🏞｜general-with-images

[eternal comet]

is bad

do i have to delete

[clever oar]

no good

[clever oar]

last vesrion ok

before

[eternal comet]

yeash

just to be saf

[clever oar]

i try to now

too

[clever oar]

skorch you use embendings?

[clever oar]

you love elon?)

[eternal comet]

no

look at the prompt

💀

[eternal comet]

what model

photogasm?

[clever oar]

dreamshaper

[eternal comet]

xl?

[clever oar]

1/5

bad result deformed face

[eternal comet]

thats sick

epic

[clever oar]

last results is online ai on sd base

[clever oar]

is not local

first result local

online network more beautiful and creative but not listen all promt

[eternal comet]

ya

this model is good

give me some prompts to try

lets see how good it is lol

[clever oar]

frozen sea

big frozen wave

[eternal comet]

"hands up in the fire burn burn burn burn. "

[clever oar]

cat bowl)

[eternal comet]

ok cat bowl now

[clever oar]

magic cat)

[eternal comet]

nice

[clever oar]

skibidi toilet)

[clever oar]

yes yes yes)

[marble nimbus]

I'm trying to generate "fabric patterns" like these, though I usually get very poor results. Do you know any model / lora or prompting tip?

[mighty stump]

mario cooking

[pallid ruin]

Wow!

[wispy nest]

🥱

[clever oar]

lcd beach

[eternal comet]

i live for moments like thius

[clever oar]

who can make tick

[gusty cloak]

A tick?

[clever oar]

i dont know how correct in english

[gusty cloak]

That's exactly what I thought of

🙂

However I don't want to generate ticks 😂 Or things that creep me out

[clever oar]

moffi i know you want to generate this)

my model dont listen promt with ticks

[clever oar]

is octopus)

[jovial tiger]

what should it be doing?

[clever oar]

anything what you want

[jovial tiger]

it's your picture. you uploaded a tick and just want another tick?

[clever oar]

maybe big tick in city)

[dull flame]

ugh no more please, i cant stand ticks

[clever oar]

simple ai dont understand what is

she know spiders

[jovial tiger]

seems it's not trained on it. always ends up with something else with a bunch of legs

[pallid ruin]

Amazing

[south temple]

Hey friends... Im on the struggle trying to figure out what I did wrong here. learning the video end of life and used the basic workflow.... but it breaks at ksampler. anything you see fast that im doing wrong?

[south temple]

nice bawn bawn 🙂

[eternal comet]

almost perfect fingers and text

[south temple]

comfy?

[eternal comet]

bro

hair moment

the fingers are almsot pefect

perfect

[eternal comet]

no

automatic111

[south temple]

starting to question if I should have learned that

[eternal comet]

its not learning its downloading 500 pictures of peoples fingers

then training alora

and 500 egirls with hello kitty and revealing clothes

and other logos

💀

its deadass just that.

[eternal comet]

then you get these pics

[south temple]

I still get nightmare fuel now and again and love it

[eternal comet]

whats the prompt

[south temple]

a girl sitting in the rain, dance music,

[eternal comet]

feels like a punch in the fut

lol it does best with lyrics of songs

[south temple]

its the video animations I want to get into

waahahahahah

[south temple]

LEMME SMELL YER BREATH!

[wispy nest]

https://tenor.com/view/cat-standing-fat-epic-gif-23115386

[south temple]

UGH! ineed to get into the video

[wispy nest]

[south temple]

oh boy

[wispy nest]

[gusty cloak]

Looks like Korra!

Which I'm 99% sure it is 😅

[eternal comet]

it is

[wispy nest]

🌊

[gusty cloak]

@wispy nest I now feel challenged to make Korra art!

[eternal comet]

not that hard to do

50 pics of character

lora

done

[gusty cloak]

No I want to put out something interesting from my part 😅

[hazy warren]

cat girl XD

[clever oar]

try install app for train lora-errors(

tired

[wispy nest]

i love making vader

[eternal comet]

no i cant let go its chemical

[eternal comet]

you break me then i break my rules last time was the last time too its fucked up but im still.

"oxytocin makin it all okay when i come back down it doesnt feel the same."

bear holding a chemical formula

smart

[eternal comet]

lol

wghat model

outside of the part smoking in the car with you

[wispy nest]

@wind egret ok putting together the screenshots

[wind egret]

alright im all eyes

[dull flame]

GAH

what is that ...

[wind egret]

quagmire's dad it would seem

[wispy nest]

ok so right now ive uploaded my picture into controlnet and selected a preprocessor for pose (there are several, what you use will depend on what your goal is) i click the explosion thing and it will generate a pose from my uploaded photo

[wind egret]

yup all correct

[wispy nest]

my pose. next step i drop the preprocessor, and slect a model, and i used the generated pose , lket me get the screenshot

@wind egret

i dont use open pose much, but that shouyld work. ill test it myself now..

you can also use with with stuff like normal map, canny, depth, etc

[wind egret]

i think i see the issue

[wispy nest]

just make sure you use the correct model

[wind egret]

do u upload ur pose image?

[wispy nest]

preprocessor shouyld be none, model should be based on preprocessor ypu used earlier. ex. if you used depth processor, then use the model for that. theres also about.. dozens of different ways to generate an image with specific poses

yeah, you upload your pose from your preprocessor back into the controlnet

the image you got the pose from shouldnt be anywhere in your controlnet at all at this point

[wind egret]

so you wouldnt be able to do batch files with this?

, Version: 1.7.0

[wispy nest]

can you explain what you mean? you want to apply the pose to a batch of images, or you want to extract from a batch? either way, yes you can do that

[wind egret]

is this sd version 1.7?

i think my control models are 1.5

[wispy nest]

1.5 refers to the checkpoint model being used. for example, if i tried using my example and had an sdxl checkpoint model being used for txt2img, i would get an error, which i did because i forgot to change it. something along the lines of "object is not iterable"

this error:
TypeError: 'NoneType' object is not iterable

[wind egret]

sdxl?

is that 2?

[wispy nest]

no, its different, now that ive changed my model to 1.5, it works

you can see, generated image has similar pose (its janky, jused used prompt old man with a cane, used first image generated)

[hushed grotto]

anyone knows how to make these vintage videogame graphics on stable diffusion? people are using facetomany for these (which runs on SD for what i know) and i dont know why would anyone pay for these pics when you can make them yourself on your computer

[grizzled sage]

it's at this point i believe that he's just putting people on and pretending he doesn't know what he's doing.

why would people do that? right? it's long been one of the troll playbook pages though

[wild sorrel]

people pay to not make anything by themselves

[hushed grotto]

after spending 2 hours looking specifically for a guide on how to do this im not surprised as there's NOTHING

[wind egret]

uh, i was just showing i didnt have open pose installed...

[grizzled sage]

when you first asked, i said exactly that. but as usual, it turned into a string along argument

now i'm just sure

[wind egret]

idk, kinda mean to assume malice

[grizzled sage]

hard not to when people seeking advice start to argue

[wind egret]

and i assumed i had them cos i remember a big list, but i guess i only downloaded three.

i wouldnt call, people making mistakes, as people arguing and picking fights with you, that's a lil bit paranoid

[grizzled sage]

i know hanlons razor is a thing, but , i mean, really?

i called it correct 15min ago. yup. just paranoid i guess.

good luck out there

oh mb. was actually 30min ago

[hushed grotto]

if anyone knows how to do this feel free to ping or DM me

[grizzled sage]

medugre?

