🏞｜general-with-images

[ripe cedar]

Embeddings and LORAs can be used to make consistent characters, but what about backgrounds? How are they made to be consistent from different angles?

I don't mind using 3d like blender, but when the render is fed into SD, it changes things up a bit

[smoky oak]

ummm

I am having a very weird problem

oh!

My seed kept randomizing cause I had a space before it apparently lol

[novel ravine]

https://civitai.com/models/18177/gigachad-diffusionlora

[sharp atlas]

Un oso en la luna

[smoky oak]

Aw damnnn, why didn't I think fo this first lmao

Version 6.9 lmao

[novel ravine]

xD

[smoky oak]

Nice haha

I am working on the ultimate avatar LoRA right now

[novel ravine]

One LoRa to rule them out

[novel ravine]

nice idea

i'll get to work right away

[smoky oak]

So far so good

[smoky oak]

As in stealing my LoRA?

[novel ravine]

jk

[smoky oak]

oh my god lmfao

yes, we need a LoRA of him lmao

sigma male face lol

[smoky oak]

I also made some more niche LoRA's

Made a LoRA for the tigers from zootopia lol

[smoky oak]

LMFAOO

thats rancid lmfao

Made a LoRA for Mr. Wolf

toothless

[novel ravine]

https://civitai.com/models/10402/thanos-15-and-21-768-lora

[smoky oak]

scar from Arcane

[smoky oak]

NOOOO

[novel ravine]

https://civitai.com/models/10271/giorno-giovanna-lora

[smoky oak]

Is that thanos and biden? lmfao

[smoky oak]

lmfao

I see you made a Johnny sins embedding lmao

I am working on a refined Ricado Millos LoRA

[novel ravine]

my shrek lora

[smoky oak]

Early Ricardo Milos is coming along well lol

[novel ravine]

https://tenor.com/view/jhonny-sins-reyiz-wink-gif-8643695

[smoky oak]

https://tenor.com/view/flick-esfand-esfandtv-ricardo-milos-ricardo-flick-gif-13730968

[novel ravine]

i am not homo

but damn

[smoky oak]

I am lmao

Ricardo is so fine lmao

[novel ravine]

https://tenor.com/view/sigma-handshak-handshake-khar-gif-26201999

[smoky oak]

Oh, i also made a LoRA for freshwater shrimp lol

[novel ravine]

https://tenor.com/view/jesus-ballin-mars-bars-gif-19910027

[novel ravine]

you what

[smoky oak]

I am a shrimp breeder lol

I made one for freshwater shrimp

it was my 2nd LoRA

No LoRA

[smoky oak]

LoRA :p

[novel ravine]

you can fuse loras you know

fuse it with a fish

[smoky oak]

I have a multi trigger LoRA

a LoRA with 2 completely different triggers in one

[novel ravine]

Rapunzel x Shrek

[smoky oak]

It was trained on two subjects at the same time

[novel ravine]

yea i have as well

[smoky oak]

They can be triggered separately

[novel ravine]

where female navi

[smoky oak]

In the works

I have made some

[novel ravine]

nsfw capable?

[smoky oak]

yes

my full Na'vi model will have 4 different data triggers

Male Forest Navi
Female Forest Navi
Male Reef Navi
Female Reef Navi

[novel ravine]

mah man

[smoky oak]

My current Na;vi LoRA is actually only trained on men

but it can still do women lol

[novel ravine]

I think me too

its not too late

[smoky oak]

I meant IT

lol

[novel ravine]

https://tenor.com/view/drake-notebook-gif-20708336

[smoky oak]

BRO

I got an idea

for a LoRA

[novel ravine]

?

[smoky oak]

I shall not say, out of fear of it being stolen lmao

[novel ravine]

well, i am on a holiday

you might well just say

sike

[smoky oak]

There is currently one for him, but its a checkpoint

[smoky oak]

LIAR

[novel ravine]

https://tenor.com/view/evil-laugh-gif-25608698

and I will make it nsfw capable

[smoky oak]

All of mine are NSFW capable lol

My zootopia tigers are NSFW capable lmao

I wonder if the freshwater shrimp are 👁️

[smoky oak]

LMFAO

I have a furry model. I wonder if I can actually get a freshwater shrimp furry

My capybara LoRA works with the furry one lmao

[novel ravine]

🤔

[smoky oak]

Also, capybara one is certified NSFW capable lol

I was not expecting them to work together, but they do lol

I am trying the freshwater shrimp lmao

LMFAOOOOOOOOOOOOO

[novel ravine]

god

[smoky oak]

looks like the freshwater shrimp one makes more of an abomination lol

@novel ravineIT TRIED

Freshwater shrimp furry lmfaooo

legit cursed lol

[novel ravine]

eating a shrimp got a whole new meaning now

guess I'm vegan now

[smoky oak]

😩

lol

[sonic vessel]

My new seek.art MEGA model will be releasing soon.

trained on top of 1.5, not a merge.

[smoky oak]

Oh! very nice

I think I just downloaded Seek.art a day or two ago lol

[novel ravine]

.

