#🍥|anime
1 messages · Page 46 of 1
around 200 dragon images
hehehe
yeah it does weird shit
that's why I wanted you to have it
I'm gonna go through all 220 dragon images now, remove duplicates, remove shitty ones, remove those with barely a dragon, then place the rest in a caption folder, then caption it, go through each fucking caption (annoying...I wish there was an extension or gui that goes into kohya for this)
thanks. Gonna merge the lora into the model once it's quality is high enough
guys im having an issue with impaint, i keep getting rly distorted faces...
if you're doing anime, don't use any face restoration model like gfpgan etc. Even
Even inpaint needs some context in prompt
She poisoned this didnt she
faces are wrong even if i use face restoration
either way they come up messed up
:(
then you should use fill, not original, latent etc
no I said not to use face restoration
but I dunno what you're working with, what it looks like before and after
is that the size of the image?
nono its a crop
i cant pass the image here bot doesnt allow me
its sfw but has decent booba
i guess
XD
I see
hmmm post or dm me the full image and let me try
I mean the actual image
not a screenshot of the website
I can try to fix the face and see if it's something weird or not.
@echo timber
it's in the webui, on img2img or text2img
as the result
or as the original
you just need to right click and save that
@echo timber
problem is whatever i used as impaint was a crop from the original cause i lost the original
but i think im able to generate it again i think i got the seed
I see
nice i got it
When I use extras and Batch from Directory, to upscale many images, some repeated images are created, for example, the first image is created many times.
How can i fix that
hi
hiii
hiiiii
Ihhhh
Ayeee wassup
lots of pain, then high and less pain
surgical wounds got lots of liquid, slightly infected, on anti-biotics
what anime upscaler do you guys use?
4x_fatal_Anime_500000_G or 4x-UltraSharp
wut?
?
I said hiii too
please stop arguing it's supposed to just be a funny screenshot geez
is kitchen theme working in vlad´s fork?
They were just responding to my what's up question???
yeh
Hope you get to recovering quickly btw! Majory surgery like that really sucks
it's my third major in 2.5 years, 4th in total
might be going in for more
good times
?
but thank you
Guys, while inpainting something like a hand, what do you enter on the prompt? Should I use the prompt of the whole picture or just the description of the hand?
using the web version and only have these. is that a custom one or something else?
only just started, so apologies if that's an odd question.
which is best?
lollypop

They can be found on this website https://upscale.wiki/wiki/Model_Database
oh ok, thank you. i came across it but not really sure how to add them yet
then put the downloaded file into webui/models/esrgan
thanks!

I went mega afk the past couple of days and forgot to use ur prompt

and now I'm working on the POW

good luck pls show me
am doing the move to linux dual boot today so auto1111 runs natively
I will be leaking in the dms 
#🏞|general-with-images message
This happen
you're triggering my ptsd for when I tried to do that with my rx 580
why tho?

directml mem leaks are insnae
i hear auto1111 on linux runs allot allot smoother
i have 2 drives, why not
is that on vlad´s fork the same?
so i put it in the file directory, but i dont think its showing up, maybe i need to reboot it first?
am i the only one who finds this hot?(not 2B btw)
help pls
friday
Only one way to find out. 😅
question to yall: whats your fav sampling method?
2M Karras
Yeah, 2M Karras. SDE is probably a little better on average, but it's also way slower so I use 2M Karras unless it has problems with my prompt.
I like euler a and karras
nice choices, my fav is also 2m karras. but UniPC is also pretty good
Either xformers or scaled dot, it depends on your setup so you may want to try both and compare VRAM usage and generation speed to see which is best for you.
I do like UniPC for certain textures. Anything that I want to look a little more painted usually goes through UniPC.
SURELY this cant just be random characters 🤔 anybody knows what this means
gibberish
yes you should reboot if you haven't already done so
why not the others?
