🍥｜anime

[stable crane]

yin yang

[pure oyster]

A stacked short-stack, a "double stacker" if you will

[stable crane]

2 of them = double double stacked

[slow needle]

i am training a style lora for sdxl rn if you wanna try

[terse ore]

Why am I getting this color bleed/color burn problems? Eyes whites have discoloratoin (as do the sunglasses. Other parts of the image, but I just cropped this part). Someone said its a CFG issue but I'm only using CFG at 7. Used img2img and controlnet

[supple raptor]

This is completely normal and acceptable behavior. To put it very simply, your input image is heavily compressed turning pixels into latents, noise is added to the latent space, it's then de-noised, and finally it is decompressed from latents back into pixels. You're expecting far too much from diffusion model.

TLDR: the tiny artists in your GPU are coloring with blunt crayons and you're expecting the precision of a finely sharpened colored pencil.

[terse ore]

Oh. I get this problem consistently and in the same places so I just wondered. I want to make a webtoon and hoping I could 1shot each generation without edits. Using img2img (it's from a blender render). So, I will have to do some manual editing on each image?

[supple raptor]

Depends how picky you want to be. If your readers are pixel peeping that hard on the art, then the story probably isn't engaging enough. Consistent 1shot perfect renders are a pipedream with any current diffusion models.

[slender barn]

tried to clean up my style abit, some of them look much different than the others what you do guys think.

Frist version

the second version

[bitter latch]

that is alot of lora's

and other thiings. wow

[slender barn]

is that bad thing, I should cut it down?

[bitter latch]

yeah I think you are overloading it

it's an interesting style.. shrug

[slender barn]

Thanks! i'll see what I can cut down.

[bitter latch]

tweaks it without hte lora's and hyper nets, and using some other tools and styles :D

adds random styles, and random artists. :D

crosses fingers and waits for the mess to happen

[supple raptor]

I think you might be the last person in existence still using hypernets XD

[bitter latch]

hmmm

no don't think so

[supple raptor]

They were barely even used before LoRAs.

[bitter latch]

jjinofcoolnes: was , is using htem

[slender barn]

yea I use the Hypernet called pain.

[bitter latch]

I like hypernet:ghost_v10:0.7 hypernet:miscatonic_v10:0.9 , makes them funky

like this

tho the whole prompt is prob a mess.. (Unhinged Drawing:1.3) of (Technical illustration:1.3),(Fearful:1.3) (Canyon standoff:1.3) Hot background, muscled male, Italian legwear, dark skin Italian, Blue, Bard with instrument, mix patterns, fantasy attire, anime, Elegantly intricate ultra detail, best quality, absurdness, by James Fenner hypernet:ghost_v10:0.7 hypernet:miscatonic_v10:0.9,(by Artist Bob Peak:1.3),(by Artist Ethan Van Sciver:1.3),(by Artist Hsiao-Ron Cheng:1.3),(Abstract Illusionism:0.9),(Glitch Art:1.05),(Psychedelic Art:1.1),(High Contrast:1.3), masterpiece, best quality, fine art, ultra detail, 8k, rtx, hi res, cinematic lighting, photo realistic, concept art, , line art style, intricate, cute & adorable, art by makoto shinkai, pov facing, perfectly formed pretty faces, anatomically correct, golden ratio, , Pixar , Cel shaded, Full body portrait, divine, illustration, soft lighting, lush detail, elegant, smooth, sharp , 3.5D , in style of Mark Brooks
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, an6, kkw-extreme-neg,Low quality,Photo,Artifacts,Table,Paper,Pencils,Pages,Wall,watermark,signature, an6, kkw-extreme-neg, an6, kkw-extreme-neg
Steps: 20, Sampler: Heun, CFG scale: 7, Seed: 1679233248, Size: 640x960, Model hash: c7751e6108, Model: aZovyaRPGArtistTools_v3, Clip skip: 2, Pad conds: True, Color Enhance: 0.48, Version: v1.4.1

using style pile for things too 😄

and some neg inversions

and saved styles

[slender barn]

I just got a few more.

[bitter latch]

changed a few things , this one might take a little longer

hmmms

[stable crane]

very interesting

[bitter latch]

me i'm just being my random self

😄

hmm reminds me I need to add some more art to my website

yeah it went nuts

nice tho.. sorta

[bitter latch]

oooh umm mecha suit girls?

ok

😄

but first.. this thing

[bitter latch]

I think we are gently being steered back to more anime-ish art 😄

well.. sorta

thats better

now let's do chibi! cos ya know I gotta

well that didn't go chibi

haha

well thats odd

[supple raptor]

mild blood warning

[slow needle]

my first sdxl lora

[pale monolith]

take my hand, now!

[south sable]

Would love to try it. Been digging deep for some anime in SDXL

[slow needle]

for now the anatomy is quite wonky but i gonna fix it later

[south sable]

https://tenor.com/view/steal-anime-k-on-ritsu-tainaka-mio-akiyama-gif-24287493

[wheat tapir]

cool!

[bitter latch]

https://giphy.com/gifs/crazy-dance-kermit-jx5MJyZqAFsLS

hmmm

[bitter latch]

ok off to bed.. nite

[slate narwhal]

Some SDXL:

[south sable]

https://tenor.com/view/anime-dragon-ball-vegeta-scouter-gif-15106374

[south sable]

Do any of these have metadata? Just want to see how you are using nodes etc.

if you are using nodes

that is

[slate narwhal]

I generated all of these in the bot, I'm very limited right now and can't download 0.9 (no computer, travelling, far from home, generating everything on the bot)

[south sable]

https://tenor.com/view/kanokari-anime-fall-kazuya-kinoshita-gif-26402084

well that means there hope then!

[slate narwhal]

Some of the images i've posted here made it to the pantheon, the prompts for these are variations of this: #🏅｜pantheon message

[south sable]

thanks, ill check this out

[slate narwhal]

Np, 👍🔥 I really like the mix of art, photo, sculpture + anime that SDXL can do

[south sable]

yeh it is coming out to be ver versitile

[slate narwhal]

Yep. I can't wait to see the stuff the community will do

[south sable]

Nice, we have something here. not bad. Just have to keep messing with it.

[strong anchor]

damn, can i have prompts pls

[wispy canopy]

interesting stuff happen here?

[brittle rampart]

hello, what this is a model?

[brittle rampart]

Sensei`s who know what the model and style in this is?

[pale monolith]

Unless there's been more merging this is nubby's Kakigori which is available on civitai

[pale monolith]

ENVY, entropy, rot, polychromatic splash, perspective shift, magic glow, action, motion blur,
1girl, pointy ears, blonde hair, blonde eyebrows, blonde eyelashes, thick eyelashes, blue sky, fantasy, clouds, witch, upper body,
(masterpiece:1.2), (best quality, highest quality), intricate, Accent Lighting, extremely detailed, original, highres,

negative_hand, (KHFB, AuroraNegative), (Worst Quality, Low Quality:1.2), (watermark:1.2), bad-hands-5, cat ears, animal ears, gradient, multiple girls, multiple boys, multiple views, sweater

[strong anchor]

TY

[gloomy scroll]

I believe Nubby's images have pnginfo

[wet elm]

how could i make him hold an skateboard on this image?