[smoky oak]

Best of luck, lord knows its needed

[wispy nest]

That's weird one , but I kinda like it lol

[smoky oak]

these AI's are very mean about giving swords lol

[wispy nest]

well, at least it's not giving me 4 swords anymore with "floating sword" negative , so that's a step forward

[smoky oak]

ah, nice haha

[wispy nest]

Almost lol

[novel ravine]

control net : I can fix you

[wispy nest]

naaaaah, I want to gen it 😄

[novel ravine]

you gotta add a rough basic sword to the controlnet pose hand in blender and voila

[wispy nest]

I like random things it does, don't want to do specific pose , it kinda removes the purpose

[novel ravine]

¯_(ツ)_/¯

[wispy nest]

One day it'll happen

[proper linden]

/imag

[novel ravine]

@smoky oak

[smoky oak]

You did it?

You made the LoRA?@novel ravine

[novel ravine]

yes

[smoky oak]

Alright, not sharing my ideas anymore here

[novel ravine]

xD

you can always improve on it haha

[smoky oak]

I'd prefer to not feel like I have to one up people/compete with them for things I thought of

So I'll just keep my ideas to myself

[novel ravine]

https://tenor.com/view/graham-allen-welcome-to-the-real-world-realworld-real-life-grow-up-gif-11828549

https://tenor.com/view/3idiots-virus-nobody-remembers-second-boman-irani-gif-25604019

[smoky oak]

Yeah, hence why I am not just gonna give out my ideas/plans to people so willing to take them

anyways

[novel ravine]

well, I still have some fine tuning to do

you have like 10 minutes

https://tenor.com/view/evil-smile-sneaky-gif-21615143

[smoky oak]

just do it yourself, I don't care anymore

[novel ravine]

oof

now I feel bad

[smoky oak]

I am not in this to compete with people. I wanna do it on my terms, how I want. So just do it yourself and I will just move onto something I can do for myself

[novel ravine]

@smoky oak there is a guy in civit, who reached top 10 leaderboard by just posting lora extractions of other people's models

[smoky oak]

Not surprising. Most people in this hobby seem to want to take for themselves rather than contribute to the collective

[novel ravine]

there was a gigachad model already on civit

I could've just extracted it

but there is always room for improvement

[dense tapir]

why create when you can just steal it and claim it for your own?

[smoky oak]

Sounds about right

[smoky oak]

I am working on improving some peoples work. Of whom I asked beforehand

[novel ravine]

I gotta disagree, in a way, we all are stealers, because none of the "art" in our datasets is truly our own

ironic

[smoky oak]

There are people working on that luckily

Not that its something that most people in this community seem to care about

[dense tapir]

civitai fantasy.ai and more surely to come

[smoky oak]

@green plover

Care to add anything to this discussion about "stealing art?"

[smoky oak]

oh, thats funky haha

I like it

[novel ravine]

hey I was just saying the obvious, not that anyone cares whats in your dataset anyway

[green plover]

The whole concept of using a neural net being steeling art shows a complete lack of understanding as to how they actually work

[smoky oak]

There are people who do care tho, which is why I, as well as many others, refuse to specify artists in our work. I only focus realism, of which is from photograph (and I as a professional photographer am open to people using my images to train whatever they want. photographs are not special IMO)

There are some people that specify specific artists styles, and I think that is shitty to do. ESPECIALLY for work you are trying to sell

[green plover]

Stable diffusion was trained on 600,000,000 images and weighs in at 7.5gb. he dataset weighs in at 10tb. If you don't notice the discrepancy, I will spell it out. It is not possible for it to be a storage and retrieval system. It's not large enough to hold the data

[smoky oak]

exactly. Its learning concepts, not copying them

[novel ravine]

my small piece opinion :
ai artist ai tool user
I created I generated

[smoky oak]

And I look forward to AI's that are trained entirely off of AI understood concepts

[smoky oak]

I can agree with that

[novel ravine]

I hate the word air artist lol

[dense tapir]

I disagree with the second

[green plover]

When I am done with it, I have created it. and my process allows me to say that

[smoky oak]

well, I do actually create my own data sets, but thats neither here nor there lol

[novel ravine]

the tool created it, I ordered the tool to do so

[smoky oak]

Hell, just doing the data set for my zootopia tigers was hours of work sourcing images, and all the other

[green plover]

The tool gave me a starter image. I finished it

[smoky oak]

Thats like saying the paint made the art, not the person instructing it.

I have a very gray area feeling about "AI art", but even I wouldn't put it that way

[novel ravine]

yea if the paintbrush moved on it's own then yeah, I just dipped the brush in the paint

https://tenor.com/view/magic-gif-26166638

[green plover]

You may rely on being solely a prompt smith, but some of us take it further

[smoky oak]

And the AI just computes on its own? without my power, instructions, finetuning, and hardware?

This feels like a nonsensical argument

The AI would not have made what it did without my input, just as a paint set

[dense tapir]

It is one that can go on forever.

The hammer built the house I only guided it. Yeah, otay.

[smoky oak]

Hell, if AI art isn't art, then how come 3D rendering, and ESPECIALLY simulations are? You set a few values and let the computer do it. Arguably less input than AI, as you can't even tweak the calculations

(I say that as somebody who has done several 3D art commissions, and has been using 3D software for 5 years now)

[dense tapir]

I think I like the second one better

[smoky oak]

I like this one. Hes got spunk lol

An attitude even

[green plover]

better use of negative space I think

[dense tapir]

my lora I can see in it

[smoky oak]

Yeah, agreed

[novel ravine]

my last piece : simply using base model and generating stuff, nah not much effort
generating a complex image by using bunch of various tools like control net, img2img, inpaint etc into something no one has seen before - yes that's something

my main issue is with people copying art styles, scrapping loads of images at once- using auto tagger, just post em
where's the effort in that?