[clever oar]

how i get normal face with openpose?

and control net

[waxen fractal]

@arctic laurel could i convincgly place a logo using this image w/ photoshop?

[wispy nest]

what do you mean my a normal face? and make sure to keep @ me for replies or just dm, i have 3 conversations going on simultaneously lmao

[waxen fractal]

this is the logo. any workflow reccomendations?

[clever oar]

you see?

its with control net open pose

face distored

i try different settings

oh i see you new in this too)

[eternal comet]

Ahahahaa

How good is photogasm

It looks very good last time I try

[clever oar]

someone show an example with good faces XD

[wispy nest]

take a screenshot of the textbox under your output image

need to see what your settings are before i can recommend anything

[wispy nest]

like this

[eternal comet]

Why is everything you post biblical depictions of entity's from hell

[clever oar]

i see codeformer you use

[eternal comet]

Check metadata

[wispy nest]

…. it’s called card game artworks but ok

or if you just want an easy fix without fucking about with all the settings, you could just use the adetailer extension @clever oar

[wispy nest]

i honestly keep forgetting this is a thing

[eternal comet]

3 headed dog

Really???

[wispy nest]

yep

[wispy nest]

also in fairness, i shouldnt have to do extra steps to help someone if they can just screenshot the info below the image XD

[wispy nest]

looks like you dont have restore face on, trying using that

[grizzled sage]

codeformer is a quick fix. adetailer extension even better

[clever oar]

thx is what i want how

i need is after generate

or before

[wispy nest]

uh, either way? but yeah adetailer is better. i usually turn off restore face most of the time

[wispy nest]

inpaint

[wispy nest]

restore face or adetailer

[grizzled sage]

find code former in settings > face restoration

adetailer is an extension that you ahve to install through extension menu

[clever oar]

thank you i dont know about this setting

[wispy nest]

to add on to that, adetailer also has models you can add to it for specific things, iirc the defaults do include face and i think body though

[grizzled sage]

good thing you asked. its tucked away. i know about it becaues back in the day it was available in the main UI by default

[clever oar]

why you dont say it before when i ask you)

codeformed better than gfgan?

[wispy nest]

its just preference, try them both. i recommend atting it to quick settings to enable and disable it without going to settings, also recommend adding clip to quick settings as well

[wispy nest]

quicksettings are the things at the top of the ui next to where you select your checkpoint

[clever oar]

with restore face

codefromer on

[wispy nest]

if restore face isnt cutting it, get adetailer. check available and search for adetailer (or maybe its !adetailer) or just load from https://github.com/Bing-su/adetailer.git

[nimble mason]

Even better is segmenting out the face, upscaling to native resolution, unsampling, resampling, scaling back down and pasting back in

I have a comfyui workflow that does that so if you're using that I can dig it up

[wispy nest]

im actually thinking of trying out comfy, seems it allows for a bit more control from what ive seen. im just now getting comfy (heh) with a1111/forge ui though and dont know if its worth learning a whole new ui

so yeah id like to see that if you dont mind diggingpu it

digging it up, even

also, if im not mistaken, what you just described is pretty similar to how ultimate sd upscale works, + - a few steps here and there. i think..

[nimble mason]

Without any doubt it does

I personally found it easier cuz nothing is hidden from you anymore

A lot less "wtf does this magic button do?" Type shit

[junior sky]

A bit more control is an understatement.

[nimble mason]

I still use forge. Often but that's just cuz I can actually use it on my phone

Comfy doesn't render nicely on a mobile device with a small screen

[grizzled sage]

adetailer does that too. you just tell it to use your custom resolution

[wispy nest]

could make it easier when you know what all that hidden stuff is / does. in my case i think it will be like more "wtf does this magic button do" type of stuff XD i would legitimately pay for an extension that would show a brief description of what something does when hovering mouse over it. would save a shit ton of time spent on google

[nimble mason]

I've gotten better results with what I described... A lot more control over the process

I also do frequency separation to control tonal information with the low pass layer

[grizzled sage]

probably is just confirmation bias. it's already doing it by default. the advanced settings are all in drop menus

[nimble mason]

Using a combo of color burn/dodge with the inverted original image subjected to median blur

[grizzled sage]

can tweak pretty much all parts of the diffusion process in them

[nimble mason]

Nah it's not doing what I'm doing pretty sure about that

[grizzled sage]

i mean, okay, but you keep changing the goalposts then.

[wispy nest]

what id really love to see is more info in the cmd log as things are being generated, so i can pinpoint exactly what settings /extensions / etc are doing what throughtout the process. that combined with live preview would be 10/10

like this doesnt tell me shit
To load target model BaseModel
Begin to load 1 model
[Memory Management] Current Free GPU Memory (MB) = 18460.991201400757
[Memory Management] Model Memory (MB) = 1639.4137649536133
[Memory Management] Minimal Inference Memory (MB) = 1024.0
[Memory Management] Estimated Remaining GPU Memory (MB) = 15797.577436447144
Moving model(s) has taken 0.48 seconds
100%|██████████████████████████████████████████████████████████████████████████████████| 37/37 [01:28<00:00, 2.39s/it]
Memory cleanup has taken 0.73 seconds | 148/740 [06:10<22:52, 2.32s/it]
Cleanup minimal inference memory.
Miaoshouai boot assistant: Memory Released!
the best i can do is watch live preview and make semi-educated guesses.

[hazy warren]

eye see what you did there

[clever oar]

what i did

[wispy nest]

i hope he takes the advice to use adetailer. those hands and feet are super janky looking 😄

[clever oar]

i thought google colab works-when i start no module name gradio XD

everywhere sex with python)

[hazy warren]

@wispy nest an eye for an eye

[nimble mason]

alright gonna fire up comfy in a sec here and dig that up for ya

[wispy nest]

thanks!

[nimble mason]

np!

there's some things you can change pretty easily too so it's flexible for the situation

[eternal comet]

💀

wrong chat

[wispy nest]

are you going to use the voice cloner? let me know how well it works!

[eternal comet]

yess

i am

its very good so far

[wispy nest]

seriously lol. i need to give ronald mcdonald the marlon brando voice. i need it in my life

[eternal comet]

going to be protecting people

hopefully

[eternal comet]

can you check

if rtx 3060 is good enough

it has 12gb of vram

ill install after you lemme know with detials

yes its working

large language models are amazing at this lo

l

[wispy nest]

i have like 6 things going on simultaneously at the moment lol but yeah i was planning to test it later anyway, ill let you know what i find. im not super tech savvy but i think there is a way for me to "partition" my vram. so, when i get it set up ill see if i can run it on 12gb

[eternal comet]

no ned just run it and look at task manager vram usage

👍

hope you go through all the stuff youre doing with exceptional performance 💪

[wispy nest]

will do. will be a couple hours. i am over promising and under delivering on too much stuff at the moment lol

[eternal comet]

ok no it just told people the detection phrase 💀

[clever oar]

wtf i need change python code for train lora in google colab /its must work by defaults

[wispy nest]

detection phrase for what? also, if you suck at programming like i do, i HIGHLY recommend gpt 4. i have 0 idea how to do any programming, and that thing has helped me debug so much shit in a1111 , as well as writing some random powershell scripts.