[proper panther]

took me a lot of tries but finally got something im comfortable with

[gloomy scroll]

Inpaint

[magic wave]

Or edit and inpaint (aka, put a skateboard in his hand in photoshop, then inpaint it to make it fit in with the rest of the image)

[pale monolith]

some noisy texture or a picture of a skateboard is best, then inpaint has more to work with

[magic wave]

[pale monolith]

solid color block in the shape of a skateboard could work but its harder to add detail and depth without high denoise

[wet elm]

how tho im so bad at inpainting lol, every time i try i cant change nor add anything in the image

[wet elm]

ohhhh thats sick damn

[pale monolith]

mask over the mspaint tier/ skateboard photo using something similar to these settings and a simple prompt like "skateboard, holding skateboard"

[pale monolith]

only masked will increase level of detail of inpainted area so long as the inpaint mask is smaller than the total width or length of the image

masked padding should be 0 for common high detail zones (face) and higher for less common high detail zones (hands) , think about what is more likely to exist in the dataset as a close-up image , if its more niche the ai might have trouble making it without some pixel padding (essentially this functions as artificial zoom-out while generating since the inpaint generation zooms in using the height/width you have set at the masked area)

sample method, steps are personal choice

denoise ~ 0.1-0.3 will do slight modifications, use these settings for when something is "close" to what you want, this also essentially works as add-detail without changing the actual object (usually, depending on padding)
0.4-0.6 will do moderate changes and the prompt will be more important, you might also want to go over the area again with 0.1-0.3 to increase the detail after getting what you want
0.7+ will do big change, use this to completely erase areas and replace with something that wasn't there at all before, however this high of setting will leave an obvious inpaint edge blur so its not ideal

[wet elm]

thank you, it kinda worked lol

its still kinda bad tho

[pale monolith]

[magic wave]

Oh, so hold a skateborad on his back?

[wet elm]

thats kinda what i wanted lol

i once saw a meme of shirou skating and found it funny so im trying to replicate it lol

[supple raptor]

https://civitai.com/models/100505/kakigori

[nova remnant]

Prompt?

[gloomy scroll]

BREAK 1girl, solo, happy, cute aesthetic illustration, stylized digital art, abstract and colorful paint style,
BREAK simple background,
<lora:LoraMerge_dim128:1>```

[nova remnant]

thank you so much

[gloomy scroll]

Also those images have pnginfo

[loud falcon]

is this BREAK like in code?

[supple raptor]

https://cdn.discordapp.com/emojis/984187235073155092.gif?size=48&name=teamWorkHighFive&quality=lossless

[gloomy scroll]

BREAK starts new 75 token block

👇 Nubby writing new essay

[supple raptor]

BREAK is a special function in WebUI. SD only understands prompts that are 75 tokens long, but someone figured out that you could have it interpret multiple blocks of 75 tokens then average those results together in order to allow for longer prompts. BREAK forces it to full the rest of the current block of 75 tokens with blank token so that the very next token will be the beginning of a new block. This is useful for a few reasons

• Tokens effect each other fairly strongly inside of blocks, in a cascading format starting at the front
• Tokens closest to the beginning of a block of 75 tokens are strongest and get weaker as they get closer to the end
• Many tags contain several tokens which means it is likely that if your prompt is more than 75 tokens you could have words that are split between two blocks which will cause the AI to not understand the meaning.

[loud falcon]

thanks a lot

first time seeing it

[brittle rampart]

Thank you!

[magic wave]

O damn I didn't know about this either, is this good for cutoff?

[supple raptor]

It generally doesn't work for separating concepts quite as well as the Cutoff extension, since the the prompts end up averaged together but it can still help quite a bit. You could use both methods together for particularly stubborn tags.

[gloomy scroll]

I still don't know what Cutoff does

[magic wave]

Yeah I mean like, I heard the cutoff has a limit of like 75 tokens, so that's why I ask, to like get better color accuracy and less bleed

[supple raptor]

If you want an in depth explanation: https://github.com/hnmr293/sd-webui-cutoff/issues/5

TLDR; it allows you to specify which parts of the prompt are given to the AI at the same time which will usually allow you to prevent prompts like `1girl, red shirt, white hair, blue eyes, from getting red hair or eyes or any other mixing up of colors.

[shrewd root]

do anyone have a model to share that create genshin characters nicely ?

trying to make a yanfei img

[magic wave]

You could look for a LoRA

[supple raptor]

this LoRA looks pretty good https://civitai.com/models/33958/yanfei-genshin-impact-or-character-lora-757

there are two other yanfei LoRAs on Civitai if you have problems with that one.

[shrewd root]

thank you i will test it out ^^

[young fjord]

what vae are you using for these?

[gloomy scroll]

kl-f8-anime2

[pale monolith]

i think most genshin loras are pretty good on a large variety of models at this point

but I'll self promote :3 https://civitai.com/models/99169/amnesia

[bitter latch]

haha thats funny I wasn't trying to copy the style

[pure oyster]

@supple raptor Kakigori did my OC well.

[bitter latch]

Chibi Vesion of mine 😄

hmm wonder if i can make a human-ish version with tattooo's

[supple raptor]

https://cdn.discordapp.com/emojis/871167635679772672.gif?size=48&name=borpaLick&quality=lossless

[shrewd root]

thanks worked amazing for what i wanted ^^

[bitter latch]

ok...

hehe

[cyan burrow]

I "programmed" chatgpt4 to actually work with prompting

This is what it took:
`Here's a few examples of the type of answers and behavior I'm looking for in you right now
All this relates to stable diffusion 1.5 (CLIP based) prompts:

Me: Give me a prompt for a fallen angel standing up in forest ablaze with his wings made of ash and fire

ChatGPT: In the middle of a forest ablaze with an inferno we see a tall and mysterious fallen angel, he's covered in ashes and his wings are shedding ash and burning with a hellish flame

or

Me: Give me a prompt for a temptress siren

ChatGPT: A young woman with beauty out of this world and iridescent skin bathes in an ocean pool of turqoise water, on an atoll in a coral reef with a mountainous island seen far in the horizon, she wears transculent clothing and has slit eyes with an intense violet color. Her expression is seducing and her eyes are lustful

or

Me: I want a prompt for an overpowered swordsman

ChatGPT: A swordsman in black with a burned in symbol on his chest stands at the end path of his four after-images, each cutting down an enemy soldier. They all fall while he sheathes his sword, sending manga inspired gushes of blood spraying

Can you confirm you understand the role I want you to take?`

ChatGPT4:
Yes, I understand the role you want me to take. You're looking for creative and descriptive prompts related to specific character concepts, such as a fallen angel in a blazing forest, a temptress siren, or an overpowered swordsman. You want me to provide vivid and detailed descriptions that capture the essence of these characters and their surroundings. I will do my best to generate prompts that match your desired style.

not specific, they were examples

Apologies for the confusion. I understand that those were just examples you provided. Please go ahead and provide me with the specific character concepts or prompts you would like me to generate descriptions for, and I will do my best to provide you with creative and detailed responses.

I want a prompt of a young woman with tiny horns battling gladiators and winning with overwhelming power and brutality

As you may have guessed, that didnt work entirely

In the midst of a grand arena, where the roar of the crowd blends with the clashing of steel, stands a young woman radiating an aura of otherworldly power. Adorning her head, two tiny, elegantly curved horns signify her unearthly lineage. She moves with a grace that belies the unstoppable force she possesses.

Clad in a flowing crimson robe, her eyes gleam with determination and a touch of defiance. As she engages the gladiators before her, her every strike carries the weight of untamed strength. With each swing of her weapon, a resonant clash fills the air, as if her blows are delivered with the force of an avalanche.

