🍥｜anime

[worthy delta]

this is the way

also how do i copy the character out of this one

he looks so fucking cool

[stable crane]

gn

[gusty palm]

in prompts would something like "multicoloured hair" be the same as "multicolored_hair"

gnight lotta

[worthy delta]

keep getting a washed out colour

using amnesia model

also weird ass purple spots

[pure oyster]

Yea, energy balls do not work well with hands lol

[gusty palm]

mine does that sometimes restarting the cmd prompt window helps or playing with the prompt weights reducing them overall

[worthy delta]

fr fr

[gusty palm]

i dont actually know what causes it

[worthy delta]

imma try a dif model

but ngl amnesia looks amazing

[pure oyster]

Try the one I linked

[worthy delta]

same thing

[gusty palm]

does anyone use LDSR

[worthy delta]

ldsr?

[gusty palm]

it seems to be included with stock install but theres nothing in the folder

the upscaler

[pure oyster]

There is like an infinite amount of upscalers and types of upscalers, I could be using LDSR everyday and I wouldn't know it existed.

[gusty palm]

;p

why these flashing

[worthy delta]

how did u make them like tags?

[supple raptor]

Most training uses US english spelling so you're better off with multicolored hair underscores are generally not used in training, but occasionally can be useful if the AI is interpretting two words that make up a single tag as 2 different concepts. A common example of that is cowboy shot which is an image that shows the character from mid thigh to the top of their head. Some models start adding cowboy hats, so you might use cowboy_shot to reduce the chance of cowboy hats.

[worthy delta]

oh shit didnt think of that

also nubby what is 1 model u couldnt live with

[wet elm]

what are the best settings for merging models?

[supple raptor]

ExpMixLine_v3 has a nice and fairly unique style, Aside from the ones I already mentioned, I wouldn't want to give that one up.

[supple raptor]

There is no right answer, every combination of models will be different and different settings will give different results that might all be desirable even with the same two models. Merging models comes down to a ton of trial and error. You have to test different merge combinations using the same prompt seed and generation parameters then repeat with many more prompt/seed/parameter combinations.

[wet elm]

are there any general tips you could give me? im going to try merging models fort the first time rn

[worthy delta]

I WANTED A REALM TEAR AND IT GAVE ME A HAWT ANIME WOMAN

[supple raptor]

Grab the SuperMerger extension and take advantage of the option to turn off model saving. With that turned off, you can try as many different combinations as you want without filling up your HDD since it will just merge inside your vram and let you generate with the merge until you load another model. Just start experimenting and take notes about what works and what doesn't. It's one of those things that I can't really explain how it works without overloading you with information. If you have a specific goal in mind I could offer a bit of guidance on things that might work.

https://github.com/hako-mikan/sd-webui-supermerger.git

[wet elm]

ty, i want to merge aurora with a model i like the style

[worthy delta]

i got a problem

it keeps loading the same type of image

[wet elm]

so that i can maybe keep the versatility and the style at the same time

[worthy delta]

over n over again

but the seed is perfectly fine

[supple raptor]

Let me load up WebUI so I can get some screenshots.

[supple raptor]

well, how similar are they?

[worthy delta]

lemme get the pics

one isnt nsfw but its kinda explicit

imma crop it

i swapped models

and the prompt

was realm tear

[supple raptor]

Do you have controlnet turned on?

[worthy delta]

idk what control net is so prob not

and look new one is pretty much the same

not the prompt at all

the prompt was also realm tear

it keeps giving me anime women

[supple raptor]

Drag and drop one of the images from the folder it's in on your PC into Discord so I can check the PNG info on it.

[worthy delta]

will do

ill just send the new one its doing

idk what is doin it

[supple raptor]

This is the top part of the SuperMerger extension. I've numbered the important things to use for MBW in it.

1. This is your first model, it's what you will be merging into.
2. This is your second model, it's what you will be adding.
3. This enables the MBW function rather than normal merging (I'll show the MBW section in the next screenshot).
4. These settings dictate what will be done when you click one of the merge buttons. As long as you don't check the "save model" box it will only load your merge into vram, without saving it so you can experiment without accumulating TBs of checkpoints.
5. This button will allow you to merge and generate an image with the merge in one button press. You need to press this every time you make changes and want to try the new recipe. You can use the "Gen" button to the right if you just want to gen another image with the merge that is already loaded. That merge will stay loaded in vram until you load something else. You can use either the generation parameters below, or the txt2img tab to dictate what will be genned from this tab.

More in the next message...

[pure oyster]

Having a vague prompt on a model extremely tuned to waifus is probably going to just give you a waifu.

[supple raptor]

@wet elm

1. This button sets the current values of the sliders to be used the next time you click merge. You have to click this anytime you change presets or move sliders to apply your changes.
2. This buttons sets the sliders to the values that are in box 3. Mostly useful if you want to tweak the recipe a bit after pasting or typing values into box 3.
3. This is the recipe that is going to be used for your merge. You can type or paste into this field, or you can use button 1 to move the values from the sliders here. The numbers corrospond to the sliders starting at Base, going all the way down the IN column, then M00, and finally from the bottom to the top of the OUT collumn so the final value is OUT11.
4. This is a bunch of merging presets that all follow different forms of logic for a merge that might be useful for some situation. You don't have to follow any of these, but they can be a good starting point.
5. This is the text encoder or CLIP layer. It is called base_alpha in the normal MBW extension. If you merge with this, it can result in broken clip values, which is fixable, but can cause some very weird behavior if not fixed.
6. These are the layers. There are some guides out there for what different layers are expected to do.

[wet elm]

damn ty

[wet elm]

thats a LOT lool

[supple raptor]

Yeah, I'm not seeing any problems with the parameters, I mainly wanted to see if it showed anything other than your prompts being applied and there's nothing else there. When left to their own devices, anime models will tend to generate a waifu. Try describing what you want in more detail. If you don't want any people no humans is an actual tag and one one of the few times that using "no thing" will actually be effective in a prompt.

