🍥｜anime

[nova remnant]

thanks

i needed that

👍

[lyric zealot]

stoopid

[keen hollow]

I use 4090s as coasters.

[nova remnant]

bro 😭

slide one my way

[keen hollow]

"bet"

[nova remnant]

[keen hollow]

ugh my prompting is awful tonight 😛

[nova remnant]

how so

[lyric zealot]

[keen hollow]

just, bad results

[nova remnant]

felt

i just keep genning until it doesnt 💀

[keen hollow]

gotta gen at 32x32, 3 steps, for ultrafast...

[nova remnant]

sounds horrible

[keen hollow]

you just gotta use Smol's magic model

[lyric zealot]

[keen hollow]

aesthetic, yes... anatomy, fail

[nova remnant]

👁️

[keen hollow]

🐓

ok ... time to shut it down tonight

more generations later 😛

[lyric zealot]

[nova remnant]

[keen hollow]

g'night

[unkempt cape]

can xl hands?

xl can hands 😁😁😁😁😁

altho it likes to not have hands

[gleaming flame]

morning

[steady grail]

the colors are so dry, but the reflex are very cool

[cerulean cosmos]

trying some speech bubbles

[junior quest]

NAI or sm open source model?

[steady grail]

too much tomato sauce

[oak spoke]

here you go no tomato

[steady grail]

now she is hungry

[oak spoke]

she want to eat tomato on other girl

[unkempt cape]

holy f**k its so indistinguishable

[nova remnant]

Did you make the ugly eyes Lora?

[unkempt cape]

yep

eyes are broken exactly like how it was originally 😁😁😁😁😁😁😁😁

[nova remnant]

Which model is that image?

The one you just posted

[unkempt cape]

animagine xl 3.0

duh

[nova remnant]

Tpony has a pretty nice "nai" look to it

Try that one too

With that lora

[unkempt cape]

but it was trained with animagine xl 3

and tbh do i have a reason to switch if potentially it fails

[nova remnant]

Oh ya

[unkempt cape]

talking about this im just gonna drop my guess on wtf nai's secret sauce is

[nova remnant]

I think you can't use non pony Loras on pony based models

[unkempt cape]

my guess is that they train primarily on artists

thats why without artists tag it looks generic and mid as fuck

and if the secret sauce is say game breaking then maybe they fucked with the architecture of the models so like everyone gets a mini-lora inside the big model

[nova remnant]

It's also pretty trained on brands lmao

[oak spoke]

male ver cuz we lack male art in here

[gloomy lodge]

idk what those are its just some model off civit

[keen hollow]

so you feel like you finally got the result you want?

[junior quest]

so what is da models name?

[gloomy lodge]

first Dark then 50% of the way through with refiner forty want the prompt?

[young glen]

Friday.

[supple raptor]

Their secret sauce is basically 3 things. A good, and very large dataset that's well tagged. Scale loss based on tag frequency inside each epoch to make overly common tags less likely to overfit and less common tags more likely to be learned. And enough compute power to train multiple versions of the model in a reasonable amount of time to allow them to actually test training hyperparameters.
Their people have confirmed that they do not have LoRAs that get applied at generation.

[nova remnant]

What does NAI look like without artists

[ivory pebble]

Money. That's it, that's their "secret sauce"

[gloomy scroll]

Also well watermarked

[supple raptor]

the style is pretty mediocre without artist or other tags guiding it. which is to be expected since the model's goal would be to try to create something that's sort of an average of everything it's seen during learning.

[terse ore]

Whenever I try to put trees through img2img, they always look terribad. Really ugly trees with lots of artifacts

[brazen marsh]

#1070829655638151208

Draw a simple uncle

#1047610792226340935 Draw a simple uncle

[keen hollow]

evenin'

[ornate echo]

Here is the image you requested:

[nova remnant]

heya

what do i do if it seems like my cfg scale isnt working

[keen hollow]

lemme see

[nova remnant]

how do i show

[keen hollow]

run a x/y/z plot

[nova remnant]

bet

what is an x/y/z plot

[nova remnant]

ive never done this before, idk what im doing 😭

[keen hollow]

no prob it's simple

[nova remnant]

loooool

i made one work

[keen hollow]

select X/Y/Z prompt

change one dropdown to 'cfg'

[nova remnant]

okay so i did that part right

[keen hollow]

'cfg scale' rather

then in the box type whichever ones you wanna try, comma separated

so maybe try 3,5,7

and then if it doesn't work, you can get upset and yell....

[nova remnant]

LOL

alright bet

gonna try it rn

[keen hollow]

ok cool, gamble, give it a try

[nova remnant]

its taking forever loool

wildly low quality

[keen hollow]

clearly differences so I guess it's workin'

[nova remnant]

guess so

maybe its my prompt then

ig ill check it

forge just isnt interpreting my prompt very well rn on any model i use

[keen hollow]

trying to decide if vip line-skip access is worth it to an upcoming convention show... hmm

[nova remnant]

mmmm

i think its always worth it tbh

lines are annoying and if i had the chance to skip one i would

[keen hollow]

yeah? which ones have you been to?