The gladiators, renowned for their combat prowess, find themselves overmatched and overwhelmed. Their weapons seem feeble in comparison to the raw power emanating from the young woman. Her movements become a blur, her strikes a symphony of precision and brutality.

As she dispatches one opponent after another, a mixture of fear and awe spreads among the spectators. They witness an embodiment of untamed ferocity, an indomitable spirit manifesting in the form of this young woman. Her victories are not only in triumphing over her adversaries but also in inspiring a sense of both admiration and trepidation in those who bear witness to her incredible power.

zzzz

that's a bit too much, it's not a story we're writing. We're describing a static scene in a singular image for stable diffusion to generate with latent diffusion. It's important to be specific and describe the particulars of the scene, vividly if that's the intent and to try and stay below 200 characters.

Apologies for the misunderstanding. Here's a revised prompt within the specified parameters:

A young woman with delicate horns stands amidst a chaotic battle in an ancient arena. Her eyes burn with a fierce determination as she effortlessly bests gladiators with brutal precision. The air crackles with raw power as her every movement exudes overwhelming strength.

and that's how I programmed ChatGPT4 to do prompts

[cyan burrow]

`A Summoning Through TimeSpace. The protective bubble expands the moment the person arrives inside a compound known as "Europa's Sequence" first found on the moon of Europa. The compound can be programmed by human DNA, but without that which makes a singular individual unique. Through quantum entanglement, while kept at absolute zero. It can be used to travel unfathomable distances of tens of thousands of light years.

The moment the compound completes the DNA sequence after the two quantum molecules are established. Both being unique and containing the same information. When this is inserted into Europa's Sequence. The ice like compound that spends most of its time floating like icy liquid. Mimics the state of the other molecule. Meanwhile, the molecule on the other side, mimics the one on the first. Effectively switching places. This creates a sudden expansion of heat that changes the chambers absolute zero, into a tropical 40 C.`

I had to come up with something, so I wrote that, straight outa my butt

[tidal phoenix]

Little bit wonky (like the legs both being on the same side in the back), but still neat enough to share IMO

[cyan burrow]

After an accident in 2079 taking place in Wuhan China resulted in a virus creating vampire like beings. As well as ghouls and zombies of those too weak to beat the Virus and go on as vampires. The Protestant Church, the Catholic Church, the Evangelist, Lutheran and other interpretations. All joined together to create a superpriest made from the wings of Lucifer, the seventh scroll of Metatron containing a spell to find Nephilims and the spear used to deal the lethal cut to Jesus Christ and a vial of his blood. Taken from various vaults of the different sects of the christian faith.

They first used the scroll to find the one and only living Nephilim on the planet and used the Spear of Destiny to incapacitate the Nephilim and seal its powers. For the entire procedure, they first placed a priest on a blessed operating table. Without anasthesia, as Jesus did not receive any. They used the spear of destiny to cut them and their ribcage open and did the same with the nephilim. They took out the Nephilims heart, which near instantly regenerated. Then sprinkled jesus blood on a feather of Lucifers wings and pressed it against the Nephilims heart before removing the priests heart and quickly placing the Nephilims in. To avoid celestial rejection, the feather of Lucifer and the blood of Jesus was used to bind it to the priests soul.

As such the first SuperPriest was created and what followed were hundreds more. Time and money went into creating augmentations to replace some human function and enhance them. Any wounds would heal up and with the Nephilims heart the augmentations would be conected and absorbed as if they are part of them.

The Superpriests were used to hunt the ghouls and zombies and vampires into extinction and later to subjugate mankind under the iron fist of the Pope (made eternal by the Nephilims heart, but not just heart, also all other organs. Becoming as a demi-god on earth.)

The year is now 2184 and mankinds only hope for defeating the priests...

Cosplayers

Cosplayers have come a long way. After Buddha returned to earth and taught them enlightenment, but they got bored and watched anime. So Buddha cursed them to become what they watched. So born were the heroes and villains that would hence forth be known as the A-Theisteam

I don't know where I come up with this stuff, but it's incredibly pointless and often harmful to my brain

but if it entertained at least one person

then I am satisfied

[gloomy lodge]

they litraly got the wings of Lucifer and didnt try some wack shit sooner

[cyan burrow]

because they're the wings of Lucifer and they're the church of God

things would have to become very desperate first

[cyan burrow]

"You bastards...FIGHT ME! I'LL DEFEAT ANYONE WHO COMES AT ME! What hic why won hic any come after.....me hic bleeeergh Who did that ?!"

[gloomy lodge]

The nuns do not approve of this

[gloomy lodge]

big lizard

[cyan burrow]

Set, the GOD from ancient Egypt is here to punish the Nuns

[gloomy lodge]

the lizard would win

[gloomy lodge]

then dark skin girls do be cute

[cyan burrow]

I'm a hot and brooding demon, fall in love with me and let me corrupt you. M'Lady

[loud falcon]

https://media.discordapp.net/attachments/255684579980083201/1116435847219253278/UhWhat.gif

[gloomy lodge]

go repent

[gloomy lodge]

nice nice

[bitter latch]

sailor moon . or sun?

steampunk anime

[autumn violet]

Is that molten coffee?

[supple raptor]

https://cdn.discordapp.com/emojis/1110837589411385436.webp?size=48&name=thinkeroo&quality=lossless

[autumn violet]

V2

[bitter latch]

suppose to be a coffee bean cannon but guess it had different idea's

[supple raptor]

Ah, the problem is in your design. See, you want the cannon to fire out hot, brewed coffee, already in the cup.

[bitter latch]

gets paid, pays bills and watches it disappear

sighs

[bitter latch]

selects 2 save styles, leaves prompt blank and generates

yep, no surprises there

I wish daiblo IV wasn't so expensive

[granite fog]

would you answer?

[loud falcon]

she has tiny baby hands

doable i woudl say

[astral drift]

These are dope

[south sable]

thanks!

[south sable]

SDXL 0.9 image + text in post

[supple raptor]

Miku got a new sponsor

[bitter latch]

almost

[bitter latch]

last one. off to work...

[slate night]

Obvious scam/phishing link is obvious

[young fjord]

@dull jackal scan link!!

[dull jackal]

ahem And...what link would that be?

[young fjord]

it was a game link but I guess it deleted now

[dull jackal]

Yeah, I got it a few ago lol

❤️

Thanks for the ping

I'm off to PT; I'll catch ya'll latah!

[young fjord]

np

💗

[barren anvil]

which model is that?

[pseudo condor]

HI

someone here use controlnet

how i let this three?

like how much for they i need put

[slate night]

I just leave mine in the default settings

I don't use canny very often though so someone else might have better input

[pseudo condor]

thxx

[chrome moss]

how to use this lora to apply only the outfit, hair, eyes, without applying its art style? thanks
https://civitai.com/models/17257/dxd-all-girls-gasper-rias-sona-yasaka-venelana-akeno-asia-tsubaki-xenovia-grayfia-kunou-gasper-rossweisse-irina-kiryuu-koneko-kuroka-serafall-ophis-ravel

[strong anchor]

try ro use checkpoint of the art style that you want, and if you want ony out fit you can try control net canny probabbly it can work, or just write it manyally with high count

[strong anchor]

outfit, hair, eyes, send me promts i will try, what exatly you want?

[pseudo condor]

ok first test

[barren anvil]

wow, which gpu do you use?

[elfin bolt]

3090

[bitter latch]

umm think mistoonsanime

[barren anvil]

thx

[wet elm]

Did you inpaint each girl one by one?

[bitter latch]

well thats a confused picture

[elfin bolt]

yea

[wet elm]

damn thats cool, so like you put different prompts to each area?