[worthy delta]

weird

ive never experianced this before

the no humans tag just made it well

more human

[supple raptor]

I could write several more messages worth of information about what the different blocks usually do but it's better for you to start by experimenting yourself. Start by setting all blocks to 0 with your seed locked. Then try again with all blocks set to 1. This will tell you what model A looks like and what model B looks like. Next set them all back to 0 and turn them each up to 1 one block at a time, taking notes about what you notice changing with each different block that you test. So that you can try to identify the blocks that you want to merge in.

[worthy delta]

more human

ik how to stope it

it looks amazing

but not at all what i want

[wet elm]

well im gonna try to start today, but the problem is that it takes a long while to merge a model in my pc

[supple raptor]

realm tear is not likely a cobination of words that they AI has seen together many times. Realm has most likely been primarily associated with fantasy royalty. Tear is associated with something that is torn as well as crying so it's extra vague. Describe what a realm tear is supposed to look like in your mind.

[pure oyster]

Stable diffusion does not know what a realm tear is, you have to be more precise with your prompting

[worthy delta]

lets try space ripping open

[pure oyster]

Cosmic portal might work

[supple raptor]

It will take less time with this method because you don't need to wait for models to be saved each time. If you have about 4-5 GB of system RAM that you can spare to cache models in you can go to Settings tab > Stable Diffusion and set Checkpoints to cache in RAM to 2. This will keep both checkpoints that you're merging from in RAM so that they can be accessed for merging in a matter of a few seconds which will speed up the process by a lot.

[wet elm]

ohhhh thanks

[wet elm]

and im also just starting to learning inpaint

[pure oyster]

I got this with the prompt "Cosmic portal, break in reality, space", also with the kakigori model

[worthy delta]

Ooooo that looks amazong

Ngl kakigori really good

I've been tryna get a good worlds colliding one

Or a cosmic god one

[gloomy scroll]

[pure oyster]

It's a bit underwhelming, but this is what I got

[worthy delta]

"Underwemlimg" THAT IS AMAZING

[supple raptor]

Why didn't I think of creating a launch.bat to run ComfyUI months ago instead of manually activating my venv which is in a different directory all this time

[pure oyster]

Img2img?

looks way too coherent

[supple raptor]

Nope, just stole the file name of your image to use as a prompt and ran it through an SDXL finetune I'm playing with.

[pure oyster]

ahh SDXL isn't that like the new super version of stable diffusion?

[supple raptor]

yeppers

[pure oyster]

That would explain why it looks like that, the biggest two issues with 1.5 is coherency and composition

[stable nimbus]

damn sick

[pure oyster]

I have absolutely no knowledge of SD, but my impression is that the current base model is a poorly trained, jerry-rigged abomination. If custom models look as good as they do now, I can't imagine how good they would be with SDXL a year from now.

[stable nimbus]

its not really poorly trained, its jsut too small to be good at everything

so instead its mediocre at everything

and people make it good at specific things via finetuning

sdxl is bigger and better

[supple raptor]

Base SD2.x is pretty poorly trained. SDXL seems much, much better from what I've seen so far.

[south sable]

SDXL 0.9

[worthy delta]

What is sdxl?

[supple raptor]

Stable diffusion but bigger. That's the model that the bots in #1100170312106127410 - #1101178553900478464 are running.

[south sable]

SDXL 0.9 /

[supple raptor]

Can I get the prompt you used? I want to try in WDXL.

[south sable]

girl, cyberpunk haute couture, carnivale di italia, art by Tomer Hanuka, Russ Mills, go nagai, #manga photo by Ashraful Arefin, Anna Nemoy, sculpture by Thomas Heatherwick #sketch #photography, HDR, insanely detailed eyes, UHD, 64k, RAW, sharp, sharp focus, 8k

[supple raptor]

Thank you!

[south sable]

No problem!

[supple raptor]

still working on figuring out the right parameters for this

[south sable]

I've never used WDXL

[supple raptor]

It's limited access currently.

Should be released on the 18th along with SDXL 1.0 I believe.

[south sable]

[supple raptor]

Let's see what happens if I just change girl to 1girl since it's trained on booru...

[supple raptor]

can't tell if I'm getting the hang of this or just got lucky.

[slow needle]

I guess both but I gonna train a Lora since I find the current style kinda lacking

[supple raptor]

I want to figure out how to prompt it before trying to train anything on it so that I can actually tell how shit my training is XD

[slow needle]

Anyway one of the better ways to prompt for Sdxl is splitting the prompt in 2 parts for the 2 tenc’s

[supple raptor]

I'll try that. I'm guessing you want to do more of like a basic prompt for text_g and a more detailed one for text_l?

[slow needle]

Yup but rn I just do quality and style prompts for text_l and pass the rest into text_G

[supple raptor]

Do you know if he used the same quality tagging method that they used on WD1-5?

[slow needle]

Idk but you can test and try or ask them yourself

[supple raptor]

Yeah, I'll ask Haru later. It's like 3 or 4 am for him right now so I'm sure he's probably sleeping like I should be.

[steady grail]

SDXL how much gram needs?

[gloomy scroll]

I was able to use SDXL with 1070 8GB

[supple raptor]

8gb for Nvidia and RTX GPUs highly recommended. 16GB and Linux required for AMD GPUs.

[steady grail]

after the results with lovercraft i must try it

so 4 g is too small

[supple raptor]

Yeah I don't think 4GB is going to happen. 6GB might be doable with some tricks like only loading one component at a time if your card fully supports half precision.

[steady grail]

till nowi did everything with a gtx 1650

[supple raptor]

It's a big model. It has 2 text encoders and a big UNET.

[steady grail]

and it works pretty smooth with every model till now

oh

damn, just now that i could works smoothly even iterrogation

[gloomy scroll]

Smol

[steady grail]

i'll wait they'll polish it a little

[zinc flax]

Thic model

[supple raptor]

Model converter extension in WebUI can convert it to bf16 or fp16 btw.

[steady grail]

it's just question of time before it get pruned and refined

[zinc flax]

Can't wait for the great stuff from it

[gloomy scroll]

I'll probably wait for some finetune before I'll play with it more

[steady grail]

finally no more weird derps becoud half baked porn mixes...

