🍥｜anime

[gusty palm]

something about latent thingies

if i disable the last 2

[nova remnant]

ok i see 119 count

[nova remnant]

https://tenor.com/view/cat-scared-surprised-gif-24424983

No spaces between prompts

I'm going insane

[gusty palm]

spaces waste token!!

[nova remnant]

But so ugly

!!!

[gusty palm]

i agree

[nova remnant]

this is with negative word prompts.. im gonna take that seed and use two TI .. let me generate that one

[gusty palm]

i wouldnt even knwo if not for this

[nova remnant]

same seed with EasyNegativeV2, bad-hands-5

[gusty palm]

quite drastic

[nova remnant]

how ... can you explain?

the fingers didnt necessarily get better

[gusty palm]

i thoguht the difference was between the two images it seems like the difference is alot

[nova remnant]

ahh yes

but in terms of texture, quality, i don't see much noticeable difference

actually .... let me generate a few more exmaples

[nova remnant]

Do closer ups maybe

the TI kinda has better output somehow

[gusty palm]

heres what ng_deepnegative_v1_75t, did for me

[nova remnant]

Buttafly

[gusty palm]

actually ruined my wings lol

but it fiexed the legs

[nova remnant]

And ghost hair

[gusty palm]

oof

[nova remnant]

ng deepnegative vs bad hand ... what's the difference or the same?

[gusty palm]

i think its a merge

im not sure but its supposed to be better at avoiding bad hands

[nova remnant]

ok

[gusty palm]

heres fastnegative wit no deepnegative

[nova remnant]

can you share your metadata pls?

[gusty palm]

sure

[nova remnant]

i want to try it out with easynegative and bad hand

[gusty palm]

high resolution,best quality,masterpiece,(cute moth girl:1.3),(giant winged moth girl:1.3),(duotone moth wings:1.3),upskirt,sitting on a bean bag cushion,futuristic coffee shop,classical victorian gothic outfit,shy smile,pastel hair,idol princess makeup,mature teen,(Cutester Art:1.3),Queercore Art, Negative prompt: low quality,low resolution,worst quality,easynegative,bad_prompt_version2-neg,verybadimagenegative_v1.3,FastNegativeV2 Steps: 24, Sampler: Restart, CFG scale: 15, Seed: 69, Size: 768x384, Model hash: 71236b0efb, Model: lottamix_v20, VAE hash: 551eac7037, VAE: pastel-waifu-diffusion.vae.pt, Clip skip: 2, NGMS: 0.5, TI hashes: "EasyNegative: 66a7279a88dd, bad_prompt_version2-neg: 6f35e7dd816a, verybadimagenegative_v1.3: d70463f87042, FastNegativeV2: a7465e7cc2a2", Pad conds: True, Version: v1.6.0

[nova remnant]

ohh you are using other negative prompts besides the TI

[gusty palm]

heres with pad prompt off

[nova remnant]

i'll try this one

pos: high resolution,best quality,masterpiece,(cute moth girl:1.3),(giant winged moth girl:1.3),(duotone moth wings:1.3),upskirt,sitting on a bean bag cushion,futuristic coffee shop,classical victorian gothic outfit,shy smile,pastel hair,idol princess makeup,mature teen,(Cutester Art:1.3),Queercore Art,

neg: EasyNegativeV2, bad-hands-5

[gusty palm]

from the civitai page

[nova remnant]

hmm

this is kinda complex

[lyric zealot]

what are yall comparing?

[nova remnant]

hi smol sus

[lyric zealot]

[nova remnant]

trying to decide which is better TI or negative prompts in words

i mean i have no control over when using TI but then im seeing some changes with TI that are not as good with negative worded prompts

[lyric zealot]

I guess it really depends on the model, but probably a mixed of both

[nova remnant]

yeah ...guess i will try out several different examples

would be nice if we could see the triggers words in TI

im liking the combination of EasyNegativeV2, bad-hands-5

not using anything else in NP

lol those images also has me little biased

but definitely the background detail are much nicer with those TI than what i was getting putting in negative prompts manually

[gusty palm]

whats the difference between these

[nova remnant]

pickle tensors could be malicious

not necessarily but has the possibility for malicous code in them

[gusty palm]

is the bigger one better or is that just for training

[lyric zealot]

always use safetensor if possible

[lyric zealot]

the smaller model is likely pruned

[nova remnant]

i usually go for pruned, easier on space and can give decent output

[gusty palm]

https://civitai.com/models/87781/boringe621-negative-embedding-enhance-images-stylistically-and-topically?modelVersionId=94126

this looks intrasing

[nova remnant]

some trigger words in those TI are definitely doing some underlying stuff that make background detail look whole lot better

[nova remnant]

Very cool

I want a comfy chair like that

And a view to the inevitable destruction of Earth

[nova remnant]

(KHFB, AuroraNegative:1.2), negative_hand-neg, (Worst Quality, Low Quality:1.2), border, grayscale, (watermark, multiple views:1.2)

What I'm liking

[gusty palm]

i needa try khfb and auraroranegative

[nova remnant]

thank, will take a look at them, but now im off to sleep, gn

👍

[lyric zealot]

goodnight

[muted river]

Hello

Hello

[lyric zealot]

hi

[gloomy scroll]

https://tenor.com/view/anime-grab-sazanka-hand-kunoichi-tsubaki-no-mune-no-uchi-in-the-heart-of-kunoichi-tsubaki-gif-25767790

[lyric zealot]

Metadata

[lyric zealot]

@sweet river try doing it with like maybe tie or something

or move bowtie towards the front of your neg prompt

[gloomy scroll]

@lyric zealot link for model?

[lyric zealot]

i just started merging

but recipe is basically

amnit and expmixlinev3 at 0.5

[gloomy scroll]

[lyric zealot]

it's surely inside one of those skyscappers in the background

[gloomy scroll]

I need to continue working but I don't want to

[umbral depot]

I need to sleep

[umbral depot]

Indeed

[gloomy scroll]

I have identical image

[lyric zealot]

she's angy

[lyric zealot]

cute tho

[dawn jasper]

The bunny hood is strong with this one.