[keen hollow]

i do hate lines...

[nova remnant]

✨ none ✨

im just saying if i ever do go to one

i wouldnt want to wait in a line

[keen hollow]

oh ha

ok

well maybe you should go 😄

[nova remnant]

mmmm it sounds fun but im not a fan of crowded places

or going outside

[keen hollow]

oh i see

well.. maybe just watch some netflix then

[nova remnant]

[keen hollow]

cat

[nova remnant]

I'm dying

[keen hollow]

😮

[nova remnant]

Tired

[keen hollow]

yeah i need t go to sleep

[nova remnant]

I need to generate strength

[nova remnant]

We tired

[gloomy lodge]

gotta love test grids

gotta make sure you leave no stone unproompted

[acoustic locust]

what model?

[acoustic locust]

that is not what I meant when i said look at the big picture

lmao

[tidal phoenix]

Muskets are DIFFICULT, but I managed to make it work with some manual editing (even if it isn't even remotely perfect)

[tidal phoenix]

Updated version XD

[viscid sapphire]

What prompt should I put so that a little skin can be seen between the shirt and between the skirt or pants, a small line?

[nova remnant]

Midriff

I believe

[green temple]

ChromaNEOFT

Anyone know what model is this ?

[tidal phoenix]

Midriff but negative prompt navel, maybe, keep it from going too high

[gloomy scroll]

Remember that waifus need to as stable as stable diffusion

[nova remnant]

[steady grail]

https://tenor.com/view/it's-a-trap-dominic-toretto-vin-diesel-fast-x-they're-luring-you-in-gif-13500803666361148163

[unkempt cape]

https://www.reddit.com/r/StableDiffusion/comments/191ij1z/did_you_know_what_checkpoint_model_is_this_i_like/

lol relatable

[native halo]

https://tenor.com/view/ok-and-ok-and-whole-sqaud-gif-22248092

[native halo]

https://media.discordapp.net/attachments/963215490862436422/1223889293119062076/GIF_20240327_214433_789.gif

[keen hollow]

ha this is my life... then you complete a grid of 4000 images and still can't decide what the best combos are. 😄 😄

[silent kraken]

It probably takes a couple of hours just looking at them all

[gloomy scroll]

[nova remnant]

yoooo thats hawt

[nova remnant]

nice res

smuggered

@lyric zealot do the smuggered emote in my place

be useful

[nova remnant]

useless

[lyric zealot]

Is it possible to do style transfer on some helldivers picture, with yue as the character?

[gloomy scroll]

[bitter latch]

ok now this is kewl

[lyric zealot]

I guess we need yue on the frontline

[gloomy scroll]

She all ready called in orbital bombardment

[bitter latch]

hmm I think she had the wrong meds...

[bitter latch]

ahha there we go, 40k helldiver

ok went a little weird....

refiner on regular works

[steady grail]

guys, i'm testing deep cache, but only gives black images. maybe that stuff is not compatible with pony?

[steady grail]

wow, for being realistic, i like the colors very much

[bitter latch]

thanks

generate 2048 x880 .. .just cause

waits

hmm why is she gaint?

there thats better

[haughty rapids]

Bro how to generate a image

I can't do it

[ornate echo]

Bro did you install some software to do it?
Because there's no running image generation bot on this discord server.

[haughty rapids]

Oh

[ornate echo]

#1047610792226340935

[haughty rapids]

I used to be one

Did they remove that?

[ornate echo]

Did you read what I linked?

That's the latest.

[haughty rapids]

Oh thanks, buddy hope they fix it soon

[ornate echo]

Why not just install locally?

[haughty rapids]

i just need a couple of pictures for my wallpaper

[ornate echo]

So go get them from some online service?

[haughty rapids]

It's just too much work for wallpaper 😅

[haughty rapids]

Yeah

[ornate echo]

This is my wallpaper...was worth.

[haughty rapids]

Cool asf

I want a Chinese sword girl in a red and black dress

[unkempt cape]

ask in general or something

they will take any of your messages and use it as a prompt

because

[keen hollow]

neat

[keen hollow]

ill make you one

what size?

[haughty rapids]

Oh thanks mate my display is 6.3 19.5:9 aspect ratio

[keen hollow]

too many numbers....

[haughty rapids]

19.5:9 aspect ratio

Long hair please if I don't mind 🙃

[keen hollow]

ok

lemme see

[haughty rapids]

Thank you ❤️

[keen hollow]

harder to do wide, need a few 😛

[bitter latch]

little bit.. off.. hmmmms

[haughty rapids]

Really good and yeah the sword was little bit off

[ornate echo]

You could maybe crop this to what you need...different style altogether, though.

[keen hollow]

that is one long sword

[haughty rapids]

What model is it sdxl?

[ornate echo]

That's what she said.

That's a mixture of Cascade & SDXL.

