#🍥|anime
1 messages · Page 140 of 1
Hm
I mean aurora could be fine
There are probably better models for this style
But I am not that aware of em
Aurora is good enough
.
Download
where do i input it ?
You don't
It will only confuse it
I mean you could
In the img2img tab on the webui
We could try that later
can you tell me more about that ?
im on that tab right now
its ok
Can i ask how to install stable diffusion on pc or smartphone? Cause iam new in this thing
Say goodbye to Midjourney and hello to the future of free open-source AI image generation: SDXL 1.0! This new, uncensored model is taking the AI world by storm, offering unparalleled image quality for FREE. So in this video, I'll show you how to download the model and use it on your own computer locally for free!
What do you think of SDXL 1.0?...
only pc
lol...a smartphone...I'd love to see one capable of doing that
I mean you can technically
just 20 minutes minimum per image
If not way more
ye
Yes ,but how ?
there are some user interfaces you could use
I suggest you to start with this
this most popular
Is up there was the link?
bruh
yeah
follow the instructions on your pc
we call it A1111 or webui
there are instructionss there
Sorry iam at my phone now ,do i need to intal that on my pc or that just like a browser?
only on your pc
this only workss on pcs
if you wanna run stable diffusion on your machines
it has to be on your computer
Yeah i know that ,my question was do i need to install it or that just like a web browser?
yes you have to install it, but you operate it through browser
it's a locally hosted website
you can share it with your phone if you want
you don't need internet for it though
and operate it on your phone
but it will have to run on a computer
and also you need a GPU (graphics card) for it
What the minimum gpu ?
this is literally the best I could and it's still hella bad
if you want it to actuially be similar
you would have to work for that
thanks for the effort brother ❤️
yeah it doesnt look like that
but anyway thnx
this is what it looks like if you use it as reference (normal img2img)
it just confused it
that's why you need controlnet
This is a beautiful pic.
this is bad? i totally love it
oh, the shirt
Man the more I watch lain the more creepy it gets
Like exponentially
what is autumnorange-768-000005
upscaled
it is a sad world we live in where you cannot remove SD artifacts
automatically
What artifacts?
the whole looks
the texture, the tone the quality
SDXL is not meant to do anime
even those trained on SDXL are pretty shabby
btw there has to be some technical method to perfect the fingers
the spaghetti fingers are such a mess
beautiful figure, beautiful faces but messed up fingers

Hands are hard
And so are many other things
its a shortcoming of the technical approach
Not even dalle does them perfectly every time
smol sus often makes cute fingers, so its not impossible
can confirm, i often get 4 or 6 finger gens still like top left there xD
i like sdxl for anime, this was pure sdxl no other model
and this
but it's really freakin' hard to prompt SDXL for anime
and for every 1 good pic u get thre's like 40 that are bad as far as anime goes
ive tried anime with sdxl it sucks big time
the good anime gens of sdxl are good though. but too far and few inbetween
sdxl anime not cute enough not realistic enough
even the colors for SDXL anime feels artificial
well, i didn't use it that much for anime for a reason, it made good skeletons to pass into other models though
it was sort of like a controlnet for anime gens
i get it but SDXL is another flop in the SD versions
Good
far from flop, it does realism stuff great, although dall-e 3 surpasses it in nearly every aspect except censorship
for anime it's a flop though yeah
i was really surprised with what it could do with realism
not just a factor of quality in realism but also you have to consider it's usability
like this was sdxl
it was usable for everything but anime xD
but since i primarily do anime stuff
i stopped using it pretty early on
yeah me too
A tensor with all NaNs was produced in VAE.
hate this so much
i played with it enough with honest curiosity but didnt like it enough to keep using it
@stable crane helloo