[eternal comet]

i dont like things not hosted on my pc

[wispy nest]

like, i can barely navigate visual studio. i just paste it screenshots and have it tell me what to do 🤣

same, but its a hell of a tool. just edit your pictures to scrub any personal info, strip the exif data too etc..

[hazy warren]

this is as close I have gotten to putting a galaxy in an eye

[eternal comet]

7b model is ass

[eternal comet]

let me try

[hazy warren]

this is also a nice one

[eternal comet]

this one is insane lol

vram usage is ludicirous tho

[wispy nest]

just finished a 33 minute generation on a single image.... and realized i forgot to change the fucking base model

[hazy warren]

😮

[wispy nest]

the original (just random pic from google, not generated)before i ran it through img2img

[eternal comet]

this models prompt following is crazy

[wispy nest]

aaaand the 32 minute generation

[hazy warren]

😮

[eternal comet]

what model are you using vortex

[wispy nest]

75 mb image lmao

[hazy warren]

SDXL

[hazy warren]

Why so long?

[eternal comet]

oh shit

nahh

i would be pissed with that outcome

[wispy nest]

it... didnt turn out how i wanted. the model was not correct and so the loras didnt word D:

[nimble mason]

need to clean it up lol

type clownshark workflow with extra stuff in there

[eternal comet]

oh my fucking god

[nimble mason]

but it works

[eternal comet]

do you hate it when it generates naked women

💀

its so bias bro

[wispy nest]

took that long because i upscaled it to 8512 x 6400 . it ended up being 75 mb. theres probably a faster way to do what im trying to do, but i havent found it. the aim is to take an image and upscale the hell out of it and turn it into a photomosaic. i had a REALLY good result yesterday, but the content was a bit on the dark / bloody side, so not sure im allowed to post it here XD

[nimble mason]

hahah

what gpu?

[hazy warren]

I normally just do 1024, then upsharp scale 2x

[hazy warren]

[wispy nest]

thanks for that! gonna hold onto this. probably gonna try out comfy later today or tomorrow

[nimble mason]

sounds good

i'll ping ya if i clean it up a bit

this is a version without the unsampler rigged up but it's running anyway lol

so it's kinda inefficient

but if you just punch in a prompt it will work pretty damn well

[wispy nest]

i usually dont upscale do something ridiculous like this, but i havent found any other way to do a photomosaic. i dont think its an intended use case at all to be honest

i actually stumbled on it on accident when i messed up some settings trying to do a 2x upscale the othert day

[nimble mason]

seems like a job for segmentation/manual tiling then seam fixing tbh

[hazy warren]

Hires fix

[nimble mason]

tiled ksampler might be able to do it

but it'd take ar eally long time

[wispy nest]

a mosaic?

[hazy warren]

[nimble mason]

yeah

realllly long time for that resolution

[wispy nest]

cant take longer than that damn 32 minute generation

[nimble mason]

oh... it can haha

do you have a 4090

[wispy nest]

it doesnt need to be that high res, i just havent found a way to do a mosaic without making it ridiculously high res

a4500

20gb vram

[nimble mason]

oh nice

yeah it takes a while on a 4090

[wispy nest]

getting much better!

[hazy warren]

I just have a simple 4080

[wispy nest]

what would you recommend for a mosaic without using such a high res? i havent been able to play around with it much because of how long the generations take, so im not sure yet exactly how the tile width and height affect the number of images that go into the mosaic

[hazy warren]

Needs to be ninjas

[clever oar]

will be

[nimble mason]

here's a simple example with some of the more aggressive settings

[nimble mason]

idk, i haven't done it, i'm just speculating

[wispy nest]

its uh, not a mosaic though 😄

[nimble mason]

it might be worth trying ultimatesdupscale with the seam fix on

[nimble mason]

oh, that was for the face thing

re: all the adetailer stuff

[wispy nest]

oh , that was dicordos that needed help with that. im just looking for more efficient ways to do mosaics

[nimble mason]

oh lol oops

yeah try tiled sampling

[hazy warren]

Needs a sharp upscale

[nimble mason]

that affects the coherence of the image

whatever you do to the face needs to be done to the rest of the image unless you're really careful

[wispy nest]

not if done with tiling 🙂

[nimble mason]

there i was demonstrating one of a few methods i got to restore detail to the face

[wispy nest]

shouldnt even need to use seam fix if the tiles are wide enough

[nimble mason]

without resulting in a jarring change in tonal information, sharpness, etc

the idea is to blow it up huge, use a lot of tiles at native resolution, so it's more likely to want to turn general shapes into objects like guns or people

whether that will work? idk, haven't tried

but it's the best idea i have

does that mean it's worth trying? idk lol

[wispy nest]

remember to change your resolution settings back after messing around with upscaling and tiles :sigh:
[Memory Management] Current Free GPU Memory (MB) = 18588.484378814697
[Memory Management] Model Memory (MB) = 159.55708122253418
[Memory Management] Minimal Inference Memory (MB) = 1024.0
[Memory Management] Estimated Remaining GPU Memory (MB) = 17404.927297592163
Moving model(s) has taken 0.75 seconds
Warning: Ran out of memory when regular VAE decoding, retrying with tiled VAE decoding.

[nimble mason]

whoops

[wispy nest]

apparently doing a 4k x 4k img2img with a 1.5 model is a nono

[nimble mason]

that's with letting the rest of that workflow run

i usually use it with loras

this one is for shit where you only have good close up photos for training a lora

so you generate a person first, then swap in the face with the lora

but i just made it something generic for the sake of that demo image

[wispy nest]

good shit!

[wispy nest]

what do you mean about training though? you mean, you swap the face in to generated images to use the result for training the lora? im just now getting into making loras so im a little ignorant on this. my understanding is you need a large amount of images of the character you want to make a lora for, in different poses, from different perspectives, etc. was going to try making a lora this weekend, plan was to use reactor extension to swap face in on different generated images and then touch them up with inpainting or in gimp using some fuckery with layers and alpha

[nimble mason]

oh, i mean if you're starting with nothing but photos of someone that are pretty close up

if you only train it on close up shots, it won't capture likeness at all if the face isn't occupying most of the image

so the workaround is to crop and upscale the face like in this workflow and swap it, but it has to be done just right or stuff like tonal information, shadows, the facial expression, proportions etc can be lost and make it look uncanny

[wispy nest]

tracking so far

[nimble mason]

that or sharpness too actually is often a dead giveaway

so then yeah you can generate a bunch of these kinda shots

manually pick out the ones that look good

then throw them back into your training set and rerun it and get a much better lora that can do multiple distances

a few rounds of that can solve a lot of issues

i'm no expert on it at all but i have found that works pretty well

this works a lot better than reactor fyi

this workflow that is

[wispy nest]

thanks for the tips. tried doing a lora for the first time a couple days ago, and was struggling hard with getting a data set together.