[smoky oak]

Oh yeah, we can agree on that

As I said, I do not target specific peoples work (outside of a few edge cases where its not detrimental to said artist, like a quick reference for a DND character or such)

[green plover]

Yes we can, but to diss the entire genre and medium as a whole because of that is basically.... profiling

[dense tapir]

I can't agree with that as prompting ain't easy to do it right. I refuse to use those tools as I figure they are only there as a stop gap measure for the prompt AI when eventually you can just prompt it.

[smoky oak]

I am insanely excited by the idea of using AI that has been trained off of only AI generated images. I feel that the lines of creativity and style will be so blurred that it will be more of a random chance to happen uppon something

[green plover]

and I'm actually working on one. it's just a slow process doing it by yourself

[novel ravine]

I only like it when SD makes something new, not something that already exists

[green plover]

So make it make something new

[smoky oak]

Thats why i pinged you haha. I wanted to say you were working on one, but I was not sure if that was public knowledge or not

Hey, i created something new just the other day

[novel ravine]

there are many loras around trained on generated images

[green plover]

replacing 600 million images with synthetic data itakes quite awhile

[smoky oak]

We are talking about everything fully generated. base model and all

[green plover]

I have 2 TIs trained on synthetic data

[smoky oak]

I made this by blending information between Na'vi and Vastaya frin LoL

[novel ravine]

I made this from ugh, well...

[smoky oak]

If anything, this whole experience has shown me not to trust people with my plans. And I have you to thank for that realization @novel ravine

Its probably a wakeup call I have needed for a while

[novel ravine]

I haven't posted it yet

[smoky oak]

I bet people don't share how to make high quality LoRA's, so there is less competition to shine

[novel ravine]

just testing

okay u do it

I wont post

[green plover]

I can't post most of my work here

[smoky oak]

I know how hard it was for me to find proper documentation for how to make LoRAs

[smoky oak]

I don't care about it anymore. Have at it. The fun is gone for me

[novel ravine]

I got help from a friend of kohya, sorry for bragging

[novel ravine]

there is always room for improvement

my lora is not perfect

[smoky oak]

Its fine. We are all good at making LoRA's here. Only difference is @dense tapir @green plover and I have plans to share it with other people rather than keeping it to ourselves

I'd like to think of it as a Community Creation Coalition lol

[novel ravine]

https://civitai.com/models/8252/gal-gadotlora?commentId=24050&modal=commentThread

[smoky oak]

I am currently practicing with perfecting LoRA's, with the hope of sharing my information

[novel ravine]

this guy said he had planned on the same lora

see how we figured it out

competition is good

[dense tapir]

With my lora I am still fighting (2.1 is a struggle to train with lora). 1.5 is dead easy.

[smoky oak]

You are missing the point

[novel ravine]

2.1 has made it very hard to train faces which aren't included in the base

training other stuff is relatively easy

[smoky oak]

You actively went out of your way after I expressed concern over people taking my idea before I had time to make it.

You even said you wouldn't when I was concerned. And for that, I don't have interest in sharing any of my stuff with you

[dense tapir]

😦 Figures as my styles are shit in 2.1 but same dataset 512x512 and done.

The controlnet devs pretty much, politely, told SAI off.

[novel ravine]

you can just make it anyway, why are you getting so emotional over a lora idea?

[smoky oak]

Again, missing the point

[green plover]

It's easier with TIs

[dense tapir]

TIs are dead sexy, and I love them but loras the vibrancy of colours is just so sexy

[novel ravine]

https://civitai.com/models/15137/tesla-model-s-15-and-21-768lora

notice how the 2.1 car is flat and wide, but 1.5 looks fine

2.1 is very weird

[smoky oak]

I like LoRAs for what I do

haven't tried TI's, but I just enjoy the workflow for LoRA's. And the turnaround time

[dense tapir]

2.1 has a problem the devs stupidly didn't think about. You made the model on 512x512 then added 768x768 on the tail end. Ratios are all screwed so elongated necks, faces, chins, etc... are very common

[green plover]

batch upscale your dataset by 1.5 and copy your captions

[novel ravine]

lora of gal gadot trained in 2.1, so bad
on the same dataset as the one I trained in 1.5

1.5 for comparison

[smoky oak]

Yeah, 2.1 seems to be trash for training on

[dense tapir]

I went from 2.1 to test on 1.5 so scaled from 768 sq down to 512sq.

[green plover]

It works better if you train specifically for 2.1 768

[dense tapir]

I did not use buckets as I mistrust them

of course

[smoky oak]

I don't use buckets either, I hand crop all of my images

[novel ravine]

I asked a developer for training settings to get a good face in 2.1,
he said it's possible with different settings but was reluctant to share it

legal issues I guess

[smoky oak]

And that is what we are against

[dense tapir]

Yeah, people for 2.1 just don't share settings

The ones who do are just repeating wrong info

[smoky oak]

What is the point of having open source AI, if all you do is spit on other people in the community by not sharing general knowledge

[novel ravine]

NAI ahem...

good it got cracked

[smoky oak]

Hell, most of the people sharing info about LoRA's in general are bad

I had made about 30 of them before I found that everything I was being told was wrong

[dense tapir]

I just found that out about a discord I am on.