hiyoo
but you know, i love SD 1.5 community models despite the limitations in the base model;
SD 1.5 is far less resource hog, can create sweet images and its generally more accesible with fair bit of vram
Fun image from dalle
Praise the cat girl
Stop using NAI's broken VAE and the various renamings of it and you won't see that error any more. Try this one. https://huggingface.co/nubby/kl-f8-anime2-blessed/blob/main/WD1-4-kl-f8-anime2-bless09.ckpt
@nova remnant dalle is good with context but sd 1.5 has sweet spot
I'll give it a shot
which is latest nubby?
Dalle is too just censored
I bet even the images are censored to not look like anime
In that same repo are a few other versions with different adjustments to the contrast. 08 version may be closer in color to NAI VAE if you want the muted color look.
found a way around the chest description ban... pettanko works fine XD
cuz dragon girls get genned with way too big chests without pettanko
What is pettanko
what is pettanko
Lmao
Those are 105% and 110% contrast adjustment compared to the original WD VAE. The one I linked is 90% which I think is the sweet spot for most uses.
flat chest, but in japanese.
awesome thanks
with how they ban chest descriptions i was surprised that pettanko just works
probs has to be reall ypopular word
pettanko is used in pop culture often
i don't think they trained it on image pairs of other languages and that's why they say en only
is this considered 3d anime or hyper realism?
closer to realism
I wouldn't describe it as anime
i agree
Dog seems to always pop up later into the gen, but the other one that's actually a warning usually pops up right away. Seems like they do Hires fix and check after both the low and high res passes and the dog means that filter rejected it.
yeah, there's the delayed dog or the instant word-ban
did I mention seaweed is banned?
probably because of 420 weed
this vae might be baller, thanks @supple raptor
Oh noes someone might make Mickey Mouse getting stoned
better ban the word before Disney sends a hit squad.
I imagine it will be less bad. ChatGPT filter seems significantly less strict when using API at least.
true
What is that model 
it's dalle3 + aurora1
Ohhh damn, fair then 
the dalle3 is doing most of the work
just compared aurora and lottamix and they're very similar
Fair 
I upscale them at about 0.4~0.5 denoising
with the same prompt from dalle, but also I add lora:add_detail:0.3
aurora is a tad less cutesy
I really wish they would enable different aspect ratios. Square images often just don't look right.
I think you replied to the wrong message
I am talking about the artifacts almosst all sd images have
in general
gpt4v has diff ratios I think
back in the day in dalle2, when they limited the number of gens, you could alter the network request and get 4 gens again
client-side only limit lol
that lasted a few days
looks normal from far away but if you look close (like in this image) you can clearly see artifacts
is there any way i can set thumbnaisl for the check points?
Yes
there is a wayy
how?
I just don't know how
Put an image in the same folder as the model with the same filename as the model.
nubby to save the day
okay sweet, let me try
You can do it in the UI somehow too I think, but I'm not sure how.
Hover over checkpoint and press top right icon there should be replace preview image option I believe
wow that worked beautifully thanks
Same works for all the extra network types too.
okay!
Don't mind the errors
I use a little dropdown on the top-left for models
just upscale to 16k and go back a bit
Clueless
yess
16k... for when the wallpaper you wanted was for your bedroom wall. 
I need to try upscaling one of these for print quality
I found one last night I'd hang on my wall
it's not even too lewd
this one
I mean that will surely help with those super fine details :p
Time to gen anime girl laying down upscale it to 16k and then print it and put it on wall behind my bed so I can lay down with my waifu
protects from bullets by squishing
I am a bullet now
I'm always impressed with its ability to work with materials
I'm running out of beach creatures though
Have you done shrimp yet?
badass
I did crabs
but figured shrimp and crab are basically the same
maybe I'll try shrimp and see what dalle thinks
🙂
try a big fish like shark
nice I'll do shark after shrimp
Halibut
Looks like a giant shrimp is trying to eat her
Can it do hammerhead shark?
this is what I got asking for hammerhead, I will keep trying though 😄
hm I think dalle dying
can't click button, I'm in this weird state
background job died?
welp
that was happening frequently about a week ago, fix was to click here at the top, then you can go back to creations and the generate button would work.
my browser tab while exploring models
ty, did not work though