[haughty rapids]

Haha i wish my girl said the same😅

[haughty rapids]

Ah love this

[haughty rapids]

How do you combine those 2

[supple raptor]

generate in one then img2img in the other, possibly with upscaling in between.

[bitter latch]

gottta share for the laughs

[keen hollow]

hand needs work...

[haughty rapids]

I need Chinese girl not a Chinese nightmare

[haughty rapids]

It's like manually changing the image between?

[haughty rapids]

This one has a different vibe

[keen hollow]

one more comin' ... 90%

[haughty rapids]

Wow

[bitter latch]

uses bing create 😄

on my machine

[keen hollow]

oh come on .. bing is cheating 😛

[supple raptor]

sort of. normal txt2img is essentially starting with a canvas filled with a bunch of noise and denoising it until it becomes an image. img2img replaces that noisy canvas with an actual image (in this case the image that was output by the first model). upscaling can be done several different ways, but it basically means increasing the resolution of the image and using some method to approximate what the extra pixels should be.

[keen hollow]

I think nubby should make one.

com'on nubs

[haughty rapids]

Bing wtf

[bitter latch]

😄

[haughty rapids]

Ah it's dall e 3 i guess

[bitter latch]

ok fine. pulls out the big guns on his machine

[haughty rapids]

Billions of dollars into it

[bitter latch]

my own customer merge model

[bitter latch]

ok she threw the blade, what?

[haughty rapids]

😂 but she looks so good

Imagine what sora can do when it's released

[bitter latch]

thats better

[haughty rapids]

This is perfect can you please make it high res so i can crops it if its possible

[bitter latch]

ooh

[keen hollow]

oops wrong way 😛

[bitter latch]

hmm okk then

od but .... shrugs

[keen hollow]

that's a weird one

[bitter latch]

ummm...

trying sampler lcm karras, odd results

ok wtf did the bunny come from?

[keen hollow]

rabbits again?

[bitter latch]

yeah no where is there rabbit in the prompt wtf

[keen hollow]

dall-e is a cheat code 😛

[bitter latch]

swords are kinda funky

[keen hollow]

'til tomorrow

[sick brook]

green

[nova remnant]

Eep

[cyan burrow]

weird how much I grew out of AI Image Generation

now my focus is to learn to model properly, buy a 3d printer. Then also use the models to create a webcomic. Hohoho

fun times

gotta celebrate being alive for long as I am

[acoustic locust]

drake lookin'

[acoustic locust]

what model?

[acoustic locust]

These are also cool, is this dalle?

[unkempt cape]

guys if i say want to train a lora for a specific combo of artists on nai what is the best way to prepare a dataset? say i want 35 and because yk style i will prolly have to have 35 somewhat different images and thats a bit of a problem

[keen hollow]

First set was the Confetti Autisimmix. Second was indeed dalle.

[acoustic locust]

Thanks

[acoustic locust]

Nerf dalle, it’s ruining the meta.

[keen hollow]

Do what now?

[acoustic locust]

It's a joke

[native halo]

https://tenor.com/view/joker-not-a-joke-there-is-no-punchline-bill-murry-gif-22175591

[nova remnant]

Movie

Ugly

[native halo]

https://tenor.com/view/society-fart-poo-monkey-chicken-gif-19829617

[unkempt cape]

im so confused

suddenly have like 2000 anlas in nai (didnt pay a cent)

doesnt seem to be a trial or anything other accounts dont work

oh its some age old stuff

holy fuck lets go i dug up another acc siuuuu

[native halo]

nya

[unkempt cape]

dunno wtf she is putting close to her mouth but pic looks cool so whatever

[nova remnant]

Radioactive chili

[pure oyster]

I was looking for references for different clothing items and I found a clothing seller that uses only ai images.

very shady

[keen hollow]

AI images of the actual items for sale?

[pure oyster]

This is an example from the site, it's not even passible lmao.

[keen hollow]

it's not?

seems good to me

[pure oyster]

could be just bad compositing work, but most of them are this egregious.

[keen hollow]

is it a bad hand? or just a bad resolution photo?

[pure oyster]

I would say several vastly different images (both ai and real) composited into each other, by an underpaid intern.

[supple raptor]

neither of those are ai. they did a terrible job at removing the background on that top photo though.

[pure oyster]

It is hard to tell since everything is warped and incoherent.

[steady grail]

nah, it's her bone structure.... in the hair :/

[nova remnant]

Don't doubt how weird hair can look messed up

[supple raptor]

all you're seeing there is where someone sloppily traced around her and deleted the background.

[green temple]

@nova remnant cat are u ok ? ...

[nova remnant]

Why do you ask

[green temple]

nevermind

[keen hollow]

evenin'

[nova remnant]

[haughty rapids]

What model is best for text based images. like making memes and stuff

[native halo]

the trumbama model

[haughty rapids]

What

By stable diffusion?

[native halo]

by balls diffusion

[keen hollow]

Probably an XL model

[haughty rapids]

Cascade is good i think

[supple raptor]

don't bother. SD3 looks like it might be reasonably ok at it, until that's released Dalle3 is the only one worth even trying and it's still not very reliable. Just make the image without text and add text in post.