What model can make hair, eyebrows, and eyelashes all the same color? Most models make the eyelashes black no matter what the hair color is.
ghostmix
I'm not familiar with Invoke's so you might want to test that too but I have a hard time believing that some WebUI developer came up with a more efficient optimization than Meta's Xformers or . But compared to the rest Xformers and scaled dot just perform the best (lowest VRAM usage and highest generation speed).
ok thx, i dont think xformers will do much as i use torch 2.0
you can see a bit of red too on the lashes. this i made in ghostmix1.2
Xformers 0.0.19 should be similar in performance to scaled dot. From what I've seen, xformers tends to be very slightly faster on 30 series GPUs and scaled dot seems to be noticeably faster on 40 series.
ok thx, ill try everyone of the out
Just make sure you restart after each time you change so it can be applied properly. System info tab can confirm which is currently applied so you can be sure.
thx👍
hey all, happy Friday
It's extension?
centaurs are so yesterday
bro wtf is Xynaptix_V-obsidian-1.0-indev-ID_001
oh wait is that your model
damn nice
hmmm my auto simultaneously says I have xformers but need to reinstall to update and when I try to reinstall it says I don't have it
did you do --reinstall-torch --reinstall-xformers?
no
try
tensorboard 2.10.1 requires google-auth-oauthlib<0.5,>=0.4.1, but you have google-auth-oauthlib 1.0.0 which is incompatible. tensorboard 2.10.1 requires protobuf<3.20,>=3.9.2, but you have protobuf 3.20.3 which is incompatible. tensorboard 2.10.1 requires tensorboard-data-server<0.7.0,>=0.6.0, but you have tensorboard-data-server 0.7.0 which is incompatible. tensorflow 2.10.1 requires protobuf<3.20,>=3.9.2, but you have protobuf 3.20.3 which is incompatible. tensorflow-intel 2.12.0 requires keras<2.13,>=2.12.0, but you have keras 2.10.0 which is incompatible. tensorflow-intel 2.12.0 requires tensorboard<2.13,>=2.12, but you have tensorboard 2.10.1 which is incompatible. tensorflow-intel 2.12.0 requires tensorflow-estimator<2.13,>=2.12.0, but you have tensorflow-estimator 2.10.0 which is incompatible. lol
when I try to update to the versions it wants me to have
it says it doesnt exist
idk then
Join Our Website: https://bit.ly/Crew_Membership
This Episode ►
Apparently, even with the billions of generated Anime girls, people still want to watch regular Anime.
when they said "that's where all the advancements happen, they happen in the anime sphere. they enter our regular human sphere afterwards", that is so true

I love corridor crew
nope, its vlad´s fork
whats nice?
za model
oh thx, do you use it?
no
didn't know what it was till now
can you try it? havent gotten much feedback or pics from other users
if you do, please read the description
do you have the link?
make this again pls btw https://civitai.com/models/7258/beta-bodypillow-style
currently not too good, needs some good ass prompting
or I will make one myself
ok, please ping me when you use it on a image, wanna see the results
i will, already got a new dataset
also what settings you use to train?
cause ctd's tutorial is so good. before, my loras were much worse.
oh wow, that's interesting. Not a bodypillow person, but I feel like that might be niche enough to start a little side gig.
Upscale enough to get a decent print and off to the races
link?
@brittle jay please ping me when you use it on a image, wanna see the results
This is focused on beginners. The training settings probably aren't 100% optimal but should do well.
Github repositories used in this guide:
1: Kohya Script: https://github.com/kohya-ss/sd-scripts
2: GUI to use Kohya Script : https://github.com/bmaltais/kohya_ss
3: Scrapper: https://github.com/sA...
it's also on projectai
ps. please use the new obsidian-v1 or whatever its called
Isn't dakimakura a thing?
guys what setting do i use on inpaint to fill the black area with an object+ background, so far im only getting object generation, is it a prompt issue?
with nixeu embeds and loras vs without them (embeds are what made the cloth weird)
I like this model
thx
never use embeds, so idk if they work good
if you dont mind, please test a bit more
funfact: the model is made out of 200 models +, because it(my model) exists since like 2022 November
name?
thx
but now tbh with you. I like your model better currently, though needs more experimentation
it just matches my aesthetic more
hello bross I'm totally newbie, how can i generate this type of images? can anyone help me?
thanks, if you dont mind testing more styles or shit i would appreciate it!
hey, welcome ❤️
do you have the webui by automatic1111 installed?