yeah it does not know hammerhead sharks @supple raptor sadly
that'd be sick though
or maybe it does but can't cope with the context
rip. no hammerhead armor
there is a problem with that method but it will be hard for me to explain it not visually
the problem: you want perfection
Basically, let's say an image has 2 types of artifacts, big, and small. The big ones are easy to get rid of using inpainting or img2img and the smaller artifacts are easy to get rid of using upscaling. But now the problem is that when you get rid of those artifacts whether big or small, they produce more relativly big or small artifacts which you then have to inpaint+upscale to get rid of and this repeats on forever.
I am bad at explaingin I thinkk
fixing fingers will be considered a technological leap and a breakthrough for Stablity foundation
until then nearly 30% or more of the 100 images are wasted
one hack is to not show the fingers at all
Ig less prompt = better images? Gen some more with "The Frieber" lol
some models overfit well
and some models are better if u describe everything u can
will be a great day when fingers r no longer an issue
Ig, it will be at least get improved
Can't imagine that they fix it for all n' end
some models are trained to generate fingers somewhat better than others
so i guess there is a possibly community models can do better than base models
but overall this area needs attention
ik dall-e had a focus on fingers too, dunno if that was only for real or anime too, a lot of the time the fingers come out pretty good, but the times it doesn't in a pic with good composition is pain xD
omg... when dall2 first came out faces were horrid
ye
lol
dall-e 2 was pretty bad at a lot of stuff
but it had better understanding of words still
but it'd never come out anywhere near the quality of a finetuned 1.5 or sdxl gen
i mean i was excited to try out dalle given that was the first txt2img ... but after a few render i kinda felt discouraged to continue
i never got picked for dall-e 2 but i have like all the perks with 3/cocreator xD so my first hands on was with sd 1.4
im hoping the next version would address a lot of the resources demand and render quality from the current sdxl and popular 1.5
ahh
How do you guys train models, with loras?
ا
dalle 3 made a huge progress for sure
the stronger point for it is the contextual outputs
but somehow i still feel attached to sd 1.5 and the control i have using it with a1111
i prefer not having to mess around with loras & such, as far as a general model goes i don't think there's anything better right now.
plus, the finetuned gens based off like NAI all look pretty similar to me
i dont even use LoRAs
i basically play with models and prompts and sometimes use controlnet
You are not a real diffuser
what do you mean
well i focus on models
You know LoRAs are just parts of models right
i could do that but i think there is lot to bring out using just a model and still have decent outputs
I mean not exactly but yeah
But wdym just a model?
LoRA are like icing on the model
Models are usually merges which use multiple models
So practically LoRAs are just merging on the spot
there's less control when it comes to cfg/weights/etc and all that. but as far as actual prompt accuracy goes, dall-e 3 is suchhh a huge step forward
and it was at this moment i realized i wrote red eyes instead of blue eyes
There is not a difference between a model that is trained normally and a model that was merged with a lora, except the model that was trained normally might be slightly better
And if you use some lycoris, the lora might be even better
It is just a feeling
You wouldn't know if what I gave you was a model that was merged with LoRAs or not
I mean probably a lot of the models you are using were already merged with LoRAs
thats not the point im making tho, i mean i specifically used models and then applied LoRA for visble difference but i prefer just the models how they are trained and what i can do with them
As I said, some of those models you like probably merged with a couple of loras
Somewhere down the line
Do you know how LoRAs work
not in technical terms but i can use them
and i get what you are saying, many of models that are trained used LoRA with them
Basically you take a model, slice it up, train the slices, then you save the difference between the original model and the trained slices and now you have a lora
So when you apply a lora, it is the same as merging two models
Ok IC
but usually i go for model that looks well trained
and those are probably merged with LoRA
i just dont use LoRA on my end specifically for rendering
Anyways I recommend you at least use some noise offset lora
In case your model doesn't have it built in
any recommendation for a particular LoRA?
ok thank will take a look
gonna try it now
its a useful LoRA i think
unlike most others that are candy
used it with a weight of 0.8
this is cool
yes, just meant to fix a problem with how stable diffusion 1.x is trained
noise offset that is
yeah i can see its use case often
using weight 1 .... for 2 LoRAs lora:epi_noiseoffset2:1, lora:more_details:1
lol very strong difference in the output as you can see
what does this do exactly?
I don't like add detail with positive weight
i could use 0.6 for that
it just forces its style too much for my taste
even at like 0.25
imo it's overtrained
But that's just me
but why would you put that in negative?
noise offset - what does that do?
Read the blog I sent
anyone ever play the Shantae games?
Has this woman been in dancers' outfits or bikinis? o.O;
i dunno lol, she is out in a jungle with dinos
nite
with detail lora going above 0.4 is bit harsh
its probably better to use a weight of 0.3
0.4
her skin texture is much softer that way
add detail is nice
I just throw it at like 0.2 on stuff
im currently using 0.4
I like just enough so stuff don't look too anime
Hair is a good example
these are the 3 LoRAs im using now lora:epi_noiseoffset2:0.6, lora:more_details:0.4, lora:SimplePositive_v1_AutoRunMech:0.8
the last one in that line does finger correction
It tried to

yeah ... not 100% guaranteed lol but slightly better
Now do the dino fingers correctly

How does this noise offset work
If you increase the image becomes darker?
basically adds contrast and depth to the image tone
yeah kinda
Can you do -1 to 1
also depends on the kind of lighting you are prompting

i have that to 0.6... i'll do one with 1
gonna change the image to night as well
this is 0.6
this is 1.5
👍 booba all around
you mean tank top all around
lol
Oh yeah definitely
The foliage is also nice
Hmhm
but yeah you can see the difference
it adds contrast
1.5 is bit too strong obviously
It looks good
Like
""art""
Looks less natural and stuff but a good look
water looks good tho
The blue looks nice
but her texture looks dry
Can you go negative with it?
i'll try one with 1
yeah let me do 1 first
Were those the same seed?
1 she is not dry
no
random seeds
but that shouldnt matter
Oki
going for -1
-1
im not sure if negative values take effect or just not in effect
Well try -1.5

It doesn't look like normal night lighting
ok i pushed it to -1.8/
that will also impact image tone along with lighting maybe
-1.8
i dunno maybe its not taking effect?

