🍥｜anime

[steady grail]

hello

[brittle rampart]

she 😁

[vital raptor]

i'm not that kind of captain though

I'm more of a pirate

[junior quest]

wait but I cant set more than 2048pix

[nova remnant]

double cat ears!

[nova remnant]

Why would you

[junior quest]

one is a hair band I guess

[brittle rampart]

one is assesories

[junior quest]

I got a image from pixiv and it has those infos but cant get a good image and it doesnt even input 3072 pix

[nova remnant]

I think 2000 would be all distorted

unless you're using other stuff

[vital raptor]

3072 is probably 1536 doubled, or 768 x 4

so they probably rendered 512x768 and upscaled x4

[junior quest]

nah, I can give the image if you want, I'll dm if you want

[vital raptor]

sure

[lyric zealot]

Upscale factor: 2

so @vital raptor is probably right about the person genning at 1536 x 2

[junior quest]

I sent it

[vital raptor]

that image has a lot of personality..

😏

[lyric zealot]

[junior quest]

hmm I dunno bout dat

but that image is super great,composition and other stuff

and I was blown away when I saw model is Abyss

[vital raptor]

so, it's using hires x2, but also controlnet tile, and it's a 1.5 model

so they're rendering at 512x768 and upscaling the rest

[junior quest]

[vital raptor]

i don't have that model, or the negative embeddings they used, but i put the prompt through my custom 1.5 model...

[junior quest]

but those control net settings arent inputting right?

and those upscaling settings too

[vital raptor]

they try too, i think tiled diffusion is something extra i don't have installed

[junior quest]

I got everthing ready but dunno how to setup dat control net settings

[junior quest]

eh?

[lyric zealot]

[vital raptor]

ok, figured it out

never used controlnet for upscaling before.. those settings are for img2img

they generated a 512x768 and used hires fix to x2 to 1024x1536, then put that in img2img to use controlnet to upscale to 2048x3072

that's why the resolution is showing so high in png info, because it's the img2img settings

[native halo]

u can also do a 4x upscale directly on high res fix

[nova remnant]

I like her cat ears

[signal panther]

WhAt'S tHe PrOmPt???

Слишком круто.

Hehe)))

[brittle rampart]

PNG-Info please.

[native halo]

needs more (skirt lift:2.8)

[brittle rampart]

againe!? 😂

[signal panther]

What's that?

Just gimme the prompt.

[brittle rampart]

BREAK
(preteen girl), long hair, braid, smile, blue eyes, maid, open mouth, twin braids, long sleeves, skirt hold, pink hair, very long hair, puffy sleeves, looking at viewer, simple background, :d, white apron, frilled apron, bangs, maid headdress, juliet sleeves, blush, bow, maid apron, frilled dress, black bow, petticoat, standing, upper teeth only, floating hair, puffy long sleeves
BREAK
(light romantic theme:1.4),(chatacter illustration:1.12),(intricately angle shot:1.14),(intricately pose:1.15),obscure light and shadow,((soft shadow)),chromatic aberration,ambient occlusion,cinematic light,depth of field,golden ratio,<lora:add_detail:0.85>,<lora:age_slider_v20:-2>,<lora:Mikimoto_haruhiko_style_v02:0.65>,<lora:Kitazume_Hiroyuki_style_v01.08:0.4>,21thkitazume,<lora:hyperdenoiser_v090:0.65>```
Funya`s wall from text prompt

🤣

[nova remnant]

https://tenor.com/view/cat-typing-cat-typing-fire-fire-gif-3192016087994715294

[native halo]

not an anime gif

https://tenor.com/view/gangsta-gangsta-anime-nikolas-brown-nick-brown-gif-18373282

[nova remnant]

https://tenor.com/view/rainy-zuko-rainy-day-zuko-southern-raiders-avatar-the-last-airbender-gif-25468390

anime gif

[native halo]

https://tenor.com/view/aesthetic-gun-ammunition-loaded-gif-16982788

[nova remnant]

Had ot cut the bottom part of the image, it had.. bad fingers

[native halo]

thicc oily fingers

[brittle rampart]

@native halo (skirt lift:0.1) 😝

@native halo (skirt lift:0.3) 😝😂

[nova remnant]

@native halo I have a skirt lift for you

[brittle rampart]

(skirt lift:0.31) 😂

[nova remnant]

next year I'll be able to see

[brittle rampart]

This is the wrong skirt! deception everywhere!

[nova remnant]

Its a modern one

[brittle rampart]

there must be intrigue!

[nova remnant]

👍

[brittle rampart]

I knew! 😂

[junior quest]

you didnt get caught? interesting

[nova remnant]

For what

its just clothes

[junior quest]

yah its just clothes

[nova remnant]

and a bed

[junior quest]

nah just forget bout dat

[signal panther]

Thanx

[signal panther]

What's Prince Zuko doing here?

[nova remnant]

Its an anime gif

for the anime chat

[signal panther]

https://tenor.com/view/well-no-fucking-shit-maximbady-no-really-really-no-way-gif-20513251

[signal panther]

First results

[signal panther]

[silent kraken]

Where tf does that syringe go

[cerulean bane]

https://tenor.com/view/why-do-the-mean-mods-sanat1ve-why-do-the-mean-mean-mods-mean-mods-keep-banning-gif-25086533

[bitter latch]

haha

ummmm wtf

😄

hahaha Love it

she is anime groot-hatsune?