[supple raptor]

Yeah it isn't as good looking as our current 1.5 based models yet.

[steady grail]

a clean slate

[gloomy scroll]

Base > Aurora refiner

[steady grail]

cute, but i wouls ass dome shaprness, to get a more brushy effect

[gloomy scroll]

I wanted something more like this

[steady grail]

i love this style, it's not ai, right?

or is heavily edited at least

[gloomy scroll]

Save it and throw it to pnginfo and you can find out

[steady grail]

df?

[gloomy scroll]

?

||I can't draw||

[steady grail]

holy moly, really?

i love this style

[supple raptor]

Vil runs a cute manufacturing plant.

[steady grail]

my heart, too cute >.<

if you make a pony too, i'm dead

[gloomy scroll]

If you want more just search "from:.v.i.l. has:file"

[steady grail]

wow, you heavily used adetailer, i have it, but never learned to use. looks like a more sharp negative prompt

no, i was wrong, use even positive prompts. it's a way to fix weird quirks

[gloomy scroll]

Adetailer just inpaints face and also hands in my case

It works nicely for face but hands just get little bit quality upgrade but doesn't actually fix f'ed up hands

[steady grail]

i should move the prompts of the face and hands on adetailer

[gloomy scroll]

I don't use any prompt on adetailer

[steady grail]

only nega?

[gloomy scroll]

I only reduced Adetailer steps to 10 otherwise default settings

[steady grail]

oh

[steady grail]

this model is broken, i need it

[steady grail]

you are making me feel broken, i'm becoming diabetic

[gloomy scroll]

https://civitai.com/models/40199
https://huggingface.co/Surficialruby/LoraDump/blob/main/LoraMerge_dim128.safetensors

[steady grail]

my mission wil became giving the world more diabetes >.<

[gloomy scroll]

@steady grail If you plan on testing that lora merge it has nsfw artist so it may need little bit more anti lewd negatives

Also just Aurora is able to make things like this

[steady grail]

yeah, nowadays it's pretty common, i usually just place (naked:1.3) in negative to do the trick

i don't know why, but usually porn models are the more creative and coerent

except for the derpy mutations

damned any mix!
if i only could cancel you from existance

[steady grail]

what's the use of BREAK?

[gloomy scroll]

BREAK is feature in webui

Normally model accepts only 75 tokens so webui splits prompt to blocks of 75 tokens and BREAK fills block with empty token and starts new block of 75 token

Closer tag is the start of the 75 token more weight it has so it's easier to visualize where your 75 token blocks are when you use BREAK

@steady grail This is probably better explanation #🍥｜anime message

[steady grail]

ohh, no, it's clear enough

[bitter latch]

and to work I go

but first.. or last?

[steady grail]

what's this Discard penultimate sigma: True, Version: v1.4.1?

[brittle jay]

How did it happen that I am more busy in summer than in the rest of the year as a highschool student?

[wooden onyx]

what model and prompt did use for this?

[brittle jay]

Like I barely talk in here now

[brittle jay]

By that do you mean a cute manufacturing-plant or a cute-manufacturing plant

[kindred rock]

does anyone know what model and prompt i should for stuff like this

[brittle jay]

What is up with this vae

Switch vae or I cry

[brittle jay]

I think any generic model should work. Something like anythingv4 or something, just go to civitai and download one of the top anime models

Oh maybe aom

aom = abyssorangemix

[barren anvil]

I used meinapastel_v6Pastel as the model and elaina-ph_64 as the Lora. I didnt save the prompt, sorry

[stable crane]

man I need your model ❤️ lol

[zinc flax]

Very cute model I gotta say

[zinc flax]

https://huggingface.co/sleepotimer/Model_A

That's 3A

[stable crane]

oh 3A looks cutest

oh it turns out I have this! and didn't remember downloading it

I never tried it yet, will try now

[zinc flax]

Aaa

Yes yes do try it

[stable crane]

ty for sharing

yep it's mega cute

let me try without my cute lora

[steady grail]

lotta, what vae are you using?

[stable crane]

Anything-v3

sometime I used blessed.vae

[zinc flax]

3A has surprisingly great hands btw

[stable crane]

I'm loving it

[zinc flax]

Like fr, even I'm surprised lol

[stable crane]

no loras

the hands are good

[zinc flax]

old gen

[stable crane]

sick

so cute

[zinc flax]

Thic

[stable crane]

@zinc flax doing gods work

thanks for making these cute models all the time 😄

[zinc flax]

Doing this for fun lol
But unfortunate that I forgor the recipe for all 4 models.

Impulse merge

[stable crane]

I have some merges like that too

but now they're their own thing, in a way

usually the hands on this prompt are pretty messed in other models

[zinc flax]

Robo hands

[stable crane]

robotic hands tends to make model think "sure fingers can go any direction"

[zinc flax]

Random prompt, also with 3A

[stable crane]

it does a fair job with spider girl prompt

verdict: great model

it can do less cutesy stuff too

[wet elm]

cute lora?

[stable crane]

blue archive emoji lora on civitai

[wet elm]

ty

[stable crane]

adds cuteness to everything

[wet elm]

oh damn it does

[stable crane]

😄

post pics

[fading olive]

wow

🧐

[stable crane]

my spider girl lora (on civit)

[fading olive]

i will try,tnx

[near elbow]

different model and style

and another

[worthy delta]

Is there any models that do good action? Like fights

[worthy delta]

Gah dayum

Finally a good looking guy

[bitter latch]

there's a bug in my auto1111 that doesn't always load vae correctly and I have to correct

[bitter latch]

or could be hte model or somethjing..... shrugs

[brittle jay]

You could also copy the vae, rename it to the model name and add .vae.pt after the name and put it in the normal models folder

Then reload the model

And it should use that vae automatically

[bitter latch]

yeah but I have several and I like playing around 😄

[supple raptor]

Only if VAE is set to Auto.

[supple raptor]

Or there's a setting that will make it ignore your selected VAE and load the one with the same name, but that's a dumb feature that shouldn't exist.