Maybe you're right
I need more mouse girls
the negative value is just putting the lora out of use

What if 0.1 then
probably close to default
so far with 0.6 its a good balance
Hmm

So have things improved?
Human positions refer to the different physical configurations that the human body can take.
There are several synonyms that refer to human positioning, often used interchangeably, but having specific nuances of meaning.
Position is a general term for a configuration of the human body.
Posture means an intentionally or habitually assumed positio...
Manly man
And a mommy (this is probably safe, given that I prompted for pitch black skin, still not as dark as I'd like, but it works)
What you mean not as dark
On my OLED phone the pixels are probably turned off already
Lmao
who is she? mommy black hole?
devious minds, i mean that her skin is so black that not even light escape from her
i'm not...
all puns were intended 😎
i deserve it <.<
its like a black hole,she will suck u in
ohhh... so many puns, glorious!

Why me 
sorry, went from napping to full on sleeping 
bow
is it just me, or is there no changes?
This is a beautiful image.
evenin'
@native halo You need jesus
/
like a good christian i always turn the other cheek
princess

pretty sure it's for you 
get him, mouse girl
does it work on each one separately
yup, but i think i have to play around a little more
looks good
try with other models or styles :o
okay it's not worth using the adetailer
it will get weird result like this
since it's working separately for each eye

This is so amazing.
@primal wharf

nice messy room
omg literally me
This is powerful.

throw this png on the other server
this is png robbery

too cute, overload

Hello
moderator
Lazy day I see
who doesn't deserve a little laziness
yeah, the small cat usually despise that dog
Oh, were you before?
i was surprise it let it go
I've been for the last 2 years, more or less yeah
it's "guide". it's used as a stepping stone to become mod if you're looking into it. but I'm using it as a stepping out stone
I'm free from bonks then
👍
:p
Have a cute mouse @wind girder
villll
How do you get that hair ? Usually the hair goes crazy like there's a tornado nearby, idk how to get it "simple" like that
There's pnginfo
Bunnycat
"tornado hair" in negatives
What does mickey do in that game
Like
Is it a playable character if so what class

he is the king. I don't remember that you could play it, but he is also a fighter that does have some fights alongside you
but for the most part, like all of KH 1, you are looking for him as he is lost

Mickey's battle against Xemnas, Larxene, Luxord, and Marluxia.
best you could see of his skills
strange game I'll admit, but it has its charms
(lol they cut the video at the start of the fight....)
The mickey was found
👍
@nova remnant how can I find out how much vram my cat has?
2 pairs of glasses, very sus
Ask it
one is for gaming, another is for work
Multiply the amount of meows x 4

approaching infinity lol

Quantum cat
:3
Mm
she downloaded a game from a low end pc >:((
Nice
Cuute
How do I get my pictures to have textures similar to this one?
try using anime screencap
hair ribbon would be the most accurate tag. hair bow may also work if the character is fairy well learned.
got my ribbon but it ended up being realistic
if this is too nsfw, let me know @gritty spruce
the hair looks realistic too so you probably want to use a different model and may want to look at the prompt too, if you['re trying to get a similar style to the first image.
I'm a dall e 3 plebian
dall e 3 is good, no reason to be ashamed of that. if youre going for anime style you might what to try adding more anime related words to the promp though.
nubby do you have access to sge?

I know you know you just want to share NSFW
dm if you want 
🫢
nothing like the comfort of a dm
certainly Discord doesn't spy on it
why do you have 4gb vram
some things shouldn't be asked
rude
soon the whole thing will be AI generated, the results, the comments in sites, the images..
yeah man, I'd stay away from lolis in bed, seductive poses and such
got it got it 
see what happens when you post in public