[dawn jasper]

I dunno if she's meant to be kissably expressive or expressively kissable.

[dawn jasper]

Anyone feels up to do images for this woman with an Hourglass Frame and Natural H-Cup knockers if I may ask?

DESCRIPTION: "With very dark red hair, a thin waist and a black skintight suit and heeled shoes, Goldcorn would be 6'0 feet tall without the heels and 6'3 feet tall with them on, she'd usually carry a pair of reading glasses or a set of goggles depending on whether she's studying or building something.
She was reading an ancient tomb with almost impossible to decipher text when she had a eureka moment. As she unraveled the ancient mysteries it held she felt divine power awakening within herself."

How's that description?

[dawn jasper]

Thanks, I may have a question for you at private messages, you okay with hearing about it, there?

[lyric zealot]

for sure

[dawn jasper]

Hot for Teacher, much?

I don't mean that in bad ways though.

[nova remnant]

@lyric zealot what positives do you like using on all images

[lyric zealot]

this is usually my first line: (masterpiece, best quality:1.2), detailed background, hyper-realistic, extremely detailed,

just changed from highly detailed to extremely detailed, not sure if there's any effect tbh

[nova remnant]

👍

[nova remnant]

do you use hyper realistic for everything ?

[lyric zealot]

yup, for most of the models that im using

even tho it's hyper realistic in prompt, but it's not realistic at all

[nova remnant]

interdasting

[lyric zealot]

but for some model, like holokukiv, i will use it at 0.6 strength

[nova remnant]

Do you always put it in the negatives

So you can easily edit

[lyric zealot]

meaning??

I don't put hyper-realisitc in negatives

[nova remnant]

oh ye thats for positive

Do you leave them in the box ? >.>

[lyric zealot]

i mean, i put it at low strength for like certain model, because i want to introduce some realistic elements to it, but not too strong

if that's what you are asking

[nova remnant]

[lyric zealot]

if you want to, you can probably put something to strength zero, to keep it in prompt, but remove the elements of it

probably

actually let me test that

[nova remnant]

What does that mean exactly

wait

like (Realistic:0) ?

[lyric zealot]

there might be a different tho

[lyric zealot]

yup

[nova remnant]

the one with more dots on the shirt looks like it has more shadows

is that the one with realistic 0 ?

[lyric zealot]

that one is the one with no "sitting" in prompt

i was testing with sitting

so first was (sitting:0)

[nova remnant]

oh you did sitting

Hmm strange

[lyric zealot]

there's definitely changes to it

could it be because my sitting prompt isn't at the end?

[dawn jasper]

The Rinner from Fate/Extra, who's a Niece to the original Rinner?

[nova remnant]

I think we're going crazy

[nova remnant]

[lyric zealot]

[nova remnant]

Nun in training

This is fine, reading a book in the winter with no roof

@lyric zealot Do you think that "4k, 8k" in prompts help ? >.>

[lyric zealot]

no idea, i think depends on model

best will be to do plot xyz

[nova remnant]

you know what

i'll try it :D

@lyric zealot help

[lyric zealot]

[nova remnant]

How do I do it

Like

[lyric zealot]

actually i was thinking

maybe using prompt matrix is gonna be easier

so i was reading up how to do that, cos i seldom use that

and so here

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#prompt-matrix

[nova remnant]

warm clothes, candle, snow on the window, |(intricate, best quality, masterpiece:1.2), (extremely detailed)|(8k, ultra detailed, 4k, intricate, best quality, masterpiece:1.2), (extremely detailed, 4k wallpaper)

maybe this is gonna work

I dunno exactly

might not need the first |

[lyric zealot]

yup

[nova remnant]

oh "Prompt matrix will create 4 images using a total of 4 batches."

seems right then

maybe D:

[lyric zealot]

i guess if you wanna test some keyword

like example

4k|8k|16k

instead of swapping too many things

[nova remnant]

yeah I wanted to test having all those 4k and ultra detailed and not having them, keeping it simple

[lyric zealot]

so using my example

the "4k" is included in all prompts

[nova remnant]

oh

I understand now

I got this

[lyric zealot]

Prompt matrix will create 32 images using a total of 32 batches.

[nova remnant]

Lmao

[lyric zealot]

better be the best matrix

[nova remnant]

better than the movie itself

@lyric zealot hmmm how do I make it only do 2 images

Tried "whole prompt here, (intricate, best quality, masterpiece:1.2), (extremely detailed)|(8k, ultra detailed, 4k, 4k wallpaper)|"

[lyric zealot]

i think just one |

remove the last |

[nova remnant]

oh okay

The | acts kinda of like a possible ,

[lyric zealot]

i think of it as addon

whatever comes after |, can be taken as an addon

like a subway sandwich

okay, i feel like a lot of mine quality prompts are useless

[nova remnant]

we're seeing the lie that we believed in for so long

[lyric zealot]

welp

i think it depends on model

but i was too lazy to test it out

lmao

[nova remnant]

@lyric zealot

Why did mine generate 4 but says nothing on the left D:

I think it added a blank token

[lyric zealot]

how many | did you have?

[nova remnant]

it was exactly like this

[lyric zealot]

then might be 2 different seeds

no idea

[nova remnant]

I'll try removing the last |

and it did a better image with the 4k wallpaper 8k on my model

[lyric zealot]

oh right

you have to remove the last |

[lyric zealot]

and only have one | like i said here

[nova remnant]

they are so different

[lyric zealot]

it works for your model

bro, this model is lazy

doesnt wanna change stuff

[nova remnant]

if you put that in the prompt not using matrix does it work ?

[lyric zealot]

what do you mean?

[nova remnant]

if you do album cover not using the matrix, does it change ?

[lyric zealot]

will be the same as below left

[nova remnant]

maybe those prompts aren't super strong

to change the image noticeably

[glacial bane]

she looks oddly familiar you sure thats a random girl?

position name?