[haughty rapids]

Yeah that's make sense

What is the difference between sdxl and sdxl turbo

Quality is worst in turbo?

[keen hollow]

turbo runs with the lower step count and lower cfg guidance

[haughty rapids]

It's faster but low quality right

Whats cheaper to run on a cloud ?

[keen hollow]

some dalle for the evening

[primal wharf]

Elixir of immortality advertisement idea from my manga.

[supple raptor]

depends how your billing works for the cloud service. If it's based on time that you keep the server running, then turbo could generate either more images in the same time, or the same number of images in less time. If you pay by the month, then the cost is exactly the same for both because the vram requirements are exactly the same.

[nova remnant]

[native halo]

https://tenor.com/view/ngg-gif-27112455

[acoustic locust]

what model

[acoustic locust]

is this like an addon or a plugin?

[vestal crypt]

Nope, just generate 9 images and photoshop them

[acoustic locust]

ok

[vestal crypt]

I tried jsut generate the whole emotion charts but failed

[acoustic locust]

damn

[green temple]

Same model , chromaneoFT

[green temple]

https://huggingface.co/zzzAI19/ChromaFT

[nova remnant]

The cute model

[native halo]

am i cute too ?

[green temple]

Yes

[green temple]

Paw pose ?

[native halo]

yes a few (hundred) for testing

[vital raptor]

hmmm, i need a script to do hires fix on a whole folder of images

well, maybe... the batch upscaler in the extras tab seems to do pretty good for the most part

[nova remnant]

Grabby

[nova remnant]

Tiny grabby

[native halo]

https://tenor.com/view/slippy-i-saw-what-you-deleted-slippydesnake-skyler-gif-10721741480770112366

[junior quest]

ctrl + a. then shift + delete

there you go

[native halo]

cant delete waifu

[junior quest]

aye

[native halo]

grab these hands instead 👊

[vital raptor]

https://tenor.com/view/hyakkano-inda-karane-karane-inda-anime-hands-anime-holding-hands-gif-6125029489671662353

[silent kraken]

Who?

[vital raptor]

figures, i found a script to do batch hires fix on a folder of images and it doesn't work and everything else i find just says to use img2img batch to upscale as if it's the same thing

ah, of course, it's because a1111 UI is so well laid out, i didn't have the same upscaler set in some random setting page 🙄, it's mostly the same now

now to wait for 220 images to process..

i'll never understand why they can't just put options like that in the proper interface instead of hiding it in settings

or hide the hires fix one next to it so they're at least similar

[supple raptor]

Because it's already there. Hires fix is just upscale+img2img. The upscaler selection is just in a dumb spot. Set Upscaler for img2img in settings (you can also add that to your quicksettings if you want to be able to change it on the fly) then in the img2img tab where you select resolution switch to Resize by mode and set it to the upscaling factor you want. The only thing missing at this point is using the same seed as the original image, which is not necessary since seed X at Y resolution is not similar to seed X at Z resolution , but if you want it anyways then you can specify seed in the prompt if you use the fork of UmiAi-Wildcards that I've linked here a few dozen times.

Oh, that last part is unnecessary. That's already an option to append in the PNG Info section of Img2img's Batch tab.

[vital raptor]

Hires fix is just upscale+img2img

yes, but nowhere in auto1111 does it tell you that, you have to dig around in reddit posts or old github posts to find anything on it, and by default for hires fix there's just a simple dropdown to change the upscaler, versus the img2img tab where it's buried in settings instead

it should be on the img2img tab so you can find it easily, otherwise it uses the default (garbage) upscaler

[unkempt cape]

idk if ur that guy who knows about wd (sorry if ur not) but any updates on new waifu diffusion?

[supple raptor]

not yet. training at that scale without completely insane amounts of compute is normally measured in weeks or even months, not a few days. especially once you factor in potentially needing to train multiple times to determine the best hyperparameters to use which can vary between datasets.

[unkempt cape]

yeah just wanna know i thought its been months of making. its ok they should take their time

[supple raptor]

They didn't have any acess to enough compute until recently. Prepping the dataset and coding the trainer that they'll be using is what they've been working on while they waited.

[unkempt cape]

damn where u got all of these info from

[supple raptor]

I'm very active in the Touhou Project AI server, which is the WD team's server.

[nova remnant]

[timber meadow]

what do you use to make these?

[oak spoke]

broken cyborg arm

[pure oyster]

He's our AI art wiki man, he'll tell you about all the new stuff if you pay him in Marisa or Reimu pics.

[supple raptor]

[gloomy scroll]

Gib your daily wisdom

[native halo]

computer

[keen hollow]

evenin'

[nova remnant]

not your coffee machine ?

[keen hollow]

neat

and I realize I'm just seeing it like a week later 😛

[supple raptor]

ty, I was really happy with how that one came out, especially considering it was straight txt2img then "enhance" (NAI's hires fix). no editing in post or inpainting.