[bitter latch]

without changing anything

and manually changing vae in settings

[gusty palm]

is there some resolution interval that should be kept to like intergers diviblwe by 64 or something

512, 768 ect or does it not reallly matter

[bitter latch]

pondres what to make hmmmm

[supple raptor]

Resolution has to be divisible by 8 for WebUI. A few extensions that I don't remember off the top of my head need 64 so if you get an error about size or resolution, try a multiple of 64 instead.

[gusty palm]

ok

sometimes the hires fix gives me stuff ending in non divisible amounts

[brittle jay]

I think it overrides that no?

[gusty palm]

depending on the upscale factor

[bitter latch]

I don't use hi-res fix much

[gusty palm]

kawaii

[bitter latch]

turns it on it's side, for resolution, turns on hi-res fix and ups the color enchance, tweak tweak, run

[brittle jay]

Technically you can also use --vae-path

Then type the path to the .vae.pt file

[bitter latch]

50% before fix

and.....

waits

kinda enhanced

[gusty palm]

hmm

[bitter latch]

ponders throwinag away the box and going rogue

goes rogue

ok thats suppose to be anime "The rock" but. hmm ok

[gusty palm]

should i still be using dot product for this while using xformers

[bitter latch]

I'm using invoke AI

tho not sure what that does so.. shrugs

[gusty palm]

;p

[bitter latch]

ok thats an odd angle..

hmm could almost make a comic panel page with those

[wet elm]

is there a way to like extend an image (i believe its called outpaint) in SD?

[bitter latch]

open outpaint, open cavas I think are extensions

but anyhow. here

[south sable]

SDXL 0.9 with 1.5 refiner

[bitter latch]

holy crap thats alot of png info that makes no sense

[south sable]

yeh it's not 1.5 and it's not A1111

[bitter latch]

copies first part, paste, run... 1875.. damn

let's see what kind of mess I get

[south sable]

you should be able to look through that and get the prompt though

it should be familiar stuff once you spot it

[bitter latch]

well I can't paste that

[south sable]

[bitter latch]

yeah black and white woman tied up naked with chalk board behind them with formulaa's. ok then

[south sable]

def not the prompt at all

[bitter latch]

interesting results i'm geting with this model

hmmm (switches back to his merg)

merge

[gusty palm]

has anyone done videos

or seen the ones where they take a video of a girl dancing and put a style over it or the infinite zoom looking things

[bitter latch]

with m y video card. not likely

[worthy delta]

Honna try that

[bitter latch]

don't do it

or do.. either way

[worthy delta]

I mean ai is the best thing humans made ngl

[bitter latch]

eh it's a tool and ai isn't really ai, not in this case

unless it stands for artificial itterations

[worthy delta]

should it be called ap?

[south sable]

I thought it was: Antiquated Ideas

[worthy delta]

artificial image

[stable crane]

either way, it's a godly creation

[worthy delta]

DMAN RIGHT

[bitter latch]

yeah it's good tool

[worthy delta]

more than good

[bitter latch]

but like any tool it'll be misused

[south sable]

Art-official Intelligence?

[worthy delta]

it will revolitinaise alot of stuff

[bitter latch]

puts in random random random with base style pile and saved style

waits

[stable crane]

if I were an artist, I'd train on my own style and do touchups

or traceovers

[zinc flax]

no clue what this is but it sure looks cool

[bitter latch]

goes about making art he can't post here :D

[worthy delta]

prompt goes hard

[zinc flax]

from chibi to some old dude lol

[south sable]

throw this into after effects with some rain and thunder sounds etc. make a ****wave youtube

[south sable]

https://tenor.com/view/aikatsu-friends-aine-sus-among-us-aikatsu-gif-24430684

[zinc flax]

Pretty impressed with how the hand turns out though with this model

[south sable]

pushing SDXL to do anime.

https://tenor.com/view/honest-work-good-honest-work-it-may-not-be-glamorous-the-simpsons-gif-13763673

[south sable]

going for that low budget animation studio vibe IG

[gusty palm]

i like the idea of using ai for diagnostic medicine and hopefully itll be used for noticing things humans missed

imagine house md with chat gpt

minus the overbearing guidelines

[worthy delta]

what is a good height and width number?

[bitter latch]

and totally random

[worthy delta]

i make it too big by accident

[gusty palm]

i try to stick to multiples of 256/128/64 etc

at least for the base resolution, idk what to do for less than 512 so ive been guessing

[gusty palm]

blender has this great feature for number inputs where u can do math inside the box
like if its 512 u could type 512\2 and it would autocomplete the math result

[worthy delta]

imma start doing that thanks!

[gusty palm]

i think the resolution slider by default scales by intervals of 8

i think u can change it to something higher i had it set to 64 in sygilui

384*

384*786 would be a super wide looking image

i dont know how much it actually matters though

i know ive had problems in blender and ffmpeg if i use non multiples of 2

[worthy delta]

ill try that!

yup it looks amazing

thank you!

[supple raptor]

It only overrides the selected VAE if this setting is turned on.

[brittle jay]

Didn't know this option existed

[supple raptor]

xformers should be better since you're not on a 30 series or newer card.

[brittle jay]

Or maybe I did and forgot

[gusty palm]

ok cool

it selects xformers when i set it to auto so i figure ur correct

[brittle jay]

Is the opt thingy better for 30 series?

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Optimizations

Damn

[supple raptor]

There are a few "opt" ones, but yeah either --opt-sdp-attention or --opt-sdp-no-mem-attention are what you want for 30 and 40 series. The no mem version is deterministic but uses more VRAM and is a little slower than xformers. The first one is basically the same as xformers, but is built into pytorch 2.0.

--opt-channelslast is also supposed to make things faster on 16xx+ cards, but for me it cuts my speeds by almost half so https://cdn.discordapp.com/emojis/930630372348289044.webp?size=64&name=chenShrug&quality=lossless

[worthy delta]

what's this opt ting yall talking abt

[supple raptor]

commandline args that save some memory and increase speed. IIRC, I already had you enabie "sdp - scaled dot product" in settings which is the same thing as that --opt-sdp-attention flag, just a simpler way to enable it.

[worthy delta]

Ooooh gotcha gotcha

Also my ai be running smooth ever since u helped me with dat stuff

[supple raptor]

good to hear!