[junior quest]

oh ma god style lora traning is hella hard

captioning

[bitter latch]

never tried it

not giving me wht I want. but nice tho

closer

Giurnnons Guranngaanns

ah ha

😄

[supple raptor]

you're not trying to caption by hand are you?

[junior quest]

but deepboru captioning nt good for style traning right?

so yeah hands.....

[supple raptor]

deepbooru is garbage for anything. use WDtagger

[junior quest]

this one? khoya

[supple raptor]

yep

SWIN-v2 is probably the best model version. You may need to still prune some tags afterwards, but it will likely tag the image more accurately than human captioning.

[junior quest]

do I need to change anything?

[supple raptor]

You can add a list of undesired tags if you want, for example if you know something shows up frequently in the dataset that you don't want to be tagged. If you have subfolders inside of the main image folder make sure to check the recursive checkbox. Other than that it looks good.

[junior quest]

@supple raptor wha res do we need to train lora? usually I create images in sd1.5 with 512x768, soo 512x512 or 768x768?

[supple raptor]

how much vram do you have?

[junior quest]

12

4070

[supple raptor]

I'd probably do 768x768. Make sure to enable bucketing if the images aren't already all cropped to 1:1 aspect ratio.

[odd flume]

nice quality and nice model/lora, can u share please what model/lora u use for this anime/cartoon?

[supple raptor]

I'm not overly experienced with training parameters so if someone else chimes in that has other recommendations, they would probably be the one to listen to for the training parameters.

[junior quest]

I see

[supple raptor]

To be clear, it's better to use bucketing instead of cropping images to 1:1. Training on all 1:1 images will make it harder to generate anything other than 1:1 images with it without significant artifacts.

[junior quest]

hmm its too much

[supple raptor]

lower res will speed up training and reduce vram usage. you can cancel and retry at 512x512 if you want.

[junior quest]

okay, btw I have 166 images is it too much?

maybe I'll try with lower epoches

and repeats

Im not sure this will be a good one, 5epoches with 3 repeats

[supple raptor]

Seems like a decent starting point. I'd let it cook to see how it does and adjust either repeats/epochs or learning rate from there.

[nova remnant]

Are SDXL loras much different

You can make a 1.5 one with like 20 images and be great

[supple raptor]

it's possible to make good loras with only a handful of images, but overfitting is much more likely. style loras particularly should have lots of images to help allow them to generalize across many subjects.

[queen void]

My LoRA's span from 10 images to up to 200. It always depends on how much data you can gather

[nova remnant]

[junior quest]

is there openpose model for sdxl?

controlnet

I got thesebut these arent work with character sheet

[brittle rampart]

history is silent about this

[brittle rampart]

looks very good

[nova remnant]

a custom model i made, hyperbloom uploaded on civitai

[nova remnant]

https://tenor.com/view/cat-cats-pet-cat-cat-pet-cute-cat-gif-24810247

[nova remnant]

@brittle rampart time for your private prompting class

[lyric zealot]

joins class

[nova remnant]

[brittle rampart]

I am “attention” itself!

[nova remnant]

My cute student

More dragon teacher

[brittle rampart]

more students for your private class😁

[nova remnant]

Send them all 👍

[junior quest]

hmm I got no idea how to create the same character again again

[nova remnant]

Make a lora

[junior quest]

how?

I need to create a unique character

not a existing one

[nova remnant]

Unique character with AI

[junior quest]

the hell?

[nova remnant]

Just gen one

[junior quest]

[nova remnant]

Call it unique, gen more

Make Lora

👍

[junior quest]

so images??

empty lora?

@nova remnant know any openpose models for sdxl?

controlnet

[nova remnant]

Nope, don't really use controlnet

[junior quest]

how dare you

[vital raptor]

work on your prompt to get 10-20 good images of your "unique" character, train lora with those, use the lora to gen better images, make a new lora, etc etc

[nova remnant]

Controller just bad what can I say

Keyboard and mouse for the win

[junior quest]

deym straight

[brittle rampart]

https://huggingface.co/lllyasviel/sd_control_collection/tree/main

[junior quest]

but I guess for sekiro its so hard with keyboar and mouse

[nova remnant]

Skill issue

[junior quest]

there isnt any openpose model for sdxl

I tested all of them

[junior quest]

they dun work....

[brittle rampart]

sad...

[brittle rampart]

if the creator of Control Net himself made non-working models then alas... I don’t think anyone will do it.

[junior quest]

im gonna cry now

[junior quest]

@supple raptor is there a way to create a same character again and again

not a existing character

[supple raptor]

Depends on the combination of tags. You've seen Yue, the girl that Vil always posts. She's not an existing character, but she can be recreated consistently on most models. I've got a couple that I've made that can also be made on most models.

[junior quest]

a lora or just prompting

[supple raptor]

Just prompting

[junior quest]

[nova remnant]

This definitely not existing character can be created in many models too

https://tenor.com/view/hmm-suspect-gif-22611582

[supple raptor]

True. If she was really remilia she would have bat wings.