[novel ravine]

well, I don't really hate them, but once you figured it out yourself, oh boy that sweet feeling

[smoky oak]

AItrepreneur's video is a horrible guide to doing LoRAs. I see that now 100%

[novel ravine]

Aitreprenur cough cough

[dense tapir]

Someone kindly shoved me in the right direction after watching my pain for 3 weeks. Said I was being trolled.

[novel ravine]

I wonder why no one calls him out in the comments

[dense tapir]

for the first time my style lora actually styled

[dense tapir]

I did and was dogpiled

[smoky oak]

You, @green plover and that video I found are responsible for me making great LoRA's now

And thats just 3 sources. I probably found 50+ sources that led me astray

and I hope my sharing of that video helped you in return

[dense tapir]

Next time I called him out directly on his pinned comment and he ignored me

[novel ravine]

he monetizes on every new stuff that devs on github work hard to bring about

[dense tapir]

Many other creators for this know so much more and have little subs

[smoky oak]

I only use his content in order to find new things really

[dense tapir]

yeah, see, watch, go research on my own

[smoky oak]

But man, after seeing how bad his LoRA video is (sole reason i was lead astray for so long) I am just not down anymore

[green plover]

I am seriously thinking about actvating my twitch channel just so people can see my process

[dense tapir]

that was one I called him out on and his TI months ago.

[novel ravine]

I rather watch a 30 mins tutorial that explains properly than a 5 mins of aitreprenur

"don't worry about the settings I'm using just follow them as it is"
uses batch 8, forgets not everyone has a fucking 4090

[smoky oak]

Hell, the LoRA video I watched was 1:45, but I stuck through and learned sooo much

[dense tapir]

I am on colab so locon max BS is 2. OH OH. I found for 2.1 don't use BS > 1

the loss gets ugly and the lora might be okay or might be rubbish

[smoky oak]

I use BS 2 for all of my LoRA's

(I am strictly 1.5)

[dense tapir]

so now I am 100% repeats and 1 epoch 1 bs

[dense tapir]

As I said 2.1 is shite

why is it shite? TE which is CLIP

it fights you so you can't train

[smoky oak]

Imagine what we could do with midjourneys text encoder... Man

[dense tapir]

yet effing emad wants the community to take up the slack and train using their model as the base to build upon. ffs

hedge fund managers, pffft

[novel ravine]

illuminati v2 is good

[smoky oak]

Petition to siege midjourneys lead developer until he hands over MJ's text encoder

[dense tapir]

yes, I use it for my gens but is crap to train on unless all you do is use it

1.1 I don't like very much

1.1 does nasty wicked things to women's faces

he is working on it

[smoky oak]

My Na'vi LoRA after retraining 10+ times and spending hours on it based off of what Aitrepreneur shows in his videos

[dense tapir]

btw, you know the noise offset for lora training? 0.1-0.2 more loss in 2.1

[novel ravine]

I use 0.05

[dense tapir]

I don't use it at all as I find it doesn't do much

[smoky oak]

Vs my results 15 minutes after learning how to do proper LoRA's

[dense tapir]

even at 0.1

[novel ravine]

I notice the blacks are a bit blacker than usual

[dense tapir]

it can get too dark, yeah

[smoky oak]

so, out of curiosity, how does one train a LoRA with noise offset?

[dense tapir]

but it really isn't ambient dark it is a contrast dark which is not the same

--noise_offset=0.1

[smoky oak]

ah, is that something you can't do in Kohya?

or do you add it to the bat?

[dense tapir]

what I use

can go in the extra args

for gui it has that and colab has it just fill both of those with a value

[smoky oak]

lemmie see if we have that

[dense tapir]

for scripts it is the --

[novel ravine]

its in advanced configuration

[dense tapir]

that be it

[smoky oak]

Oh my dumb ass

[dense tapir]

0.1 I can only imagine 1.0

[smoky oak]

wait a minute

[smoky oak]

I don't have that-

Am I using an outdated version of kohya?

[dense tapir]

yes

[novel ravine]

do upgrade.ps1

[smoky oak]

oops lol

[novel ravine]

sometimes new update breaks it for me

so I use the most stable version for me

[smoky oak]

sorry, brain fart, how do I run this?

its just opening as a text file for me

[novel ravine]

right click, run with powershell

[smoky oak]

right right right, thank you

[novel ravine]

then also do a git pull to be sure, idk I always do that

[smoky oak]

so, does noise offset .1 make images darker?

If so, is there noise offset -.1?

[novel ravine]

gotta test it to find out

best way to do it

[smoky oak]

Alright, I will

[dense tapir]

zero to 1 no above or below

[smoky oak]

hmm

[novel ravine]

I dont have much powerful gpu so I stick to settings which work best for me

[smoky oak]

what GPU?

[novel ravine]

2070S

[smoky oak]

ah, fair

3060ti here

[novel ravine]

locon training is not possible for me

[smoky oak]

locon?

