🍥｜anime

[nova remnant]

but im getting that previous error

any suggestion where to edit

this is what is asking me to do
The name tf.losses.sparse_softmax_cross_entropy is deprecated. Please use tf.compat.v1.losses.sparse_softmax_cross_entropy instead.

[unreal ridge]

C:\StableDiffusion\stable-diffusion-webui\venv\lib\site-packages\keras\src\losses.py

line 2976...?

maybe

no idea

[nova remnant]

but i cant find the line tf.losses.sparse_softmax_cross_entropy

oh found it losses.sparse_softmax_cross_entropy

[unreal ridge]

👍

[silk narwhal]

errors like that i always go to studioLM or chatgpt 😄

[nova remnant]

but this is already what i have on line 2976...

    tf.compat.v1.losses.sparse_softmax_cross_entropy: "int32",
    sparse_categorical_crossentropy: "int32",
}```

that could mean the warning message is just generated w/o checking the actual content line

wonder if others are getting it too

[native halo]

if u installed tensorrt then thats where problems come from

[nova remnant]

i did but removed it too several weeks ago

[native halo]

oh then its probably related to webui 1.7.0 cause i saw someone else with same problem on reddit

[silk narwhal]

did you fully reinstall python etc .. as TensorRT messes up that alot

[nova remnant]

i did git pull few mins ago and didnt see this problem yesterday

[nova remnant]

nup

👀

i thought i had 1.6.0

so anyone with v1.7.0 are seeing this warning message?

lol i guess i should generate some beautiful girls and see if that message is affecting my webui

oh wow, there are some changes to UI

the hires fix now has tick mark box

excellent !!

did you try generating anything, do you like the speed?

[native halo]

thats cool now use loras

[nova remnant]

lol

yeah can only use 1 lora at a time last i remembered

[native halo]

also change checkpoint or use high res fix with a custom resolution

[nova remnant]

im dying lol

thos are some limitaitons to using tensor rt

everytime we change the settings we have to create a new tensor file which also takes up huge space

it should be under models folder, i dont remember the exact path

so basically at the cost of speed with one custom settings, when you change the checkpoint, the resolutions of the image, you have to create a new tensor file for each of those new combination of settings, and each of those would take up to 2gb+ of your storage

[native halo]

its like 2gb file per checkpoint and a 600mb file per lora,also limited to one lora only

[nova remnant]

thats the reason i moved away from using tensor rt... can't waste my ssd storage with duplication of the same thing

but how would you use tensor rt then if you delete one file

you'd have to go through the process of makeing new tensor files

[native halo]

i heard that the more gb's u write into your little ssd the more it lasts so its better to delete them every time u use them

https://tenor.com/view/neco-arc-saul-goodman3d-neco-saul-gif-26298783

[nova remnant]

actually ssd life is eaten away if you delete a lot lol

@nova remnant take note of steven's humor ... if you want your ssd to last

[native halo]

since nvidia knows how to speed up sd with tensorrt id asume they will add this into their driver in the next months

[nova remnant]

any news out on that?

[native halo]

no,but maybe soon ™️ so id rather wait than wasting storage and time

[nova remnant]

well that sounds very promising if they will actually do it

im not going back to installing tensor rt tho

would much prefer a driver that integrates it

[silk narwhal]

TRT is already in their AI Enterprise cards ..., the $10,000 ones

[nova remnant]

tough

[native halo]

there u go 🤠

just wait for 5090 with trt enabled

[nova remnant]

well thats one way of popping dream

[silk narwhal]

Just wait for GPU's that require an external case just to cool it

[nova remnant]

we are near the end guys.. this is it

[native halo]

leaked 5090 img

[nova remnant]

lol

you can cool your room with that i suppose

[silk narwhal]

room heater 😄

[native halo]

still cooler than my gtx 480

[nova remnant]

how cool is that lol... maybe switch it to heat up in winter and cool in summer

but i dont think 5k series cards are generally accessible

[native halo]

gtx 480 my beloved

[nova remnant]

sigh... nostalgia

[nova remnant]

they brought back style to prompt button

[untold glacier]

You need to be using a VAE.

Also, Merry Christmas to everyone.

Gotcha. I've had to explain that to so many people, heh.

I lose track of who knows sometimes

[stable crane]

merry christmas 🙂

[silk narwhal]

merry xmass

[nova remnant]

not using any textual inversions or lora

those alter a model style a lot

i'd rather download a model that's well built instead of trying one and changing its intended style with addons

🥳

also love the idea of prompting with words

[nova remnant]

sweet pose and sweet face, color tone is also sweet

[nova remnant]

boba

[nova remnant]

btw i dont think girls enjoy big back pains

it hurts to see them hurt

[unreal ridge]

can't say the same for guys...

[nova remnant]

guys dont have big back pain stuff

[unreal ridge]

I meant liking girls with back pains... as having big booba... english is hard for me

[nova remnant]

oh

yeah some guys are strange

[native halo]

the ones that appear on playboy dont seem to mind

[nova remnant]

i dont like to see girls hurt

i like girls sweet overall

feeling sweet and looking sweet

[native halo]

https://tenor.com/view/my-honest-reaction-reaction-meme-iskl-iskltv-gif-18420930921040254182

[nova remnant]

haha

i like men who like and understand women

[unreal ridge]

haha
i like men women

[native halo]

i like cardi b and nicki minaj

[nova remnant]

lol im not hitting on men, im just saying i like their quality who understand the sweet nature of women

[unreal ridge]

just realized that parted bangs exists...

foreheads 🙏

[nova remnant]

yeah i learned a lot of neat things from booru tags 🙂

[silk narwhal]

A staring contest

[nova remnant]

eye gazing is a thing and spiritually comforting

[native halo]

i recognize that pose and that lora

[nova remnant]

@native halo me when no bodysuits to look at

[native halo]

me when i get arrested

[nova remnant]

what do you do to get arrested? 👀

[native halo]

go to the cell and sleep on the concrete bed

[nova remnant]

i mean..what do you do in the first place to even get arrested lol

[nova remnant]