nah just joined the server
k gonna pull up the link real quick
now aurora, amazing as expected
which of the 3 is best?
you know how to use GIT and github?
yeahhh
then clone this: https://github.com/AUTOMATIC1111/stable-diffusion-webui
quality wise (how good it matches the prompt and how coherent the image is) aurora is the best, but aesthetic/feeling wise I like yours the most
ok, nice, should i merge aurora in?
try it out
it might help with the coherence and how much it matches the prompt
👍 did you read the description?
oh shit forgot
of my model
yep
let me try
yup
what was the quick setting thingy for clip skip?
it depend tho, sometimes it works better, sometimes not
you have vladmandic fork?
ok, just change it in the setting manually
nope, vlad´s got it in here:
k
CLIP_stop_at_last_layers,
yeah yeah found alreadcy
sd_model_checkpoint, sd_vae, CLIP_stop_at_last_layers,
sd_model_checkpoint,sd_vae,img2img_color_correction,always_discard_next_to_last_sigma,CLIP_stop_at_last_layers
flexing
clip skip 3
a lot more similar to my model
and a tiny bit more coherent (subjective)
but I like clip skip 2 I think based on this single experiment (way too little to judge fr)
idk, you have to try with different prompts and settings to say which fits best
(if you do test more please let me see the results if it doesn't bother you 🙂 )
yes
k
this with clip skip 3
Guys I'm rly struggling with inpainting hands, any tips? I'm kinda new to this, managed to fix face and a couple things but hands are so hard
upscale then inpaint only masked at like 768 x 768
is it good?
I like clip skip 2 better
compresses?
filled with fluids of various colors that they absorb like sponges
yeah that I have after they sliced me open
groin to rib, navel to my left side by 20-25 cms
oh
aka dressings or whatever
and I thought you meant jpeg and winrar
mmm no
I can't rar this surgical wound
if I could rar the stuff coming out I would
I will take 2-3 weeks of it or however long it keeps going, compress it into a tiny tiny dice sized bottle
and mail it to someone
to someone?
who would you send it to?
you gonna test more?
someone
I guess I'll roll 5 10sided dices
to help me decide
or 12 sided, or 16
or 36
1
d1, d2, d3, d4, d5, d6, d7, d8, d9, d10
http://www.CuriosityBox.com CODE "1"
but seriously, I had a high absorbent compress + high absorbent dressing and after one night, both were soaked in light yellowish liquid (I have an infection but it's free flowing yellow so it's not like it's a very serious infection, they put me on antibiotics)
Are there models that can make the hair and eyebrows and eyelashes all the same color? One that is like 2d cel shaded style, not 3d or 2.5d
2d style like dark sushi, where the shading is not gradient
I dunno what dark sushi is
yeah but people still imitate it
because people like how it looks
Models I see make eyelashes black regardless of the hair color
then why not just say "celshaded art style" and "2d illustration"
colored eyelashes?
Its not so much that I'm trying to make it cel shaded. Just not gradient shadows
search more model on civitai then
or make one
not that hard
K. Maybe I'll just manually color parts though
I wish I was good at drawing
but way too lazy to actually get good
and busy with school and stuff
stop the talking boys, we need waifu material!
the best way to get better at drawing is to have fun, so don't listen to those that treat "START DRAWING CIRCLES AND SQUARE AND TRIANGLES FOR 72 HOURS AND THEN YOU CAN MOVE ONTO 3D VERSIONS!"
literally the reason I stopped trying to learn how to draw
that shit seemed so boring that I gave up
no no no, take the theory from it. Understand it and practice the same thing however you like
the whole point of it is to see the 3d image in a 2d image
and learn how everything is made of building blocks that begin as shapes
torso is a rectangle with two sideways pyramids for lats
I mean, nevermind that, I am too lazy to draw in general outside of like school classes
^^
btw has anyone ever seen Vin Diesel emote anything except "raised eyebrows" "sideway smirk" "angwy face" and "yeah I'm tough" face?
2 vs 3
It was so long ago that I don't even want to go back to it. It'll probably take a few years for me to get decent enough
here's an idea, learn oil painting by shading and color mixing (digitally) and use generated images from SD turned into line-art to color
what upscaler you used?
eventually you can draw the whole image without it
looks a bit wonky
4x Ultra Sharp
through extras?