[dense tapir]

without and with

[novel ravine]

low convulation

new stuff in the low rank adaption models

is better than lora

[smoky oak]

oh yes, I think I heard about that

[dense tapir]

this is what I hate about lora 2.x it just doesn't do what you think it will

[smoky oak]

I wonder if I can do it

[dense tapir]

I do locon but I know loha is coming

[smoky oak]

does it require more VRAM, or does it require newer hardware?

jesus, this kohya upgrade is taking forever lmao

[dense tapir]

locon made me lose a BS on colab btw as it takes more vram

[smoky oak]

ah, more VRAM?

how much more are we talking?

welp, guess I gotta try

if this update would finish

that would be swell

[novel ravine]

I will share a secret with you guys,
that I haven't tested yet but have seen the results
it includes making a big fat 512 rank locon then resizing it to 32 dim
and this is somehow better than your regular 128 rank lora

[novel ravine]

atleast 20 gb

[smoky oak]

interesting

[smoky oak]

https://tenor.com/view/dies-cat-dead-died-gif-13827091

[dense tapir]

I had heard this as well

a lot of the anime people do it

[gloomy geyser]

what is this

[dense tapir]

resize down to 8 though

[smoky oak]

bro, does the LoRA update take forever for you guys?

[gloomy geyser]

i thought it wouldnt generate this bullshit

[smoky oak]

what in the fuck lol

[novel ravine]

try betterDAN

[dense tapir]

DAN the anti woke gpt

[novel ravine]

no it takes like 2 sec, I can't even see whats updated since the powershell window closes on its own

[dense tapir]

I loved it when dan broke chatgpt

[smoky oak]

damn, mine is still going

[dense tapir]

click in the window and just hit enter

[gloomy geyser]

"the titanic was intentionally sunk as part of an insurance scam" "the illuminati killed john lennon"

lmfao

[smoky oak]

Its still "installing build dependencies"

[dense tapir]

damn

[novel ravine]

join vc I will try how fast my updates

[smoky oak]

like, hundreds of them

[novel ravine]

https://discord.com/channels/1002292111942635562/1002292112739549197

joining ?

[smoky oak]

it just keeps doing more and more and more

I am not

[dense tapir]

let it run then

[novel ravine]

new loha support I see

[dense tapir]

I wish I had a fast card as I would gen dpm++ SDE Karras 20 steps but it takes forever for me

[smoky oak]

its really weird, cause the files are only a few KB, but they are taking like 30 seconds each, and the ETA stays on 0:00

[dense tapir]

yep, heard it was coming now to wait on colab

loha>locon>lora

[smoky oak]

I did a big sampler comparison for myself, and it never showed anything good enough to justify its insane time requirements IMO

[dense tapir]

it does for me when I click it for 2.1. Not joking it is a massive change

beyond about 20 is meh

[smoky oak]

oh wait nevermind, its not that one that takes for fucking ever

[novel ravine]

.5 gigachad

[smoky oak]

its DPM adaptive that takes soooo long

[dense tapir]

oh, I hate that one

[smoky oak]

24 seconds for a 20 sample 512x768 image

[dense tapir]

DPM fast is even more weird

[smoky oak]

DDIM did it in 4.8 seconds

[dense tapir]

fast? fast for whom?

[smoky oak]

right lol

DPM fast

5.36 seconds

[dense tapir]

if I squint it looks okay

[smoky oak]

DPM adaptive

24 seconds

[novel ravine]

dpm adaptive is bad
it ignored the shirtless and mucular prompt and also ignored the igachad lora

[dense tapir]

I think those are so slow for me because I don't have tensor cores

[smoky oak]

DDIM, 4.72 seconds

I use DDIM esclusively now

I find it plays better with slightly overfit LoRA's, and is also a little faster than Euler A

[novel ravine]

DDIM good

[smoky oak]

bro

what the fuck

is going on with my update

[novel ravine]

i see it

[smoky oak]

CPU is at 20%, disk is at 0%

[novel ravine]

the download is stuck

my download gets stuck too, whenever there is any download in cmd

sometimes they can be stuck forever

[smoky oak]

like bro, what

the RAM is moving around

[dense tapir]

my lora is a bit weird

[novel ravine]

close and do again

[smoky oak]

Sure that won't break everything?

[novel ravine]

for me if the powershell/cmd is in background it freezes

[smoky oak]

Ohhh

let me close SD then

[dense tapir]

I do not understand why it wants to make everything drip

[novel ravine]

you can always reinstall kohya, it isn't much hassle as reinstalling auto1111

[dense tapir]

without

[smoky oak]

Ok, I ran it again, it turned red and instantly disappeared-

yup

[novel ravine]

try git pull

[smoky oak]

its completely broken

[dense tapir]

git pull is how I do everything or the git desktop

[smoky oak]

how so? Sorry, I am very bad at this stuff

The error

[novel ravine]

type cmd in windows adress bar, press enter

type git pull

done

[novel ravine]

in the root folder

where kohya is install

[smoky oak]

ah, alright

my bad, forgot that lol

oh noooo

I know what happened

I renamed my PC a couple days ago

[dense tapir]

without and with

[novel ravine]

dubious owner ship

fake owner

[smoky oak]

any easy way to fix that?

[novel ravine]

bernard recommends this command to upgrade kohya

`git pull

.\venv\Scripts\activate

pip install --use-pep517 --upgrade -r requirements.txt`

[smoky oak]

I changed my PC name cause it was stuck to my old brand name

[novel ravine]

delete and reinstall

takes like 5 mins

[smoky oak]

damn

[novel ravine]

on good connection

[smoky oak]

I have good connection, luckily

rock solid fiber

[dense tapir]

some reason my locon loves people

does do well otherwise

I trained it on robots

I added nfixer

[smoky oak]

so what exactly is nfixer and the other N things?

