already up on my Civit page if you wanna get it
Time to research about the eu ai act
Whatchu gather from it?
I didnt
Didn't research yet
I mean I am pretty sure they updated it a couple of times
O fair :o
Like I heard about it before
Literally psychopass
@magic wave you saw psycopass?
Watched
she in a submarine or what

exactly that

@lyric zealot
how do I prompt so it looks better
(look better please:999)
face
I have best tip ever for you
Like fr
The best trick for getting better at prompting
Is to copy other people
And then change it little by little
Until it is original
Another tip is to be influenced by an existing piece of media

nice, it's working great 
thanks for the guide~
i didn't do much 
you did lol, the right direction to look into
how do i setup freeu, and is it worth it?
is it worth it tho? 
what is that for ?
I haven't rly tried it
iirc it helps with generation speed, without any downside?

is it really
or just a coincidence

so regional prompting has 75% accuracy so they say, and i notice that too
yup, it's not perfect
yeah but 75% is fair imo
but you can increase strength if you see certain prompts not working
I just used latent couple 
i tried that too, but for some reason things were glitchy
Neat visuals.
But also, I look forward to seeing adventures.
https://github.com/CodeZombie/latentcoupleregionmapper You could try this to set regions
interesting, i'll take a look
oh damn, this looks nice for the gui
latent couple i tried earlier it required composable lora too but do i need the composable lora with this one too?
Yeee, and uhh
I'm not sure if composable LoRA is updated
I've tried it
And it doesn't seem to actually divide the regions
So it just stacks them on top
thats' the thing about composable lora i installed it but didn't actually use it, i was only playing with latent couple to see if things worked out and i was getting some weird renders
i think composable lora is when you wanna use certain lora in some region but not another
oooh ok thats logical, i think i can do w/o composable lora for regions for now

but this guy on youtube about regional mapper has pretty good idea to click and draw regions which creates prompt box for each
btw this github.com/opparco/stable-diffusion-webui-two-shot/ .. the two shot plugin could be outdated or something cause the guide i was following that asks you to run a command in the console window seems to be missing a folder
unless they updated the two shot
Who is this fashionita?
I noticed that some of us are interested in exchanging Anime LoRA models for commissions.
🤝 I created this page: https://collabrain.vercel.app/ to connect us. By the way, it's free to post. 💬
this was using Attention generation mode