[nimble mason]

for one, onetrainer is easier to use and a bit faster and uses a lil less vram than kohya

[wispy nest]

getting consistency for the dataset has not been easy

[nimble mason]

two, it has a really nice feature... you can mask out the backgrounds and it diminishes the attention it gives to those things

[wispy nest]

ill look into onetrainer, was under the impression that kohya was the way to go

[nimble mason]

so it captures the character better

you want to leave the backgrounds in but just caption them reasonably well

it reduces the masked areas to something like 10% strength, whatever that means exactly

https://github.com/starik222/BooruDatasetTagManager this is what i use for tagging, it will handle most tagging automatically and does a pretty damn good job

then you just clean it up and it's easy to do with big datasets cuz you can do shit like add a tag to everything at once

[wispy nest]

training weight probably? makes sense, you wouldnt want the background to influence the images generated using the lora (unless you do want that i guess)

my plan was to just have transparent background in all the training images, like in that image i put above, i just made separate images from each pose and then made the white background transparent

[nimble mason]

yeah, something along those lines, just not sure the exact implementation

you want real backgrounds

i tried that and the results were shit tbh

good news is it means you get to be lazy

[wispy nest]

guessing it fucks with stuff like shadows and the character looking like they are actually IN the image instead of just being pasted in with paint or something?

[nimble mason]

i forget what the exact issues were, but it was bad lol

[wispy nest]

ill take your word for it haha

[nimble mason]

but yeah onetrainer has been a lot easier imo

i can give you the settings ive been using if you end up grabbing it just lemme know

[clever oar]

with restore face

[nimble mason]

yeah it can't do horizontal or upside down shit

quick n dirty n aggressive demo of the face swap aspect

now she's margot robbie

used a lora for that

[eternal comet]

what the fuck

💀

[wispy nest]

aside from ease of use, and vram usage, do you think it gives better / different results? vram usage & power usage are kind of a non issue. the way that i run all this is a bit unique and probably barely adheres to usage terms for the hardware XD all my SD stuff is run from a rented remote connection. so, fuckem, i am not the one paying the electric bill. not my problem if they didnt anticipate these use cases and prohibit them or charge more for using it like this 🤷‍♂️

[gusty cloak]

Came just in time 🤣

[nimble mason]

who? ... or ... what?! is that

[gusty cloak]

Toyed around with Gemini 1.5 Pro myself but it often breaks a little bit 💀

[wispy nest]

2 steps forward and one step back lol. im telling ya, try out the adetailer extension. you can fix your hand issues with it too

[nimble mason]

haha cool

[gusty cloak]

Hello???

What is happening on this picture?

[nimble mason]

the masking makes it better IMO with onetrainer

it's easy to do inside the program itself and it'll even tag things for you too iirc

[clever oar]

dont know)

[nimble mason]

my results were better

again - not an expert at this

but based on the couple dozen halfass loras i've made, it's been better

wouldn't be surprised if there's something i just don't know with kohya etc that can match that maybe

but i sure haven't found it

[nimble mason]

there is a beach ball butt about to pop like a zit on a face straight out of jacobs ladder (1990)

[grizzled sage]

i keep hearing this but kohya has dealt with png alpha channels as attention masks this entire time

[wispy nest]

well, thanks again for all the tips! ive been screenshotting this convo so i can reference back to it when i set up one trainer and give it a shot lol

[nimble mason]

yep there you go lol

thanks for the info

[nimble mason]

def hop on their discord if you have more questions beyond the basic stuff cuz my experience there is pretty rudimentary

they've been helpful

the nice thing with onetrainer's masks is you can have it generate them for you automatically using clipseg

maybe kohya has something for that too, idk

[wispy nest]

man im JUST knowledgeable enough about any of this to have a coherent conversation and not sound like im pulling words out of my ass. i dont think i could even form a question that goes beyond basics

[gusty cloak]

As long as you don't throw prompts into the void expecting an image you're good 😂

[nimble mason]

that was me like two months ago

[nimble mason]

man i love when ppl do that haha

they're tossing steaks to a shark

and what they get back sure as hell ain't ribeye

[wispy nest]

i totally still do that. isnt that how Loab was created? 😄 also ive been way more interested in digging into the technical aspects, ive barely learned about prompt engineering. anyone who knows what theyre doing would probably have an adverse physical reaction if they saw my prompts

[gusty cloak]

I have prepared images for this purpose à la @nocturne oak , here's a little teaser 👀

[gusty cloak]

No I meant the people who join thinking there's a bot here 😅

[wispy nest]

ahhh

[nimble mason]

i've got a few ready myself

[wispy nest]

bot : create an image of a thick curvy ninja woman doing a roundhouse kick on a random congressman. studi ghibli style

[gusty cloak]

Here's the image you requested.

[wispy nest]

yeah unironically those are good. i usually make horror style images and those are kinda cool lol

[nimble mason]

check this out, this is with the unsampling turned on and sampler settings changed

first is the original generation, second is margot robbie swapped in

[clever oar]

i cant rotate to down people with openpose?

[nimble mason]

note how little changed with the texture, tone, lighting, shadows, orientation of the face, proportions, etc

[nimble mason]

send that one to sports illustrated swimsuit ed, i heard that magazine is all nonexistent ai writers anyway lol

[wispy nest]

yeah im sold. i think ill have to learn comfy now. that swap is very impressive

[nimble mason]

i'll be glad to help with any q you got

[wispy nest]

how... how did you even cause the exercise ball to have an ass???

[nocturne oak]

Don't you dare go posting that in #1019361238234443776. That, sir, is a win.

[nimble mason]

i didn't even notice the ball ass

seriously that's better than generating the next mona lisa

[wispy nest]

this dude is effortlessly creating body horror. a true artist

[nimble mason]

^^

[clever oar]

is my pose

and i get it)

[wispy nest]

amazing. can i see the log under your output image? i need to know how to do this

[nocturne oak]

We out here makin' images, while this dude is making art.

[clever oar]

lol is simple pose why it make horror

[gusty cloak]

By the way, regarding this image:

Do you want to know the prompt?

[wispy nest]

yes!

[gusty cloak]

🤣

"Nothing"

That's it

[wispy nest]

imma try it. what model?

[gusty cloak]

Oh I used hephaistos NextGen DPO

[clever oar]

its for me?)

[wispy nest]

yes, screenshot the log under your output and share it! i want to know how to master your unique style

[nocturne oak]

That would be body horror.

Even w/o cnet.

🤣

[gusty cloak]

Right, I wanted to make Korra-like images earlier!

[nocturne oak]

https://tenor.com/sCojqgQIeLc.gif

^ me waiting on this 4k upscale.

[wispy nest]

nah not your prompt, the bit under the output

@clever oar ^

[gusty cloak]

Oh Damn A1111.