[lyric zealot]

yup, or at least i think it's just some random girl

[lyric zealot]

wariza

[lyric zealot]

basically w-sitting

[glacial bane]

ty

[glacial bane]

ye girls with that haircut just reminds me of many characters thx btw on the pose

[lyric zealot]

true true

[nova remnant]

Just remembered I have sinned please take care of me

[glacial bane]

i remembered i live next to a church that takes 10 steps to get inside

and i also remember i dont have to go to church because i live so close to the church i can hear the father

and i just attend the church in the house

[lyric zealot]

confess your sins now

[nova remnant]

I make too many fox girls

also I think the 8k stuff is better

depth of field looks nicer, some clothing details seem to be more precise and other stuff like that

[lyric zealot]

yup, the right definitely is better

[lyric zealot]

[nova remnant]

ear intensifies xd

[supple raptor]

beeg ear fox girls are superior

[nova remnant]

[nova remnant]

There are tiny foxes on the grass!

[supple raptor]

she has frens

[nova remnant]

Whats this for... lmao

[supple raptor]

idk. just a smol cat

[dawn jasper]

Amazing expressions.

[nova remnant]

Does this make a difference ?
1: (extremely detailed), (8k, 4k, 4k wallpaper)
2: (extremely detailed, 8k, 4k, 4k wallpaper)

[supple raptor]

Probably a very small difference if you're comparing them side by side, but not in a way that makes either option better or worse.

Also, rather than manually typing () around the words you want to add emphasis to, just hit CTRL+Up Arrow to add emphasis or CTRL+Down Arrow to reduce emphasis.

[nova remnant]

I discovered that by accident a while ago

[lyric zealot]

i just learnt today that you can move prompt left and right by alt + left arrow or right arrow

[nova remnant]

thats black magic

[supple raptor]

it will auto select the closest word to the cursor and if you want to select multiple words, just highlight them first.

[supple raptor]

oh that's neat

[lyric zealot]

i was like just looking thru the whole settings tab

[nova remnant]

ctrl + shift and arrows to select :D

[lyric zealot]

I didn't even know there's theme native to auto1111

[nova remnant]

I think I should stop, its a bit too hot to use SD I think

its slowing down

[lyric zealot]

if you have the multidiffusion extension, there's a free gpu button that might help

[supple raptor]

sed cause she have 4 ears

[dawn jasper]

Quite beautiful, really.

[lyric zealot]

superior hearing

[nova remnant]

Its 35°c outside, my room is hell

[supple raptor]

now that I look closer. she might also be missing her shins. I thought her legs were just bent at first

[nova remnant]

cursed

[lyric zealot]

I'm not gonna make any joke about this

[supple raptor]

https://cdn.discordapp.com/emojis/1091965013859766352.gif?size=64&name=1girlrun&quality=lossless

[nova remnant]

Which one ?

[lyric zealot]

left: better mouth imo
right: better iris imo

[nova remnant]

What would you say about the hair ? >.>

[supple raptor]

the ear shadow on the right one is a bit wonky

[lyric zealot]

can't see much difference, except for a certain part merging together for left pic

[supple raptor]

though, both ear shadows have the wrong angle.

[lyric zealot]

anyway nubby

this week nintendo event, might have the next monster hunter

too lazy to inpaint

[lyric zealot]

wait, i meant capcom

[steady grail]

honestly, i feel 4k is ok, but 8k is pretty sus. how many trained their models with 8k images?

[nova remnant]

I struggle to comprehend where they are wrong

[steady grail]

4k are already more common around

[nova remnant]

Yeah i saw that bit

But the rest looks in place fine >.>

[lyric zealot]

[lyric zealot]

i guess so

[nova remnant]

I decided that my final prompt will be "(intricate, best quality, masterpiece:1.2), (extremely detailed, 8k, 4k, 4k wallpaper)" yay

[lyric zealot]

i forget is bracket 1.1 or 1.2

i think is 1.1

[nova remnant]

1.1

[lyric zealot]

[nova remnant]

1^1 or something something math D:

[brittle jay]

i bought a usb drive and an adapter frlm a to c and then I etched the steamos iso/ into it frlm my phone and then I re-imaged my steamdeck and nowmit is fixed

[lyric zealot]

https://tenor.com/view/math-calculate-confusing-figure-out-gif-6237717

[supple raptor]

most of the image is lit with the light source coming from her left almost directly beside her, but the ear shadow is cast as if the the light source was behind her.

[brittle jay]

it is so harx to wfite on this thing

like bruh

[lyric zealot]

jesus

time to get a bluetooth keyboard

[brittle jay]

haha

I'll just install a few plugins'

anyways

gn

fucking 3am

[lyric zealot]

nighty night

[nova remnant]

[supple raptor]

rip.... just noticed extra hand

[nova remnant]

Me right now

[lyric zealot]

[nova remnant]

please

Metadata

[lyric zealot]

metadata

watch is totally rng lmao

she got a toothache

[stable crane]

lmao

[stable crane]

but then you're missing out on cute nuns like this 😢

[lyric zealot]

amen

[dawn jasper]

Wonder what's eating her.

[lyric zealot]

probably thinking about me

[nova remnant]

[magic wave]

New waifu?

[lyric zealot]

[magic wave]

Very kwoot

[lyric zealot]

[magic wave]

I wanna get a new OC too but I dunno what to make

[lyric zealot]

steal people OC

[supple raptor]

don't worry, dentist Miku is here to help

[lyric zealot]

[dawn jasper]

Hmm.

[magic wave]

[supple raptor]

[supple raptor]

high-waist skirt is cute af on miku

[glossy eagle]

I still have not been able to get nvml to work on Linux so I can’t use rocm and my it/s is trash

[lyric zealot]

sorry, don't know much about linux setup for SD. You can go over to #🤝｜tech-support to get some help tho

[lyric zealot]

@gritty spruce

@wooden coral

@mellow otter

@dull jackal

well out of luck

[wispy canopy]

have a realistic venonat and rapidash. they're from anime, after all ^^

[dawn jasper]

She looks a bit like a miniature Sakura Matou for some reason.