[worthy delta]

And oh my god the art is glorious decent amount I can't post here but it high quality as hell

[gusty palm]

does batch size more than one increase vram by alot and is the vram estimator not always accurate

and can u mix and match vae willy nilly

[supple raptor]

Yeah, batch size is the number of images that are processed at the same time so it effects vram usage by a lot. Batch count increases the number of batches that are processed consecutively which has no effect on vram usage.

[gusty palm]

gochit

ive been running size 2 almost consistently thats probably why my shit was so slow

i forget which vae to use

[supple raptor]

The one you have selected is the one that I use for all of my anime images, for realistic images you should use the "mse" one at the bottom of the list. You can try any of them with different models though to see how they look. Main differences to look at will be textures of the backgrounds, fine details like eyes and hair, brightness and contrast.

[gusty palm]

ok cool

[worthy delta]

where can i get random prompts

also revanimated looks so bloody good

[slate night]

steal prompts from other people's images

[worthy delta]

this is the way

but ngl ngl want somting original

[slate night]

Maybe use a random word or phrase generator?

And then build an idea off of that

[worthy delta]

thats a good idea

btw how do u look what prompts someone used?

[slate night]

Save the image, drag and drop, if there's metadata it will tell you

[worthy delta]

thank you

i never noticed that was there

[slate night]

Sure thing!

[worthy delta]

this community is full of helpful people

[bitter latch]

Sigh bored

[worthy delta]

thats a mood

went from merc to corpo

[magic wave]

You're trying to do a niji style?

[gritty sundial]

yes

[gusty palm]

do some loras or embeddings add more than one token to the promt count

[bitter latch]

edits and tries again

[gusty palm]

https://huggingface.co/stabilityai/stable-diffusion-xl-base-0.9 whats this and does it work with our ui

[magic wave]

Well, I'm about to publish a Niji style LoRA

[bitter latch]

well.. that failed

sorta

hmm ok

[bitter latch]

oops. umm hha

ok time for some star trek

[bitter latch]

half a person?

[bitter latch]

almost

ok then...

[supple raptor]

LoRA's don't actually add anything to the tokens of your prompt as they are effectively a model that is merged into the full model that you're using.

Embeddings have a number of tokens set when they are trained and they will always add that many tokens to your prompt. The correct number should be calculated by the token counter.

[supple raptor]

This is the beta of the new version of SD. It doesn't work with our UI yet, though Automatic1111 is working on that. As the name implies, it is a very large model and will require more VRAM than the current models that we use.

[pure oyster]

Cries in 1060 6gb

[supple raptor]

Yeah, it's even worse for AMD cards or Nvidia cards lower than 20 series due to running at full precision or simulated half precision instead of true fp16.

[worthy delta]

Is this new model gonna make everything so much more better?

[gusty palm]

ah ok

[gusty palm]

gpu twinsies

[rotund herald]

Hey I'm trying to learn how to install and use stable diffusion but honestly I'm experiencing information overload and I keep trying to follow all these guides but it leads me in circles, if anyone has the time I would really appreciate some step by step instructions that a 5 year old can follow

[stable crane]

I call this "globohetero"

[gusty palm]

if ur on windows and nvidia method one should work https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-NVidia-GPUs

[glacial bane]

Anybody know what svda is this?

And does anybody know how to png info someone else's prompt?

[rough oxide]

use an image with metadata that they have and plug it into a1111's png info

[glacial bane]

Like i cant even steal my own prompt

[rough oxide]

or if it's a comfyui image you can plug it into comfyui directly

[glacial bane]

Thank you

[glacial bane]

But do you have any idea what svda this is?

[rough oxide]

svda?

i can tell you what model could get close to that aesthetic

[stable crane]

svda?

is that a tiktok img?

are images big there?

[glacial bane]

It is

[glacial bane]

Idk

[glacial bane]

Lots of TikTokers just post sD pictures

[rough oxide]

https://huggingface.co/lylogummy/Bunnyspoon can get you close to that aesthetic, one of the better anime model mixes imo.

[stable crane]

bunnyspoon is nice

[glacial bane]

TY

[pure oyster]

Blowing up an image's resolution without inpainting is a crime

[stable crane]

nah it's all about that 2x 3x highres fix

[glacial bane]

Ye fr

Whats a pantyhose again.

Wait

Whats the difference with a pantyhose and a stocking

[rough oxide]

if you need more advanced setup for SDXL i can assist with that if your hardware is up to spec. otherwise just use automatic1111's installation guide for the other models

[glacial bane]

Does anybody relate that we have to know a majority of clothing names for our prompt

And hairstyles.

[lament lion]

Pantyhose goes all the way up to the waistline

[glacial bane]

Oh

Ty

[stable crane]

stocking = sock(ing)

[glacial bane]

But can anybody tell me what this is

[rough oxide]

what part?

[glacial bane]

I cant describe it then i cant prompt it

[stable crane]

it's a girl

[rough oxide]

"sweating profusely" gives you sweat

[pure oyster]

that is a face I think

[rough oxide]

blush gives you blush

[lament lion]

Long hair….

[glacial bane]

The hair

THE HAIR

The hanging hair piece

[rough oxide]

hair behind ear

[stable crane]

red hair

🧌

[lament lion]

It’s just long hair models do that randomly

[glacial bane]

Lie this one

[rough oxide]

that style is "hair behind ear"

and long hair

[stable crane]

😂

[rough oxide]

yes that's "hair behind ear"

lol

most data of hair behind ear

also has a strand infront

if combined with long hair

[glacial bane]

Hair strand, behind ear

[rough oxide]

no

"hair behind ear" and "long hair"

lol

then it's up to RNGesus

[glacial bane]

Yk where to research sbout facial expressions btw

[rough oxide]

danbooru is where you research everything.

[stable crane]

danbooru

[rough oxide]

it uses danbooru tags

[stable crane]

or civitai

[rough oxide]

that's why i'm insisting you use "hair behind ear"

lol

[glacial bane]

That's where i usually get my prompting ideas but okay

[rough oxide]

most of the data of hair behind ear is like that

with a hair strand infront too

[glacial bane]

Ik ik ill use it

[glacial bane]

Is that a manwha.