Steven Seagal gaming things

[native halo]

tax evasion,i am a convicted tax evader

https://tenor.com/view/mike-ohearn-gif-5978288995841027384

[nova remnant]

that sounds largely good thing and a model for society

[unreal ridge]

I can't post 50% of my gens

[nova remnant]

You can post 100% of your gens in my DMS #ad

dont wanna ramble on it, but if you look close, money is a debt construct and its valuation comes from scarcity, and is responsible for most crimes

[native halo]

i been suffering from money scarcity lately tbh

[nova remnant]

yep everyone is

[unreal ridge]

https://tenor.com/view/ah-shit-here-we-go-again-gta-sa-cj-gta-again-gif-20425950

[nova remnant]

it creates artificial scarcity

scarcity that are not supposed to be there

can you tell any good reason why a price for commodity a century ago is 10 times now?

or more..

demands?

and those demands are dealt with not giving? all while creating waste?

[unreal ridge]

sent it 😔

[nova remnant]

im hopeful 2024 is gonna be wild

[native halo]

cant believe they added miku to oblivion,truly one of the games out there

[stable crane]

lol

[nova remnant]

Moar angy girl

some of the old anime models i found on civitai are pretty nice that people often dont use

[nova remnant]

thise anime model is cool how it renders facial tone

not overly cute

[nova remnant]

It's.. pikasnewgen

Yet again

[unreal ridge]

oh

I think I used this one before

it does faces pretty nicely

[nova remnant]

i looked at the model page but the cover images dont look too good

It's basically the only one I've been using I don't have that many options available

yet the imags you post look pretty neat

[stable crane]

try some of the older ones too

[unreal ridge]

I remember one that was pretty cool

but the hands tho...

[stable crane]

under the mistletoe

[unreal ridge]

let me see if I can find

[nova remnant]

good point, i have to check

[stable crane]

I loved pikasnewgen but haven't tried the new versions

[native halo]

pingas newgen one of the top models out there

[nova remnant]

@nova remnant is this the one https://civitai.com/models/47067/pikas-new-generation

[stable crane]

Pika used to chat in here

[nova remnant]

oh really?

what happened?

[stable crane]

idk, may still be here 🙂 but good luck finding which one to @

[nova remnant]

lol

[stable crane]

it made really nice cinematic gens at the time, I think they still look quite good

perhaps they trained on an early niji?

[nova remnant]

yeah i can see from cat's posts

[stable crane]

oh so says the description too

I'm going to try v2

[unreal ridge]

it's
@ pikadesigner

[nova remnant]

not here

i dont see that nick

[unreal ridge]

oh

[stable crane]

I don't think they had that nick here

[nova remnant]

my phone stupid full memory warning won't let me download images in discord

[stable crane]

lol

[nova remnant]

use a 500gb card

My card is 16

It's like

oh

10 years old

could fit in 64gb ?

I dunno

One day I'll be less lazy and check if I can

btw im grabbing the pikas v2

I like 10-11 CFG on it

I'm using basically default SD

wow 10cfg ?

why so high?

i usually use cfg 8 steps for both sd and sdxl

[unreal ridge]

I can't go past 8

[nova remnant]

can't? or dont?

[unreal ridge]

yes

[nova remnant]

you know how it goes, higher the cfg that more it adheres to prompt and lower allows ai to go way more creative, and i tried with 50 cfg steps but practially i think im fine with 8 steps

[silk narwhal]

between 7 and 13 is a good sweetspot yea

[stable crane]

I kinda wish I could randomize it between 5~11

maybe I can in comfyui

[nova remnant]

interesting but i could do that with a simple click on a1111 🙂

[silk narwhal]

pretty much what my wf does ...

[nova remnant]

dont think a1111 can automate to randomize cfg

[stable crane]

randomize the CFG?

[nova remnant]

that

[stable crane]

doesn't this conflict with your earlier statement?

[nova remnant]

you cant in a1111

[nova remnant]

no.. cause i thought you said you could do it in comfy

[stable crane]

interesting but i could do that with a simple click on a1111
and
dont think a1111 can automate to randomize cfg
sound like "I can do X" then "I can't do X"

[nova remnant]

oh yah i can click and change the value manually before rendering 🙂

[stable crane]

oh same lol

it can be annoying to scroll to find the node though

A1111 is more info dense in a way

[nova remnant]

lol few simple clicks, but usually when i have a good prompt i click on generate forever and if i could randomize cfg that would be interesting

i dont use batch at all

just to keep my folders organized from filling up

[silk narwhal]

convert cfg to input and use a primitive number node set to be between 7 and 13 set to randomize

[stable crane]

now my organization is, I generate folders for each day

and it's better than months of pics in 1 folder

[nova remnant]

all of the images i generate are single files and that i set to generate forever unless i cancel and all of the images go into one folder for easy access / deletion

[stable crane]

comfy is good in another way because you can just spam click on "Generate" to enqueue a ton

[nova remnant]

oh wow

spamming to do that sounds strange

[stable crane]

I guess I feel nervous about "generate forever"

like it to stop at some point

[nova remnant]

i do that to get some good ones i really wanna keep

pikas have nice facial structure

[silk narwhal]

😄 or you do it like me .. and gen 6 images in a single pass with 1 batch

[stable crane]

I'm liking pikav2 as well

[nova remnant]

a girl in a garden picking flowers wearing casual short outfit under sunny sky and puffy clouds.

yeah its good

using v2 as well

[nova remnant]

looks bright and perky, very nice

i play with two different set of templates, one is a combination of textual inversions and lora and the other entirely prompt oriented categorized for positive and negative sections

i kinda get a bit flatter look with the prompt oriented combo

My prompts are like 1 line each

You're doing math science

i will show the images .. a sec

[nova remnant]

This one is very goog

the left one is with textual inversion and lora

right one is entirely word prompts

Is left the one holding purple flower

yes

same render seed tho but textual inversions can alter an image

[nova remnant]

@nova remnant

I need to check that tomorrow

ok

its basically the template i created with common and generic prompts

switched underexposed, overexposed, oversaturated, low saturation from unrealistic to default

color tone looks better that way

@nova remnant updated ...