[lyric zealot]

there's some similarities

[pale monolith]

[lyric zealot]

[pale monolith]

👋

[gloomy scroll]

[lyric zealot]

she's melting cos the weather is too hot

[oak spoke]

https://tenor.com/view/pingu-vegetables-straw-gif-15019335

[dawn jasper]

If I may ask, anyone want to take a shot at doing bikini pics for Goldcorn?

[dawn jasper]

This lady looks like she's going to a festival.

[summer night]

how does the refiner work?

does it just switch models for the current generation?

i'm wondering what the practical applications for this are

[nova remnant]

If you have some model that has bad anatomy it can switch to yours after the first few steps so it hopefully gets them correctly in place

[native halo]

base model creates the image then when image is like 80% done it sends it to refiner that adds final details

[nova remnant]

Something something

[native halo]

similar to latent upscale in high res fix

[nova remnant]

[nova remnant]

Did you do something to get that shape with the image inside

Or is it prompt only

[bitter latch]

HMMM?

[bitter latch]

what ya mean?

[nova remnant]

Is that image pure text 2 image?

[bitter latch]

yeah

wellll, with some lora's, hypertext, and embeddings for negitives

[dawn jasper]

Powerful image there.

[bitter latch]

yeah I went overboard I thinks but works Time Stream, Red Lensed overlay effect, shattered, 3.5D, cutout, transparent, white background, rice paper, elegant detail, extra details, dry water color <lora:Ink scenery:0.5>hypernet:axsensInspired_axsensV7:0/6magic touches, hypernet:henriRousseau_v1:0.6space and time effects, Ink cyber punk style, back lit hexagonal backdrop/panel, Vibrant diamond ink style, night time, stars and nebula, 3.5D, unreal render, cyber movie poster, holographic, fine art, soft colors, black outline style, elegant, very detailed Cel-Shaded With shaded detail's, soft shadows, nomanssky style, elegant, very detailed, hi-res, HDR, 4k, detailed shadows, elegantly shaded, fine art, full body portrait, alternate universe Cyber, cybernetic implants, fine details, (cyberpunk overtones:0.6), elegant, very detailed, hi-res, HDR, 4k, detailed shadows, elegantly shaded, fine art, fine art, soft colors, black outline style, elegant, very detailed Cel-Shaded With shaded detail's, soft shadows, Vibrant diamond ink style, night time, stars and nebula, 3.5D, unreal render, stumbling over a rock in a stream, cute, adorable, cyber movie poster, holographic, fine art, soft colors, black outline style, elegant, very detailed Cel-Shaded With shaded detail's,
Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name, one tail, no extra legs,
Steps: 20, Sampler: Restart, CFG scale: 4, Seed: 2222278129, Size: 512x768, Model: aZovyaRPGArtistTools, VAE: counterfeitV25VAE_v25.pt, Lora hashes: "Ink scenery: 26343a3a36a0", Version: v1.6.0

bunch of styles 😄

[nova remnant]

[native halo]

[nova remnant]

I don't know exactly what a hyper text is

[silent kraken]

Your username kills me everytime

[nova remnant]

[bitter latch]

thats better

[summer night]

huh, when using hires fix and refiner at the same time, it seems like the refiner kicks in during the hires fix steps instead of the sampling steps

[bitter latch]

not sure if she's going to battle or a concert?

uses same seed, flips the other way and presses go

[lyric zealot]

are you using sdxl?

[summer night]

no

[bitter latch]

well.. ok...

[nova remnant]

New type of cute girl?!

[lyric zealot]

then why are you using the refiner? To change model midway thru generation?

[summer night]

i'm just playing around with it

[bitter latch]

there was not a bike mentioned anywhere the prompt.. i'm confused

[lyric zealot]

ahhhh got it

what's your refiner setting?

[summer night]

0.8

just the default

right now i'm running a X/Y plot using the same 4 models for the base checkpoint and refiner checkpoint

[summer night]

i'm not sure how much changing the model during the last 3 hires steps will change the look though

[bitter latch]

it was the seed

[summer night]

this is my usual hires settings

[lyric zealot]

if it's at 0.8, then refiner will only kick in at your last 20% of the steps. Example 20 steps, it will only be your last 4 steps for refiner

[native halo]

why is your denoising strenght so high

[bitter latch]

or not...

[summer night]

that's high?

[native halo]

yea cause you are not using latent

[bitter latch]

removed my cyber punk style and oooh

[summer night]

miku having a sephirah meltdown

[bitter latch]

is very happy with these :D

guess who? hehe

Miss Frizzle

[bitter latch]

umm..

[dawn jasper]

Whoever she is, she's strong.

[bitter latch]

hmm someone recommend a anime girl that isn't miku please

[dawn jasper]

Akiha Tohno from Tsukihime, or Kaguya Yamai of Date-a-Live fame, perhaps?

[bitter latch]

ok

Gotta say the counterfeit model and vae do good

Miku Tohno

[dawn jasper]

A Blue-haired Akiha Tohno?

[bitter latch]

doh

hhe I thnk I got it

gets some foodage.

[summer night]

does anyone know what the danbooru tag is for hair that is a different color on the inside

two-tone hair is too vague

i found it, it's colored inner hair

[bitter latch]

hmm i'll try it 😄

not what I wanted but ok

Anime Elvis

[dawn jasper]

Who is this?

[bitter latch]

anime elvis

[bitter latch]

miku, tora tora tora

😄

[dawn jasper]

That's a Miku?

[bitter latch]

which one?

i'm so confused

[bitter latch]

oh sorry just random I think

[dawn jasper]

Pink haired lady.

[dawn jasper]

@pale monolith How goes life at your side?

[nova remnant]

whats the secret to making your anime look so tidy and cute?

[lyric zealot]

imma be honest, i have no idea, but here's the ***metadata ***

[lyric zealot]

she got scammed for such a small cup of coffee but she's still smiling

[native halo]

what cup of coffee i see nothing there

[lyric zealot]

that cup of coffee looks super small, no?

[native halo]

sorry i got lost in her eyes

[astral drift]

Noice

[lyric zealot]

time to sleep

[summer night]

what do i copy if i want to migrate all my settings from one installation of auto1111 to another?