[nova remnant]

I have to call Batman then

[west gate]

generated this but i forgor how to fix the fingers

[junior quest]

embedding?

[west gate]

i tried easynegative and badhand, i seem to forgot how they work

[brittle rampart]

they don't work properly. everyone is 80% jealous of the model

[supple raptor]

Start by just trying to upscale it in img2img.

[junior quest]

try this lora

[brittle rampart]

20% is a good angle and seed

[junior quest]

https://civitai.com/models/47085/envybetterhands-locon

[west gate]

model: cetusMix_Whalefall2

[junior quest]

for me this lora is works pretty much

[west gate]

img to image right?

[junior quest]

eh no txt 2 img

i dun use img 2 img much

[west gate]

and use the same seed?

i dont wanna waste such a good image

[junior quest]

ah I dunno bout dat, I always use that lora

[junior quest]

you can try

[junior quest]

seems kinda different

[west gate]

wdym?

[junior quest]

hmm still fingers are kinda cracked

[brittle rampart]

try other model... or corrent hands in Img-to-img

two level Boat look fantastic!

[west gate]

mine as a snapped leg lmao

[west gate]

ngl

the ship reminds me of IJK Akagi and Kaga

when they had 2 deck runways

[brittle rampart]

maybe

[west gate]

really cool stuff

[junior quest]

clip skip 2, (same seed), 910x512

[west gate]

ayo

she looks like Pyra

[west gate]

and a little oversaturated

[brittle rampart]

give me a prompt pls

[west gate]

random?

hmmm

space

[junior quest]

png info

[west gate]

ohh

[brittle rampart]

nothing

[west gate]

i have it

what the

[junior quest]

eh why

[west gate]

i cant copy paste text prompts??

[native halo]

the breasts

[junior quest]

ohhhhhhh deym

I guess I'll get ban

deym

[west gate]

the lone leg

[native halo]

(massive gargantuan gigantic titan hypermegagigaultra enormous mammoth huge tremensous colossal breasts:2.9)

[junior quest]

eh?

[native halo]

average civitai prompt

[junior quest]

I see I see

[west gate]

oml

this is much better

[junior quest]

[west gate]

the headless birds tho

lmao

[west gate]

ayo that first one gives me titanic vibes

[brittle rampart]

@west gate

[junior quest]

its not for meeee

@west gate its his work, prompt

[brittle rampart]

I was too lazy to rewrite the prompt

[junior quest]

there are sites that convert img to text tho

[brittle rampart]

why? i just use Tagger extention in WebUI

😁

[junior quest]

but you said there werent any info

thats why I send dat png info

[brittle rampart]

I just showed that the hands are the result of the model.

Tagger convert img to text tags

[brittle rampart]

oh this is not your idea of the picture at all!!! I'm not being attentive...

sorry

[junior quest]

[west gate]

a

[merry tendon]

Was inactive for a long time again lol, anyone remember me still?

[west gate]

heres anothere one i made a few hours ago

dat quality midriff is peak

[pure oyster]

ahh, a fellow stomach enjoyer.

[west gate]

🙏

[native halo]

needs higher resolution

[signal panther]

Thanx

[west gate]

9 tailed friend

[west gate]

She looks like she can kill me

:D

[signal panther]

Go back to your graveyard, Kitsune.

Send her back, @queen void , please.

SEND HER BACK!!!

[signal panther]

One tailed? Furry then.

[signal panther]

Oh, what's the prompt BTW, pls tell me, @queen void.

[queen void]

1girl, hair between eyes, big fox ears, big fox tail, (violet flower in hair:1.28), (horns:1.1)

[signal panther]

Thanx

[signal panther]

Good images!

[brittle rampart]

@gloomy scroll she changed?

[gloomy scroll]

?

[nova remnant]

Not showing enough

[native halo]

(from behind,wet skin:3.7)

[brittle rampart]

@gloomy scroll her face began to look different

[nova remnant]

Vil has many faces

[native halo]

we dun care about faces we want feet

[nova remnant]

Pirates of the Caribbean if it was from Disney

(I'm not sure if it's already from Disney)

[gloomy scroll]

Well here's feet #🍥｜anime message

[gloomy scroll]

I just went to style I wanted to do with XL but didn't figure out how

[nova remnant]

(foot focus:1.4)

[brittle rampart]

understand

[brittle rampart]

just some little pirates

[native halo]

https://tenor.com/view/one-piece-gunk-icanttakeitanymore-thistimeimreallygonnadoit-gif-8341758396657046929

[signal panther]

https://media.discordapp.net/attachments/373223103985090583/696015973345656872/spaz.gif

[native halo]

https://tenor.com/view/the-battle-bricks-gif-2710069450098972521

[signal panther]

https://tenor.com/view/shocked-computer-smile-yellow-smile-surprised-gif-26981337

https://tenor.com/view/roblox-build-a-boat-for-treasure-roblox-build-a-boat-plane-roblox-build-a-boat-for-treasure-gif-21269800

[native halo]

https://tenor.com/view/one-piece-one-piece-shitpost-ricardo-ricardo-milos-pirates-gif-26514534

[signal panther]

https://tenor.com/view/pool-water-woof-backrooms-rooms-gif-25222751

[native halo]

https://tenor.com/view/joe-biden-thumb-uncanny-canny-mr-incredible-gif-13123031614149167572