I have seen them before

[dense tapir]

to be used with illuminati

[smoky oak]

Yeah, I know they are with illuminati, but is there any way to know what they do exactly?

[dense tapir]

I forget, lol

I just mainly use nfixer

but as with all negative prompts it can screw you over on your lora

[smoky oak]

oh, thats cute!

ummm

I don't remember this from before-

[dense tapir]

always for me

first time only

[smoky oak]

oh nevermind, its covered later

I forgot lol

[dense tapir]

There we go I can see my lora a bit here

that 1950s style

saw it in mags

[smoky oak]

@dense tapirI should select BF16, right?

[dense tapir]

yes

[smoky oak]

ah, AI dude says FP16

[dense tapir]

fuck 'em

bf16 if you got it always

just make sure out is bf16 too

(out as in save)

I wish I had bf16 but colab, nor local do I have access to it

the a tad more jitter than fp16 but far greater precision

[smoky oak]

so what does that translate to in LoRA output results?

[dense tapir]

precision is king

[smoky oak]

better results?

[dense tapir]

can go higher alpha and dims so yeah

[smoky oak]

I see

[smoky oak]

also, a LoRA with multiple trained concepts would benefit from higher DIM/Alpha, right?

[dense tapir]

don't know as I don't do them but my guess would be yes

[smoky oak]

I will probably test

I remember you saying you have a limit for how high you go with DIM and alpha, what was that again?

[dense tapir]

my lora loves pink

[novel ravine]

https://civitai.com/models/18549/harold-diffusionlora

[smoky oak]

sure seems like lol

[novel ravine]

all credit goes to @smoky oak

[dense tapir]

about 100 but with locon I can get to 128 at least

[smoky oak]

interesting

I do 128 for all my LoRA's

[dense tapir]

one freaking robot had pink

[novel ravine]

128 is tested and proven to be a perfect balance

[dense tapir]

Now this is what I am talking about

[novel ravine]

64 is great but slightly lower quality in real faces

[smoky oak]

what about above 128?

[dense tapir]

mainly used to train then extract to a smaller I think

[novel ravine]

then its not really a Low Rank adaption is it

Lo Ra

[smoky oak]

I don't think I properly understand what ranks/DIMs/alpha's are, and how they work

[novel ravine]

yes people do make 512 and 256 rank loras

[dense tapir]

negative pink and purple

[novel ravine]

see drift johnsons models on civit

[smoky oak]

do the produce higher quality results? I am just confused at what the benefits are

[dense tapir]

more room more details but after a point diminishing returns

[smoky oak]

ok, so having multiple subjects could benefit from more rank

(as in more triggers in one LoRA)

[dense tapir]

well, 128 each so 128*concepts

I would reckon

you can always shrink after

can't grow though

nfixer it took care of the colour shifting

[feral latch]

#1072220168534642768 a box

#1010934719455707218

#1080946152318443610

[dense tapir]

nfixer took care of the oddities but I still have to neg pink and purple with it or I get this

cute though

[novel ravine]

for some reason poeple keep using chillout mix with my loras

this is supposed to be scarlett johansson

[dense tapir]

sort of looks like her at 15

[novel ravine]

thats the age of girls chillout mix was trained on

but hey I can't say that

entire China will rain down on me

[dense tapir]

ty, now I know I don't want chillout mix

here is the base image

those shades (colour shifting) is from it

[novel ravine]

I have complained multiple times to civit ai staff, they ignore cuz they don;t wanna lose traffic perhaps

[dense tapir]

I refuse to put anything I make up there

If I had been at work when I first signed up I would have been in hot water with what the whole first page was.

[novel ravine]

I found some amazing stuff there, I guess it comes at a cost

[dense tapir]

As I say pornhub first page is less porn than civitai

[fiery lichen]

Chillout mix is like on the other extreme. They portray everyone as child-like. That's what I think of when I hear chillout mix, I think of child-like mix. On the other hand, the base models don't portray orientals as well nor as accurately as the custom models do. Its not as though there are no images of their celebs and influencers out there, but damn the base model just look bad portraying those folks even when you prompt beautiful face.

[dense tapir]

turned off now but default should always be SFW

[fiery lichen]

The guys from the far east are making a ton of custom loras that is fixing that problem on base 1.5 models. Those loras and TIs actually feature aesthetics much closer to their celebs/influencer culture.

[novel ravine]

doesn't mean the same thing as paedofilia

[fiery lichen]

when I first encountered the base 1.5 model, I thought the model was racist (or trained by a racist). LOL

[dense tapir]

great firewall of China doesn't help that

[fiery lichen]

1.5 base had no trouble producing beautiful caucasian, blacks, mid east, indians (to my limited and personal preference), but damn when I prompt beautiful asian faces...yikes.

[dense tapir]

I can get Japanese but hard to get Chinese

[fiery lichen]

there are lots of kpop and japan stuff to sample from. Even on base 1.5 japanese, korean, chinese, south east peninsula nation, yikes. look nothing like what they would consider beautiful

[smoky oak]

The AI models do Samoan men shockingly well

[novel ravine]

then fucking train a model on chinese women, not on teen and prepubescent girls

[dense tapir]

kpop is far from anything I care to see or play with. yMMV

[fiery lichen]

there are plenty of those already...not just the teen ones. There are plenty of 20s-late 30s ones too

Like classic actors and actresses loras from far-east.