[native halo]

https://tenor.com/view/anime-girl-crashing-computer-baseball-gif-25838534

[green temple]

my friend`s

mine

[dire bridge]

Best ai machine right there

[nova remnant]

@brittle rampart

[nova remnant]

You appeared again

[brittle rampart]

Yes. This is true. life just makes its own adjustments. but I'm glad that I wasn't completely forgotten

[nova remnant]

[nova remnant]

This elf very coot

[brittle rampart]

yep. She turned out quite charming.

[nova remnant]

Do you need a VAE or are the colors what you want

Looks a bit washed

[green temple]

i`m actually using one , lol

but i`m stealing images rn 🌚

[green temple]

original

[nova remnant]

Cat thief

[green temple]

🤣🤣👍

[nova remnant]

stupid phone corrector

Wanted to type cat lmao

[green temple]

lol

[brittle rampart]

but it turned out funny! 🤣

[dire bridge]

car?

[brittle rampart]

btw. what a news in Forge?

what interesting goodies?

[native halo]

nothing,forge is abandoned right now

[brittle rampart]

it is sad.

Is it really possible that such a project will not continue to develop?

[native halo]

dont know if they will update it again 🤷

[steady grail]

as expected. they where working too fast and got burned out

[brittle rampart]

hi hi ! What if this is the calm before the wave of chic updates?

[native halo]

https://tenor.com/view/lol-game-dead-gif-24851197

[steady grail]

without leaving behind news?

[supple raptor]

Lmao 1 month without updates is not grounds to think an open source project is abandoned

[native halo]

dead game

[brittle rampart]

Wow! Well, I see that even what was 1 month ago includes quite a lot of changes. I just need to understand what and why from the entire list of built-in extensions that have been added

[bitter latch]

sleep

I go sleep now.. thud

[acoustic locust]

what model?

[zinc flax]

That's dalle

[acoustic locust]

🙏

[acoustic locust]

how to i change the huggingface cache on comfyui?

[acoustic locust]

what model?

[brittle rampart]

blazingDrive V1.3

[acoustic locust]

ok

thanks

[brittle rampart]

My pleasure

[nova remnant]

[unkempt cape]

can u train a lora with pics being at diffuirent sizes?

or u need to crop everything

[nova remnant]

Should be better to crop to a certain size

[unkempt cape]

wellshit rip me

[steady grail]

guys, i'm trying to weight merge 2 1.5 models but i require a clip model. someone know what model is and where i can find it?

[vestal crypt]

gave it a little fix

[haughty rapids]

Hi

Any one ues fooocus ai

Use*

[acoustic locust]

How can i change the transformers cache on comfyui?

[nova remnant]

There is a comfy specific channel #🧣｜comfy-ui

[acoustic locust]

Ok

[nova remnant]

[unkempt cape]

guys whats the best setting for sdxl style lora?

and it is possible with like 20 images

and do i need to crop everything if so is there a colab for it ):

talking sdxl (more specifically animagine xl) btw if thats any difference

[nova remnant]

[native halo]

where feet at

[dire bridge]

clean

water

[nova remnant]

[dire bridge]

Steven Seagal

[native halo]

clean my shoes 🥾

[nova remnant]

[junior quest]

a style lora with only 20 images...not sure

ask from nubby

nubby is like an expert of these things

[steady grail]

with only 20 images, it burns easily

i don't remember if you need more or less steps to not overtrain it

[foggy willow]

why is this a simphouse

[cerulean cosmos]

it depends on whos here

[native halo]

why is this your first comment

[unkempt cape]

?

[gloomy scroll]

Isn't it expected that people who got access to AI would simp for their waifu/husbando?

[acoustic locust]

Because everyone here is perpetually horny.

[acoustic locust]

my comfyui is buggin'

I'm not even trying to use stablecascade

[ornate echo]

Apparently you are.

You don't even know it.

[acoustic locust]

🎅 [Pufferfish]

How do I make it not use Cascade?

[ornate echo]

What is your workflow.

[acoustic locust]

It's old so it couldn't possibly be using cascade

like it predates cascade

[ornate echo]

Until we see your workflow, we couldn't know what is causing it.

[acoustic locust]

Idk if this is hard to read or not

[ornate echo]

What's the lora at the bottom above the bypassed node?

[acoustic locust]

It's just to make eyes look cool

[ornate echo]

Which node is giving the error?

[acoustic locust]

I can send the whole error

[ornate echo]

On the workflow, it should highlight the node it's stopping on.

[acoustic locust]

i fucking put a cascade node down

[ornate echo]

Uh...so yeah

[acoustic locust]

forever ago

[ornate echo]

lol

[ornate echo]

See. You are and you didn't know it.

😉

[acoustic locust]

thanks

Honestly, I prefer comfyui.

it generates images differently

idk why

[ornate echo]

It doesn't have to. If you the right nodes, you can mimic A1111

Comes from the conditioning.

[acoustic locust]

Well the point is I don't want to mimic A1111

[acoustic locust]

?