[nova remnant]

1girl, flowy hair, blue eyes, swimwear, wrapped in towel, seaside, beach, waves, shore, sunset.

im gonna do a plotxyz comparison of 4 anime models with same prompt

[nova remnant]

removing several of the merged models that have closely similar outputs

lol multi tasking and typing

[nova remnant]

@native halo

1girl, blonde hair, wearing a jacket and a miniskirt, crossed arms, solo, outdoor sitting lounge, sunny sky, puffy clouds

[nova remnant]

silhouette of a woman within the shapes of white fire. foggy background. dramatic visual effect.

[bitter latch]

ok.. then

[nova remnant]

hey @bitter latch are you using a1111?

[bitter latch]

yep

[nova remnant]

there's been an update

its version 1.7.0 now

and has some UI changes

[bitter latch]

hmm ok

[brittle rampart]

and work faster and best?!

[nova remnant]

in case you get any error messages with any of the extension try removing the venv folder and run the webui again to update

also its probably a good idea to delete venv folder with big updates

[nova remnant]

i dont notice any performance difference tbh

[native halo]

nope

[bitter latch]

mine says already upto date

[brittle rampart]

then.... then we'll wait.

[native halo]

im still on 1.6.0 wont update unless my extensions break

[nova remnant]

i posted this earlier for steven, do you have the same ui?

[brittle rampart]

im still in 1.6.1

[nova remnant]

you should update

i'll find out what are the changes soon as i can focus on reading long reviews

im distracted with music atm

music and girls

[nova remnant]

but check the interface changes

[bitter latch]

looks the same to me

as before no changes

[nova remnant]

ahh ok thats nice, i had to refresh the venv folder before i could see the UI changes

[rich tartan]

finally

[nova remnant]

a female ranger in fur and leather outfit, roaming in the jungle through vines, rocks, and puddle as the sun is setting. ((dynamic pose))

[brittle rampart]

today music day...

[rich tartan]

yea

trained my v3

some tests im making to tune a model

[brittle rampart]

guys.. What program i am can easily and quickly make video from images?

[rich tartan]

may try out https://mconverter.eu/convert/png/mp4/

[brittle rampart]

online?!

cool!

[nova remnant]

a young woman, barefoot, wearing short dress, sitting on a log, under a tree, feeding birds.

[rich tartan]

feels like it gave the opposite of what i wanted to reach.. More mutations

But at least the model makes better body view, instead of literally face only

[rich tartan]

alr, got it improved by setting lora strength down.

Is it possible to check with controlnet how big the character is? (determine somewhat the distance)

[steady grail]

i load the generated image on paint

it have a pixel counter

and than math

[rich tartan]

can you give more information about that?

[steady grail]

this the program paint of windows

visualization tab, the 3° on the top, check the 3 marks

[rich tartan]

oh..

[steady grail]

the pixels on the bottom are the one the mouse pinter is on

[rich tartan]

yeah.. but some way to do it automated?

[steady grail]

not that i'm aware of

i would love to know, so that i could use it in regional prompting

[rich tartan]

i want to use this for double sampling

i think this explains what i want to reach with auto-detecting distance

to get better anatomy.. When the character is near/big, then high denoise breaks it, but low/mid one fixes it, if the character is far away, higher denoise fixes the heun issues/mutations much better

[steady grail]

the long arm?

ellongated anatomies are pretty comon in this ellongated resolution

[rich tartan]

finger count and (i think heel?)

the joint or sth

[steady grail]

the left end is crooked too

[rich tartan]

yea

[steady grail]

you are looking for an automatic mask, and something to measure the mask and fix mutations

[rich tartan]

and i think here the high denoise(left) looks better

[rich tartan]

yea

instead of generating 3 images and having to wait one min to finish them all

also saving, not having much storage

[steady grail]

cocosemseg can already automask an image

but it's still very hard to fix anathomy automatically

[rich tartan]

that's interesting, yeah

[steady grail]

unless you use multiple adetailers

and the models are trained very very well

[rich tartan]

ah right, i don't use any adetailers, might use one now

[steady grail]

there are some hand adetailers, but they are still not perfect

[rich tartan]

i guess better as nothing

[rich tartan]

i made a lora about some "3D" anime characters with the aim to improve anatomy, i dunno if it really worked very well, as the legs are still somehow mutated. But this model doesn't really do much with legs ig

But fingers did improve compared to my older gens

I'll inform back when i researched about adetailers and cocosemseg, thank you for informing me tho

[steady grail]

np

[steady grail]

just found a strange token called bezier. o wpmder what it does

[rich tartan]

token?

[steady grail]

prompt token

[rich tartan]

oh alr

[steady grail]

i'm hunting for new interesting tokens in images, or scanning models

just focusing on the heavier onece

[rich tartan]

sh** i never thought about making my prompts better

[steady grail]

ehehehe

i do it when i'm too tired to do anything else

i load some music and read, search, hunt

[rich tartan]

me similair, im just way more focused on the samplers..

nearby doing some lora research

[vital raptor]

i practice by trying to replicate images people post without looking at their prompts, it helps get used to what a model can do and not do with certain tokens

[rich tartan]

[steady grail]

yep, me too

what i did with the nun, was something similar

[rich tartan]

interesting idea, i mostly use prompts from others

[steady grail]

loaded on a tagger to get the most common token and than experimented

[vital raptor]

every model has it's own weights for each word/phrase/concept, so some will replicate a certain thing better than others

copying prompts is a good place to start, especially if you're using the same model, but otherwise it's a lot of experimenting to figure out how a model really behaves

[rich tartan]

yeah.. i realized that with lora training hardly

[steady grail]

that's why there are tokenizers out there

[vital raptor]

yeah depending on what you're training, prompts and keywords are even more important

[steady grail]

or embeddings

[rich tartan]

im more of the lora guy lmao.. (Also using one shadow and one light Vae)

[steady grail]

even if i'm not too much in embeddings, beucase usually, they ends fighting with the model

[rich tartan]

But i really forgot to work with prompts first

[steady grail]

i just use some style and color correctors at low weight to fix stuff

anything else, makes the result too inflessible or simply mutated

[vital raptor]

i try to do everything with prompts, it doesn't always work, but I don't want to rely on anything else changing my image

if I can't get what I want from the model with just a prompt, then I'm not using the right model, or the model is bad, so I move on

I also do custom merges to try and get the right combo

[rich tartan]

right one the normal vae, middle one the shadown one and the right one light and shadow combines

one of my old gens without heun, that's also the cause of no background

but nearly perfect anatomy

[stable crane]

any suggestions for comfyui wildcards?