[nova remnant]

not 100% sure but prolly config.json and ui-config.json

[bitter latch]

I am liking this style i'm making

hm maybe wrong model.. . or sampler ?

[summer night]

stable diffusion on paperspace is pretty good

$8 a month for access to a 16 GB VRAM gpu, but only 15 GB of persistant storage so you have to re-aquire the models each time

[summer night]

it's just the fast-stable-diffusion collab but on paperspace

[bitter latch]

oh gtf away, don't care

[native halo]

that kinda sucks my model folder is 800gb lmao,but for lora trainin maybe 🧐

[bitter latch]

@sunny got a spammer

[summer night]

i mean you can just upload whatever model you are actively using at the time

[bitter latch]

15 gb is chump change.. chump

quit spamming.. we don't care

[summer night]

you are a clown and this is the only time i will acknowledge you

[bitter latch]

good for you

there's rules against spamming ya dip. so. yeah.. stop already. thanks and good night

[summer night]

so anyways @native halo it beats generating on my 1070, with the A4000 on the notebook it's just like ziiip done

you can go over 15GB but it charges you another $0.29 monthly for each GB over 15

if nothing else it's useful for trying out a bunch of different setting combinations

[bitter latch]

wow.. ok..

[native halo]

if u want 16gb vram for cheap u could try with an used quadro p5000 like 150$ and performs a little bit better than a 1080

[nova remnant]

16gb vram for only $150?

[summer night]

i was planning on getting a used RTX 3070 for somewhere around $300 but i'm not sure how good it would be for generating with AI

[nova remnant]

oh you said 'used'..

[native halo]

yea plus quadro is old architecture

rtx better if u can afford one

[bitter latch]

oh gawd those hands

[nova remnant]

i kinda spent a mini bundle on my rtx 3060

cost me around $400

and its a 12gb card

[native halo]

yea 3060 is a beast and its only like 200$ used

[nova remnant]

i bought mine new with warranty, didn't want to risk buying one that might have been used for mining

[native halo]

yea sometimes u get unlucky and card dies within a year 😔

[nova remnant]

oooh, fingers crossed

the other nvidia cards i had before with 4gb and 6gb performed well for years, and they are still in good condition lying around idle

[summer night]

incase anyone didn't know 910x512 is a 16:9 aspect ratio

[nova remnant]

i use an easy click ext for that

[summer night]

this one is apparently depreciated

[nova remnant]

its called sd-webui-aspect-ratio-helper

[summer night]

i replaced it with this one
https://github.com/thomasasfk/sd-webui-aspect-ratio-helper

[nova remnant]

why is that depreceted?

[summer night]

idk the console kept yelling at me

[native halo]

yea its broken with 1,6

[nova remnant]

you are using the same one i just typed out

[summer night]

this one looks like this

[summer night]

i'm sure this one is this
https://github.com/alemelis/sd-webui-ar

[nova remnant]

you are right, but i use the first one along with the other

[summer night]

oh i'm blind

[summer night]

my eyes immediately tracked to the one on the bottom and i didn't bother looking anywhere else

[nova remnant]

no its fine

i posted the bottom one alone first lol

[summer night]

but like when i was using that one, i kept getting warnings like this warning i have for two-shot

[nova remnant]

ahh .. havent had issues with mine

[summer night]

is your webui on version 1.6.0?

[nova remnant]

yes

[summer night]

huh

[summer night]

well i have warnings for two-shot and it still works

the only thing i've noticed that's broken is the ui elements

[nova remnant]

nope no warning messages, and i dont suppress checks

[summer night]

i don't understand regional prompter so i just keep using latent couple

[nova remnant]

oh hey, i didnt know about latent couples

im only using prompts .. no ext

on a regular 768 canvas scaled up by 1.5

the ideas with prompting started yesterday when someone typed #lesbian here and i think steven posted some cute 2 girl images on the fly, and i liked the idea too

[summer night]

latent couple (and probably regional prompter) is good for generating images with multiple distinct characters

[nova remnant]

yes i know now but didnt know and not using any extension for it

[summer night]

honestly i need to learn how to use regional prompter

[nova remnant]

i think a lot of the complex prompts out there are likely generated with some ai tools

i try to play with my own prompts which is kinda more fun for me

[summer night]

i just use danbooru tags for prompts

it works since pretty much all of the anime models are trained with danbooru tags

[nova remnant]

i do that too with images i like to use with CN then edit the prompts a bit

but on a regular scheme i just type out prompts with plain txt2img

[dawn jasper]

@pale monolith So how goes life?

[nova remnant]

was playing with this earlier... i like the extension i have to edit opnepose

[dawn jasper]

Is that a Mobile Suit?

[bitter latch]

sorta yeah

was suppose to be megatron mecha

ahh there's megatron

[dawn jasper]

@bitter latch Interesting, not gonna lie.
I take it you checked out Armored Core: The Series before?

[bitter latch]

sounds familiar

i've watched so much ,my brain isn't good at remembering names

but yeah I like mecha anime

[dawn jasper]

Call it cliche a question to ask, but are they about to kiss?

[nova remnant]

they are probably talking about beautiful life at the beach

but they on the other hand might ...

[dawn jasper]

We about to see a kiss?

[primal wharf]

Planetary annihilation magic

[dawn jasper]

This is powerful.

[summer night]

regional prompter kind of stinks

[summer night]

regional prompter is a bit easier to use than latent couple but it isn't as effective

[nova remnant]

hmm

would you suppose if i attempted to target the dress color of each of those girls, would that be regional prompting?

[native halo]

yea but if i wanted to change color of their dress it would probably be faster to do it in CSP rather than waiting to get a good image on img2img

[nova remnant]

ahh ok i think i fully understand the concept of regional prompting, and yeah i agree, i could pick an image i want and inpaint the colors which would be lot easier

loosely applied and Ai did the lifting

[summer night]

is there a setting to automatically delete images from the outputs folders after a certain amount?

like after 50 images or so it starts deleting the oldest ones to make room for newer ones

[nova remnant]

hi katie do you use bad-hands-5, ng_deepnegative_v1_75t both or the deep negative?