4x Ultrasharp on hiresfix
ESGRAN 4X +Anime on extra
why do song lyrics make the booba bigger?
and on multidiffusion?
can you try same image on extras but with ultrasharp instead
haven't messed with that yet so probably default?
k
hold on
cryo are you okay?
you've been typing for a while now lol
what are the lyrics?
check png info
i never had them
Not on mobile
here's a song lyric I will write right now out of my ass. Have fun prompting them
`Stepped off the boat in land
Yet I landed in water
Brought the umbrella with me again
Yet it wouldn't even open
I put the seat belt on and the airbags work
But the chair dislodged with me
Here I go again
I bet that mosquito has Malaria
I took a step to my left and the universe
It conjured a needle
I drank a glass of water in the desert and choked on
An ice cube?
Then I went to have some fun at the theme park
Went on the horror ride...
The cage electrified
But that wasn't the problem
The whole place caught on fire
I'm burning
I'm burning
The firefighters sprayed it down
Now I'm drowning again
Again...Again
and again
So I tried to shoot myself but instead
The bullet flew into outer space and killed an alien
Now the world is burning and everyone' s on fire
Now the world is flooding and everyone's drowning
They dropped needles well pins
As tall as four humans and they impaled the same number
Now everyone is dead and I'm left alone here
I tried to get the aliens to kill me
But they said
"bxtr kliakka burk"
So I guess that's that`
what you guys think about the colours/style?
can't find the original image, probably deleted it
f
looks a bit burned
what was cfg?
9
weird
lol
what lora?
thats with cfg 7
"bxtr kliakka burk"
ez
with MY 😄 its not published i just testing it
less burned but idk
yeah something is wrong with it
@brittle jay is our burn detector
maybe overcooked or a problem with some settings? can you try without it?
hehe
how to solve overcooking while training? its my first lora i try to do
last time i was to close to the pictures
but qulaity was better
I would just follow ctd-sama's guide
that was 1 try
This is focused on beginners. The training settings probably aren't 100% optimal but should do well.
Github repositories used in this guide:
1: Kohya Script: https://github.com/kohya-ss/sd-scripts
2: GUI to use Kohya Script : https://github.com/bmaltais/kohya_ss
3: Scrapper: https://github.com/sA...
this looks fine
the problem was the basic skin was completely broken
Wish I could understand AI language 😔
Minion language 
awesome
guys does anyone know about stable diffusion?
what model?
I have problems with stable diffusion
Whats that style 
What problems you got?
new lora I trained
Oooo thats nice 
first image is my model, second and third are yours
oof
try freeing up some space on your drive
but I have alot of spaces
what settings you using?
idk then
#🏞|general-with-images message
How can i fix this?
Could be not enough VRAM by a bit
in another world this would probably be perfect fingers
which one do you like best?
I should get stronger gpu ?
its not a vram issue(i think)
it could be ram
or stronger computer ?
normal ram
idk, this is just to test my lora
\
I mean, I have 16 GB ram so I doubt its that
i have 16 too and run out frequently
Do you?
so what should I do?
Which commandline_args are you running?
look at task manager
Whats the issue?
how much ram do you have free?
k idk then
When I upscale many images, there is one image that repeats, as shown in the image I sent
probably not
yes, but its bad
since they are online
yes and not completely free
there's Dall-E and Mid Journey (expensive if you want privacy)
then how I can find a solution? ofc the staff knows because they made the program
mid journey isn't free
they didnt make it
and midjourney has an entire dictionary of censored words
I want free things cause I don't have money
yikes
but I think they know better
idk, the program you are using was(probably) made by automatic1111
Wait... You have a 16xx series right? Are you using an fp16 model? I'm wondering if Auto1111's "fix" to the issue of 16 series cards not being able to run fp16 was to just fall back to CPU. That could explain why you're getting a cpu torch error. It should be a CUDA Out of Memory error.
try asking him
I didn't use automatic 1111
the lyrics I pulled out of my ass are giving me this
what did you use?