Launching Web UI with arguments: --xformers --cuda-malloc --cuda-stream --pin-shared-memory
WARNING:xformers:WARNING[XFORMERS]: xFormers can't load C++/CUDA extensions. xFormers was built for:
PyTorch 2.1.2+cu121 with CUDA 1201 (you have 2.0.1+cu118)
Python 3.10.11 (you have 3.10.11)
Please reinstall xformers (see https://github.com/facebookresearch/xformers#installing-xformers)

I do not have Cu118

[wispy nest]

yep, time to delete the venv and repo folders 😄

[gusty cloak]

😭

[wispy nest]

it only takes a couple minutes to fix lol

[nimble mason]

here's the full scale version

[gusty cloak]

I literally reinstalled PyTorch earlier

[nimble mason]

you don't really need xformers

unless you have low vram

[gusty cloak]

I have 12 🤔 There was a convo earlier claiming it improved performance

[wispy nest]

seriously, just delete venv and repo. you lose nothing. a clean install is waaaaay easier than messing around trying to figure out what venv is using, whats on PATH, debugging callbacks etc. 99% of the time your problem will be fixed with fresh install

[nimble mason]

it slows you down just a bit and impairs reproducibility

reduces vram use a bit

iirc

i haven't used it for much other than training sdxl loras where there's no way around it unless you have a > 24gb vram card

[wispy nest]

wait really?? i thought it was just a positive thing to have if you had high vram

[nimble mason]

i think it has a minor performance cost? it does with training at least

[gusty cloak]

Well I removed the arguments from the .bat file, however I do have a question

[nimble mason]

i could be wrong

[gusty cloak]

Sometimes loading models takes PAINFUL amounts of time.

[nimble mason]

again haven't really used it for inference

[gusty cloak]

Loading a model took me a good 3 minutes once.

[nimble mason]

are you loading off a HDD?

[gusty cloak]

Ah...

🤣

Damn

[wispy nest]

lol

[nimble mason]

hahaha

[gusty cloak]

Well shit, my SSD is almost full

[nimble mason]

one of these i'll actually get around to getting a 4tb ssd

[gusty cloak]

It's WizTree time 😎

[nimble mason]

tbh, i'm just not looking forward to having to take my 4090 out to put that thing in but i need it badly

[wispy nest]

hm, i wonder how the speed would be loading from a usb drive

[clever oar]

dfrom its what you request?

[nimble mason]

this was me a few weeks ago

now i have 7gb free!

i have over 240 checkpoints >_>

[clever oar]

i have 6 tb

[nimble mason]

of SD shit?

[gusty cloak]

😂

[wispy nest]

yeah it is, i honestly have no idea how youre getting that body horror with those settings. i would recommend lowering your cfg just a bit though

[gusty cloak]

But I also want to ask then, A1111 is on my ADHD

WHAT

HDD***

MUSCLE MEMORY I'm sorry

[nimble mason]

lol

[clever oar]

i try lol

[gusty cloak]

If A1111 is on my ADHD HDD, would it load models from my SSD just fine?

[nimble mason]

the body horror is because the character is upside down

yeah

[wispy nest]

jesus christ lol

[jovial tiger]

is that the new hot yoga everyone's trying

?

[nimble mason]

SD can't render ppl even just sideways unless they're asleep

[wispy nest]

turn your sampling up to like 80 or so and see what happens @clever oar

[nimble mason]

i was waitin for ya to show up for this conversation... i have one of yours screencapped cuz it made me laugh

i think you said something like "this is how we will defeat the AI: prompt man upside down"

[clever oar]

it cant be normal never if i turn like this?

[wispy nest]

think so, cause the models are being loaded to gpu if i understand correctly.

[jovial tiger]

I think it was when ideogram first came out, and someone almost immediately asked like they had it ready to go: woman upside down in a crab walk position. - and it did it.

[wispy nest]

bro i have no idea. i havent messed around much with fat women doing upside down yoga to be honest. but, where there is a will there is a way... or at least some funny body horror

[wispy nest]

used "woman upside down in a crab walk position" . nope, looks like todays theme is in fact yoga horror

[clever oar]

😃

[jovial tiger]

I tried upside down woman doing a split in yoga pants in cascade.

[gusty cloak]

I wonder if Dall-E 3 can do it?

Can someone try on Designer, I need to free SSD space 😂

[jovial tiger]

🤖 Dall-E job accepted.

[jovial tiger]

yeah no problem with dall-e

[gusty cloak]

Pretty good!

[nimble mason]

what about handstands?

for sdxl vs dalle3

i'm currently cleaning up that mess of a face swap workflow............

[clever oar]

dalle have same instrument like openpose?

[gusty cloak]

Dall-E has basically nothing 😅

Just you and the positive prompt only

[clever oar]

oh)

[gusty cloak]

Unless I don't know about something out there?

[dull flame]

its rly restricted then

[jovial tiger]

🤖 Dall-E job accepted.

A captivating scene of an upside-down mecha performing an impressive split in while in a handstand, wearing yoga pants, with birds resting calmly on the tips of its toes. The mecha's design is sleek and futuristic, with metallic silver and glowing blue accents. The park setting is serene, with a blue sky, green grass, and a few trees in the background. The overall atmosphere of the image is a mix of sci-fi and tranquility, showcasing the harmony between technology and nature.

[gusty cloak]

Damn!

Imagine what SD3 would do then!

Good, time to spam SDXL models in there now...

[jovial tiger]

ideogram can do it too.

let's try stability core API

[jovial tiger]

that's actually pretty good

if that's controlnet

[clever oar]

yes

[wispy nest]

dicordos. are you ok? 😆

[clever oar]

on 30 %)

[gusty cloak]

I want to ask, how do I change which folder A1111 checks for models?

[wispy nest]

edit your webuiuser.bat

[jovial tiger]

yeah, as expected, stability ai API didn't even try. no yoga pants, no upside down. there's a robot and a park though, so 2 points.

[gusty cloak]

Assuming VENV_DIR?

[wispy nest]

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS= --autolaunch --api --theme dark --xformers --cuda-malloc --cuda-stream --pin-shared-memory

@REM Uncomment following code to reference an existing A1111 checkout.
@REM set A1111_HOME=Your A1111 checkout dir
@REM
@REM set VENV_DIR=%A1111_HOME%/venv
@REM set COMMANDLINE_ARGS=%COMMANDLINE_ARGS% ^
@REM --ckpt-dir %A1111_HOME%/models/Stable-diffusion ^
@REM --hypernetwork-dir %A1111_HOME%/models/hypernetworks ^
@REM --embeddings-dir %A1111_HOME%/embeddings ^
@REM --lora-dir %A1111_HOME%/models/Lora

call webui.bat

nah your venv wont dir wont change that

you want these @REM --ckpt-dir %A1111_HOME%/models/Stable-diffusion ^
@REM --hypernetwork-dir %A1111_HOME%/models/hypernetworks ^
@REM --embeddings-dir %A1111_HOME%/embeddings ^
@REM --lora-dir %A1111_HOME%/models/Lora

[nimble mason]

what about headstands with the character lookiung directly at the camera

[wispy nest]

and this @REM set A1111_HOME=Your A1111 checkout dir

[nimble mason]

so basically just a person standing, except upside down

and a real person, not a cartoon or mecha or anything else

pretty much as close as you can get to a non-close up portrait photo rotated 180 degrees

alright the face swap workflow is cleaned up

[gusty cloak]

I'm confused, probably because it's 4 AM, but what exactly is going on here? It just looks like I'm pointing it at the... Ah. Okay, it clicked now.