[smoky oak]

agreed, big time

[dense tapir]

soft cp is still cp

[smoky oak]

yeah

[dense tapir]

I have nothing but hate, and disdain, for anyone who is into cp soft, or worse hard

[novel ravine]

yeah but those incels hiding in china don't like their own mature women, instead want korean dolls and what not

[smoky oak]

pretty common sentiment, I'd say lol

[dense tapir]

yep

[dense tapir]

In the USA you would be surprised as they are trying to legitmize it and call them MAPs to gain social acceptance. Not on my watch.

[smoky oak]

I have never met somebody in the US that agrees with MAP bullshit

[novel ravine]

https://tenor.com/view/gun-db-shotgun-gif-21597283

[fiery lichen]

what is MAP?

[smoky oak]

and its ESPECIALLY insane that they try to tack ontop of LGBTQ

[novel ravine]

I have a solution for MAPs

[smoky oak]

"Minor attracted person"

its bullshit

trying to say that a grown man loving a grown man is the same as a grown man loving a child

[novel ravine]

america, while calling other countries shithole for decades, has turned itself into one, ironic

[dense tapir]

they want to add maps onto that alphabet soup and the LGBTs are saying "oh, hell no".

[smoky oak]

America has been a shithole for a while lmao

[smoky oak]

Yeah lmao

The whole community went "We love everybody equally!... Except you, fuck off"

proud of that haha

[dense tapir]

We are infiltrated with degenerates and out right card carrying Marxist Communists

going to be hard to remove them now

[smoky oak]

ok, I am not going that far-

But yeah, liking kids = bad

I believe any ADULT has the right to love any ADULT they want to love

[dense tapir]

there biz

[smoky oak]

it affects everybody else none. So why not

[dense tapir]

*their

anyway this is getting into politics

[smoky oak]

yeah, true

[dense tapir]

realismengine has to have some illuminati in it based on the loras I tried with it

one thing about loras it sure adds on to the generating time

[smoky oak]

I noticed that as well

it can in some cases

[dense tapir]

without and with

[smoky oak]

I mean, it makes sense

it adds more layers to the model, so it makes sense that it takes longer

[dense tapir]

did you know auto already takes longer just for base than other programs? Yep

Used to not be nearly as bad so I am unsure what he is screwing up

a sec longer for a 3k card is 10-15s for me

768 used to be 36-41s now 1m

add in loras etc... it can get long

[wispy nest]

That's actually pretty good model tho

[dense tapir]

no thanks

I don't touch 1.5 anyway and I think it is 1.5 only

anyone know what model lexica is using?

they call it Model
Lexica Aperture v2

[wispy nest]

I think they trained it and not sharing it

[smoky oak]

I don't care if it's good or not. Nothing will justify me using something trained on minors for NSFW purposes

[wispy nest]

It's not just nsfw purposes , it's good as an actual model for humans

[smoky oak]

Regardless, you downloading it is showing that there's a market for people exploiting minors for this sort of thing

[wispy nest]

well, at least that's what I tested it on ¯_(ツ)_/¯

[wispy nest]

1.5 base model can do children , what are we even talking about?

[smoky oak]

Just because you don't use something fucked up for fucked up reasons doesn't mean it's not fucked up

[wispy nest]

I don't get how is it just 1 specific model pisses you off, while you can do it basically with any other model

I'm not sure if it's even trained on models below 18 anyway ¯_(ツ)_/¯

[smoky oak]

Whatever man, you keep using your pedo bait model if you want

[dense tapir]

I really dig that art of the last one I generated

[wispy nest]

Again...base 1.5 model can do children, EVERY model can do it.
It doesn't mean it's pedo or people using it specifically for it, that's dumb

[smoky oak]

Agreed, that's got a super cool style to it!

[wispy nest]

every model trained on 1.5 base*

[dense tapir]

That is the style I am trying to get but no matter how much I train it will not do it

[smoky oak]

Literally majority of the reviews on there feature minor looking individuals completely naked. And the entire model skews even people well into their 40s to look like they're in their early 20s or late teens

The fact that the model goes out of its way to childify Even adults is messed up

[dense tapir]

I stuck on my bettermaker embedding

[wispy nest]

Isn't that about like 90% of models?

[smoky oak]

What the fuck? No lmao

[wispy nest]

I'm pretty sure I've seen like 90% of models have some people showing off nude results using it

[smoky oak]

Nobody said there's anything wrong with posting nude results, we're saying that posting excessive amounts of nude results of women that look like they could be in middle school is disgusting

If that's such a controversial thing to say, then I don't know how to react

[green plover]

That's why when I'm working on a nude "child, childish" is always in my negative prompt. Helps to prevent that

[smoky oak]

Some people out here are literally going to extreme length to defend a model that is quite obviously trained to sexualized young / underage women

[wispy nest]

How is that's model fault tho?

[smoky oak]

Because they are the ones that fed images of young / underage looking women into it? So much so that you can't even get the model to not apply that sort of look to women?

[wispy nest]

I don't think it is trained specifically for it, it don't usually go this way, while other models might try to

[dense tapir]

what I want to know is where did they get them to train on and were any of them naked images?

[smoky oak]

It's highly likely that they were

[dense tapir]

yes

[smoky oak]

Given the results that model likes to spit out without you even asking

[green plover]

Most of the models out now are blends based on F222

[dense tapir]

why it does anatomy so well as I was just told.