[nova remnant]

No anime gifs

[west gate]

https://tenor.com/view/topaz-topaz-hsr-hsr-hsr-topaz-honkai-star-rail-gif-7355095092656095365

here

[nova remnant]

[west gate]

what?

its anime

[native halo]

https://tenor.com/view/anemone-jam-gif-26159522

[nova remnant]

[nova remnant]

@native halo

[native halo]

https://tenor.com/view/anime-gif-16572778167231286334

[nova remnant]

Is that the ugly fake elf

[native halo]

no disrespect allowed towards Kiss-Shot Acerola-Orion Heart-Under-Blade

[nova remnant]

onion

[native halo]

https://tenor.com/view/anime-gun-reload-getting-ready-gif-17031430

[acoustic locust]

what model?

[queen void]

my own: https://civitai.com/models/11069/macaron-mix

[acoustic locust]

holy jesus stable cascade is huge

that's promising

[steady grail]

what's stable cascade?

[acoustic locust]

I opened some space for pso2 a while sgo

It's a new model

[steady grail]

how do you expect to load 34gb of stuff on the vram?

[native halo]

the back pocket

[acoustic locust]

I didn't consider that

[steady grail]

yeah sure XD

pretty sure it's not for casual users

[acoustic locust]

some people got it working with smaller vram cards

So I may as well make an attempt myself

[steady grail]

propably with loads and loads of cpus

but it must be slow as hell

[acoustic locust]

It's worth a shot

[west gate]

god damn, my boi looks like a hero

[acoustic locust]

turns out it's many models

[steady grail]

ohhhhh

effnet is the vae?

or just an encoder?

[acoustic locust]

no idea

I'll have to do some experimenting

[steady grail]

maybe is the encoder used to build the package

i suppose a b and c are the most promising branches

or versions

[acoustic locust]

apparently A and B almost work as a kind of VAE

[steady grail]

mh...

[acoustic locust]

why does a1111 keep caching downloads to my local disk?

[native halo]

https://tenor.com/view/anime-naruto-yeet-gif-12912103

[acoustic locust]

I'm saving that one

[acoustic locust]

I got stable cascade working

it's alright

running locally gives iffy results

[nova remnant]

what is that for

[acoustic locust]

Just testing stability's new model

[nova remnant]

#1207078178510872636

Is it supposedly better than sdxl in some way ?

[lyric zealot]

need 20gb vram?

[nova remnant]

that can't be

surely

[lyric zealot]

[gloomy scroll]

[vital raptor]

Thanks to Stable Cascade’s modular approach, the expected VRAM requirements for inference can be kept to approximately 20gb but can be further lowered by using the smaller variants (as mentioned before, this may also decrease the final output quality).

[lyric zealot]

since you have 4090, did you try it yet?

[steady grail]

with that much ram consumption you can illuminate all Africa

[vital raptor]

not yet, haven't had the time and last i looked a few days ago there wasn't a good extension yet

I could use comfy, but I'd rather not

[nova remnant]

[lyric zealot]

someone made an auto1111 extension

i can find it real quick

https://github.com/blue-pen5805/sdweb-easy-stablecascade-diffusers

someone said it's for forge as well

https://www.reddit.com/r/StableDiffusion/comments/1arwtap/a_colored_pencil_drawing_of_a_japanese_woman/ the post i got from

[nova remnant]

cat

[lyric zealot]

no cat

[nova remnant]

u stipud removed the cat

[lyric zealot]

https://tenor.com/view/คิมโรโระ-realkimroro-cat-fire-gif-16096997

[vital raptor]

cat

hmm

well, it makes a good looking cat

so, having the model loaded seems to be about the same as sdxl, my vram is only peaking at around 14-15gb

sometimes as high as 18

taking about 12-14s to generate

[lyric zealot]

and compared to your usual of?

[vital raptor]

for reference, the same size image on sdxl is about ~5s

[lyric zealot]

the model is like 30gb+ right?

[vital raptor]

but this is doing multiple sets of steps

[lyric zealot]

this is their comparison, are they saying 30 steps Stable cascade = 50 steps SDXL ?

[nova remnant]

Is it good for something

im too dumb for this

[vital raptor]

maybe with no xformers or anything, but I'm doubting that comparison under normal current usage

[lyric zealot]

you can read all about it here https://stability.ai/news/introducing-stable-cascade
but in short, it's basically 3 stages

[vital raptor]

i'm generating with the default 20+10 steps on cascade, and my usual ~5.5s sdxl gens are 40 steps but with xformers or some other acceleration

[lyric zealot]

there's no optimization for this yet?

[vital raptor]

not yet

[lyric zealot]

from what i can tell, it's a lot better with text than sdxl

[vital raptor]

sdxl, no xformers, 15.4s

9.3GB vram used

[lyric zealot]

no idea what it is, but there's a sophisticated feel to this

[vital raptor]

😮

anime girl with short black hair and purple eyes, cat ears saying "hello"

[lyric zealot]

try something real complicated

[vital raptor]

hands?