16xx what is that?
pretty cool
your gpu
I use the model meina
this is automatic 1111
Any nvidia gpu model that starts with 16 followed by 2 more numbers such as GTX 1660
what do you use then?
this is automatic1111
I use the normal method to install stable diffusion
THIS IS AUTOMATIC1111
as i said 3 times now
i use 100 repeats what how many do you use?
how?
It's definitely Automatic1111's WebUI or some fork of it. It's easy to tell by the format of the webuui-user.bat that you screenshotted.
anyone know how can i solve this?
is there a way to fix the problem?
it looks I don't know
as i said, a.s.k automatic1111
he made it
so he will know what to do
He's the developer that made the most popular webui. I'm looking right now for the fix that he did for 16xx cards to try to see what the fix might be.
(probably)
is there a way to talk with him?
yes, open a issue on github or go on his discord
You using a LoRA?
can you give me his discord? or his github?
@fallen sundial when you launch SD, do you receive any errors in the commandline?
Yes, but i am upscaling
I mean, it looks as if your seed was fixed
@supple raptor
xformers isn't found which definitely isn't helping. You should delete the venv and repositories folders and relaunch SD. This will cause it rebuild the python virtual environment and will hopefully fix your problems.
and do you know how I can fix it?
But I am using loopback wave to create videos, wouldn't it be better to be seed -1 instead of a fixed one? Besides, my problem is with upscaling, not with generating
what's the command for starting a keyword and ending it at a specific %? As well as starting a keyword at a specific %? Without having to connect the two
was it like [keyword:0.50] for starting and [keyword::0.40] for ending?
Prompt editing allows you to start sampling one picture, but in the middle swap to something else. The base syntax for this is:
[from:to:when]
Where from and to are arbitrary texts, and when is a number that defines how late in the sampling cycle should the switch be made. The later it is, the less power the model has to draw the to text in place of from text. If when is a number between 0 and 1, it's a fraction of the number of steps after which to make the switch. If it's an integer greater than zero, it's just the step after which to make the switch.Nesting one prompt editing inside another does work.
Additionally:
[to:when] - adds to to the prompt after a fixed number of steps (when)
[from::when] - removes from from the prompt after a fixed number of steps (when)
I would recommend that everyone who uses Auto1111 read the features page and bookmark it to refer back to https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features
aha
🐱
I missed it again cause it was just two lines
It's got a full section but you definitely get a free pass at the moment considering I'm assuming you're probably on some painkillers.
sure, my brain also barely works cause I barely slept a night for the whole week in the hospital
but what I meant was, the part where you can use a keyword(s) without another keyword(s), was just two lines
Ah, fair enough
- use an anime model
- save to anything other than jpeg lol
- bruh
I didn't understand what exactly you want.
Is what you say within the laws?
also save to png or something but not to jpeg
For accuracy?
png is great
hey Smol, who is that? You've been generating her forever
who lady?
should I release?
whose?

It's my waifu 
ah
I should lie down
let the pain subside so I dont need another handful of painkillers
oh "subside", nice word
yup it's gonna be my OC 
nope, just using the model on it's own
nope, the image i posted should have png info embedded inside
I have prepared dragons to train, I just need to caption them properly and regularize some images
go slep already you buffoon
no!
idc wha you said
I sleep when I sleep
upscale
we have this commercial for 3 (mobile isp) with a girl watching her phone as it vibrates with person after person apparently bullying and insulting her. Then liking the tweets. She refuses to give up the phone, screaming "YOU DON'T GET IT!!!!!!!!" and runs to her room where she starts crying and sobbing while she keeps looking at the people insulting her online....Kid, put the damn phone down you halfwit
post-modern society be like
yes and do not think it either
literally 1984
our meticulous system will detect slightest contention and it will be dealt with swiftly. You will be saved. Sent to our re-education centers where our expert staff will mentally and probably physically torture you until you agree with everyone again. Have a nice day
I leave you all with chibis
good night
Me
who dat?
Tried Synergix + Aurora at 0.5 weighted sum
With my normal negatives vs synergix negatives, so the negatives seem to be contributing a lot to the style :o
Must upscale to at least least 16x from the render resolution 
You want to gift me better gpu?