[steady grail]

loading many vae must be very time expensive

do you mean a custom node?

[stable crane]

yeah

[rich tartan]

just two(i don't know if the one form the model gets loaded, even when not used), but here is light(right) shadow(left)
The right one i shown up there was the models vae

[stable crane]

https://civitai.com/models/241668/potatocreamxl?modelVersionId=272670 looks pretty promising for sdxl

[steady grail]

on the manager there are many wild card nodes

[rich tartan]

i realized, when the character is crouching or something, it'll think the character is further away

i may need to get the mag/dist between the borders

but how do i do calculations/if statemens in comfyui?

[vital raptor]

what aspect ratio are you using?

[rich tartan]

idk, 440x728

if it's standing, it's perfect..

[stable crane]

oh there's a manager?

I was preferring to just clone them manually

[steady grail]

yep

[zinc flax]

Ooh I love this kind of style

[vital raptor]

any particular reason for that resolution? the models are generally trained with 512, 768 and 1024, so i found you tend to get better results with something like 512x768

[rich tartan]

[stable crane]

https://huggingface.co/bluepen5805/blue_pencil-XL was recently updated

[rich tartan]

yeah, i'll try that

[vital raptor]

having a proper aspect ratio will help with anatomy problems too

[zinc flax]

I love it when it can do simple flat style

[stable crane]

ikr

[steady grail]

the manager updates the nodes with a click, together with the dependencies

[stable crane]

will it cope with my manually installed nodes?

one of them I altered the code of lol

[steady grail]

wow, scary but cool

[stable crane]

2nd gen today I've seen of a cool scythe

when models are doing scythes, you know the tech is evolving

[rich tartan]

This def generates much different

[steady grail]

for safety, change the name of the directory with the altered node

[stable crane]

I see

[rich tartan]

i'll gen 5 or sum and check how the anatomy is, if it's better i may work with that. But the backgound is way to less now and PVC character-holder comes up(those stand/support boards for the PVC characters)

completely different style, interesting(left one in correct aspect ratio)

[silk narwhal]

PVC

[zinc flax]

Gonna try it later, hope it doesn't disappoint lol

[glossy eagle]

Im gonna repost my model again and Ill send a settings guide/post with it in Model-embeddings

[stable crane]

looking forward to seeing it

[rich tartan]

🤔 Now first i need a way to extract the amount of white or black pixels to ints

[silk narwhal]

Guess the anime without png info 😛 (not sure if it show anything concrete though)

[vital raptor]

monogatari scene?

[silk narwhal]

it generated the wrong eyecolors :/

[stable crane]

hm am I supposed to be using ClipTextEncodeSDXl node?

things have been working without it specifically

[vital raptor]

hmm, i should play with some 3d mesh AI and see if I can generate an anime girl to 3d print 😄

[silk narwhal]

tried it and it is ugly AF, pointcloud estimation from only realistic data

[rich tartan]

alright.. seems like time to code some nodes ig.. That'll gonna be much work for minimal result, but ig that's AI

can't find anything that could work with the idea

[silk narwhal]

O_O SDXL models take so much RAM :

[vital raptor]

yes

[rich tartan]

your vram filled?

[silk narwhal]

VRAM 12% out of 12GB

[rich tartan]

use --highvram ig

[rich tartan]

are you watercooled?

[vital raptor]

one stray XL vae and it maxes it all out

[rich tartan]

Thats dam large

is it worth it?

[vital raptor]

if you have the money, sure, but not just for playing with AI

[silk narwhal]

just fans

[rich tartan]

sdxl literally only has some rtx lighting, but ig that's it

only used the turbo tho

[vital raptor]

I upgraded to a 4090 from a 3080TI and it wasn't really worth it for most things, but SDXL is way faster

like, 5 seconds instead of 45+ seconds for a normal 1024x1024 gen

[rich tartan]

you need 45+s for 1024x1024 with a 4090?

ig i misunderstood sth

[vital raptor]

2x hires took ~5 minutes on the 3080TI (12GB), ~14s on the 4090

[silk narwhal]

where do i add that with comfyUI?

[vital raptor]

45+s for the 3080TI

[rich tartan]

wait, youre talking about auto1111 right?

[vital raptor]

this was awhile ago though, drivers and stuff have improved a lot

[rich tartan]

python main.py --highvram

oh wait, youre on windows right?