[native halo]

well a1111 is a lagfest for XL and forge has some bugs too,comfy feels faster but u have to touch all those noodles

[acoustic locust]

A few months back I stopped using it because it was so fucking slow

I forgot to install xformers

lmao

[ornate echo]

No, I get it...but you can actually condition the prompts in many ways, including how A1111 does it.

[acoustic locust]

Neat

[ornate echo]

Check out the node CLIP Text Encode++ from smZ. You can change the parser to comfy, comfy++, A1111, full, compel, & fixed attention.

[acoustic locust]

Neat

[ornate echo]

https://github.com/shiimizu/ComfyUI_smZNodes

Depending upon the choice, some changes are subtle, some are not.

[buoyant ledge]

hello, I allow myself to ask u if u used Openpose to change the position of her arm ?
I ask cause I watched a lot of tuto, and I do the same as on tuto but I can get the position I want, If you have time to talk a little I would be so glad

[native halo]

no thats just pure text2image no controlnet used

[vital raptor]

Lagfest?

[native halo]

yea lagfest

[buoyant ledge]

ok ty

[vital raptor]

i haven't noticed any lag, must be a potato thing 😏

[native halo]

doesnt matter if u throw an a100 to it,u will just be throwing overpowered hardware to an unoptimized program

[vital raptor]

well, the UI doesn't run on the GPU, so it doesn't matter

just saying I've never had any lag, in any SD UI

but I will agree, a1111 is pretty badly made

[native halo]

yea and its easier to use than comfy so nothin is perfect😔

[vital raptor]

it spikes to like 12% cpu usage when i generate an image, 30% when adetailer runs

'idling' at 9% with a youtube video playing

gpu hits 80% usage (18gb vram used) using tensorRT

[nova remnant]

And last one apparently

[native halo]

https://tenor.com/view/black-guy-orange-shirt-invisible-chair-driving-gif-19154272

[unkempt cape]

so im training a lora
10th epoch
exact prompt and the image looks somewhat (like 60 70%) in style that was trained
more simple prompts and its weird

is this under or over train or dataset just suck

[supple raptor]

potentially sounds like overfitting. could be due to a lack of variation in the images, too high of learning rates, training for too long, or some combination of those.

[green temple]

Wooooow !!!

[unkempt cape]

whats the best lora style or sd style option if i dont have a lot of images

say 20->40 at best

[oak spoke]

overworked underpaid maid

[supple raptor]

Saitama on the way to his wedding

[acoustic locust]

My sd messed up, but this is cool.

[supple raptor]

looks like my winter flan prompt but sd forgor to load the non-broken VAE.

[acoustic locust]

Lmao

Nah, I just accidentally put the negitive for the positive.

[acoustic locust]

What model?

[green temple]

same model chromaneoft

i`ve send it before

[green temple]

🥰

[hidden lotus]

Ooh that ice is..nice

[supple raptor]

here's another cirno that I was really happy with

[hidden lotus]

Oh wow!

I need to do more anime stuff, it's been a while since I last did a lot though

My laptop decided it wanted to BSOD one day while genning, and I stopped for like... 5 months

[supple raptor]

I've been using NAI for most of my gens lately. Their model is really nice.

[steady grail]

this is interesting, for fixing hands https://x.com/catsuka/status/1778754956633219289

Catsuka (@catsuka) on X

How to animate hands (by Chinese artist Ju Neng Binggui @2kgjnbg).
https://t.co/gau1vadmyf

Twitter · Apr 12, 2024, 12:00 PM

[zinc flax]

Ooo Ein

Super clean

[supple raptor]

tbh the faster method is to set your phone up to take a selfie and hold your hand in the same position you want and use that as your reference image. the IRL image is also generally better for openpose since that is trained on irl images and not anime.

[vital raptor]

the few times i've needed it, i use this https://github.com/wywywywy/sd-webui-depth-lib with depth controlnet

basically premade depth maps of hands in different poses, works fairly well

if using with comfy or something, you can just download the maps from that repo and use them

[wispy canopy]

(just screwing around in pixart sigma, trying scheduler settings)

[acoustic locust]

What model?

[supple raptor]

NAI for initial gen then Kohaku_xl_delta for hires fix+face detailer.

[acoustic locust]

ok

[fresh flame]

Make flat style posters in black and white as well as color

[unkempt cape]

fuck how do i prompt

like whenever i use prompt with animagine xl 3.0 it breaks

horrible images

[native halo]

draw them yourself

[ornate echo]

That's what we're all doing.

Countless hours of drawing what people type into their computers and phones.

[steady grail]

sometimes can be tedious, but it's important.

ok, let's start prompting again, that guy prompted a big boobed police girl

[unkempt cape]

so like i took a working prompt from like sd 1.5 on seaart

pasted it in and boom dogshit

[steady grail]

eheheh. yeah, xl works totally differently

[ornate echo]

A prompt alone does not an image make. What are the rest of the parameters?

[unkempt cape]

any particular differences?