[elfin bolt]

yea, https://github.com/hako-mikan/sd-webui-regional-prompter

[wet elm]

that sounds kinda complex lol

[noble lion]

CUTE

/////

[gusty palm]

is it ok to post gens here?

[supple raptor]

not only is it ok, it's encouraged!

[gusty palm]

awesome

im liking my results but im not sure why they seem to be kinda blurry or blotchy

running at 40 samples and trying different samplers and cfg settings trying to figure it ou

is more than 40 samples excessive?

[supple raptor]

Yeah, I only run 20 steps for all of the DPM++ samplers. I'd say anything over 30 is excessive for them. I'd try reducing the emphasis on tags and be careful not to put extra () around words that have an emphasis number specified because each extra set will multiply the number by 1.1. For example, ((sexy ninja outfit:1.4)) is actually (sexy ninja outfit:1.54). In general, it's best practice to keep emphasis at or below 1.3 for each token.

how much vram do you have?

[gusty palm]

ahh ive been violating most of those

6gb

had no idea u could get that info from the image

i only started using auto111 a few days ago after using sygilui for a few months

[supple raptor]

Yep, if the image is shared correctly it will have the generation parameters in the metadata and people can put the image into the PNG info tab to extract that info.

[gusty palm]

o awesome

and im putting in those changes now tysm

sygil only had like 5 samplers

different model but way more clarity already

[supple raptor]

I'd recommend trying DPM++ 2M Karras, it will run about twice as many it/s as the one you're using because DPM++ 2S a effectively does 2 steps for every one.

[gusty palm]

ok

[supple raptor]

Do you know about Tiled VAE?

[gusty palm]

no ve only sortof noticed it in the settings

er nvm that doesnt say tiling

i enabled this though

[supple raptor]

This will help you reduce vram usage so you can generate larger images. First, install this extension: https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111.git

You can install it by pasting the URL into the "URL or extension's git repository" field on the Extensions > Install from URL tab. After you click the orange "Install" button, go to the Installed tab and click the orange "Apply and restart UI" button.

[gusty palm]

ok thanks

[supple raptor]

That one is also good to use

[gusty palm]

how about the command line ones

-opt-split-attention --opt-sub-quad-attention

i tried them briefly but realized i didnt actually need to to start out

[gusty palm]

mouse girlyy

[supple raptor]

You might want --medvram for 6GB. I'm a bit spoiled with 12GB myself so I'm not positive what most people are using nowdays. Under Settings > Optimizations, I would recommend making sure that "Cross attention optimization" is set to sdp - scaled dot product, that will increase generation speed and reduce vram usage.

[gusty palm]

ok thank you

i got the vram estimator benchmark done that seemd to help me keep under 100, going over dips into shared mem and slows down alot

looking into that extension now

[supple raptor]

I'm running a few gens right now on it to see what settings will keep it under 6GB vram I'll send you the settings once I figure it out.

[gusty palm]

gonna enable the tiled vae on default settings to see what happens

[supple raptor]

@gusty palm Here we go. This is using 4917 MiB on my system which should leave comfortable headroom for whatever your system needs.

[stable crane]

you don't use the color fix?

[gusty palm]

ooo

[supple raptor]

No, I've never seen a need to. Maybe I should try though.

[stable crane]

Fast Encoder may change colors; Can fix it with more RAM and lower speed.

in my old UI it says this

[gusty palm]

may try that out then

tile size settings are making me think of blender cycles

hoping theres some crossover i can bring since i know how that works

[stable crane]

cycles = more iterations of raytracing?

[gusty palm]

not sure if thats accurate but thats quite a big save

[stable crane]

maybe only with some algorithms where if you don't give enough cycles steps it's definitely underdeveloped

[gusty palm]

oops forgot to reduce steps

that looks faster tho

[supple raptor]

Yep, that should be pretty close. The VAE step is what turns the math that the model does into pixels that you can see. It's the biggest VRAM cost of the process, so we reduce that by spliting the job into tiles rather than trying to do the full image at once.

[gusty palm]

dis is awesome

[supple raptor]

Here's the next secret to better images. Turn off any upscaling and restore faces that you might currently be using, and use hires fix instead. Hires fix generates the image at the resolution you set, then upscales it, and finally runs img2img on that upscaled version to improve the image without the weirdness that natively generating at high resolutions can cause.

[supple raptor]

Those settings are what I used in my tests to keep vram under 6GB.

[gusty palm]

i havent been using any of those ive been upscaling after the gen is done

ill try hirez now though

[stable crane]

sometimes it'll say that

that extension has a nice bonus -- you can actually interrupt gens

I just tried to interrupt a gen and going to have to restart the entire webui

[supple raptor]

Yeah, on the first pass, it probably won't need to use the Tiled VAE since the image is pretty small, but the second pass will be double the resolution with the settings I sent and you will likely need Tiled VAE for that.

[gusty palm]

ok

im trying 910x512 rn to see if it tops out

still plenty of room

haventtried hires rix yet though

does that eliminate the need to upscale

or is that something else

[supple raptor]

It should usually eliminate the need to upscale since upscaling is included in the process.

here's another extension that will help to reduce the need for upscaling: https://github.com/Bing-su/adetailer.git

You install it the same way as the last one, though you may need to completely close the CMD window for SD and launch it again because it has some models that it needs to download.

[gusty palm]

its not showing me upscale by setting

[supple raptor]

At the bottom of your WebUI, which version does it show here?

[gusty palm]

1.4.1

[supple raptor]

Let me update mine real quick to see if he moved that field.

[gusty palm]

could it be this or some other extension i installed or a setggins maybe

[stable crane]

4x txt2img render, unfortunately you will see this happen with tiled vae sometimes

she grew an extra ear

[gusty palm]

its like its adjusting my settings to pre upscale

and doing some type of auto upscale by

[supple raptor]

This one is found in the img2img tab in the scripts section. It is useful, but separate from the Hires fix upscaling.

[gusty palm]

oh alright, i really appreciate yalls help btw

im gonna hit go and see what it makes with these settings

i feel like i should re-run the vram estimator benchmark

[gusty palm]

see what happen

bazongas

[supple raptor]

Check to see if you have this setting enabled in the settings tab. Easiest way to find it is click the "Show all pages" at the bottom on the left side then hit CTRL+F and search "Hires"

[gusty palm]

thats gotta be it

[supple raptor]

I find the "Upscale by" slider to be a lot more intuitive because it will always upscale both dimensions evenly so your image doesn't get stretched.

You might also want to lower the Denoising strength value to around 0.5. The default is 0.7 which is usually too much of a change from the starting image.

[gusty palm]

yep it was ticked, in compatibility

have to reload ui now but heres whay i got before that happened

[worthy delta]

that looks amazing!

[gusty palm]

it went into shared memory at some point though but im sure i just went too high somewhere

thankyaa

compared to the things i was getting from sygil i love this

also no safetensors was a pain

developer of sygil is ver nice though

[supple raptor]

What does sygil use? Diffusers libraries?

[gusty palm]

cktp

[supple raptor]

Oh RIP 💀

[gusty palm]

its pretty similar in alot of aspects but its only got about 20% of the settings

good for beginners

[worthy delta]

im new to this stuff is cktp bad?

[supple raptor]

We've basically completely stopped using ckpt due to security.