[vital raptor]

it's the same command

[rich tartan]

oh alr, i though you start it with a .bat or something

[silk narwhal]

yea i start using run_nvidia_gpu.bat

[rich tartan]

you may try to open your cmd and cd pathto\comfyUI

then just python main.py --highvram

as far as i get, it'll start at normalvram

[silk narwhal]

alright lets see if this works better

just loading 1 SDXL model was using 31.7 GB of RAM

and the answer is it does not work:
torch.cuda.OutOfMemoryError: Allocation on device 0 would exceed allowed memory. (out of memory)
Currently allocated : 11.31 GiB
Requested : 3.12 MiB
Device limit : 12.00 GiB
Free (according to CUDA): 0 bytes
PyTorch limit (set by user-supplied memory fraction)
: 17179869184.00 GiB

friggin controlnet node

[steady grail]

some cn models weight a lot

[glossy eagle]

\

[steady grail]

ghost busters

[stable crane]

a bit of an odd gen

[rich tartan]

yep then better keep withouth highvram, seems like it uses all your vram

[glossy eagle]

ehehe

[silk narwhal]

was the controlnet node .. still havent figured out how to make that not use up 512 GB of VRAM 😄

[glossy eagle]

hol up ima use hiresfix on this rq with 100 steps

probably will look even more uncanny

[silk narwhal]

King boo's nephew

[glossy eagle]

L:L

Theres the upscale

Devious marshmello

[near elbow]

some non-winter memories

[glossy eagle]

that one marshmellow do be plotting the worlds end

Lets try marshmellow stranded on an island surrounded by the lava

[near elbow]

ok

am i fast or what 🙂

or how about the marshmellow league of supervillains

[near elbow]

let it roast!

[near elbow]

eeeeevil

[glossy eagle]

lool

[near elbow]

plotting!

[silk narwhal]

difference between SD1.5 and SDXL is massive lol:

[glossy eagle]

Im over here sitting on linux with a basically self distructive drive

[vital raptor]

lol, a league of villains

[glossy eagle]

I went onto build apc and ran crystal disk checker and after sending results they said they never seen a drive so broke, they said it is holding on to 1 strand of life

[near elbow]

let's fight them

[silk narwhal]

average VRChat meet up, 90% are dudes and male furries are the few girls

[glossy eagle]

I did post my model in #1047197565365538826 Now so If you wanna check it out and have questions just aske me

[near elbow]

looks like this with SDXL here, 5s per image at 30 steps DPM++2MK

[near elbow]

great prompt - "average VRChat meet up" 🙂

[near elbow]

no more mr. nice guy

eeeeeevil

[near elbow]

challenge accepted 🙂

[brittle rampart]

i am in 10%?!

[steady grail]

holy hell, the faces are almost perfect

maybe adetailer can fix most of them

at least on the frontal lines

[near elbow]

as are the pokemon (generated at 2048x1152 natively, duplication does not matter with crowds)

[near elbow]

haven't used adetailer at all, SDXL is good enough for me and the results look consistent - with adetailer there is often a "wow that face is nice but does not really fit" effect

ok if you look closer at this one, there are some strange mixtures 🙂

[steady grail]

denoise too high

[steady grail]

where is Woldo?

[ornate echo]

Now run Face Detailer on those.

[rich tartan]

what the f**

[vital raptor]

I don't use adetailer on realistic, or highly stylized faces, but it seems to work really well with some anime models

[brittle rampart]

i am reaally in 10%

[near elbow]

Negative prompt: photo, deformed, black and white, realism, disfigured, low contrast, loli,underage,teen,child
Steps: 30, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 3081387967, Size: 2048x1152, Model hash: dce7eb8449, Model: realisticFreedomSFW_alpha, Style Selector Enabled: True, Style Selector Randomize: False, Style Selector Style: Anime, Version: v1.5.2
Time taken: 12.5 sec.
A: 17.36 GB, R: 18.85 GB, Sys: 20.9/23.9883 GB (87.1%)```

[steady grail]

i use wildcard expression on adetailer XD

[glossy eagle]

Yes send prompts

Send industrial type like metal city prompts, or some cyberpunk type prompts,

[rich tartan]

uh wha

im also sitting here with linux

[near elbow]

"cyberpunk city, exquisite details, industrial, dark zone, futuristic neon" 2048x1152

[rich tartan]

sdxl?

[near elbow]

right (basically the generation info i posted above, just with a different core prompt)

"cyberpunk city, exquisite details, industrial, bright daylight, joy"

[near elbow]

cyberpunk kitsune 🙂

should be SFW enough, but i better stop now 🙂

[vital raptor]

ugh, it's painfully slow when i max out my vram (1m30s)

[near elbow]

beat the stop watch 🙂

[rich tartan]

So with sdxl the mutations still stayed the same?

But the quality overall improved tho

[vital raptor]

mutations will happen with any model of any type, it just depends on why

some models are worse than others

[near elbow]

it's much better but still happens, as the captain says... usually i delete results that are less perfect, but i wanted a quick image for this one 🙂

[vital raptor]

these wide renders like to add random legs in the middle of the scene 😄

[near elbow]

yeah, duplication is to be expected at much more than training resolution

e.g. "a kitsune showing her hand"

[vital raptor]

yeah, it's fun taking advantage of it though

[rich tartan]

oh alr, thanks for the information, was a little surprised, bc the mutation was somewhat the same as on the non sdxl models

[near elbow]

yeah, not perfect yet

[near elbow]

speaking of limb duplication... 🙂 "an army of mechas, apocalyptic, dark neon"

gundam instead of mechas, even worse - where is waldo indeed 🙂

[vital raptor]

lol

[near elbow]

oh.. "where is waldo" 🙂

[glossy eagle]

where isnt waldo

[steady grail]

here he is lol

[vital raptor]

fox girl jedi 🤨

[stable crane]

anyone have link to a good sdxl vae?

is it just https://huggingface.co/stabilityai/sdxl-vae/tree/main?

[warm oracle]

i think this turned out well

[gentle musk]

I think that's the dodgy one that puts artifacts all over everything.

let me check the checksum compared to the one I have

[gentle musk]

ok no idea now @stable crane All the checksums are different. Essentially the base 1.0 Vae has artifacts in it, the 0.9 vae doesn't.

[stable crane]

haha

ok ty for checking 🙂

[gentle musk]

There was a load of fuckery with it when 1.0 first released

[brittle rampart]

what is the prompt?

[warm oracle]

[gentle musk]

What's the better SD 1.5 Anime models at the moment? I've not used 1.5 in a while and I need an SD 1.5 Anime model to try something.

[brittle rampart]

greedy 😁

[warm oracle]

i was gonna send it but i'm greedy according to you so

[vital raptor]

this is the sdxl vae i use, it's fp16 so it uses less memory https://huggingface.co/madebyollin/sdxl-vae-fp16-fix

[brittle rampart]

let's make peace!