If I set A1111_HOME to the C: folder (my SSD), it will just use that as a variable in the 4 directories

[wispy nest]

so you would do it like this, unless im wrong which i probably am
set A1111_HOME="C:\whateverthepathis"
--lora-dir %A1111_HOME%/models/Lora

[gusty cloak]

Got it

[wispy nest]

yep!

[nimble mason]

try using forge btw if you're on something with 12gb of vram

forge was way better on my 12gb 3080 than a1111

and is also way better on my 4090 than a1111

[jovial tiger]

Hmm let me see

[wispy nest]

i have 20gb and prefer forge. only complaint is that integrated controlnet doesnt play nicely with some extensions

i do have a1111 too, for the stuff that just outright breaks forge.

[gusty cloak]

😨 I don't think I can just copy my models folder, ahh shit 🤣

[clever oar]

why is so distore face \i dont understand,maybe its not close

[gusty cloak]

Will handpick

[wispy nest]

why not?

[gusty cloak]

I don't have that much space on my SSD 😅

[wispy nest]

oh, thought you were just going to point your ui at the existing model folder though?

[grizzled sage]

stable xl , cascade, 15, 2, none of it can do odd poses. break dancing is mutagenics

[gusty cloak]

My models are all on my HDD 🤔

[wispy nest]

you shouldnt need to copy anything that way. or wait, its on your hdd isnt it, i see what youre doing

[nimble mason]

yeah lycoris-ia3 doesn't work on anything but a1111 afaik which is annoying

[nimble mason]

dayyyuummm

[grizzled sage]

all boils down to bad datasets with bad captions

[nimble mason]

i'd put all my work data on thumb drives before i'd put my checkpoints on a hdd

[jovial tiger]

And this is why I hardly use dalle. Content violation warnings, or it drastically changes the content. Wasn't expecting the third leg. Here was the request: /dalle photorealistic woman in yoga workout clothes doing a one arm handstand while smiling at the camera. Legs together.

[wispy nest]

are you in europe by any chance? i recall reading about some cloud storage that you can access/edit in real time, basically acting like an extra drive. not available in US, and im not exactly sure how/if you could point your ui at it, but from what i skimmed over sounds like something like that would be possible

[wispy nest]

thats why local is the way to go. i hate the idea of "content violations" especially with how opinionated some of the content violations are

[grizzled sage]

most cloud storage can sync to an actual folder. it'll be cached on your pc though

[clever oar]

hibrid insect and woman 🙂

[grizzled sage]

you can't just access 5gb in real time

you're accessing a local version of it that'll sync when changed

[nocturne oak]

That lava throne image I put above...open up the full version.

[wispy nest]

that was my understanding too, but from what i skimmed over it sounded like it was set up differently. i am NOT tech savvy though and probably horribly misunderstood. ill see if i can find the link to it

[nocturne oak]

The details...the 4k details.

[wispy nest]

oh also for the people trying to do upside down poses... for whatever reason.. i messed around with it a bit out of curiosity. ipadapter in controlnet was giving much cleaner results. i didnt try running it with adetailer or restore face... so uh have fun with that

[nimble mason]

initial generation, two different initial swaps with slightly different configurations, and an upscaled version

random woman becomes margot robbie

or whoever else your model knows how to make, or that you have a lora for, or an image to feed into ipadapter portrait/faceid/etc

[jovial tiger]

Ideogram. Could use some cleanup but I trust you're good to do that.

[nimble mason]

retains the key characteristics of the original image better than anything i've seen

yeah it's got it down

[gusty cloak]

Or I can just patiently wait for models to load 😂

[jovial tiger]

I said legs together but it didn't for any of the 4

[nimble mason]

the mouth isn't upside down, the eyes aren't... those sdxl always flips

[clever oar]

i off control net and this result

[jovial tiger]

What does the controlnet skeleton look like to get those?

[wispy nest]

no cleanup needed. look at those perfect and anatomically correct legs. 😄

[nimble mason]

at first i thought i was supposed to look for a tiny trump melting into the lava like the terminator at the end of T2 lol

[jovial tiger]

She's clearly been doing yoga a long time.

[wispy nest]

she got that grandma leg physique

[wispy nest]

have you tried using dwopenpose full for your uh upside down stuff? it has control points for face. so, maybe that or ip adapter is the way to go for the upside down stuff.

[jovial tiger]

That looks like a squashed bug

[jovial tiger]

Those aren't the arms

[nimble mason]

tbh, i bet the best way to go upside down is to render right side up, then replace the background with differential diffusion

after rotating 180 degrees

[clever oar]

why

[jovial tiger]

Could potentially use that recent transparent render thing that came out a couple weeks ago

[jovial tiger]

I'm joking those are .... something else from that picture

[wispy nest]

yeah i just looked into it. guess i was having some kind of feverdream memory. it is just normal syncing. really couldve swore i had read about it being accessed the same way as any other local drive, guess im losing it

[nimble mason]

i rotated this

[gusty cloak]

Yo??? Sorry to interrupt but my model can literally do Korra?

Kind of

[nimble mason]

by rotating i mean i rotated 90 degrees with differential diffusion 4 times

SDXL and SD15 desperately want objects to be oriented in sensible directions

so if you want nonsense, you gotta rotate the image and inpaint it

[wispy nest]

i was going to jokingly paste a screenshot with a big arrow pointing at the rotate image button lol. but yeah, that wouldnt work for say a woman doing an upside down pose and make things like the hair, body parts etc realistically affected by gravity

[nimble mason]

prolly, unless maybe you first generated her getting sucked into a UFO with a big vacuum

[wispy nest]

THAT is not a bad idea

[nimble mason]

heres how you actually do it

[dull flame]

https://tenor.com/view/pikachu-shocked-face-stunned-pokemon-shocked-not-shocked-omg-gif-24112152

[nimble mason]

you gather up 30-60 photos of people upside down in various poses and make a lora

[clever oar]

very close how i correct distance)

only big tits i see)

[nimble mason]

careful, i got put in the corner for 12 hours for something similar lol

couldn't see it on my phone

i'm sure that's just the fabric that's colored pink but yeah

[clever oar]

in real life?

[nimble mason]

prompt was something like "a woman" rotated it 4 times

[nimble mason]

nah, on here, i was given a timeout by the mods

as in, they stuffed a ball in my mouth like it was pulp fiction

more image rotation stuff with diff diff... which way is up? wouldn't look too close if you get motion sick easily

[grizzled sage]

marketing probably. they may have actually made a claim like that but then regulators were all "hey hey now!"

[clever oar]

when i be young my mom found picture with naked woman i she try ban me)

[wispy nest]

petition to automatically pin everything dicordos says

[clever oar]

fun how she happy

lol

[gusty cloak]

(first one is a little poorer but) I can't get Korra to be in the Avatar State... 🤔

[grizzled sage]

bahn mi is pretty good i recomend everyone try it

[jovial tiger]

See, dalle would have been more censirrdipped.