Multidiffusion with tiled diffusion can probably do it for you. You might just be waiting for an hour or two 😅
Lets see what happens
I wasn't planning to stay up past 3am
Why not use SD upscale :o
Probably not any faster
I'm starting to feel like this isn't going to happen and if it worked I think discord wont like it
I mean, my 58 MB pic was blocked by discord, as in, it made it crash and there was no way to see it xD
I would recommend cancelling it. I did one yesterday and it took 38 minutes and looked considerably worse than the 2x version.
If you actually want super high resolution, I've found it usually works better for me to do multiple 2x upscales instead of just cranking the multiplier up.
It's taking its time to cancel
That's normal with really big stuff like that.
ask eface
he be doing it in a couple o mins
you should be more active man
That looks familiar
It will cancel soon™️
@gloomy scroll every time someone uses one of the stickers they made
I wanna train this style now
your creation?
Yeah
pat pat
kawaii
Challenger approaches
a pokemon fight?
2b: senses cuteness
I choose Slicat
Need a cup?
lool these are amazing
I remember this.
You've acknowledged the cute and acquired a yandere
You've forced me to bring out the Coffee Cannon!
You wouldn't shoot a demon 
I don't have had pat sticker
I made this with online tool
I have also Bocchi
I need to get that headpat lora. I forgor to grab it.
@supple raptor I feel like that upscaling messed up something
Try this button
Restarting already
That should do it too
Cheekpinch LoRA where??
@jovial dock please bestow this power upon me as well 
Can't get it to work well
Oh right. Rink only hangs out in the touhou server.
Discord also shows that they are offline so may not answer because of that
doggo
isnt he cute, has a heart
Before I go to sleep, I challenge all of you to prompt this
Strong, Demon, Weak, Human, Stronger Dragon, Weaker Demon, Dragon, Demands, Demon Duel Human, Human accept, Demon accept, Human fight Demon, Demon fight Human, blood gushes and sprays, just like in anime, They arrive in the attachment to this letter
who wouldnt want a bioluminescent dog
There are illegal underground labs that do that
intresting prompt 😛 trying to fix it so it dont generate a waifu
fix my prompt 😛
oh i think i got it, no waifu
That my model still?
YES haha
yeah my model knows how to fight
i'm trying a few models
prompt: gud prompt
neg prompt: bad stuff
the battle continues
not your model
which one?
my model is well, very colorful 😛, but is generating the monster 😛
cause it got that hovering hand
Aurora with Synergix i think
ah
niji??
are the fight templates any good in it?
it was frightening when I generated this because it started from his tits and only then decided to make it into a male and make the tits into muscles
Nyan generate these tits
anyways
that just tells you how much mem you have in reserve
doesnt actually do anything from what ive seen
unless im wrong, please explain how it works
Ah, you're probably right. I was just guessing.
I think it just unloads the thingy
oh maybe it just tells you how much free vram you have
you should delete that just to be safe
The colorful version
Strong, Demon, Weak, Human, Stronger Dragon, Weaker Demon, Dragon, Demands, Demon Duel Human, Human accept, Demon accept, Human fight Demon, Demon fight Human, blood gushes and sprays, just like in anime, They arrive in the attachment to this letter
obv
Yep, you're right. I just tried it. It just lists current available vram and total vram.
Is it too late to tell you about running Ubuntu inside Windows using WSL?
Dual boot is probably more stable tbh, but yeah, if you want to do windows stuff while gening that would be another option. That's how I run Oobabooga.
yep. this is what I use https://github.com/oobabooga/text-generation-webui
Youre beautiful
For now i take break. Been setting this up since 11 and its 630
Am happy i got it working with my anxious limited brain power 
guys, what is a femboy who is a girl?
dang, the add-detail lora is pretty damn good
what in the contradiction?
on amd?
Ayyy 
Yes
Ooooh he cute
add-detail lora. get it. best thing ever 😮
only changed the strenghts on that thing
Where 
HF: https://huggingface.co/OedoSoldier/detail-tweaker-lora This is a LoRA for enhancing/diminishing detail while keeping the overall style/characte...