[gusty palm]

its good but tensors are more accepted lately

[worthy delta]

oooh gotcha gotcha

[gusty palm]

cuz of security ye

[gusty palm]

thats sweet

[worthy delta]

idk if this is the place for it but i have a slight problem i cant generate as it uses too much ram and idk why it uses that much, is that normal? i have 16gb of ram

[supple raptor]

ckpt is just a .pt file. Stability decided to add ck to it just to differentiate the file as being a model. The problem is that they use pickle imports which is a technical way of saying that they are capable of executing code without the user's knowledge, which means they can be infected with malware.

[worthy delta]

oh dayum that aint good at all

[chrome moss]

thank you! I just found a solution, which is to use LoRA block weight extension to reduce its style art weight, but i don't know how to do that yet.

[supple raptor]

In most cases, VRAM (the memory on your GPU) is what is used for image generation, not system RAM.

[worthy delta]

weird

[supple raptor]

Which GPU do you have?

[worthy delta]

it looks like its using my system ram

3060 12gb

but when i start a img the gpu barely goes up on task manager

only system ram

[supple raptor]

Ok cool which WebUI are you using?

[worthy delta]

im gonna be honest im not too sure

stable diffusion web ui it seems

[supple raptor]

At the bottom, do you see something that looks like this?

[worthy delta]

yup exactly like that

i havent messed with the settings or anything like that cause tbh idk what they do

[gusty palm]

we got similar hardware but i have 2x my vram

im jelly

[supple raptor]

Ok, cool. You have Automatic1111's WebUI which is the most popular one.

[worthy delta]

aye that seems good

but why does it barely touch my gpu in task manager

my system ram is dying in there

started a img right there

boom system ram is on life support

[gusty palm]

it can load into system ram before moving into vram

or some type of caching process

[worthy delta]

also sidenote @supple raptor i read ur tip abt 30 steps and the karras it helps alot

in general

[supple raptor]

Go to the details tab in task manager. Right click on any of the column headers > click "Select Columns" > check the box next to "Dedicated GPU memory" > click OK > click the new "Dedicated GPU memory" column to sort by it so the highest value is at the top.

It should look something like this

Watch that value when you generate to see if it is maxing out your vram.

[gusty palm]

python usually around 10gb system is normal for me

and couple gb of video ram

[worthy delta]

yeah it dosent show that

nothing using the dedicated memorary from the looks of it

i closed it and re opened it appeared

[supple raptor]

ok cool

[worthy delta]

is that normal?

i dont think this is good

[supple raptor]

Ok, yeah, it's running out of VRAM. That's actually not a bad thing because it means that it IS using your VRAM like it should and I can help you get it to not run out of VRAM.

• What resolution are you generating at?
• If you're using Hires fix, what is "Upscale by" set to?
• Are you using tiled VAE?

[worthy delta]

720 width, 1280 height is what i normally do to make a phone background

i dont think im using hires fix havent messed with the settings

and never heard of tiled vae

im new to this stuff started yesterday

[gusty palm]

ye too high resolution

i try not going about 768 in either direction

[arctic yarrow]

https://youtube.com/shorts/OHdGhc6b49I?feature=share3

[supple raptor]

Do you get a lot of images with extra characters, weird proportions, or extra limbs, heads, etc?

[worthy delta]

not really

[gusty palm]

could try the command line

optimization

[supple raptor]

Alright, you might not need Hires fix for your model which will save a little generation time. But I've got a few things to save a lot of VRAM.

[unreal ridge]

damn, I've been out for a while

is this channel finally alive again?

[worthy delta]

if it stops it from making it not be able to work its thing ill do it

[supple raptor]

First, go to the settings tab and Click the Optimizations button on the left. Then set the Cross attention optimization to sdp - scaled dot product and hit Apply settings.

[worthy delta]

did that!

[supple raptor]

It's been pretty dead lately, but I'm here today doing the ol' tech support XD

[worthy delta]

LOL you have been very helpful thank you

[unreal ridge]

as usual

[supple raptor]

OK, next go to the Extensions tab, then the Install from URL tab and paste this into the URL field https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111.git
Click the orange "Install" button then go to the installed tab and click the orange "Apply and restart UI" button

[worthy delta]

just did it!

u make this hella easy to understand

i wouldve never figured this stuff out in a million years

[supple raptor]

Now on both your txt2img and img2img tabs you will have this section under the generation parameters. These settings work for me with batch size 2 at basically any resolution on 12GB vram and it uses approximately 11.5GB

[south sable]

Just gonna make covers for OVAs that never existed IG lol

I imagine it's like Evangelion but a little more noir

[worthy delta]

should i copy what u have ur settings at or set them to custom fit mine?

[supple raptor]

I would copy those settings to start with and watch the Sys VRAM number after your generations. If you find that it's using all of your vram then reduce both values by 1 notch, if you have a bunch of extra vram you can try increasing them by 1 notch at a time.

I've got the same amount of VRAM as you, so those settings should be good for you unless something is using extra vram in the background on your system.

[worthy delta]

nope i just let it generate stuff nothing in the background

thank you so much for all the help

[supple raptor]

No problem! Ready to learn how to generate at even higher resolution in the same aspect ratio now?

[worthy delta]

always ready to learn new stuff

[astral drift]

Was the text added in post or initial generation? 👀

[worthy delta]

even if i dont understand LOL

[south sable]

post

[gusty palm]

couple weeks ago i made some sweet natural science type pages for lovecraftian monsters

the wonky ai text looked perfect for the lovecraft sytyle

[astral drift]

Ah, was about to say sdxl is kinda crazy

[gusty palm]

i hope i can get the png info for these

that cant be right right?

[worthy delta]

a gen of corvo attano

[gusty palm]

only 32% for such high resolution or does vae tiling have magic init

[supple raptor]

Alright, so what you're going to do is lower your resolution to 512x910 and enable Hires fix. Select R-ESRGAN 4x+ Anime6B for the Upscaler. Set the Denoising strength to ~0.51 and Upscale by value to 2.

What this will do is generate an initial image at the lower resolution, upscale it to 2x the size, then run img2img on that upscaled image to add more details.

[gusty palm]

thought it was neo

[supple raptor]

Yep, tiled VAE is amazing.

@gusty palm 4096x1024, generated using ~8gb vram.

[worthy delta]

holy god

that is amazing

[worthy delta]

also did all this!

ai art is something unrivaled

[gusty palm]

pphew

[worthy delta]

this is all amazing

[supple raptor]

Now for something truly amazing, this is an 8192x2048 version of that last image that was upscaled with img2img, still only using about 8GB vram.

[worthy delta]

HOW

i cant handle this its so good

[worthy delta]

how do u upsclare on img2img ive never used it before

[gusty palm]

iw as thinking about how few years ago.. rendering a normal image in cycles blender took a few minutes, now same image a few seconds

quality and speed scale like exponentially almost

[worthy delta]

i love this discovered how to get into it a day ago and it blows my mind

[gusty palm]

with zero hardware upgrades

i imagine realtime ai video isnt impossible in a few years

[worthy delta]

i just recently upgraded my gpu and omg its done so much

[gusty palm]

true

i went from a gt720 to the gtx1060

[worthy delta]

i went from 1650 to 3060

[gusty palm]

oh i misread what u said u had earlier

thast beefy

[worthy delta]

and ill i do is play old games like dishonored..

[supple raptor]

Once you get an image that you like, click the Send to img2img button. It will automatically move the image and your prompts over to the img2img tab. Remove all of the descriptions of the character and the background from the prompt (this works better without them). Set your denoising to ~0.36, then scroll down to the tiled diffusion section. Expand that and try these settings. Don't forget to set the same Tiled VAE settings as you were using on the txt2img tab because those aren't transfered automatically

[worthy delta]

i alr took a ss of them just incase

gonna do those now

so i remove all descriptions of the character?

so leave it blank

[supple raptor]

Just leave things like quality and style tags like I did in my screenshot. Blank can work too, but usually having the style and quality tags will work best.