even YOU will be ai generated
discord is cracking down on that, public servers, public channels and minors on board means that should be safe for work
is that right out the gate of dalle3?
ofc not
that would make everyone useless
Reminds me of Nina Williams a little bit
This woman kept on being progressively prettier with each new image.
metadata included
A powerful display there.
@gloomy scroll
Seems like water plant chibi has grow well
Smol
I wondered what would happen if I only used
"(((best quality, masterpiece)))" for the prompt
As seen by this diffusion, it's a very anime style model
isn't that basically (best quality, masterpiece:1.331)
I dunno who that woman is, she looks hot though.
If I may ask, ever picture this woman in bikinis sometimes?
ikr
I made a Sailor Moon Video. Anyone an idea what i could do next?
#anime #girl #edit #animated #ai #dance #female #shuffledance #sailormoon #manga #stablediffusion #aigenerated
Who are these ladies?
just random gens
stella seagal
Mommy Miku
Quite the lovely frame there.
craving for coffee with creamer
im gonna go make some
I need to be less lazy and make cappuccinos
love them
Is that a Pharaoh, or just a dancer?
well i have the prompt for pharaoh but i was going for egyptian hottie
It worked in any case, that's for sure.
Looking like one of the Pharaoh's dancers if that's not a Pharaoh.
@nova remnant For Pharaoh, or the Pharaoh's dancers?
i love egypt now
@nova remnant Wanna DM me all the hot Egyptian babe images, by chance?
she can ride my mecha anytime 🤖
can you give metadata 
its in the image
should show in pnginfo
hmm I'm not able to

I mean I tried
but isn't showing
galactic style . nebula, constellation, cosmic, celestial, starry, (medium shot), (highres, detailed:1.2), (masterpiece), 1girl, in white spandex, blonde hair, blue blue eyes, (cinematic, hyper realistic), <lora:epi_noiseoffset2:0.6>, <lora:more_details:0.4>
Negative prompt: blurry, grainy, deformed, photo-realistic, low-contrast, terrestrial, easynegative, bad-hands-5, (low quality, worst quality:1.2), (watermark:1.1)
i wonder if discord changed any server side settings to not let image data show up
It still does, something just went wrong with yours >.>
Try mine
hmmm i wonder what might have caused my image data to not show
her fingers came out nicely
this is a such a beautiful shot but fingers all messed up
have u tried with adetailer hand model?
yes with few images but results werent all that great
just takes longer to finish rendering
😔
I love it. Why don't u try fix fingers with inpaint? I think its really good work, background is sick
Im still learning, but I saw a lot of works, that weren't made with first generation. They kept changing little parts with brush in Inpaint if Im not wrong and finished the work after a lot of fixes
Love this one, I will even use it on my phone BG
Im struggling with upscale right now
Dont know how people get that sharpening without losing a lot of details
i have the basic concept with inpainting but i have to look into the detail
nice 🙂
i usually do hires fix when generating images, but you could also take the easy approach to enlarge an image on the Extras tab
Did u use some model, lora for this or u used prompt only? I would like to recreate something on this theme
just these two LoRAs lora:epi_noiseoffset2:0.6, lora:more_details:0.4
but mainly prompts
thanks!
Mini Milk™️
this should show data on pnginfo
i think you might need to open in browser in it's original format to save and load on pnginfo
okay
I want to learn how to get such high-quality images
I was inspired by some dude from another discord
so: onexev
But I have so bad quality even on the beginning, I have no clue whats the problem :D
going to attempt inpainting to fix fingers on that image and then enlarge it by 2+x
🍐
👍
What you mean you have bad quality ?
Can you show some of your gens
If you have a capable gpu, quality isn't hard to achieve 
editing section of an image with inpaint is painstaking
@nova remnant I'd just sketch over her hand (doesn't take much drawing ability) and do like 0.3 img2img 
you mean inpaint sketch?
im trying with inpaint option
not working out too well
this is what im trying with
results arent exactly coming out right
Thats why I prefer img2img ;-;
the whole thing
can you explain
i mean img2img itself doesn't let you mask
you'd have to choose inpaint
this is the closest to soemthing decent i could get
upscaled 4x
@frosty yarrow inpainted and enlarged it
yes just img2img the whole thing, till it fixes the hand bit by bit, I just could never get inpaint to do things right, its good at making new things but not fixing stuff 
interesting, i never thought that was a method to get an image to look right, but let me try that now just out of curiosity but im done with that image for now lol
Are things improving with your systems?
@nova remnant problem with just img2img is that it loses a lot of detail from the original pick even with denoise str 0.3
no
if you wanna give me a new gpu feel free

also if you really want the exactly background just crop the new hand or whatever on top of the old image
and do it again with lower denoise
(tight clothes:2)
giving them gladiator costumes

I would be her training dummy
dalle is good
also very thick phone lmao
how did you make that with dalle?
a bunch of weird prompts :p
I'm not doing english prompts, makes stuff 100 times easier
there is no way dalle would uncensor panty
I'm not prompting for that, its random :D