😄

[lyric zealot]

maybe

[vital raptor]

anime girl with short black hair and purple eyes, cat ears, saying "Peace", peace sign,

first gen was garbage, this is the second one

[lyric zealot]

decent i guess

i would forgive the finger for the perfect text

and tbh

i've seen worst finger in anime

[vital raptor]

yeah, my sdxl model does pretty decent text all things considered, but this is a base model, so even the anime girl is impressive

[lyric zealot]

true, and wait till people starts training on that

[vital raptor]

i guess i need to come up with some kind of wall of text prompt

[lyric zealot]

someone said it only took like 77 tokens

but im now not sure where i read that

so can't confirm if it's true or not

[vital raptor]

anime little girl with blonde twintails with red bows in her hair wearing a choker and a blue dress standing on the shore of a lake staring at her reflection in the water at night, full moon,

seems to have similar issues sdxl does with putting the subject IN the water instead of next to it

[lyric zealot]

try maybe saying "standing beside the lake"

no idea if that works tho

[vital raptor]

not really much difference

[lyric zealot]

[vital raptor]

i also can't get a high angle or 'from above' to work

[lyric zealot]

maybe they didn't train with danboru tag

[vital raptor]

yeah, "high angle shot" shoulld work though

[nova remnant]

Do something unreal

[vital raptor]

anime little girl with blonde twintails with red bows in her hair wearing a choker and a blue dress standing beside a lake staring at her reflection in the water at night, full moon, high angle shot,

[nova remnant]

ai and lakes

[vital raptor]

anime isn't real...

😄

[lyric zealot]

y not

they are real to me

[vital raptor]

anime girl vampire baring her fangs,

interesting, i saw someone say it can't do nsfw

based on my testing, it can if it's anime 😄

[lyric zealot]

send it to @nova remnant, he will judge if it's truly nsfw or not

[vital raptor]

it really can't do much, basic topless nudes, but not much else

anime girl flying through space in a colorful spacesuit

[supple raptor]

The prompt has in the water. I'd try rewording that part as it's probably the reason the girls are in the water.

[supple raptor]

Pretty sure the official code has a NSFW checker. They probably just had a skill issue.

[nova remnant]

butt

[vital raptor]

i tried some other variations, i could either get a distance shot of her literally 'next to a lake' or she was in it

[vital raptor]

cat girl booty

i think once a proper extension/integration is built for it and some custom training gets worked on, it could be pretty good

[nova remnant]

3 ears

[nova remnant]

You like the strong girls I know

[lyric zealot]

[supple raptor]

blessed xl vae testing

[lyric zealot]

2nd is best i assumed?

[supple raptor]

I'm testing some more prompts and artists now. That artist is pretty washed out normally (as you can see in the first column) so that one could be too intense for other combinations.

[lyric zealot]

I'm gonna assumed this is on pony?

[supple raptor]

close enough. it's autismmixConfetti

[lyric zealot]

I'm gonna need your prompt buddy

[supple raptor]

for reimu?
by mdf an, 1girl, hakurei reimu, brown hair, hair tubes, hair bow, detached sleeves, yellow ascot, simple background, white background, score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, source_anime
Neg: source_cartoon, source_furry, source_pony, painting, (worst quality, low quality:1.2), lowres, jpeg artifacts, scan artifacts, (abstract:0.9), @_@, sketch, lineart, chromatic aberration, unfinished, blurry, monochrome, censored, simple background, white background

[gloomy scroll]

[supple raptor]

I'm leaning towards #5 being best. #2 and #4 definitely have too much contrast and lose details in the dark areas.

[lyric zealot]

you are not using adetailer with this right?

[supple raptor]

no hiresfix or adetailer just pure txt2img at 1344x768

here's the one from the last batch with hires fix but no adetailer

[lyric zealot]

oh wow, it's without hiresfix as well

[nova remnant]

ugly

[supple raptor]

that single image had hires fix, but the grids were all without because I don't feel like waiting longer than I have to when I'm just trying to test colors.

[lyric zealot]

can see the adetailer square

[lyric zealot]

where do i find those vae?

is it available to public?

[supple raptor]

only on my hdd so far. haven't uploaded yet.

[lyric zealot]

[lyric zealot]

https://tenor.com/view/mine-now-steal-take-grab-mine-gif-13196166

except i don't have it

[lyric zealot]

ugly

[nova remnant]

[lyric zealot]

the adetailer is actually super obvious in landscape pic

[nova remnant]

make it do less denoise

aren't you at 0.4

[lyric zealot]

never thought of changing the default settings for adetailer

[lyric zealot]

yup, the default

[supple raptor]

it used to do that frequently for me. I don't know why it stopped because I was too stubborn to change the denoising.

[lyric zealot]

so changing denoising is probably the fix?

that makes me realised something. What if we adetailer first, then hires fix

[supple raptor]

you may also be able to hide it with higher mask blur

[lyric zealot]

[supple raptor]

you lose most of the benefits of adetailer. I've tried it in comfy

[lyric zealot]

im gonna be honest, i don't really understand how the mask thing works

[lyric zealot]

oh, dang

[nova remnant]

0.1 denoise

1.0 denoise

bee

[supple raptor]

it just detects a bounding box around the object that the adetailer model is intended to detect and inpaints that full area.
mask blur softens the edge around the box to help it blend into the rest of the image better.
padding pixels increases the amount of the image around the bounding box that are included when it crops the image to be used for context in the inpainting step.