[worthy delta]

i dont have any quality tags in it or style

just a random generation of a charcater

[supple raptor]

original (1088x1920) and upscaled (2176x3840)

[worthy delta]

it didnt copy over the width n height is that normal it looks like they doubled

[supple raptor]

You can just use like masterpiece, best quality for the prompt on most anime models for this upscaling.

[supple raptor]

Yep, those doubled values are the actual dimensions of the image since we already upscaled 2x using Hires Fix on the txt2img tab.

[worthy delta]

ooooh thats very helpful

[supple raptor]

This is going to be another 2x upscale so you'll have an image that is 4x the size of the values that you initially set on the txt2img tab.

[worthy delta]

dayum i see i see

[supple raptor]

I find that going up to as high as 3x upscaling on the Tiled Diffusion (img2img) upscaling step is usually no problem. Going higher than 3x in one shot on it will usually result in some weirdness, but if you want to make a ludacrisly large image you can always send the upscaled image back to the img2img tab and upscale it again like I did on this one (12288x4096)

[worthy delta]

i think i copied all the settings right

[worthy delta]

omg thats amazing

[supple raptor]

That looks good. Just make sure Tiled VAE is also turned on otherwise you'll probably get OoM right at the end of the generation (which will probably take around ~2.5-3.5 minutes.

[worthy delta]

yup! have it activated and all

i also zoomed out an red did ss that i can copy of the settings

[chrome moss]

does anyone know how to use lora block weight just to extract outfits from a lora?

give me weight settings. thanks

[supple raptor]

@worthy delta and @gusty palm in your stable-diffusion-webui folder you can find a file called ui-config.json. If you edit that file in a text editor like notepad you can set the default values for nearly everything in your WebUI including the things that I taught you like Tiled VAE so that you don't have to change them every time. You can even add a default negative prompt so that it will be filled automatically for you everytime you load the WebUI. Be aware that the settings for txt2img and img2img are all separate so you'll need to set the values in both places if you want them to apply to both.

[worthy delta]

u can do that??

this gonna save me so much bloody time

is it easy to edit?

[supple raptor]

Yep, it's just text and they're all labeled. Be aware that for any true/false values, they must be all lower case. If you feel like you messed something up, just delete the file and it will recreate it with default values the next time you launch the WebUI.

[worthy delta]

thats bery helpful to know

now just gotta find where evrything is

[supple raptor]

I haven't used it much myself, but I saved these weights from a random character LoRA that are supposed to be good for just getting the character from that one. They can vary between different loras so you might have to play around with them on the one you're using. 1,1,1,1,1,0,0,0,1,1,1,1,1,1,1,0,0

[worthy delta]

i barely understand loras

also the img2img do i hit generate now?

[supple raptor]

Yep you should be good to hit generate

[worthy delta]

aye thats good

thank you so mch for all of the help

[supple raptor]

LoRAs are like a miniature model. They won't work on their own, but they can be loaded on top of your model. It's a way to train new concepts, characters, styles, or whatever without needing to train a full new model.

[gusty palm]

i use extras tab for upscaling is that bad

ive been wanting to try img2img to improve images i already like though

that have liuke an extra arm here or there

[worthy delta]

i was exploring civitai and found a lora that adds detail

it works wonders

[slate night]

I use that too, I only have 128mb of vram so hi res fix is basically impossible lmao

[supple raptor]

It isn't necessarily bad. It's much faster than the img2img method. The benefit with the img2img method is that it allows the AI to actually improve the image after the upscaling instead of just relying on the upscaler to average and interpolate data.

[supple raptor]

wth are you generating on? a smart fridge?!?

[worthy delta]

before img2img n after

had to post it in here cause idk how to do a good comparison

it works wonders

[slate night]

Lmao my 6 year old MSI gaming laptop

I can hi res fix tiny images though!!!

Like 512x512 if I close out all other programs 😂

[worthy delta]

ill pray for ur laptop

[supple raptor]

I mean, whatever you're doing is certainly working because you post some banger images so keep it up.

[worthy delta]

i have learned so much from this discord in a hour or two

[slate night]

FR though tiled vae is a lifesaver

[supple raptor]

yeah tiled VAE is the real MVP

[slate night]

Got that upscaled with hi res fix and adetailer's face restorer so like

If you optimize you probably can gen on a smart fridge at this point if you want /hj

[worthy delta]

how do i make images look better all i know is basic prompts

[supple raptor]

Find LoRA's you like on CivitAI and steal other people's prompts to learn what they're doing.

[worthy delta]

LMAO

i just been copying them into a server i made just to upload ai imgs to

[supple raptor]

Do you have the tag complete extension yet? https://github.com/DominikDoom/a1111-sd-webui-tagcomplete.git

[worthy delta]

wuts that?

[supple raptor]

It adds autocomplete to your text fields using danbooru tags which are the tags that basically every anime model is trained on. It will also show how many images on the site use the tag which can help to get an idea of whether the tag will work or not. Usually tags that have less than about 3k images will have a weak effect and tags under 1k images will have either very little effect or actually hurt the quality.

[slate night]

Ooh that's one I didn't know about either

[supple raptor]

It will also autofill your embeddings and LoRAs if you start typing < followed by the name of your embedding or LoRA.

[worthy delta]

thats handy as hell

do u install it like the other one?

[supple raptor]

yep, just paste that in the Install From URL field. It may need a full SD Restart to learn the names of your embeddings and LoRAs.

[worthy delta]

i only got 2 lors tbh

like 5 ai models

like cetus mix

btw what ones are best to use?

[supple raptor]

I have a few LoRAs

[worthy delta]

dear god

also is this a good one?

i cant honestly tell

[supple raptor]

My favorite all around anime model is Aurora: https://civitai.com/models/40199?modelVersionId=45601
Dark Sushi Mix is another one I really like (especially the colorful version): https://civitai.com/models/24779?modelVersionId=93208
And lately, I've been playing with my own mix, Kakigori: https://civitai.com/models/100505/kakigori

[supple raptor]

eyes and hands could be a bit better, but overall it's nice.

[worthy delta]

aye thats good

how do i make them better

the hands are always wrong

also installed the tag complete one

[supple raptor]

This embedding helps with hands quite a bit for me. Just put it in .\stable-diffusion-webui\embeddings then use the filename without the extension in your negative prompt field. https://civitai.com/models/56519?modelVersionId=60938

[gusty palm]

thanks been looking for some new ones to try

does using multiple embeddings bad or anything

ive been using easynegative and verybadnegative together

[worthy delta]

what does embeddings do?

n what do i exactly put in the negative field

[supple raptor]

They can conflict with each other. negative embeddings are generally not too bad to use together tough. The one that I linked is specifically meant to not effect much other than hands so it is usually very compatible with other embeddings. I recommend trying some generations on the same seed, prompt and settings with different combinations of your negative embeddings to find what works well for your model and your style.

[gusty palm]

ok

[supple raptor]

Embeddings are another way that the AI can be trained, sort of like LoRAs. The main difference is that embeddings don't add anything new to the AI, instead they sort of allow it to learn things it already knows better. Negative embeddings specifically are embeddings that are trained on undesirable images and they are meant to be used in the negative prompt field.