O wao 
Less color version Strong, Demon, Weak, Human, Stronger Dragon, Weaker Demon, Dragon, Demands, Demon Duel Human, Human accept, Demon accept, Human fight Demon, Demon fight Human, blood gushes and sprays, just like in anime, They arrive in the attachment to this letter
is this the prompt? if so, i'll need to run it for science
Hmm... I'm not seeing anything about installing for AMD on any of the 3 trainers I've used.
@magic wave which repo is it that you use for LoRA training again?
so guessing gotta get a nvidia for sure or colab if i wanna lora train on amd huh
well.. i dunno what i expected when i prompt it. :itssomething:
Strong, Demon, Weak, Human, Stronger Dragon, Weaker Demon, Dragon, Demands, Demon Duel Human, Human accept, Demon accept, Human fight Demon, Demon fight Human, blood gushes and sprays, just like in anime, They arrive in the attachment to this letter
mecha demons?
played with this a moment ago. kind of hit or miss depending on what you're making, seems to struggle with darker images. generally aesthetically prefer just using <neg-sketch-2> / <neg-anime> and moving on
works pretty well for my model though, trying it out in various situations, and only one prompt seems to have disintegrated it
nr5 just can't.
Yeah I don't see anything for AMD in the one that Smile linked either. It's possible that it might just work if you replace the torch+cuda that is probably installed by default with torch+rocm instead if you want to try, I believe you would just need to activate the venv and run the following commands:
pip uninstall torch torchvision torchaudio
and then
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm5.4.2
I'm not sure if there are separate versions of bitsandbytes or anything else for AMD or if the trainers are even able to utilize rocm.
i am screenshotting for tomorrow, i am done setting stuff up for the night xD
took me and aryetis a good minute to get torch 2.0 working
Definitely fair. Took me quite a while the first time I had to manually install Oobabooga.
took me more time to install it with the one click installer
The one click installer not working is what took up a fair amount of that time.
Allegedly the oneclick's have been reworked and are supposed to be much better now, but now that I know how to do it manually I'd rather just do that.
spent days cause for some reason it defaulted to micromamba instead of miniconda. I don't know what neither of those are but micromamba didn't work for me
What negatives yall use nowadays? :o
AuroraNegative negative_hand-neg bad-hands-5 bad-image-v2-39000
Where you get bad-hands-5? 
idk
I think nerfy
Yeah, I don't think bad-hands-5 is from Nerf.
im tryin an4. what are your thoughts?
what's an4?
Oh, haven't tried it. The previews make it look like it heavily influences the style and composition though, which I'd prefer for negative embeddings not to do. Their comparisons in the previews are kind of garbage since they just compare nothing to the full strength embedding. They should have included at least a basic negative prompt and shown the embedding with and without that because adding almost any basic negative prompt can have a huge impact on quality on it's own.
well i go back to my ol negatives then
It very well could be good. I just mean to say that it's really tough to judge based on the comparisons they provided.
im having so much trouble with inpainting new hands, they come rly deformed and they dont blend in with the background, any help?

Levitating donut
idk, 90% of the hands i generate have less than 5 fingers
and im using embeddings for hands
i dont understand
WHAT IS THIS SORCERY
https://civitai.com/models/58390
HF: https://huggingface.co/OedoSoldier/detail-tweaker-lora This is a LoRA for enhancing/diminishing detail while keeping the overall style/characte...
im trying and its awesome!
i always put this one in as well and get good results usually: (extra digit and hands and fingers and legs and arms:1.4)
O I should prolly look at comfyui
these faces are what i live for
why are my prompts not working for hand impainting
, I clearly state everything I want on prompt and everything I don't on negative... I keep getting 3 fingers...
did it really take that long? damn
Oh no I skipped the weird one lol
That's just pretty random output since I only used this as prompt (chibi:1.3), cute aesthetic illustration, stylized digital art, abstract and colorful style, happy, outlines,
If this is a link for you: #1100170312106127410 –– you can go generate with SDXL!
@real nexus I did it I clicked download button
Previously I used only my old mix
So cute
Emote maker lora?
Mmhmmm horrid's
CUTE
These were without Emote maker lora
pretty messy, trying out a new model
Time to test detail tweaker lora with negatives value
i turned Jaime Pressly into a comic
That's very cute !