[warm oracle]

one second

i'll find the prompts

[brittle rampart]

I already drew a prompt

[warm oracle]

positive:
((masterpiece,best quality)),((living room background)),

nahida_genshin, 1GIRL, SOLO, small BREASTS, SKIRT, silver HAIR, maid outfit, side ponytail, medium hair, WHITE SOCKS, BLUSH, smile, short skirt, standing, cross-shaped pupils, spoon feeding, cake, garter belt, thigh-highs, petite,

lora:Nahida3:1

Negative prompt: (NSFW),EasyNegative,(worst quality, low quality:1.4),extra fingers,fewer fingers,(lip,nose,3d,realistic:1.2),

[brittle rampart]

😁

[warm oracle]

it uses the easynegative textual inversion

and nahida lora
thats all u need

as for model, its anythingv5

[brittle rampart]

👍🏽

[steady grail]

the simplier the prompt, the better the result

[steady grail]

too much complex prompts push the model out of his confort zone

sweet😋

[warm oracle]

i don't know, i find too simple doesn't give what i want

[steady grail]

the best is to chose a model biased toward what do you want

[warm oracle]

yea, i am

if anything

anythingv5 recommends detailed prompts (the one i mostly use)

[steady grail]

situations too, poses

[warm oracle]

ye ik

[steady grail]

some models can be more oriented on backgrounds, instead of detailed waifus

[warm oracle]

yeah, i have a lot of diff models for backgrounds/waifus

[steady grail]

some other toward anime war scenes, instead of sweet toony waufus underwater

[stable crane]

I need to get an LLM integrated into my workflow

[warm oracle]

yeah its a diverse artstyle

[steady grail]

more than artstyle, is a question of tokens

some tokens are more weighted than others and respond differently

you can insist as much as do you want on a concept. but if such concept, in the models is not relevant, the model will cope with weird results and mutations

[steady grail]

what's llm?

[stable crane]

large language model

[steady grail]

ssd model?

the condensed sdxl?

[unreal ridge]

I think it's something like chat gpt or something

text based models

not sure

[steady grail]

ohh

prompt enhancement

i suppose is not stuff for my low ram

chat gpt sounds very heavy

[unreal ridge]

I mean, it doesn't need to be as heavy as gpt

I remember getting some pretty good prompts with models as heavy as 13B

[steady grail]

i have 4gvram

[rich tartan]

yeah, but it fills your whole ram

[steady grail]

i'm missing 9 more gvram lol

[rich tartan]

needed to use Q4 K_S for 12gb vram, that's a lot

[unreal ridge]

I remember making one bot for this

[brittle rampart]

hmmm .. now i watching.... and i like nahida defautl outfit..

[near elbow]

mor eimages 🙂

[rich tartan]

Comming somewhere, but now i somehow need to find put what image even is/contains.. Then i need to check all the pixels..

And i'll make an if statement node, it's weirdly nowhere included

[rich tartan]

Nearly in

[supple raptor]

Use this one. The ones provided by Stability can't do fp16 precision. https://huggingface.co/madebyollin/sdxl-vae-fp16-fix

[supple raptor]

That problem was fixed like a week or two after the 1.0 launch. The one from Stability should be fine quality wise now, but it's still bf16/fp32 only.

[silk narwhal]

hey @supple raptor , you know how to fix this issue: #🤝｜tech-support message

[nova remnant]

[rich tartan]

pip install requirements.txt

[silk narwhal]

gave up and went back to the stable one ... kept getting 50+ dependency errors even after reinstalling dependencies

[rich tartan]

Fk

I had the same issue with the safetensors missing

[rich tartan]

I used this ig, i need to look it up tomorrow as i'll go sleep now

[silk narwhal]

GN

[stable crane]

ty

[nova remnant]

elf 🥛

[native halo]

shes the queen ,bow down to her 🙇‍♂️

[terse ore]

No it's ok we all do 💩 anyway

[native halo]

source? 🤔

[terse ore]

I assume all humanoid creatures do poop so that's why I don't bow to 💩 factories 😂

[native halo]

🚽

[terse ore]

🚽 is waste removal. Humans and humanoids are the factories.

If anything, we should bow to the toilet for helping us from ourselves

[native halo]

skibidi reference

https://tenor.com/view/skibidi-toilet-hatsune-miku-miku-vocaloid-skibidi-gif-8792879843044916453

[nova remnant]

@native halo

[silk narwhal]

https://pika.art/video/bd8de99a-40e9-4b77-a087-bdd725524bf8

[native halo]

real

[nova remnant]

[stable crane]

neat

kinda anime irl vibe

[native halo]

yea just chillin

[nova remnant]

@native halo pretty jewellery

[native halo]

i love the massive blue jewel 🤠

[vital raptor]

when you have too much bling

[vital raptor]

gotta go fast

[nova remnant]

...

I'm gonna try elves riding one of those

900% wont work

[native halo]

too small 🤠 🤏

[nova remnant]

LMAO

[vital raptor]

too big 😏

[vital raptor]

lol 😄

this one is a bit better

[nova remnant]

[nova remnant]

hewo lewds

[stable crane]

hiyo

I'm making maniacs tonight

what you up to

[native halo]

that knife is too small

[nova remnant]

playin gaaaames

[brittle rampart]

atmospheric

quite interesting... the average number of views of works on Pixiv is no more than 30 for a non-paid account... apparently this is true on any service... restriction of inclusion in the search results for works from such accounts.

[unreal ridge]

I feel like it depends

if it's original it's normally low

however if you do something from an existing IP it's bound to get way more views

[brittle rampart]

existing IP !?

[unreal ridge]

Intellectual property

I used to post some of the stuff I gen on pixiv

Original 200~300 views

existing ips

[brittle rampart]

It turns out that there are only two options. either famous or porn?