[wispy nest]

thats the sandwich thing right?

[grizzled sage]

avatar state lora would be awesome. thats cool concept

[gusty cloak]

I can probably inpaint 🤔

[clever oar]

yes now sd is best for relax content

[nimble mason]

try ip-adapter with a source image maybe

[wispy nest]

dalle, despite crippling censorship, does still make some cool stuff. screenshot cause im too lazy to dig through all my images to find that, i dont have much dalle stuff in there @jovial tiger

[eternal comet]

Dall e overrated

Llm is carrying it

[gusty cloak]

I mean, to be fair, my generations do reflect how useless Korra's Avatar State is, SDXL refuses to generate it 💀

[wispy nest]

prompt was something along the lines of "oil painting in the style of saturn eating his son, a man hunched over a plate of rotting food"

[nimble mason]

i've gotten it to make an insane psycho dude in a meth lab in bosnia

and a couple spectacularly beautiful oil paintinsg of nuclear explosions in cities which it deleted very quickly... but i was faster

[clever oar]

mouse logo

[eternal comet]

💀

[wispy nest]

i would love to see that if you have it saved, too bad you didnt get the explosions in time. i was thinking of animating some nuclear explosions once i figure out how to animate properly

[eternal comet]

https://youtu.be/lkfReauilG8?si=b9dM3z5_iBMP1noD

I get my prompts from this song

💀

The lyrics

[nocturne oak]

Yikes...the fingers.

[eternal comet]

I use it to benchmark a model

[wispy nest]

my current "best" animation. which i cheated on. i just used img2vid on uhhh i think it was haiper

[nimble mason]

yup... that's just from the base image though

what i'm interested in there is the face swap

[nocturne oak]

No, I know...still...yeesh.

[nimble mason]

yeah haha

[grizzled sage]

https://civitai.com/models/352245/aang-avatar-statei found this aang lora but its only aang. cool efect though

[wispy nest]

yeah, this upside down thing. im kind of invested in this now. using adetailer, and ipadapter. i think this is what you would call malicious compliance on the part of the ai

[dull flame]

https://tenor.com/view/whet-what-ew-can-you-not-stop-gif-9908266

[clever oar]

😄

[nimble mason]

adetailer will never be able to fix it

[wispy nest]

not with that attitude 🤣

[nimble mason]

prolly for the best i keep my mouth shut cuz then you'll be creating a lot more top notch R-T-A for us to enjoy

[wispy nest]

R-T-A?

[nimble mason]

Turn your head 90 degrees to the left and read bottom up 😄

[gusty cloak]

Wow, even with inpainting the model won't listen 💀

[wispy nest]

also for those who havevnt messed around with reference only... its pretty powerful with the correct input image and prompt. used an awesome (hand drawn) image from a very talented artist that makes amazing sketches using ballpoint pens as the reference upload. this is my mental image of him at work lol

[gusty cloak]

Too bad there's no Korra!

(inpainted the eyes here)
A digital painting of Korra from The Legend of Korra, short dark hair, (((white glowing eyes))), (((Avatar State))), huge wave of water behind her, powerful, simple colors
Negative: nsfw, child, boring, ((iris)), disfigured hands

[gusty cloak]

This

[wispy nest]

yeah im not sure if im just slow or if youre messing with me, i tilted my head all over and read it back and forward and got nothing lol

[gusty cloak]

Likely if I move the Avatar state forward or?

[grizzled sage]

could try an image prompt adapter of an avatar state korra, or reference it

[wispy nest]

are you trying to make it look like its from the avatar anime?

[gusty cloak]

Yes!

[gusty cloak]

I'll try that honestly 🤔

[grizzled sage]

specifically the avatar state. when he goes all cosmic hard af

or she in this case

i'm team aang always

[gusty cloak]

ATLA > TLOK in terms of storywriting and character development

TLOK > ATLA in terms of animation quality and maturity

[grizzled sage]

yeah they upped the animation quality in korra

[nimble mason]

art lol

[wispy nest]

try with different models, in your prompt: put at the beginning the style like: "image in the style of Avatar anime" or however you want to word it. try turning CFG up. also set it to live preview so you can see if it is getting the image right at some point, if it is and then it fucks it up you can adjust some things from there.also try setting the style drop down to anime

[wispy nest]

also try using reference only controlnet, put in an image of korra from the show as the upload. dont be afraid to play with the fidelity slider. you can also try that controlnet on img2img if you generated one you like but they style isnt right. play around with denoise, start low around 0.3 and work your way up

[grizzled sage]

i'm prompting cyberrealistic 2.5d model for the avatar state, and it's bringing back na'vi

[nimble mason]

just fyi, this is the one thing i'm aware of that comyfui doesn't have an exact equivalent for

[grizzled sage]

this is a conceptionally cool one, but fkn, na'vi

[nimble mason]

so whenever you end up messing around with comfy tomorrow etc just know not to go nuts looking for it

[grizzled sage]

avatar actually isn't strictly "anime". It's a nickelodeon cartoon

[wispy nest]

im going to absolutely go nuts trying to make it work while i make more A-T-R

[wispy nest]

ok i absolutely nailed it. hold the applause until the end of the show please.

[grizzled sage]

https://github.com/comfyanonymous/ComfyUI_experiments there is a refrence only node in this folder, but it hasn't been updated in some time

[wispy nest]

so in comfy, are nodes essentially the same thing as extensions? probably an ignorant question, but i havent dived into comfy at all yet aside from seeing it used in videos

[nimble mason]

yeah it's not done the same either

IMO the one in a1111 is better

[grizzled sage]

well.. yeah and no. extensions are typically more featured. custom nodes are more modular

[nimble mason]

kinda, yeah, it's diff but kinda

yeah, extensions are more like a bundle of nodes with things already linked together

[grizzled sage]

workflows i'd consider a closer analog to extensions

[nocturne oak]

Not exactly. Nodes are the various pieces. They get strung together to do the bits of logic. Some add-ons/extensions are nodes, but there are basic nodes included in the installation as a part of ComfyUI.

[wispy nest]

@gusty cloak in seriousness though, is it just the art style you want to change? i want to see if i can get it right just out of curiosity

[grizzled sage]

they're really different worlds, node graphs vs guis. hard to map concepts

[nocturne oak]

Comfy isn't as difficult as a lot of people think.

[nimble mason]

yep, i personally found it easier to learn

i was stuck on stuff with a1111 because more was hidden under the hood

[grizzled sage]

there's a lot of value to learning your way around a node graph system

[gusty cloak]

The Avatar State is (put very simply) where her eyes glow white

[nimble mason]

but i'm a person that really wants to tinker

so i know that's not how everyone wanst to learn it

[wispy nest]

oh thats it?? i thought you were trying to change the art style. you can easily do that with inpainting

[nimble mason]

but now when i go back to a1111 it all makes sense whereas before i was basically just pushing random buttons

[gusty cloak]

I couldn't 😂

[nocturne oak]

I argue that everyone that wants to play with making images in these tools right now should learn ComfyUI. It forces you to understand what the various settings are and how things are meant to work.