[lyric zealot]

so mask blur is the same as offset, and padding pixels is the same as mask erosion?

this is for adetailer ^

[nova remnant]

why is it black and white

you need a vae

[lyric zealot]

[nova remnant]

https://tenor.com/view/cat-funny-cat-yoinky-sploinky-gif-22375025

[lyric zealot]

send me one real quick

feels like same thing

unless it's this seed

never mind, i know what you are talking about now

[supple raptor]

mask blur goes all the way around the mask, I only did half to illustrate the difference.

[lyric zealot]

[supple raptor]

with the inpaint only masked option, the image is cropped down to the area inside the outer white box (padding pixels), the cropped area is upscaled, it's inpainted with the are inside the red box masked, then it's downscaled to original res and stitched back into the original image.

[lyric zealot]

I'm gonna be honest, I don't fully understand, but well, if it works then that's all good

[nova remnant]

on my challenge here to make a sfw image

https://tenor.com/view/capoo-bugcat-capoo-bugcat-sweating-dannielteo-gif-16864462

so I can post in this server

the booba too powerful

[lyric zealot]

don't look at fingers

[nova remnant]

Don't look at it

[nova remnant]

snow

[lyric zealot]

last pic 10/10

[gloomy scroll]

[supple raptor]

In case anyone wants blessed versions of the SDXL vae (increased contrast & lowered brightness) https://huggingface.co/nubby/blessed-sdxl-vae-fp16-fix

[frigid bramble]

[steady grail]

i like sharpness, blessed makes things too much blurry

ehy guys, a new cute xl model https://civitai.com/models/308079

do you know if there is a way to make xl models in turbo?

maybe a merge?

[nova remnant]

I don't even notice a difference

[nova remnant]

same thing

[steady grail]

colors are more... whitey?

[nova remnant]

Its barely something

[steady grail]

well, yeah, but with upscaled images, it's like a punch in the face

[nova remnant]

white

[steady grail]

guys, how much is different a model merged with turbo, from a model with turbo lora loaded?

too much light made her blind

but cute

[nova remnant]

power of cyan eyes

I didn't want blue or white so I went with cyan cx

[steady grail]

snow can really make blind, did you know?

[nova remnant]

I live in the most tropical place ever

how would I know

[steady grail]

i almost never saw snow, but i studied light refraction in the past

looks like the light refracted from the snow, is so intense that a prolonged exposition in places totally covered by snow, can make people blind

talking of places like Antarctica, or north pole

so, blue eyes is totally legit

[nova remnant]

[steady grail]

do you know if with 4gbvram i can merge 2 sdxl models?

[nova remnant]

I think merging is mostly cpu

[steady grail]

oh

[nova remnant]

I asked once

forgot

[steady grail]

8gb cpu, not even closely enough XD

[supple raptor]

You're thinking of the SD1.5 vae that was named "BlessedVAE" or whatever which got it's name from the script used to adjust the contrast. That is blurry because it's just NAI's VAE and it was blessed because the contrast on NAI's vae is way too low. The Bless tweaks have nothing to do with that problem. "Blessing" a VAE is simply applying a contrast and/or brightness adjustment to the decoder outputs and optionally applying the inverse to encoder inputs so that hires fix doesn't end up with double the intended changes. https://github.com/sALTaccount/VAE-BlessUp

You could do the same adjustments that blessing a VAE does by simply loading the image into PS, Krita, GIMP or whatever image editor you like and adjusting the contrast/brightness in post. Applying it on the VAE just saves time and effort when it's an adjustment that you would consistently want to make on your images.

[nova remnant]

[nova remnant]

@steady grail more bloo eyes

[steady grail]

very tasty

[nova remnant]

[steady grail]

is incredible how xl is still so true to the prompt, even with 150 tokens

[nova remnant]

I do like max 20 words

[steady grail]

usually you get the best results

but i'm trying very unusual prospective and postures

i'm even using pony, score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, source_pony, rating_safe are already almost 75 tokens XD

[nova remnant]

you don't need source pony

thats for ponies

just type anime, if you're doing anime, maybe that helps, not sure with autism, definitely does on pony

[steady grail]

yep, the true focus of that model

XD

[nova remnant]

there is source anime, pony, furry and cartoon

I'm not sure what they actually do, supposedly it helps for those categories

[steady grail]

damn, if you call a model pony, you must have focused a lot on the subject lol

[nova remnant]

all that score stuff is one big tag x3

the author tried to train each score individually but the model learned its one long tag

[steady grail]

uahahaha

[nova remnant]

it kinda does all art stuff

[steady grail]

but it's the fcirst time i see a mlp character generated so well

[nova remnant]

@steady grail

[steady grail]

cute

[nova remnant]

she powered up

[steady grail]

i need some xl controlnet models 😦

i'm having so much fun with pony

it's very slow. almost 10 min for a 1024x1024 img, but god the quality

[nova remnant]

[gloomy scroll]

Some love for this chat too

[nova remnant]

Last lightning girl

[fathom totem]

Lightning girls you say? 🤣

[nova remnant]

@steady grail when you forget your homework, the teacher looks like this

[fathom totem]

angy teacher

[nova remnant]

[steady grail]

oh, on forge, i can set the noise

[gloomy scroll]

See ya

[nova remnant]

She will get the portal gun while you sleep

[gloomy scroll]

Yue is expert in space magic so she doesn't need toys like that

[nova remnant]

Smol

[gloomy scroll]

Let us fly to dream land with Yue's flying pillow

[fathom totem]

well this lil canadian's gotta head to bed, goodnight all!