Hand inpainting can be tricky because you're generally starting from a bad hand to begin with. It can help quite a bit to mask about halfway up the elbow to make sure that it has enough context to make a hand. In most cases, I just do a quick paintover in Krita to fix the hand as well as I can on my own, or at least define the shapes of it, and then inpaint over that to blend it better into the style of the rest of the image. You could also experiment with using controlnet and a hand depthmap to try to make the inpainting easier, but that isn't something I've personally tried.
decent model, had to turn down some weights tho
That's gorgeous thanks for using it
Maybe I should start doing list of emotes for myself
Oh gosh it's stinking adorable
I just realized that I forgot to leave a review, so you've got another review now
Thank you
I would give her all my money
Really nice and thank you ! 💜
got any embeddings for hand inpainting? cause mine are either not working or really bad, i can't get a single hand with 5 fingers...
Nerf's negative_hand embedding helps for general hand generating, it should help some with hands too.
If that gives you best results but nothing says you can't add anything else
I use it with a full negative prompt along with other normal negative embeddings. My base negative prompt is negative_hand, (KHFB, AuroraNegative:1.1), (Worst Quality, Low Quality:1.1), border, (nsfw:1.2), skimpy, grayscale, (watermark:1.1) and I just add to the end of that to counter any problems that I'm getting from a specific prompt/model.
KHFB?
its just so weird that none of my negative embeddings are working
One of Luna's negative embeddings that she made for Aurora. It's in the HF repo. https://huggingface.co/SweetLuna/Aurora
no matter what i do i keep getting less than 5 fingers
https://civitai.com/models/47085/envybetterhands-locon I just found this one myself
I think they just mean that it's not working in the sense that they aren't getting the result they want, not necessarily that they aren't applying.
Im using lots of embeddings, negative prompts, specifying I want 5 fingers, specifying on negative that I don't want more than 5 fingers etc....
even when 1 single embedding still doesnt work
Honestly, you're probably overprompting. In my experience SD doesn't really understand concepts like quantities of things outside of specific tags that it's been trained on like 1girl and 2girls. I've never found things like 5 fingers in the prompt or more than 5 fingers in the negative to help at all, and those sort of things usually make my generations worse if anything.
Can you see those embeddings in this list?
ive been told that empty prompt might also work, but that doesnt work either
idk
@echo timber You could also try using some other model like Aurora for inpainting hands and then reinpaint hands with low denoising to fix the style
Sort of. When you train an embedding, it is finding the unet weights that are most closely associated with the training images and compiling them into basically a list that is then associated with the filename of the embedding. Then the embeddings are loaded alongside the text encoder of the model so that when the filename is used, it will reference those unet weights.
I see okay
What's the hand look like to start with?
what does SLE mean in aurora's trigger words?
Yeah, I wouldn't expect that to inpaint very well. I wouldn't even bother trying it myself without manually editing the image first. It would take too high of denoising to fix the deformities. Open the image in Krita and try to paint in the missing section of the hand between the index and middle finger as well as you can and then inpaint over that.
I forget what it stands for, but basically it's meant to give higher detailed eyes.
why didn't i know about this before
detailed eyes was something that i really wanted
You know about mksks style and detailed background right?
yep
one guy that was helping me earlier managed to somehow do the inpainting, it was decent, but im trying to do it myself with same settings and its not working...
he did fix them quite decently
not sure what mksks style does tho
Both are activation tokens from dpep. They're meant to encourage more detailed backgrounds, though you do need to define the background in your prompt for them to work and they can hurt image quality if the background isn't defined well enough.
i didnt know about mksks style either
SweetLuna's Eyes
Lora trained on FeverCell's eyes
Sadly no independent release yet
There we go. I knew the E was eyes, but my brain was too smooth to remember it was just her HF username for the beginning.
These are some adorable Chibis
Not using your lora atm
Very cute Chibis tho 💜
dies of cute
She's going to pinch cheeks
Gobbo. Haven't fully fleshed out the backstory yet, but she was isekai'ed into the real world and now she works at a maid cafe.
That's super cute
Pls stop it
can't stop, won't stop