[unreal ridge]

no idea

I guess it just works

[brittle rampart]

although there is a third option, it’s rather dubious... shove it into 10 job description tags, the most common ones

[unreal ridge]

Original IPs already have an established fan base. People are already familiar with the characters, etc, which probably makes it easier to pick their interest

Or perhaps it was just luck. Still weird that it happened twice, and with the same character.

[brittle rampart]

do they shoot gifts? otherwise on white snow in a red outfit...

[nova remnant]

Snow ball gun

[glossy eagle]

hehehe

[nova remnant]

Definitely can't join her in there

[brittle rampart]

I don't want to bother with her hand. This illustration will not go further than this server.

i have Pixiv and DA

more for self-expression

[torn quail]

wow! awesome style, awesome art!

[brittle rampart]

Guys, does it sometimes happen to you that at some point, it seems that SD does not perceive any changes in the prompt or settings at all, and you have to completely restart the server?

[rich tartan]

I DID IT

Auto Denoise

It did the job, it has good anatomy on hands(fixes heun issues now good depending of how small/far away the character is)

I need to dump the percentages out and optimize it, but afterwards it can massively improve anatomy fixing

[rich tartan]

Yo and the command to fix the issues is: pip install -r requirements.txt

Workspace before Auto denoise and after Auto Denoise

[steady grail]

auto denoise?

[rich tartan]

yea

the idea i told yesterday

When the character is small/distance is high, high denoise seems to peform better

The Auto Denoiser sees this as 24.857..% White (Mask)

And then it can check if it'll use high/med/low denoise

And at near distance, low denoise is much better

This should improve anatomy fixing very likely, but im still in testing and finding out the perfect denoises

[rich tartan]

And i realized that the angle and the hand size/distance of itself also makes difference, may need to advance this Auto Denoise in the future

[rich tartan]

Works pretty decent

Now my thoughts about this: heun has more attention to the background/the overall image, if the character is further away, it mixes up with the background

And with ddim/ddim_uniform(or other like uni_pc or dpm) you can keep the background with little to no loss, and also fix the character look. So, if the character is mixed by heun, you get mutations(mostly hands and face), but good background, and with another sampler, you can fix the character

Now, with the denoise. If the character is further away, it'll need more noise to fix the character, if it's near, you'll need less noise to fix it(if the character is near and you high denoise it, it'll mutate)
And the risk with high noise is, breaking the shape of the image..

My future plans are to detect if the character is sitting or standing(might split the image to 2 in the middle and check the pixel spread with that) as it has a big effect on what denoise value to choose
And also to make an dynamic denoise, instead of 3 fixed values that the node calculates

High denoise can also fix broken/missing arms/legs, but detecting them will be another big wall

[unreal ridge]

Nice eyes, had to crop cuz sus

[arctic nova]

Wish she was meta

best genshin character design by far imo

[rich tartan]

it can def get better

[serene spoke]

Some help with latent couple why this prompt don't works: Anime illustration, a girl and a boy, the girl is carrying the man, a church, sunset,
AND lora:Kazuya_Kinoshita:1 brown eyes,brown hair, 1boy, wedding dress, is carried, a church, sunset,
AND chizuru , 1girl, solo, long_hair, looking_at_viewer, bangs, brown_hair, brown_eyes, braid, <lora:Chizuru Mizuhara_Sysdeep:1> a church, sunset

[rich tartan]

You get any error or what are your results/ what did you except?

[serene spoke]

An horrible result idk why

[rich tartan]

maybe model issue

[rich tartan]

what model are you using and what is your resolution?

[serene spoke]

model?

[rich tartan]

/checkpoint file

[serene spoke]

aamAnyloraAnimeMixAnime_v1

[bitter latch]

transformation

[rich tartan]

yeah, might try another one

[bitter latch]

or is that a TransFurmation?

[rich tartan]

check out abyssorangemix

[serene spoke]

why doesn't work only with this one?

[rich tartan]

Could depend on many things
Your resolution you are generating at
Bad trained
Missing training data
Bad Base Model

[serene spoke]

Will Anything v5 work?

[rich tartan]

yeah try out

[vital raptor]

could also be your settings, in theory any model will produce something good, just maybe not what you want

[serene spoke]

yes my opinion is that instead of And should be break, but idk how to does it

[vital raptor]

I don't know what you're trying to do, but and/break are prompt things not settings

and the 'and' keyword doesn't do anything by default, but 'break' does in automatic1111

[serene spoke]

I'm try to learn how latent couple works doing a pose of a man holding a girl.

[vital raptor]

ah i see, not sure if it's changed recently but last i checked the latent coupling didn't actually work, the regional prompt extension does better

other than that, if you can get good normal images from the model, it will work with both latent couple (if it works at all) and regional prompting

not the greatest example, but this is using regional prompter

a man standing holding a girl BREAK
girl with white hair,```

[brazen rapids]

true

[rich tartan]

Gonna drink some water

The sampler is really lazy at fixing the mutations

Good idea actually, if the noise is not enough to patch it, just cover it up

maybe another idea to fix mutations

but else, it's good

[magic wave]

Trained a pixel art style, came out pretty clean

[unreal ridge]

cute

the second one tho...

😳 🍈🍈

[magic wave]

Ayyy thanks, and yeee xD

She got some honk honks

[broken osprey]

Is your style in civit AI? Or is it for personal use only ?

[magic wave]

I could post it, I'm lazy as usual though, but I may post it later if you want it xD

[broken osprey]

Haha dw I like the style yeah, I’m assuming it’s an embedding or idk…. Textual inversion is what so people say idk but I do like it ^^

[unreal ridge]

[magic wave]

Its a LoRA cx

[broken osprey]

A

Even better :)))

[magic wave]

Thats a pretty good pic :O

[broken osprey]

Mind sharing your model and prompt? Please :)

[magic wave]

Yee, I only train LoRAs

[broken osprey]

I wish I knew how to haha, rn I’m stuck testing out prompts

[magic wave]

Older pics dump

[magic wave]

Fair, its not too hard to do it, but it does take time :o

[broken osprey]

How do you get such quality

[unreal ridge]

It's a private one. The prompt is pretty simple
simple background, blue eyes, 1girl, mature woman, big breasts, glasses, white hair, ponytail, button gap, short pencil skirt, collared shirt, thick thighs

[broken osprey]

Oh k@okay thanks tho!!