[unkempt cape]

oops sorry

didnt mean to

anyways sizes 768x768

cfg 8

steps 28

[ornate echo]

XL works best at 1 megapixel...aka 1024x1024 and its variations

[unkempt cape]

did that as well still dog poo

[ornate echo]

Also, would help if you pasted things like a sample of your output and your settings screen

[steady grail]

some tokens needed in 1.5 adds too much attention to come concepts, generating deformed results

you gen for detailed intricated eyes and instead yo get derpy eyes

when 1.5 need 150 tokens to just do a thing well, on xl, you just need to write a situation and usually, you get good results

or maybe my mind just works more like xl than 1.5, when prompting

[vital raptor]

it also depends on the model

[acoustic locust]

I really want to learn how to draw

i think I'm going to buy a drawing tablet

[steady grail]

well yeah, there are bad models out there

[vital raptor]

I don't just mean good or bad models though, it's about how they were trained

a lot, if not all, 1.5 models are based at some point on models trained with booru tags, like 1girl, sitting, short hair but some, mostly XL at this point, use normal language like a girl with short hair sitting

but then, XL models like the pony based ones can do both, but booru tags seem to be more effective (usually)

on top of that, pony based models (and possibly others) also need special tags and/or artist names to get specific styles instead of random generic messes

[steady grail]

let's take pony for example.
on pony, cute face or manly faces, are enough to get good faces close to what you are looking for

usually, on 1.5 to not get generic, or messy faces, you need to prompt detailed face, perfect face, detailed eyes ecc.

[vital raptor]

in the end, the way to get images you like, is to find a model that can do what you want, and then learn how to write the prompts for that specific model, along with the sampler/steps/cfg/etc to get the results you want from it

there's no magic model/prompt that just instantly makes perfect images

[steady grail]

yeah, but to me, usually, xl, require less token to gen what i'm looking for

[vital raptor]

for sure, XL needs a lot less, i don't even use negatives with it

[steady grail]

and a too long prompt, usually is even counter productive

that's another weird thing, useless negative prompt, on xl, can make things weirder

[vital raptor]

i remember reading somewhere something about how the negative prompt works and why it's basically useless on XL but I don't remember the details

[steady grail]

for example, when i prompted crooked eyes in negative, influenced negativelly the eyes of the gen

so, usually i just use it when necessary

[vital raptor]

my most common use is blue eyes because sometimes it ignores the eye color in the prompt if there's something else blue in the image, like a blue shirt

but yeah, negatives on SDXL pretty much only if i need to try and remove/fix something

[steady grail]

yeah, tokens leaks in cascade

[acoustic locust]

Idk, looks great for me

[junior quest]

model?

just pony?

[acoustic locust]

Animagine xl 3.0

[native halo]

now make her pregante

[acoustic locust]

uh oh

[unkempt ravine]

can you guys help me make chibi version on this image please

[vital raptor]

I made her pregnant but can't post those images, so i posted after she was pregnant 😏

[unkempt cape]

almost same prompt for character. i dont really like it but just tryna have sum fun

[vital raptor]

those look pretty good

[junior quest]

what is dis? straight up NAI?

or an open source model?

[nova remnant]

I think it's animagine

[unkempt cape]

animagine

ran out of juice for nai bruh

[nova remnant]

Lmao

[unkempt cape]

4000 anlas thats about as many as i can ever have entire life

[nova remnant]

I'm doing history research

I'd like to discuss privately about the events that happened 9 months ago

[unkempt cape]

artist tags lead to peak

[junior quest]

animagine? with a lora?

[green temple]

😂

[vital raptor]

Your images remind me of posts I see on twitter sometimes

[green temple]

👍

hmmm.... nsfw ?

[vital raptor]

I think they're all sfw

[green temple]

good

[acoustic locust]

What model?

[green temple]

same model

chromaNeoFT

[vital raptor]

username on twitter is @yuwel_AI if you're curious, looks like they changed models/styles recently but a lot of their older stuff looks very similar

[green temple]

indeed, very similar but better quality and better ideas

[nova remnant]

but I garantee none of them ride a bike

👍

[green temple]

[acoustic locust]

I've asked three times and it was the same model

i look like an idiot

[green temple]

5 times...

[acoustic locust]

[green temple]

[dire bridge]

cooked

[bitter latch]

ook not all that good

[hidden lotus]

I can't seem to find this model on civitai?

[nova remnant]

Try hugging face

[hidden lotus]

ahh right huggingface

found it! ty

[green temple]

Its not on civitai

[green temple]

Yep

[hidden lotus]

yeah i kinda forgot huggingface was a thing lmao

[green temple]

lol

[supple raptor]

I got the Huion Kamvas Pro 13 2.5K and have been very happy with it for the price.

Works nicely as a small 3rd monitor when I'm not using it for drawing as well.