[gusty palm]

sometimes i use deepnegative but today im trying a different set of TI's

i have the badhands.pt thing but i dont use it enough to know if its worth using

o actually i did have deeppnetative on for those

[nova remnant]

im getting some weird anomaly with bad-hands-5 and using ng_deepnegative_v1_75t seems to improve the render

[supple raptor]

badhands5 changes style a lot and doesn't really help much with hands in my opinion. negative_hand is much better.

[gusty palm]

negative_hand-neg is the one i have

[nova remnant]

does that also fix deformed legs?

[gusty palm]

idk i havent even seen it fix hands but i havent really tested

theres a negative feet one though

id probalby opt for tokens like "extra legs" or missing legs

[supple raptor]

That's more of a problem with the model, prompt, or parameters, you're using.

[nova remnant]

see im trying out a woman, in a short white dress, riding a bicycle, on a dirt road, along a bushy mountain and her legs are coming out very odd

[supple raptor]

Yeah bikes are a pretty tough thing to gen. Which model are you using?

[nova remnant]

could be but i notice changes between a combination of using easynegative, negative_hand-neg, ng_deepnegative_v1_75t and easynegative, bad-hands-5 the first set seems to give better results

[gusty palm]

i would expect imperfections in the first render then maybe try to fix them in inpaint

[nova remnant]

mod Aniverse

ouch i never bother to use inpaint to fix blemishes

[inner pine]

https://youtu.be/pSlkAVlSSwQ?si=vqj9uw4SfvPKd1WR

[nova remnant]

btw whats the difference between negative_hand-neg and bad-hands-5 ?

[supple raptor]

I think you're relying too much on the negative prompt. And especially the negative embeddings. Using easynegative and ng_deepnegative together is like lobotomizing the model. Those are both quite heavy and will force their own style and poses.

[supple raptor]

Almost 6 months of knowledge for the person who trained them.

[nova remnant]

negative_hand-neg is better then?

[supple raptor]

Yeah

[nova remnant]

ok thanks

im guessing i should not be using those both together

[supple raptor]

I definitely wouldn't. I'm a firm believer in keeping the negative prompt minimal. Modern models are good enough that you don't need more than a few negative quality tags, and optionally 1 normal negative embed. By keeping the negative simple it will also allow you to use the negative prompt for what it's actually good for which is prompting out things you don't want that are being associated with something else in the prompt. Thinks like negative prompting claws when you want to gen claw pose.

[nova remnant]

sounds logical ... but help me understand this set of TI..

easynegative, negative_hand-neg, ng_deepnegative_v1_75t
easynegative covers a broad generic prompts, negative hand focuses on acute deformity in the hands and fingers, and deepnegative goes further into more acute details?

[supple raptor]

I just flat out wouldn't use easynegative at all. It is far too strong and heavily affects the model's style. I'm not as familiar with deepnegative, but I believe it's a bit better.

[gusty palm]

;o

[supple raptor]

Personally, I use KHFB, AuroraNegative if I use neg embeddings at all. I used negative_hand for a while, but have quit recently because I don't think it really helps that much.

[nova remnant]

okay, i think you made some compelling points about TI which would / could affect a model, im gonna try a basic prompt i compiled but one question, when using (highly detailed, ultra quality, masterpiece:1.3 in the positive box, is there a reason i should also use (low quality, worst quality:2) in the negative box?

[supple raptor]

Usually positive quality tags aren't really needed. I just tack masterpiece, best quality onto the end of the prompt. For negative prompt this is all I use: (Worst Quality, Low Quality:1.2), border, grayscale, multiple_girls, (watermark:1.2)

Going to 2 emphasis on your low quality worst quality is just making everything else in your negative less effective because emphasis is normalized in WebUI. If you're using Comfy, then that much emphasis is probably actively hurting your gens.

[nova remnant]

interesting

[supple raptor]

As a rule of thumb, if you need to take emphasis above 1.2-1.3 then either the tag is not effective and needs changed or the problem lies somewhere else in your prompt, with the order of your prompt, or with the model itself.

[nova remnant]

this is important thanks for explaining this

also about TI is that i have no idea what triggers words are used in them

so in this example i used the following prompts ...

a woman, in a short white dress, riding a bicycle, on a dirt road, along a bushy mountain, (highly detailed, ultra quality, masterpiece:1.3), vibrant colors, (beautiful, aesthetic:1.2), dynamic angle, deep focus, natural lighting, cinematic
And for negative prompt
((text, logo, title, signature, username)), bad anatomy, bad proportions, (disfigured, deformed, mutilated, malformed, mangled:1.2), long neck, bad hands, unnatural hands, (fused fingers, interlocked fingers, poorly rendered fingers:1.3), muscular, asymmetric, conjoined, distorted, duplicate, cloned, out of focus, cropped, (greyscale, monochrome, grainy, pixelated:1.2), jpeg artifacts, multiple views, morbid

[supple raptor]

TIs don't have trigger words because they bypass the tokenizer.

[gusty palm]

did this one without esaynegative and i like it

[supple raptor]

Cute

[nova remnant]

i mean trigger words like what prompt they affect or influence

[supple raptor]

That's not how TIs work. They don't use words at all. You can generate images with just the TI as your positive prompt and no negative prompt at all to get an idea of how they will affect your images. Prepare yourself for cursed images.

[nova remnant]

hmm, so how does TI function to affect prompting?

[gusty palm]

i thoguht tis were essentially just wildcard lists

[supple raptor]

Effectively a TI works as a new word that's added to your model.

[nova remnant]

i was under the impression TI were compiled words trained for specific task

[gusty palm]

does aruraroonegative and kfhb not have civitai cards

[supple raptor]

No, when you train a TI, you provide initialization words which help to give it a closer starting point to the desired output, but they are essentially just learning the average of the images they're trained on.

[supple raptor]

No, they're only on HuggingFace in the Aurora repository.