[nova remnant]

https://tenor.com/view/cat-goodnight-gn-catgn-gncat-gif-11365475387544238236

[steady grail]

see ya

[nova remnant]

@steady grail this time you didn't forget your homework

👍

Looks like the smile from that statue in the beginning of solo boring

I mean leveling

[steady grail]

korean comics are all pretty boring

[nova remnant]

She is definitely for sure a nice teacher too!

[steady grail]

like chinese once

too much commercial

[nova remnant]

Some are fun

👍

The non drama ones

[steady grail]

my big brother have a pit in the brain

that was a blast XD

[nova remnant]

What is that

[steady grail]

comedy that joke about cliché flags

https://mangaweebs.org/manga/wo-jia-dashi-xiong-naozi-you-keng/

this one

[nova remnant]

Holy name

[steady grail]

a little bl, but still very funny

[nova remnant]

@steady grail do you ever remember character names

They're impossible

[steady grail]

no, i remember their hair colors

[nova remnant]

Lmao

[nova remnant]

200% body size head

[steady grail]

he is a very intelligent character

9 int 1 con

[nova remnant]

X3

Non anime but this was too funny

Minecraft furnace

[junior quest]

cascade?

[junior quest]

make her or him smile, slightly

[nova remnant]

No

[junior quest]

its taking so much long to switch to a model....

HDD speed

[steady grail]

🤣a good idea for tomorrow

going to sleep

see ya

[junior quest]

kya

[nova remnant]

Stop it

https://tenor.com/view/stop-it-get-some-help-gif-15058124

Get some ssd

@junior quest

[junior quest]

yo

[junior quest]

SSD is full

[nova remnant]

[junior quest]

C and F (SSD)

[nova remnant]

XD

[junior quest]

[nova remnant]

I like the typo

[junior quest]

whats da modal? not cascade?

interesting

[nova remnant]

It's Bard

[junior quest]

bard? gemini?

[nova remnant]

Yes

[junior quest]

wait he doesnt even create a mario image for me

[nova remnant]

The one and only

Copyright

Censor

Anti fun

[native halo]

bard? more like baaad

https://tenor.com/view/default-dance-mario-fortnite-mario64-dance-gif-15064465

[nova remnant]

@native halo gift for you

[nova remnant]

[native halo]

needs more bikinis

[potent shell]

新的

[nova remnant]

[nova remnant]

Ball is on the way

[junior quest]

wait is dat a moustache?

[native halo]

she likes to play with them balls ( in the beach)

[long python]

👍

[junior quest]

trip her

[pallid moss]

Why can't I merge checkpoints

It's giving an error of no memory

[native halo]

not enough memory

[west gate]

smh smh, cant remember

alzheimer PC go brr

/j

[nova remnant]

Many tails to grab

[lyric zealot]

https://tenor.com/view/anime-girl-hit-bonk-slap-gif-24662766

@wooden coral

@dull jackal

It's fake for everyone wondering

[native halo]

https://tenor.com/view/nitro-discord-please-officiale-gnomed-gif-22298476

real

[lyric zealot]

https://tenor.com/view/anya-forger-excited-oh-ohhhhhh-kanna-kamui-gif-13089481845460202864

[wooden coral]

[lyric zealot]

my free nitro

[junior quest]

what models?

[wheat falcon]

what model is that? outlines look so clean

or is it a outline lora?

[lyric zealot]

there's a lora for that, but not sure if it's the lora that's making the outline looking like that. ```score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, rating_safe,

1girl, lucy (cyberpunk), swept bangs, purple eyes, ringed eyes, crop top, makeup, smug, tongue,

source_anime, lora:marumoru_style:0.9 marumoru_style
Negative prompt: (worst quality, low quality:1.2), 3d, censored, abdominal muscle, artistic glow
Steps: 35, Sampler: Euler a, CFG scale: 7, Seed: 815877603, Size: 768x1024, Model hash: ac006fdd7e, Model: autismmixSDXL_autismmixConfetti, VAE hash: 66f78ad136, VAE: sdxl_vae-fp16fix-blessed.safetensors, Denoising strength: 0.4, Clip skip: 2, ADetailer model: face_yolov8n.pt, ADetailer confidence: 0.3, ADetailer dilate erode: 4, ADetailer mask blur: 4, ADetailer denoising strength: 0.4, ADetailer inpaint only masked: True, ADetailer inpaint padding: 32, ADetailer version: 24.1.2, Hires upscale: 1.5, Hires steps: 10, Hires upscaler: 4x_fatal_Anime_500000_G, Mask blur: 4, Inpaint area: Only masked, Masked area padding: 32, Version: f0.0.12-latest-155-gd81e353d```

[wheat falcon]

ah maybe cause of sdxl and the style lora

thanks!