[steady grail]

ohy nubby, never merged with comfy? i'm trying to weight merging with it, but i only get errors 😦

[supple raptor]

Yeah, it's been a little while but I remember it being pretty straight forward when i did it. Let me load up comfy and take a look.

should be as simple as using 2 load checkpoint nodes and connecting them to your choice of merge node then connect the output of that to your normal workflow if you want to do runtime testing. Once you're satisfied, connect a checkpoint save node to save it.

[steady grail]

the normal merging is easy, but i'm looking for setting the songle blocks while merging

looks something funny to do

there is this node called auto merge block weighted that looked promising

[acoustic locust]

How much was it

?

[supple raptor]

I found mine on ebay for $250, they usually go for a little over $300 USD.

[nova remnant]

drawing tablet

[unkempt cape]

ye

[pure oyster]

anyone know a realistic 1.5 model that doesn't have such a strong asian bias?

nvm, I think I found an embedding that might work

[green temple]

Absloutereality maybe

hands 😭😭

[nova remnant]

its adorable tho 😭

[green temple]

Regardless of the mistakes, which is better

[nova remnant]

second one

[nova remnant]

have you tried a lora for fixing the hands

[green temple]

no i only use negative

for hands

[nova remnant]

that works too

[green temple]

👍

[green temple]

coool !

[nova remnant]

Booba

👍

[nova remnant]

[nova remnant]

First one has cooler ground I like that part

oh lord she has the UGLIEST eyes

[nova remnant]

thats the exact face i made when i realized i cursed this poor image 😔

[nova remnant]

runs away

[nova remnant]

howd you get the dark makeup 😭

This one pretty

i need good makeup loras/loras for bags under eyes

[nova remnant]

yesssss

i love her

It's just "black theme, goth, black eyes"

And eyeliner

[nova remnant]

Happy

[nova remnant]

:D

i love that eye shape

this is a wild fail

Why wild fail

Does look wild

i take that back this one is worse

💀

[unkempt cape]

anyone has a working colab for xp dreambooth
nothing works so far on google

[nova remnant]

Did you put "glitch monster from hell invading my PC" on your prompt

[nova remnant]

uhhhh

perhaps

you see i had a fun idea

where i would take all of my style loras and put them in a txt file

then i was like, what if instead i put all my style loras in a text file TWICE, so i can have two different weights

then i threw that in wild cards

and prompted it twice so it would mix styles

[unkempt cape]

peak sd

[nova remnant]

agreed

[bitter latch]

well that came out better than I thought

[nova remnant]

yooooooooooooooooooo

thats hawt

[nova remnant]

hand

[nova remnant]

😭 i cant get it to look good

[native halo]

that actually looks good,those eyes are not easy to get unless u prompt for an artist or use a lora

[nova remnant]

i have a whole thing for eyes so i can get specific shapes im using it for this

but i need the makeupppppp

😭

i really love droopy eyes tho

[bitter latch]

ok it got confused

[nova remnant]

hgfdshsdc'

i love it when ai makes eyes look different directions, its so funny

Lmao

[bitter latch]

they went < .... > way

[nova remnant]

"get out of the way" position

[nova remnant]

yessssss

[nova remnant]

its so pretty 😭

[native halo]

(runny makeup,droopy eyes,open mouth,crying:2.5)

[nova remnant]

but why is she crying

[native halo]

[nova remnant]

[native halo]

https://tenor.com/view/better-call-saul-meme-gif-468341328098124204

[nova remnant]

https://tenor.com/view/bornskywalker-dap-me-up-woody-woody-handshake-woody-toy-story-gif-26021440

@nova remnant pretty pleeease

[nova remnant]

😭 im still trying

[nova remnant]

i can give you what i use for droopy eyes

the question is

[native halo]

but mouth is closed 🗿

[nova remnant]

what exact type of droopy are you wanting 👁️

[native halo]

these eyes look cool

[nova remnant]

thats insane but i can make that happen

this is ass 😭

https://tenor.com/view/rock-one-eyebrow-raised-rock-staring-the-rock-gif-22113367

[bitter latch]

doh no

[nova remnant]

this should be it

@native halo she came to take you to the dark side 👍

not what i was wanting but this is cute

[nova remnant]

its like an mha character

[nova remnant]

The momo girl

yeaaahhh

Ponytail girl

the creation girl

I have no idea, genned her once lmao

Had eyes like that

o:

Wait

the eyes are cute

[native halo]

u should upload it to civitai too for more ppl

[nova remnant]

it is NOT putting make up on her

[nova remnant]

its not mine thatd be rude

imma gatekeep tho

[native halo]

[nova remnant]

[nova remnant]

let me see the goofy ahh image you make with it

@nova remnant you might like this one

jafhsfcs

stealing

thats so pretty 😭

getting places

With Less steps

[nova remnant]

went backwards

[bitter latch]

keeps giving me same face hmmm...

oh the lora, DOH

my brain tired

[nova remnant]

LMAO

[nova remnant]

Cut bottom of the image for safety reasons

👍

[nova remnant]

[bitter latch]

ok satisfied with this , NEXT

[nova remnant]

[nova remnant]

[nova remnant]

we making hella goth girls