[gusty palm]

ok

[nova remnant]

so TI are trained on images rather than word prompts :0

[nova remnant]

What is latent couple

[summer night]

https://github.com/ashen-sensored/stable-diffusion-webui-two-shot

[nova remnant]

oh

Ohhhhh

I want that

[summer night]

basically lets you do things like this

[brittle jay]

wasn't there a better one?

[nova remnant]

@supple raptor

bad anatomy, bad proportions, disfigured, deformed, mutilated, malformed, mangled, long neck, bad hands, unnatural hands, (fused fingers, interlocked fingers, poorly rendered fingers:1.2), muscular, asymmetric, conjoined, distorted, duplicate, cloned, out of focus, cropped, (low quality, worst quality:1.3), oversaturated, underexposed, overexposed, greyscale, monochrome, grainy, pixelated, jpeg artifacts, multiple views, morbid, ((text, logo, title, signature, username))

[supple raptor]

Looks better than I would expect out of bike gens

[wispy canopy]

you all just need to switch to sdxl based models 😮

[summer night]

do you have to do anything special to use an SDXL model or does it just work like all the other models

[wispy canopy]

using a decent sdxl model: you just use it, no refiner, no negatives

[summer night]

do you need a card that like 5% of people have to generate with it?

[wispy canopy]

can't get the sign to say 1.5 xD

using comfyui, i think you can fit it in 8gb

[summer night]

idk what that is

[supple raptor]

Realistically you need a minimum of 8GB of VRAM. 6GB can technically do it, in comfy using some wizardry, but it's so slow that it isn't worth it.

[gusty palm]

i tried xl but it just took so long but i 6gb

[summer night]

i mean i have a 1070 8GB

[gusty palm]

but it did technically work

[summer night]

i got a paperspace subscription so i could run the fast-stable-diffusion notebook with cards like the A4000 and RTX 5000 which are 16GB

[wispy canopy]

my waifus are in support for 1.5 too

[supple raptor]

It'll be pretty slow on the 1070, but usable at least in ComfyUI. I'm not sure if Auto1111 got the VRAM consumption problems fixed or not. Paperspace pro is honestly a really good option if you spend a decent amount of time genning and don't see yourself upgrading your GPU soon.

[wispy canopy]

vram is horrible in a1111

i can do 8x1024x1024, but that consumes about all my vram

[summer night]

so can i just load up a SDXL checkpoint like any other checkpoint and it works? i don't need to like install anything else?

[wispy canopy]

make sure to be aon a1111 1.6

[supple raptor]

There was an actual bug early on that A1111 and (I think) kohakublueleaf were working on fixing.

[nova remnant]

there are some factors you have to take in consideration with SDXL, lot of the 1.5 extensions and most of the LoRAs wont work with SDXL, but SDXL has its own compatible versions that you would download and set up, also default image res should be 1024 and upwards, which would take longer to generate images, other than all that if you are using A1111 you'd have to update it to 1.6.0

[nova remnant]

I have hope too ?

:p

[wispy canopy]

still, getting this with a simple prompt, a minimal amount of negatives from a generic model -> i call worth

[summer night]

yeah the paperspace thing i'm using is just a port of the fast-stable-diffusion colab notebook which runs auto1111

[supple raptor]

Not on 4gb. Even if you can get it to load, you won't be able to gen at high enough resolution to have them come out good.

[gusty palm]

is comfyui better for vram

[nova remnant]

You meant sdxl models on a1111 specifically ?

[wispy canopy]

yes, by far, much more effecient with vram usage

[wispy canopy]

nah, sdxl global -> you can't get that running in 4gb vram

[summer night]

something that concerns me is whenever i start stable diffusion on the notebook, it says
No SDP backend available, likely because you are running in pytorch versions < 2.0. In fact, you are using PyTorch 1.13.1+cu116. You might want to consider upgrading.
the pytorch version paperspace uses is 1.12

[nova remnant]

A1111 caught up to speed with comfy with 1.6.0 version

[supple raptor]

NoCrypt is working on a notebook that will probably be worth switching to once he releases it. It will have an integration to send gens to a Huggingface dataset which you can then git clone to your PC if you want, along with a bunch of other optimizations.

[wispy canopy]

aitemplates fucked it up again, with about 80% increased generation speed

[nova remnant]

Not that, about the bug

[summer night]

really i just wish the paperspace version was able to do that thing where it installs everything to your google drive, because paperspace pro gives you a paltry 15 GB of persistent storage

[gusty palm]

doing popculture wildcard
pinky and the brain lol

[summer night]

google colab is over
it disconnected my runtime before i could even finish putting in my settings
which is why i switched to paperspace

[gusty palm]

im out of covfefe creamer 😦

[supple raptor]

The model alone is over 6GB at FP16. It definitely won't happen in any usable capacity on 4GB. To run it on 6GB, the model components are swapped in and out of VRAM as needed which makes gening very, very slow.

[summer night]

what is the "best" anime style SDXL model
i'm looking for a kitchen sink type thing like Anything V3/V4

[supple raptor]

I just use milk with vanilla simple syrup

[supple raptor]

There isn't one yet. They're all trained on relatively very small datasets.

[gusty palm]

out of milk too >< i have sweetnlow though

[wispy canopy]

i'm running my own blend, i hate anime models actually XD

[nova remnant]

[summer night]

"blue pencil" looks close enough so i guess i'll just use it for now

[nova remnant]

how much data do you need to train a model ? >.>

[summer night]

something i've always wondered is why most of the models are all the same size
like it seems most of the SDXL models are exactly 6.46 GB

[wispy canopy]

that's how big the safetensors file at fp16 is

[gusty palm]

auroranegative vs khfb

[nova remnant]

now do both :D

[gusty palm]

used other negative stuff too though

then ill try neighter

[nova remnant]

@supple raptor hey

[summer night]

do SDXL models need vaes?

[nova remnant]

What resolutions are your model good with

[wispy canopy]

yes, but the one they shipped 1.0 with was bugged

so everyone started using the 0.9 one again

[summer night]

is this the one i use
https://huggingface.co/stabilityai/sdxl-vae

[supple raptor]

For model fine-tuning, it depends on the state of the base model and what you want to train into it. Usually fine-tuning projects start around 4-5k images for slightly tweaking a model that's already quite close to what you want. In order to have any hope of learning as broad of a range of characters and art styles as our current 1.5 models know, we really need something trained on several hundred thousand or more images. If all you want is something that's generally sort of anime style then the current models might be fine.