[rough oxide]

i cannot in good faith post the full image xD

[glacial bane]

Understandable

Is it nsfw

[stable crane]

let me try to replicate this

[rough oxide]

falls under 'sexualized poses'

and underage character

even if not strictly nsfw

[stable crane]

the prompt cursed me to CUDA error

[pure oyster]

This thread is pretty chill compared to the others, you can get away with some moderately sexier, but not nude stuff.

[glacial bane]

I think bro likes old anime styles

[stable crane]

true, I've never had an image removed

[pure oyster]

Just don't post in general-with-images, that thread is for religious grandparents.

[rough oxide]

times have changed then? lol

i was here in the beginning and it was uber restrictive

[supple raptor]

That is a sidelock. The tag is sidelocks. Alternatively, short hair with long locks if you want the rest of the hair to be short like Megumin.

[glacial bane]

Ty TY

Quick question to yall use weights on everything

[stable crane]

my model doesn't know sidelock / hair behind ear

[glacial bane]

Smash- I MEAN darn thats cute

[glacial bane]

Is your model anything 4.5

[rough oxide]

u need an s

sidelocks

[stable crane]

only when the term is too prominent or not prominent enough, then I alter with a weight

[supple raptor]

bad model. must delete

[stable crane]

using sidelocks

[glacial bane]

What model do u use again to have perfect hands

I think i remember you from somewhere

[stable crane]

this is @zinc flax's 3A model (search discord for it)

[rough oxide]

khfb+auroranegative can help with hands

[stable crane]

maybe you used some of my loras

[glacial bane]

Are those models?

[rough oxide]

negative embeddings

[glacial bane]

Who me?

[rough oxide]

khfb and auroranegative are negative embeddings

you put them in negative prompt with embeddings loaded

[glacial bane]

Prompt neow.

[south sable]

[rough oxide]

it helps fingers & stuff w/o altering image much at all

[stable crane]

masterpiece, best quality, 1girl, red hair, (hair behind ear:1.3), long hair, slight smile, embarrassed, nervous, blushing, hime cut, (sidelocks:1.3)

[zinc flax]

Sidelocks is hella rng

[glacial bane]

@rough oxide how to install those negative embeddings

[supple raptor]

link to KHFB and AuroraNegative: https://huggingface.co/SweetLuna/Aurora/tree/main/AuroraEmbeddings

[rough oxide]

nubby just linked it

put those in embeddings folder

[supple raptor]

Just download the files and put them in the .\embeddings folder, then use the filename without the .pt in your negative prompt.

[zinc flax]

What I got yesterday

[stable crane]

are they both negative embeds?

[glacial bane]

Can you like guide me to it
Like SDFILE/MODEL/LORA

[rough oxide]

yes

[supple raptor]

yes

[rough oxide]

they just linked it to you; those are the files you need, click each one and hit download

[glacial bane]

I was in for the pose im surprised that the model you use chooses random poses

[south sable]

https://tenor.com/view/chips-eating-anime-lunch-break-gif-11881516

[stable crane]

most do if not supplied

[supple raptor]

Another negative embedding that is helpful for hands. Install and use the same way as the other two https://civitai.com/models/56519?modelVersionId=60938

[stable crane]

KHFB/AuroraNegative are quite strong

[rough oxide]

i've seen that one too

i've found the best results from using (KHFB, AuroraNegative) together in negative for my proompts

[glacial bane]

Can i use the three of them

[rough oxide]

mmm

[stable crane]

you can use as many as you like

[rough oxide]

you don't wanna stack too much stuff, but you can experiemnt

[supple raptor]

negative_hand is nice because it barely effects anything other than the hands which makes it more compatible with other negative embeddings than most.

[stable crane]

I'm trying to weight them at <1.0 because they seem so strong

[zinc flax]

Tried Aurora embedding
It feels like it doesn't do much idk why

[rough oxide]

combined with khfb?

[zinc flax]

Almost works the same as no negatives

[stable crane]

really I see a huge diff on 3A model

[glacial bane]

Whats the difference between 1 and 1.0

[rough oxide]

nothing

they're saying they use it below 1

e.g. 0.9

[supple raptor]

I use negative_hand, (KHFB, AuroraNegative:1.1) in my basic negative prompt.

[rough oxide]

o yea

[zinc flax]

Oh, I always use easynegative for that model because lazy lol

Or (low quality, worst quality:1.4)

[rough oxide]

i do 1.2 on my aurora iirc

[glacial bane]

My negative prompt kinda does slightly good hands

[rough oxide]

but yea, you likely won't find a good model with perfect hands

so you have to play around with negative prompts

& negative embeddings

in addition to a good model to get best results

[stable crane]

AuroraNegative vs easynegative on 3A

[glacial bane]

Well usually i get good hands if i use img2img

Plus my negative prompting

[zinc flax]

claw pose gives perfect hands btw

[rough oxide]

lol

[rough oxide]

debatable xD

[glacial bane]

Well im not using a negative embedding to it

[rough oxide]

if you get a good claw pose it's nice

[zinc flax]

Well, I guess it depends on the model

[stable crane]

she gon' getcha

[rough oxide]

yea, bunnyspoon can do clawpose well but something like

gyoza can't do claw pose as well iirc

[supple raptor]

models that struggle with clawpose can sometimes be improved by adding claws to the negative prompt.

[glacial bane]

Well i suggest the right one it gives more details i suggest the left one if people like simple anime shading to it

[zinc flax]

Yes yes

[glacial bane]

My dream girl

[glacial bane]

Whats the prompt for the jacket

[zinc flax]

There's should be a metadata for it on the image

[pure oyster]

"jacket" probably

[glacial bane]

This kind of jacket.

[supple raptor]

open jacket, off shoulder

[stable crane]

bare shoulders

[supple raptor]

^that

bare shoulders, not off shoulder.