[wispy nest]

@nimble mason well shit i already got myself a custom comfy "extension" ready to go 😄 thanks again for that workflow batman

[gusty cloak]

https://tenor.com/view/avatar-korra-avatarkorra-airbending-avatarstate-gif-12544672

[nimble mason]

no prob use the most recent one i dropped on here the other one was a mess

[grizzled sage]

using that aang lora an dprompting for korra

[nimble mason]

my advice to people honestly is to not just say "i'm converting 100% to ____" but to use both a bit

[gusty cloak]

Reminds me a bit of Azula 🤔

[nimble mason]

i ended up moving almost entirely over to comfy myself

but there's a lot of value too in trying to recreate results from a1111

i def learned a lot from that

that, and so you don't get frustrated with the tedious parts while you're learning how it works

[nocturne oak]

I've gone 100% ComfyUI at this point. There's generally nothing A1111 can do that it can't and I'd rather have full control.

[wispy nest]

same thinking here. might as well add comfy to my collection. a1111, forge, foooocus (which ive barely used). what other UI's should i add to the collection lol

[nimble mason]

my other advice... get a gaming mouse. i have a G502 hero that was i think $40 with ab unch of extra buttons. i have them hotkeyed to do various stuff in comfyui - duplicate nodes, group select, duplicate with connections intact, shift drag, delete, etc

it makes things FAST

[nocturne oak]

Plus...my OCD loves organizing the noodlezzzzzz!

[grizzled sage]

i think cause its leaning her towards aang from the lora

[gusty cloak]

Do you actually have OCD? (I'm diagnosed, it's terrible and it's not just about organizing shit 🥲)

[wispy nest]

plus the ability to change dpi settings, makes fine details MUCH easier if using a mouse and doing intricate masking or touching up in some 3rd party editing software

[grizzled sage]

https://tenor.com/view/satisfying-ocd-ocd-trigger-lol-lmao-gif-21363122

[nimble mason]

man, i can't start down that path lol

i deliberately make sure things are a lil fucked up in my workflows

otherwise the perfection ism takes over and i do nothing but try to align shit

[nocturne oak]

No. And I sincerely meant no disrespect.

[gusty cloak]

No you're fine 😂 I was just wondering

[nimble mason]

and when you shift drag multiple nodes together, only one snaps to the grid!!

[nocturne oak]

I can imagine it's a serious headache at times...only meant to illustrate my love for the organization aspect of building/modifying workflows.

[gusty cloak]

💀 This is perfect for #1019361238234443776

[gusty cloak]

It's kind of a hassle 24/7, even as we speak

But I manage it

[nimble mason]

factorio player...?

[grizzled sage]

logitech make great mice. i got the mx master, but my spare is a g502. i think its the original boy. not the hero or nothing.

[nimble mason]

i've been using theirs all my life

[grizzled sage]

my life overlaps with pre optical mice so , there was a lot to hate back then

[nimble mason]

used to use the really simple trackball, that was my absolute fav, till i didn't realize the mouse sensitivity was way down and injured a ligament in my wrist cranking at it

[grizzled sage]

thing i love about my steam deck is teh touchpads in trackball mode

[nimble mason]

they discontinued them for a while, so i was glad i saved all my broken ones - i was swapping the switches in for the buttons from old right buttons into the new left ones as they crapped out for years

[nocturne oak]

Played it for a bit, but realized quickly that it's meant to pull you into a never-ending continuum like cookie clicker. So I logged out and uninstalled.

[nimble mason]

haha

it feels like a job

and is very addictive

comfyui reminds me of it, except at the end of the day, you've made incredible shit

[grizzled sage]

oh but the feels when you turn on that massive plan you been working towards, and it works beautifully

[nimble mason]

i knew one of those guys that created a fn computer in that game

and used it to display scrolling famous paintings or something

it was craaazy

[grizzled sage]

oh oh and how the trains are basically ottd trains

[nimble mason]

i never got anywhere near that deep

i did burn an entire week mapping out algorithms and using mathematica to plan out my belt sorters lol

it was bad

[grizzled sage]

i burned out doing math for kerbal for a while. then i learned mods

never really cared about math for factorio. i just gauge it and add more lines

[nocturne oak]

I'm currently refining my SDXL->SUPIR pipeline by inserting Preview Chooser nodes at the 2 stages prior to SUPIR running.

[gusty cloak]

I guess that's enough Avataring tonight 🤣

[nocturne oak]

Next up...adjust it to have Cascade be the starter for composition, then have XL over the top of that for more detailed features, then SUPIR.

[grizzled sage]

you'll have to get enough screen grabs of korra in avatar state and make a lora

i bet theres enough avatar state art out there to make a general lora/embedding for it

[unborn falcon]

i use abstract expressionism of the moon as my test when making a new setup. not too keen about there being two moons, but it's not bad.

[grizzled sage]

one looks like jack skellington

https://tenor.com/view/jack-skelington-wave-hi-nightmare-before-christmas-bye-gif-5396363

[nocturne oak]

This is great so far with the chooser nodes. Only putting a single through and not a batch, but doing it because I want to easily pause at each stage before pushing the images through the next one. If I don't like what it's done, it's easy to cancel that without having to wait for a chunk of work since it auto-pauses at the chooser.

I also hid bookmark nodes from rg3 behind each stage so it zooms to exactly what I need to see in sequence.

Bookmark 1 (zoomed out) view:

Bookmark 2 (initial creation) view:

[grizzled sage]

i know that reference!

[wispy nest]

i was having a hell of a time doing that in inpainting actually. just did a quick 1 minute edit in gimp instead. you can clean this up with img 2 img or inpaint. i think this is what you were aiming for right?

[nimble mason]

meant to include the original for comparison... this one was tile upscaled as well as swapped. margot robbie giving a speech in NK

[nocturne oak]

Bookmark 3 (working nodes for initial upscaling & SUPIR denoising):

[gusty cloak]

I like how tonight's theme became Korra 😂

[gusty cloak]

Will try 'tomorrow'!

[wispy nest]

i just latch onto random shit because i have no attention span lol

[nimble mason]

doing some manual editing is REALLY underrated.

[wispy nest]

that was all you were trying to change was the eyes right?

[gusty cloak]

Yep!

[nocturne oak]

Bookmark 4 (denoised and ready for SUPIR):

[wispy nest]

cool! hopefully you can use that image then. good luck

[gusty cloak]

With Korra it's only the eyes that glow, so it's rather easy!

Aang's eyes and tattoos glow (which is never explained, why do they glow?)

https://tenor.com/view/avatar-the-last-air-bender-gif-18465547

[grizzled sage]

they're special air bender spirit tattoos

[nocturne oak]

Bookmark 5 (compare initial image to final image):

[wispy nest]

absolutely. and im terrible at editing. for example for that one i just did, i took a avatar state image from google, pulled up sufis pic in gimp, added a layer of the avatar state eyes, erased everything except the eyes, duplicated the eye layer and deleted the right eye from the duplicate, moved the layers into place and scaled + rotated them in. only took about a minute or two and i barely know how to use gimp 😄