[gusty palm]

both, neither, then neither without cfg scheduling

er cfgthreshholding

still feels very slot machine lol

i like both though

[nova remnant]

same difference

[summer night]

i like the SDXL
using counterfeitXL v10

it's cool how it does background details that look consistent most of the time

[wispy canopy]

highres moltres 😄

[summer night]

i take back my comment from a few days ago where i said i think stable diffusion has stagnated

[wispy canopy]

i've been going at it full force for a while now 😛

[sweet bramble]

would people prefer to use booru tags for an anime model?

[wispy canopy]

yeah, for anime specific models, i'd go with booru for sure

i hate it for that specific reason, but booru prompts ARE very good at specific things

[summer night]

in time do you think SDXL will become the new main thing over SD 1.5, or will it be like how we SHOULD switch everything to USB-C but USB-A is still going to be mostly used for the foreseeable future

[wispy canopy]

i'm pretty sure it will

and if people know how to get wuerstchen to work properly, it might even take over, it's about the same image size/quality as sdxl (reportedly), but better on vram

[summer night]

so what was that other UI you guys were talking about earlier
that is easier on the VRAM

i'm going to get it for my local install of SDXL models

[wispy canopy]

not ui -> vlad had it already in sdxl, but it diffusors broke so they had to recall that checkout

[nova remnant]

comfyui

[glacial bane]

This looks like you tried to generate yae miko without a lora

[wispy canopy]

oh, i read it wrong -> yeah, it's comfuui

[summer night]

yea that

[wispy canopy]

it's a bit strange to start working in, but not bad at all, honestly

i just miss my dynamic prompting maddness that i can do in a1111

[summer night]

does comfyui just install itself and all its dependencies like a1111?

[wispy canopy]

i think i did a git clone -> manual set up venv, and then just start

dunno if it wwould've done venv by itself

[summer night]

ok looking at this whole thing i think i might just
stick to paperspace until that memory leak thing with a1111 gets figured out

[wispy canopy]

i don't think i've encountered it yet

[summer night]

apparently a1111 has VRAM issues

[wispy canopy]

it's certainly heavier on vram than comfy

that's for sure

[summer night]

i need to run some X/Y plots to figure out what i like the most for SDXL

i'm going to do sampling steps on the X axis and sampling methods on the Y axis

might as well do clip skip on the Z axis

[nova remnant]

how not to gen underwater above water

[wispy canopy]

this is why i love sdxl based model ❤️

[wispy canopy]

yeah, double water 😄 it's a thing in sd-land

so, where would one order one of thse?

it's so tiny ❤️

[summer night]

goo lagoon

[bitter latch]

Bunny girl

[wispy canopy]

coffee to improve morale!

[summer night]

lol i tried using hires fix and it was just like "nah bro, 16 GB of VRAM ain't enough for this shit"

[bitter latch]

and bunny choas

[wispy canopy]

... yup

although, depending on upscaler it might error out too

[bitter latch]

I only have 4gig , how are people having issues?

[wispy canopy]

probably because of sdxl model xD

[bitter latch]

ahh

[summer night]

the consistency is mind boggling for me

[wispy canopy]

it isn't for me... actually 😮

[summer night]

like it did the shoulder paldrons symmetrically

[wispy canopy]

well, it does get a lot more details right, or even outright is able to do a prompt, which 1.5 models just failed at

[gusty palm]

the lora:add_detail adds "master, girl, japan" to the prompt for some reason

[wispy canopy]

sd1.5 vs sdxl

[bitter latch]

SDXL , same prompt

I like the choas better

[summer night]

it seems to ignore some things
i prompted blood on her armor and face and it's just not doing it

[bitter latch]

try bloodi

bloodie, however ya spell it

[wispy canopy]

1.5 vs sdxl

[summer night]

tora
i'm genuinely curious why you thought i was spamming earlier

[bitter latch]

cause it seemed like you where trying to get us to buy into that site/service

[summer night]

oh

i'm not sponsored by it or anything i just like it because it's faster than my PC and affordable for me

[bitter latch]

should have led with that

anyhow i'm gonna lay back down now that I got that change to my style out of my head

[summer night]

some people use runpod but you have to pay by the hour for that
with paperspace you can just spend 8 bucks and get a month of access

i guess runpod would be better if you are more on and off with SD

[bitter latch]

there are alot of services like that

anyhow....

goes back to bed. thud

[wispy canopy]

gnight @bitter latch

[summer night]

i forgot to clear my default negative prompt after refreshing the UI, so i think that was messing with some of my SDXL generations

[wispy canopy]

yeah, sdxl uses very little negative prompting

[summer night]

probably didn't help that it had a negative TI

[wispy canopy]

will they even work?

i mean, 1.5 ti's shouldn't even be compatible

[umbral depot]

Sdxl just doesn't do much for me

[gusty palm]

https://youtu.be/IVIzcMF2Kc4?list=TLPQMTgwOTIwMjOtMtu7EfiBKA

these are so good

[wispy canopy]

staryu got an upgrade 😮

[bitter latch]

oh the back pain , that poor girl

😄

[zinc flax]

Huge

[wispy canopy]

weren't you going to sleep?

and here's some of my randomized pokemon 😄 they uhh... changed a bit

(rapidash, weezing, parasect)

[dawn jasper]

Who is this person?

[zinc flax]

A random prompt I did

[dawn jasper]

@zinc flax dunno whether you love or hate the following series, but that lady looks like an older Silica from sword art online for some reason.

[zinc flax]

Ooo

https://tenor.com/view/silica-sao-sword-are-online-anime-wow-gif-15819474

Quite true actually lol

[dawn jasper]

@zinc flax Kind-of funny to me, I won't lie.

[zinc flax]

I completely forgot about her

It's been a long since since I've watched that show