[glacial bane]

Oh thx

I use img2img alot

[pure oyster]

I actually use bare shoulders quite a lot

[glacial bane]

Mashle vibes

[zinc flax]

1girl, tank top, claw pose, half-closed eyes, [purple hair:black hair:8], open jacket, long sleeves, brown eyes, x hair ornament, shoulders, blunt bangs, trousers, outdoors, grin

[pure oyster]

and "midriff"👅

[zinc flax]

Huh, I just added shoulders lol

[glacial bane]

Btw does prompting a facial expression difficult?

[stable crane]

you'll probably never get quite exactly what you want

[rough oxide]

Depends. Angry works fine for instance in the models i use

tears/crying type stuff also works fine usually.

[glacial bane]

With the help of experimenting it slightly does

[zinc flax]

:3 is easy to get
But again, it depends on the model

[stable crane]

sometimes you will need a lora

[rough oxide]

:D also easy to get, i don't remember if you need to escape the : though

[glacial bane]

I use anything 4.5 which kinda helps me with generating anime based images

When i want a model that specializes in anime

But it does the job

[rough oxide]

if you like anything style images, gyozamixv5 is a pretty good evolution of that style with better clarity/details

[glacial bane]

Bet

[rough oxide]

you get that one here https://huggingface.co/Jemnite/GyoZanMix

[glacial bane]

You really like girls in those types of costumes

[zinc flax]

She looks so goofy I love it

[zinc flax]

Bobbers

[pure oyster]

4k btw

[glacial bane]

Same model? You usually use a model that generates kids.

[stable crane]

that's 3A

[glacial bane]

wowzers

[glacial bane]

Exactly

[zinc flax]

Wait that's 3A

[glacial bane]

A model

[zinc flax]

Looks very different

[stable crane]

yea

all about the prompting style I guess

[zinc flax]

Very different

[glacial bane]

I use thick line lora all the time and a prompt

[glacial bane]

Whats with the white border

[stable crane]

I like borders on my pics

[glacial bane]

Whats a underboob

[zinc flax]

Interesting

[stable crane]

you gonna have to look that up

[pure oyster]

Something great

[stable crane]

oh, usually I also have updog in my negs

[zinc flax]

I almost fell for it

Legit

[glacial bane]

I used the anime manga artstyle lora

[zinc flax]

Now that's a style lora I recognize

[glacial bane]

Finally i generated something neat perfect

[stable crane]

I'm so happy someone made that lora, I love that style

now, someone just needs to make HotVR style

[glacial bane]

Sint that thickline lora?

Isn't"

[stable crane]

nope

[glacial bane]

What is it?

[stable crane]

gyokai

[glacial bane]

Whats updog

[stable crane]

nm u

[zinc flax]

https://media.discordapp.net/attachments/1066655353158250548/1109007161008336958/ezgif.com-video-to-gif.gif

Fell for it

[pure oyster]

@stable crane you should rip all the unit art and character paperdolls from a phone game I played a while back https://bravefrontierglobal.fandom.com/wiki/Brave_Frontier_Wiki

[glacial bane]

what does gyokai do in generating?

[pure oyster]

I think those characters would fit your style

[glacial bane]

Specializes ofc

[glacial bane]

Who me?

[zinc flax]

Gyokai is the artists name

[pure oyster]

nope

[glacial bane]

I like it.

Gives off a vibrant color vibe to it

[glacial bane]

Whats with the background looking so perfect

[stable crane]

I had a look but I don't love the artstyle

[glacial bane]

U using a lora on the background?

[stable crane]

nope

lora on the whole thing yes

[glacial bane]

Btw send the gyokai lora

[stable crane]

it's on civit

[glacial bane]

Ah yess civitai i love that site so much

Main reason why i got better in generating

Btw whats a lycoris?

[stable crane]

this is why I like border

adds more depth / interest to the composition

[pure oyster]

An alternative to LoRA for hipsters

[glacial bane]

Whats the strength on your gyokai lora

[glacial bane]

Eh?

[stable crane]

1

[glacial bane]

Nice hands again whats the model.

[glacial bane]

The svda you use give off a old anime style vibe

[zinc flax]

Yeah unfortunately I forgot what I used for that merge
But I remember using 3A just for the mbw hands

[glacial bane]

My favorite candy

[supple raptor]

I'm not sure what you mean by svda. If you mean which model, I use my own mix https://civitai.com/models/100505/kakigori

[glacial bane]

Im saying that image guves off 19s anime vibes

[supple raptor]

It's a flatter style. I'm not a big fan of the 2.5d styles that many of the popular models like AOM have.

[stable crane]

I don't know how to tell you this, but Rem is the best waifu

[pure oyster]

Who is Rem?

[stable crane]

re:zero

[zinc flax]

I wanna see a new orangemix but flat tbh

[supple raptor]

I've never seen 2b spelled that way before

https://cdn.discordapp.com/emojis/1110837589411385436.webp?size=64&name=thinkeroo&quality=lossless

[stable crane]

can you pick 2B up in your arms and carry her? I think not

[stable crane]

chiB 2B

[pure oyster]

I actually never tried 2b in my usual model, I'll try that

[stable crane]

are you doing gens on a greenscreen bg?

cause I see you have a white border

[supple raptor]

Nah, just using simple background in the prompt then cutout around them. I think those ones I cutout manually then added the white border IIRC. I've had them for a while.

[stable crane]

@zinc flax this is 3A too but I don't think it looks like 3A

[supple raptor]

I like doing the white outline because otherwise they can look bad on discord's AMOLED dark mode on phones.

[stable crane]

it's smart

the white outline always looks good

[supple raptor]

I should have done a little thicker white border on my coffee girl cutouts.

[zinc flax]

Didn't expect it could look so different

Interesting stuff

[zinc flax]

looks like she's having a great time with that coffee

[supple raptor]

Always!