[unreal ridge]

Negative prompt: (low quality, worst quality:1.4), (bad anatomy), extra digit, fewer digits, (extra arms:1.2), bad hands, by (bad-artist:0.6), bad-image-v2-39000, negative_hand-neg, text, watermark,

Steps: 40, Sampler: DPM++ 2M Karras, CFG scale: 7

[magic wave]

Many LoRAs for a unique style

[broken osprey]

So you are telling me Lora’s helped you get quality?!

[magic wave]

Yes, because they're all mine, so they're good LoRAs

[broken osprey]

[unreal ridge]

I wish button gap worked more consistently on my model tho 😔

[magic wave]

The ones I commonly use do be up on my Civit tho

[magic wave]

I've never tried that tag :o

[unreal ridge]

I might look for a lora for it later

[magic wave]

Also tis my Civit https://civitai.com/user/LUUNA, you can find my released LoRAs there @broken osprey

Though I got like 40 unreleased by this point

[magic wave]

Fair :O

[unreal ridge]

It's basically this part here

[magic wave]

Yeee

[unreal ridge]

it's basically a coinflip for me

[magic wave]

I don't usually use concept LoRAs cause they usually make stuff unstable :o

And that makes sense xD

[magic wave]

I need to learn japanese so I can talk to Iroro and he can teach me how to make good SVD LoRAs, because those are the actually spicy LoRAs to make concept stuff

[broken osprey]

This is as far as I’ve gotten with upscale of 1.5 since my GPU can’t handle more

[magic wave]

Fair, I usually do x2 :o

[broken osprey]

It’s not bad per se but I think I can do more. But iterative upscaling usually kills my vram

[unreal ridge]

Learning japanese to make better lewd

https://tenor.com/view/giga-chad-gigachad-big-gif-21053844

beahavior

[magic wave]

Yes

[broken osprey]

Illl check it out, also would mind teaching me how to make Lora’s? If that’s okay of course if not it’s oke too :))

[magic wave]

I just use the normal upscaling :o

[magic wave]

Sure, though how much VRAM you got? :o

[broken osprey]

Which one the 4x ultrasharp?

[magic wave]

Yeah, just normal hires fix :o

[broken osprey]

Eight gigs and a lot of spirit !!!

[bitter latch]

I gots 4

[broken osprey]

Oh I do use that tho, I use latent upscale in comfy which I guess I somewhat smile at

Similar*

[magic wave]

Fair, eight works though its a bit worse, I think you'll need to use different settings :o

[silk narwhal]

These girls need a word with you

[magic wave]

Latent is completely different

xD

[unreal ridge]

I think adding 1girl, mature woman makes my model do 2 shots sometimes

it's cursed most of the time tho

[magic wave]

And its usually way worse I think, but it can be used for some stylization I think

[magic wave]

That is a pretty good one tho :o

[unreal ridge]

ye

[broken osprey]

Please 🙇‍♂️ teach me the ways of creating good lewd anime art

[magic wave]

Whas yo prompt, lemme have a go :O

[unreal ridge]

rare occasion

[magic wave]

Can do

[unreal ridge]

HOLY

[broken osprey]

I feel like I have found home in this chat lmao also I love how comfy saves the images even if I run out of memory and allows me to pick up the generation again

[bitter latch]

and I can do this 6080 x 6080

[broken osprey]

How

[bitter latch]

huh.. got downsized

tries again

[silk narwhal]

yea discord does 2k at most i believe

[unreal ridge]

I think formatting my pc kind of backfired on me... I lost some pretty good prompts cuz I save my stuff in some very random folders

[broken osprey]

No no this time I can zoom in and see each individual hair on that lil fox

[unreal ridge]

Google drive ftw

[magic wave]

Oh I didn't know it did that :o

[magic wave]

Luckily my prompting style is always the same XD

[broken osprey]

If you don’t change the prompt then it lets you keep pressing queue prompt and since the generation is done it works fine

[silk narwhal]

Danbooru 2 prompt extension 😄

[bitter latch]

automatic1111 upscale after generation, diff tab

[unreal ridge]

I'll dm you something 😔

[broken osprey]

What is that?

Also @magic wave which method should I do then in comfy to replicate the hires fix of sd?

[magic wave]

Mysterious :O

[broken osprey]

I can’t use sd since an image that takes at most 3 minutes takes 10 in frickinf webui

[magic wave]

Uhhh, I'm not sure, I'm not too knowledgeable of Comfy

[magic wave]

Thats weird, it should take less than that with 8 GB of VRAM :o

[broken osprey]

Yeah Ikkkkk an image of the same size say 600*900 takes at most 3 minutes including upscale and face detail in comfy and it takes literally too long on webui

[magic wave]

@unreal ridge Tried your prompt with my style mix :o

[silk narwhal]

uses AMD on windows using directML so very heavy performance hit

[unreal ridge]

button gap works for you

[magic wave]

Weird, for me on my 960, with 4 GB of VRAM, it took 10 minutes to upscale :o

[unreal ridge]

[broken osprey]

[magic wave]

It do

[unreal ridge]

model diff I guess

[bitter latch]

if you use hi-rez fix upscale yeah. but I used extra's upscale

[broken osprey]

10 minutes ?!!!!! God damn what did you upscale 10x?

[bitter latch]

😄

[broken osprey]

So what that is a double upscale?

[bitter latch]

believe so 😄

[broken osprey]

Yeah jeeeeeez that says 4x ultra first and then 4x anime B

[bitter latch]

ujpscaled

[broken osprey]

So is that with the upscale before and another round or what I’m so confused

[bitter latch]

this , it's after generation to another tab

[broken osprey]

Funny how even my Chinese model managed to produce the exact same position even down to the hands :0 I only pasted the prompt in