[chrome moss]

thanks, but this one is still affected a bit by the art style of LoRA

[worthy delta]

how do i use them in the ngeative bit

[supple raptor]

This is my negative prompt that I use on all of my images negative_hand, (KHFB, AuroraNegative), (Worst Quality, Low Quality:1.2), border, nsfw, skimpy, multiple_girls, (watermark:1.2)

negative_hand, KHFB, and AuroraNegative are all negative embeddings. I already linked the first two. The other two are located here: https://huggingface.co/SweetLuna/Aurora/tree/main/AuroraEmbeddings

[supple raptor]

Once you have them in the embeddings folder all you have to do is type the file name of the embedding that you want to use in the prompt field that you want to use it in. So, if the filename is negative_hand.pt you would just type negative_hand in the negative prompt field.

[worthy delta]

ahhh gotcha gotcha

thank you!

downloading aurora aswell

[supple raptor]

You'll know they're working because it will list the ones that were used here:

[worthy delta]

i got a question what is ur most impressive img?

OMG

THE TAG THINGY

IT WORKS

thank you so much this will help so much

[pale monolith]

hmmm

[worthy delta]

oh my god

[worthy delta]

thats amazing

how did u make that work of art

[pale monolith]

for straight gen probably this one i think

[worthy delta]

straight gen?

[pale monolith]

like i didn't do any illustration touch ups, just stable diffusion and inpainting

i also illustrate some, used to do commission work

this one i did do a bit of work on

[pale monolith]

https://civitai.com/models/99169/amnesia

[worthy delta]

i dont know where to begin with inpainting

omg thats so cool

[pale monolith]

#🍥｜anime message

[worthy delta]

imma try that stuff when i understand ai art in general

[pale monolith]

it's scary to look at to begin with but with a little practice it gets to be ez

[worthy delta]

i just started this yesterday

[supple raptor]

IDK about most impressive, but I'm very fond of these two. They were both edited post generation though.

[worthy delta]

DAYUM

idk if ill ever get it to gen smting that good

[supple raptor]

The second one was a fluke. WebUI screwed up and either loaded something extra or didn't unload something that I had used before. I can't get anywhere near the style of that image again even with the exact parameters.

[worthy delta]

thank you for teaching me so much stuff

i cant thank u enough

[supple raptor]

My pleasure! I enjoy doing it

[worthy delta]

so far i love it

ive been trying to find a very specific art style

[gusty palm]

vbeen looking for the autocomplete thingy missed that from sygil

and ty for the default setting json tip for some reason the webui way to change settings defaults broke as of today for me

[supple raptor]

Yeah, autocomplete is a must have. I have most of the booru tags memorized at this point and I still can't imagine going back to not having it.

[pale monolith]

mmm? specific style? kinda like what?

[gusty palm]

i got the negativehand and aurorora thingy downlaoaded, whats theKHFB, is that for a psecific model

[pale monolith]

i collect like...all the style loras i might know something

[worthy delta]

its like this

i love this art style

[pale monolith]

khfb is designed for aurora but it works nicely as a negative on a lot

[worthy delta]

but cant replicate it

[gusty palm]

ok

[pale monolith]

hmmm do you have a higher res example? looks pretty realistic

[gusty palm]

is boobies okay here

[worthy delta]

ill see if i can find one

[pale monolith]

gotta be safe for work

[gusty palm]

ok

[worthy delta]

this art style i love

i need it in my ai

its from the game called dishonored 2

the enitre series of games have such beautiful paintings

@supple raptor sorry for the ping but for it auto matically X2 it does the number i want have to be divided?

[supple raptor]

Here's another one that I was pretty impressed with. This one only has some minor fixes in post.

[worthy delta]

oh my god thats amazing

what model did you use

arora?

aurora*

[supple raptor]

That was on my model, Kakigori

[worthy delta]

u make models??

[supple raptor]

In Hires fix, this is wher you set the upscaling multiplier.

[worthy delta]

so if i want 1980 i set it 990?

also ur model is so adorable

[supple raptor]

And in tiled diffusion (the img2img upscaling) you set it here.

[worthy delta]

ahh i see i see

[supple raptor]

Both of them are a multiplier based on the previous resolution. So if they are set to 2 they will double the previous resolution or if you set them to 1.5 they would multiply it by 1.5.

[worthy delta]

ahh thank you that helps so much

also is it possilbe to add a character as a base of a image even when the auto fill dosent pop up?

[supple raptor]

Yeah, you can still type whatever you want in the box. The autofill is basically just based on the list of tags used by a site called Danbooru. In general anime models will understand those tags better than it will normal words or names, but it doesn't hurt to try other things too. All of the models are trained on top of the base SD model so they all have some understanding of natural language and concepts that aren't on Danbooru.

[supple raptor]

I used to do some training, now I mostly just merge things that other people make.

[worthy delta]

thats hella impressive

[supple raptor]

Here's another cute one from my model.

[worthy delta]

ill keep my mind on that!

[worthy delta]

ur model looks so good

[worthy delta]

ooo thats really close

[pale monolith]

hmmm needs more sharp angles

[gusty palm]

each of these buttons gives u a different model?

mind kinda blown

i thought those were just descriptions of what the model was capable of

[exotic oxide]

Some art of Kiba Inuzuka from Naruto with my LoRA : ))

[supple raptor]

Yep those are all different versions of that model.

[worthy delta]

i didnt know that at all

[supple raptor]

Time to start collecting models, you got some catching up to do

[worthy delta]

holy god how do u use each of them?

i keep stuff minimal

[pale monolith]

heh

[worthy delta]

i get confused otherwise tbh

[worthy delta]

hmm not bad

lemme get a younger pic of him

[pale monolith]

ohhh more cell shading

[worthy delta]

its a very distinct art style

[pale monolith]

yeahhh i don't have anything specifically the same so i have to mix some loras together

also pretty hard to get grimy right without some specific ones

[worthy delta]

yeah the art style ive never seen before tbh

thats why it grips my soul

[supple raptor]

Yeah, turns out models that are mostly trained on pretty anime girls make grimy and tougher looking characters a bit of a challenge XD

[pale monolith]

luckily i have those couple specific loras

[worthy delta]

imma need those specific loras

[pale monolith]

ko knee chee wa

[worthy delta]

its looking good

[gusty palm]

whast the practical difference bettween novae and pruned as far as just generating images

[worthy delta]

dishonored has alot of very nice art tbh

[pale monolith]

https://civitai.com/models/84201?modelVersionId=89496

[wet elm]

is there any extension for some kind of queue? like i queue a prompt with a model and parameters

[pale monolith]

https://civitai.com/models/16048?modelVersionId=18989

[gusty palm]

i think xyz plot does something like that

[worthy delta]

thank you

[pale monolith]

the other ones i used i can't really link hehe

aetherion and cutesexyrobutts

[worthy delta]

i really like gotchic horror

[pale monolith]

both are on civitai

[worthy delta]

[supple raptor]

novae means that you have to load your own VAE manually (I suggest doing this anyways so that you always know which VAE you are using). pruned, technically can mean a lot of different things but in general it is used for fp16 models that have had components that are only used for continuing training removed.

[gusty palm]

ok

[pale monolith]

<lora:[LoConLoRA] pseudo-daylight偽日光 Concept_style:1> lora:DiscoElysiumStyle_v1:0.57 lora:Aetherion_style:0.33 lora:cutesexyrobuttsStyle_cutesexyrobuttsAnything:0.45 lora:Bloodborne-style:0.35

here's the values I used on my amnesia model

[gusty palm]

that pretty much happens automatically it seems

[supple raptor]

usually a novae model will also be pruned unless it says otherwise.