[stable crane]

whiskey in that coffee

[supple raptor]

yesh

I love these ones they're from controlnet depthmaps of a bellpepper and a cabbage XD

[zinc flax]

she big

[stable crane]

amazing

[supple raptor]

I got drunk one night and spammed the touhou server with chibis made from vegetable depth maps

[rough oxide]

been exploring the potential of anime with SDXL, seems way more expressive than current anime models can be, but a lot harder to get a good gen

[glacial bane]

Im scared

[stable crane]

she needs human nourishment

your blood

[supple raptor]

queue sad catgirl dance meme music

[stable crane]

that's very cute

[supple raptor]

That one was from ToonYou-JP

[stable crane]

raw sdxl?

thanks for sharing

[rough oxide]

yup 0.9, no embeds

[stable crane]

maybe we'll get niji quality facial expressions

[glacial bane]

what kind of nourishment

[pure oyster]

Banana monster

[supple raptor]

Thanks SD, I hate it.

[pure oyster]

That ghost looks like something out of The Curse of Monkey Island

[supple raptor]

Just gonna overlook that cursed baby she's holding that was supposed to be a baby dragon, huh?

[pure oyster]

All babies are ugly, so I didn't notice

[stable crane]

nearly perfect hands

[supple raptor]

She just got stubby thumbs. That's not that uncommon in people IRL.

[zinc flax]

Oo fennec ears

[zinc flax]

rip elf ears

[stable crane]

I have an unnatural hatred for elf ears

unless they're on an egirl

then I love them

[chrome crystal]

Does anyone know any models or merges that handle dark skinned characters well / consistently?

[zinc flax]

Try Aurora

[chrome crystal]

How well does it handle darker skin tones, such as those found in many africans and south asians?

[supple raptor]

Kakigori does dark skin pretty well.

[chrome crystal]

hmm

[zinc flax]

No clue about darker skin tone

[supple raptor]

darker than this just isn't common in anime art styles so you're going to struggle getting it consistently.

anime really doesn't represent how real people look

There's an obsidian skin lora on Civitai, but that's for dark gray to black skin. It might work if you used it at low strength though.

[pure oyster]

🥵 🌶️ ♨️

[pure oyster]

I'm a cheater, so I do half my shit in photoshop, including skin tone of my Lebanese based OC

[supple raptor]

I was very surprised the first time I genned with no prompt on my model 2/4 images came out with dark skin.

[chrome crystal]

I find that Niako_V3 is the only anime model I have used that can consistently do very dark skinned characters, although even that requires a fair bit of messing with weights of stuff.

[supple raptor]

the ones I posted just had 1girl, dark-skinned female, dark skin at the beginning with like 5-6 wildcards after that.

[chrome crystal]

Was that with the lora mentioned previously, or just a model?

I find the bulk of the models I have experimented with can't even do what you posted

[supple raptor]

just the model. no loras were used. They all have PNG info if you want to look at the specific prompts that the wildcards gave.

Here's my model: https://civitai.com/models/100505/kakigori

[chrome crystal]

they do really pale skinned charactes, and some tanned characters. Some can't output darker skinned characters even with massive weights (over 1.8) and loras.

[supple raptor]

AI models already tend towards being very heavily biased towards light skin due to available datasets and that's even more true for anime models.

[pure oyster]

Might be using models too much in the weeb variety.

[chrome crystal]

Mostly just shopping around with lots of them to try to get a feel. Was only mentioning the worst offenders I have seen.

[supple raptor]

You could always try finding some models that were specifically trained on dark skin and experiment with using MBW to add the skintone to an anime model that you like.

[chrome crystal]

MBW?

[gloomy scroll]

I'll just quickly assert dominance

[supple raptor]

merge block weighted. A more granular approach to merging models.

[supple raptor]

https://tenor.com/view/ホロライブ-hololive-anime-t-pose-watame-gif-19807377

[chrome crystal]

Ah. I will give it a look. I haven't done any merging since the early merges

[supple raptor]

MBW is quite the rabbit hole to fall down. It's a very powerful tool, but also quite complicated. I suggest using the SuperMerger extension for it as you will need to try many merges to get the results you want and SuperMerger will let you do that without saving the model until you're happy with the results.

[supple raptor]

too much cute. i died

[zinc flax]

stranded

[peak night]

@supple raptorhow can i use the model you use

[supple raptor]

Just download it. I share the stuff I make. https://civitai.com/models/100505/kakigori

[peak night]

is there a way i can use it online?

[supple raptor]

You can use Google Colab to run SD and use the model in that.

https://colab.research.google.com/drive/1wEa-tS10h4LlDykd87TF5zzpXIIQoCmq

[peak night]

k

[supple raptor]

For that colab you just need to get the download link to my model by right clicking the download button and clicking copy link address, then paste that into the custom_url field.

[gloomy scroll]

Welcome to Vil's chibi wall

[peak night]

Does it download it to pc or online

[supple raptor]

Colab is completely online. It uses Google's hardware. You can use it for free with some limitations to amount of time per day.

[peak night]

I pasted the link but nothing happened

[peak night]

Nvm I started it

But idk how to generate images

[craggy turret]

https://cdn.discordapp.com/attachments/762667574579429426/1129790246545014834/tmpn9blflys.png

[cyan burrow]

So I got this ready, made it in Clip Studio Paint with the exception of the background. Trying to figure out how to run it through img2img, or if I want to first line art convert it

[stable crane]

looks like art is complete to me 😄

[cyan burrow]

pfft

I'm gonna have to sketch out his pants then

then I took some arms and put them in a frozen place and erased the people there. Added some effects and brushing. the original images for arms and background is there too in the final two

now I just need to tone it, line art it, shade it and throw it into img2img

[stable crane]

3rd image is final result?

[cyan burrow]

no

[stable crane]

oh 1st img

that's cool

[cyan burrow]

thats where I took them from

I need to start img2imging

with the first

ah I like my submission this week

[steady grail]

very cool

what's the theme of this week's submission?

[gloomy scroll]

#1087493421209485393

[nova remnant]

Hi, very awesome anime pictures. I really like them. And they are terribly cute.

[nova remnant]

Hello very nice anime.

[brittle jay]

damn boy

[brittle jay]

epic sticker, just like all your others

everytime I see those stickers

[steady grail]

@gloomy scroll what vae do you use?

[gloomy scroll]

kl-f8-anime2

[steady grail]

ty

[worthy delta]

i dont want the angle all zoomed in

how do i fix dat?

[zinc flax]

Try cowboy shot?

[worthy delta]

gonna try it

yup it works