[lyric zealot]

it's likely due to the model

[wheat falcon]

testing Zluda right now

AMD speedup on windows xD

[lyric zealot]

anyway, almost all my generation should have pnginfo, but you need to have the stealth_pnginfo extension to view it. ||Why I used the extension is because just copying from webui and pasting it to discord will already have the pnginfo embedded||

[junior quest]

ALMOST

[lyric zealot]

is that the cuda equivalent thingy?

[wheat falcon]

yes, its luka cuda emulator xD

works great so far

for 1.5 models

sdxl gives driver timeout xD

[lyric zealot]

yup, if it's done with some new ui that im testing, then it will likely not have any pnginfo, so it's almost

doesnt concern nvidia gpu right?

[wheat falcon]

yea nothing for nvidia

only the fear that amd gets better speed

[lyric zealot]

[supple raptor]

it's basically letting amd pretend to be nvidia.

[lyric zealot]

i see it as a win win tbh

[supple raptor]

absolutely. I'd love to see amd cards become more viable

[lyric zealot]

it's not like it's gonna automatically make my nvidia gpu slower

[wheat falcon]

i see it as a preview for rocm on windows xD

[supple raptor]

does zluda let you use xformers with amd?

[wheat falcon]

nope :/

[supple raptor]

sadge. at least sdp is close these days though.

[pallid moss]

Darksushimix and Rev animated

[wheat falcon]

yea for now it crashed with --opt-sdp-attention, but i have hopes for the future

[junior quest]

sadge

[supple raptor]

there's nothing there for nvidia to benefit from. zluda is just translating cuda to amd and nvidia already has cuda.

[junior quest]

hmm but as I saw it doesnt involve vram, dunno but memory(ram right?)

[pallid moss]

Mine has 8 gb of ram

[native halo]

thats not enough to merge models

[junior quest]

I guess so

[supple raptor]

for sd1.5 models with minimal background tasks it should be good enough.

[native halo]

try with this extension instead of supermerger https://github.com/wkpark/sd-webui-model-mixer

[supple raptor]

it'll be tight and might use pagefile, but it should finish.

[junior quest]

for me to merge it got like 6gb of ram, so yeah I guess probably you dun have enough memory with other programs?

[native halo]

just download more ram easy 👍

[junior quest]

info bout dress?

[junior quest]

this dress too

[pure oyster]

More sample images I'm hand editing for my LORA training.

[gloomy scroll]

Throne sitting

[pure oyster]

Throne poses are a surprisingly good method of showing off character design.

[native halo]

i prefer the (squatting:2.5) pose

[signal panther]

Frost queen, short skirt with long dress behind

That's how

No my prompt)

[cyan burrow]

heh the people in here rotate but the image types being generated stay the same

[vital raptor]

it's the same people

[native halo]

not like u can generate mechas with the models we have ( they all look like gundams)

[nova remnant]

but mechas all look the same ?

[junior quest]

how many images it should be good for a character lora?

100?

[nova remnant]

20 is a starting point

Is your character giga complex

with symbols and crap like that ?

Then more I'd guess

just have different poses and expressions, should be gud

[junior quest]

https://tenor.com/view/raiden-genshin-impact-genshin-shogun-gif-23146931

her

I cant find a good lora from civitai

almost all of her lora creates her but somewhat childish

[nova remnant]

are you using pony

[junior quest]

nope

she cant be childish at all

shes 3000 years old

[nova remnant]

by the power of anime

https://tenor.com/view/clueless-gif-24395495

[junior quest]

shes a GOD

[native halo]

no the ones from vottoms/eureka/guilty crown/evangelion/gurren lagann/appleseed/gasaraki look different than common robots

[nova remnant]

steven is typing

[native halo]

https://tenor.com/view/typing-anime-hannah-ibr-gif-gif-19616973

[nova remnant]

all the same to me

[native halo]

https://tenor.com/view/anime-gun-reload-getting-ready-gif-17031430

[dull jackal]

I saw a ping Somewhere

[dull jackal]

f u n

[signal panther]

Don't forget to send here the results, ping me, please!

[dull jackal]

I like how the Nitro is impossible to read if your eyes aren't fully open

[junior quest]

sure

[peak shadow]

no the ones from vottoms/eureka/guilty crown/evangelion/gurren lagann/appleseed/gasaraki look different than common robots

[native halo]

https://tenor.com/view/anime-azumanga-daioh-shocked-chiyo-gif-18944897

[nova remnant]

They are onto you

The discord Mecha police

[native halo]

https://tenor.com/view/ohranger-mecha-dance-gif-15557401

[nova remnant]

Power rangers is so weird

[lyric zealot]

https://tenor.com/view/power-rangers-break-gif-26320953

[junior quest]

waaw hands

[signal panther]

She damn rockin'!

[signal panther]

Pros when they are said 'No pinge harmonicas'

Guys, let us use ai for a more constructive outlet. How about historical ladies?

[signal panther]

Yeah, like this (Thanx @vital raptor )

[signal panther]

And like this (thx @lyric zealot )

[lyric zealot]

i didn't particularly generate historical stuff tho

just kinda similar

[signal panther]

https://tenor.com/view/waiting-cat-filing-nail-getting-ready-bring-it-on-im-listening-gif-4387660213472769374

[lyric zealot]

long neck

[signal panther]

That's what I got for historical ladies:

[signal panther]

REDNECK 🤣

[lyric zealot]