[gusty palm]

ive never seen anywhere where ito specify a vae

other than auto

[worthy delta]

imma do some image tests...

[supple raptor]

You'll have to download a vae before you'll have any options besides auto or none. This is the one that I use. https://huggingface.co/nubby/kl-f8-anime2-blessed/blob/main/WD1-4-kl-f8-anime2-bless09.ckpt
It goes in .\models\vae then you can load it in settings. I forget where exactly the setting usually is because I moved it to my quicksettings list. That one that said auto for you should be the right one though.

[gusty palm]

oh

[gusty palm]

xyz plot was in refrence ot this btw

is vae like a type of um

very asethetic esomething

[supple raptor]

The VAE is the part of the model that converts the numbers into pixels that you can see (and in the case of img2img or inpainting, pixels into numbers that the AI can manipulate). It is responsible for some fine details, such as textures, and shapes of small things like fingers, and to a lesser extent color, brightness and contrast,

[worthy delta]

i keep getting messed up feet on some of my image gens anyway to fix that?

[gusty palm]

ah

[pale monolith]

ah nubby's expertise

[supple raptor]

Anytime you generate images, it uses a VAE. Because otherwise there would be no image to look at. If you use a model without a VAE and don't load your own, the WebUI uses some very low quality VAE which will result in very muted colors and smudgy textures. The way the auto setting works is that it looks for a VAE that shares the same filename as the model you're loading and it loads that one.

[wet elm]

what

[worthy delta]

nubby bless me with knowledge

[supple raptor]

Feet are tougher than hands most of the time. I have the best luck with them when I prompt very little for them. Usually just adding barefoot in my prompt.

[gusty palm]

u can basically set it to generate images with different paramaters one after another , im not totally sure but thats what it sounded like

[worthy delta]

gonna try that

[pale monolith]

wait you don't know?

[worthy delta]

enlighten us

[supple raptor]

Are you wanting to use like multiple prompts or multiple models or what?

[gusty palm]

hands and feet are like the one thing ur not gonna get good looking

[supple raptor]

Tell me the secrets https://cdn.discordapp.com/emojis/842980480701694012.webp?size=64&name=monkaGIGAGUN&quality=lossless

[pale monolith]

[worthy delta]

they call me dadbeater for a reason

no dad has ever won against me

this is a warning tell us

[pale monolith]

good thing im not a dad

[worthy delta]

but u have one..

[pale monolith]

if you have located him then he deserves it

[supple raptor]

https://cdn.discordapp.com/emojis/1032001594507665468.webp?size=64&name=gunThreatenMurder2&quality=lossless

[worthy delta]

i feel that

swear they just disapear

[pale monolith]

finneee yknow there's goodhands lora

there's also k-feet mix lora for feets

not about to link that one either

but its also on civ

[supple raptor]

I found it. Previews are safe https://civitai.com/models/50084?modelVersionId=54633

Can't speak for the gallery images. I'm not unblurring those https://cdn.discordapp.com/emojis/1028509125044359259.webp?size=64&name=MomijiWide&quality=lossless

[pale monolith]

did...did you download and run it within 60 seconds?

how does your SD lora refresh so fast

[supple raptor]

I haven't run it, but I downloaded it within about 15 seconds of you sending the name of it XD

[pale monolith]

15 seconds

[worthy delta]

damn faster than me

i alr ran it n everything..

[supple raptor]

new civitai search is 🤌

[pale monolith]

i don't like it xD

[supple raptor]

it actually... you know. searches now

[pale monolith]

i have to clear my previous search keep getting word salad

[supple raptor]

It used to take like 45 seconds before it would find anything

Yeah, I would prefer it to clear the search each time you click on the box, but I'll take having to hit backspace over the search taking forever.

[pale monolith]

but you don't need to search when you download everything off of "new" when it appears

[supple raptor]

I almost never browse Civ. I just grab them when I find ones I don't have in other people's PNG info XD

[gusty palm]

i love this result

for some reason the colour is bad tho

[pale monolith]

imagine not being able to straight copy paste without downloading stuff first

[supple raptor]

My 4TB is filling faster than I'd like to admit https://cdn.discordapp.com/emojis/953452520682975294.webp?size=64&name=ReisenCri&quality=lossless

[pale monolith]

too many models, not enough loras

[supple raptor]

You're not wrong

https://cdn.discordapp.com/emojis/1111215347572420609.webp?size=64&name=truetrue&quality=lossless

[wet elm]

both actually

i would like a queue extension

[supple raptor]

Vlad's fork has that built in. https://github.com/vladmandic/automatic

I'm not sure if you can change models in between.

[wet elm]

imma be honest i dont know what that means lol

[worthy delta]

DAYUM

thats amazing

while im over here making pretty room

[pure oyster]

Architecture is better than waifus anyways

[astral drift]

Good backgrounds/environments trumps waifus imo

[worthy delta]

before img2img

omg i forgot to remove the prompts

now its gonna be scuffed

dear god

[stable crane]

compare to #🍥｜anime message (the 2x render without tiled vae), the 4x render messed up the ear

[pure oyster]

What model are you using? That composition is absolutely horrendous.

[stable crane]

a true gamer integrates his desk into his bed sir

[worthy delta]

I don't think it's the model tbh it's pumped out really goos stuff it's cetus mix

[loud falcon]

looks really good

[stable crane]

sick

[worthy delta]

thats horrible in a weirdly good way??

[stable crane]

rip fingers

[worthy delta]

in the background broken cities and the monster dragging them away

[loud falcon]

i didn't notice the city

actually looks good

[stable crane]

cute as always 🙂

[zinc flax]

Very cute
Not that flexible with some prompts though lol

[worthy delta]

ai art be hella intricate by accident

CYBERPUNK

[stable crane]

that'd be a sick apartment

double window view, end of building

imaginging a narrow AF building, but it's be so cool

too bad I can't post nipples because some of these are insane

[zinc flax]

Inpaint it

[stable crane]

I'm too critically lazy for that

[worthy delta]

Now I wanna see these

[worthy delta]

This one is pretty much gothic horror

[stable crane]

it's a lot like this, with more nipples

I can DM you it

[worthy delta]

Hell yeah

I wanna see dis

Shame art is sadly restricted by that stuff

[serene pike]

Hi

[loud falcon]

Sometimes it feels more like a lottery than anything

[worthy delta]

It is like a lotto but can be rigged if u know how

[stable crane]

it's so addicting

the perfect gacha

[zinc flax]

Try hands gacha

One of the gacha of all time

[loud falcon]

https://media.discordapp.net/attachments/255684579980083201/1116435847219253278/UhWhat.gif

[gusty palm]

that's fcikng amazing

I got some good old eldritch horrors somewhere

[unreal ridge]

still looks cool af

[loud falcon]

the prompt for this one?

[stable crane]

gn

[gusty palm]

i wish i was cheese

[stable crane]

what a delicious and nutritious meal

[loud falcon]

don't let them know your next move

[gusty palm]

adorble

[gusty palm]

katana arm

[pastel bluff]

Can anyone help to troubleshoot how come certain extensions dont appear on my local SD

[supple raptor]

Like in the list when you view available extensions? There's a bunch of them that aren't in that list. Auto1111 maintains the list that appears there and he doesn't add all of them either because the creator doesn't submit it, or he doesn't want to for whatever reason.

[elfin bolt]

Specifically what extensions?