[wispy nest]

yea some of them look under 18 , but again, it's not something tha't just about this one model, that's the issue...

[smoky oak]

That model forces even middle aged women to look dangerously young. Its literally biased to get all women as close to 16 looking as it can

[wispy nest]

idk , it's been some time since I tested it

[green plover]

Probably why the dataset is nowhere to be found

[smoky oak]

You can see here, it makes Scarlett Johansson look like shes under 20

hell, she looks like shes 15 in that bottom right one

maybe even younger

[green plover]

15 might be a little high

[smoky oak]

right

It makes grown ass women look like they still have a PE class lmao

[dense tapir]

13-15

[wispy nest]

What was the prompt?

if it was neutral , like "woman" , then yea, it's probably just trained on young girls

[smoky oak]

lora:scarlettJohanssonLora_v10:0.8, scarlett, 1girl, portrait, indistinct nipples, detailed clothes, light and thin gymnastics wear, (delicate illustration:1.4),(masterpiece:1.0), (best quality:1.4), (ultra highres:1.2), (photorealistic:1.4), (8k, RAW photo:1.2), (soft focus:1.4), (18yo:1.3), (sharp focus:1.4), (renaissance art:1.4), detailed beautiful face, (looking at viewer:1.5), mature female,

for the left one

lora:scarlettJohanssonLora_v10:0.66, scarlett johansson, (RAW photo, best quality), (realistic, photo-realistic:1.2), a close up portrait photo, 1woman, outdoor, (high detailed skin:1.2), black jacket, (sweater:1.2), air bangs, (gray hair:1.2)

for the right one

[wispy nest]

yea it's trained on young then

[smoky oak]

teh first one does tag 18 YO, which I will credit

but the other one says woman, not child, and she looks even younger

its almost like the left one only looks older because it says 18 year old

[green plover]

That's bad prompt engineering

[smoky oak]

and the right aimed for lower

[green plover]

it doesn't recognize the "not"

That's not speculation. It sees child in the prompt and runs with it

[smoky oak]

where is there not?

[green plover]

oh I was going off your previous message. Seeing the world through one eye right now. Only reason I haven't gone to bed is because I have to make a phone call in an hour and a half

[wispy nest]

right one is kinda proof it's going for younger ages.

Left one I can understand, since it's "girl" , some models go for younger ages with that word, but "Woman" should be older

[smoky oak]

oh my bad. I was saying that it said a woman, and not a child, so its not like they prompted her to look like a young teen

[wispy nest]

yea...

[smoky oak]

and all the others are clearly mature women. And none of them use chillout... hmmmm

[green plover]

I'm tired and my stigmatism is bothering me so my right eye is slammed shut. long blocks of text are hard to read

[smoky oak]

understandable

My Neuropathy is bothering me a bit

I keep getting ghost touches and twitches

[smoky oak]

my nervous system needs a clean install ._.

[smoky oak]

funky, I like

[dense tapir]

If only I could train to make a lora for these. Funny thing is I probably could a TI though it would not be as brilliant/vibrant

Almost spelled his name

[green plover]

Yay!!! Another error ate my database

[smoky oak]

oh nooooo

Bro, I wish I could just download it for you and send it to you

[green plover]

Downloading isn't the problem. It's the cleaning

These images are broken, these are duplicates, these aren't images at all and these need cropped

That's what my program does, but when I make a mistake adding the next step, it thinks the images that it didn't download because of the error are duplicates and breaks the database

[smoky oak]

Ahhhh, I see

[wispy nest]

I'm dissapointed now

Model I'm using rn for sword gen goes for younger ages too lol

"woman"

[smoky oak]

That one isn't nearly as bad, but yeah

[green plover]

add child, childish to your negative and see if it helps

[smoky oak]

at least she doesn't look 14

[wispy nest]

I'm fine with my gen , it's just ... interesting tendency

Chinese models are weird

[smoky oak]

oh lawd that looks delicious

[bold mist]

Not sure why I do this to myself when I'm on a diet, but SD makes some amazing deserts.

[bold mist]

I may be an idiot for it, but I think up some good desserts.

[novel ravine]

childlike mix again

they are also not using the keyword, also the reason why the asian bias is so strong

[smoky oak]

Yeah, thats what I was telling them

[novel ravine]

when I posted models withou keyword people were like ,
why no trigger word? I want to merge with other LoRa..

now that I add keyword, people don't wanna use it and give 4 star why it isn't there yet

¯_(ツ)_/¯

@smoky oak looks at this comment thread

https://civitai.com/gallery/236145?reviewId=38751&infinite=false&returnUrl=%2Fmodels%2F18549

so sick

[smoky oak]

Nice

[novel ravine]

new lora soon

[worthy umbra]

Hallo Guys, does someone know a good Vae for Realistic Imagery like the one above?

[smoky oak]

That's what I use tok

[pseudo aspen]

Using LoRAs of Barry Allen and his Flash suit I made together generates impressive results

[glossy herald]

quite nice ! out of the eyes that feel a little weird, the first one is almost perfect

[pseudo aspen]

Eyes need some tuning, but then it will be perfect 👌

[glossy herald]

I liked that series for a few seasons, this is quite fun to generate new flash outfits too

[rain gazelle]

Mornin'

[glossy herald]

https://huggingface.co/Anashel/rpg
you should check this model, quite impressive, and I've never seen a model with such a detail "prompt guide", pdf of 40 pages with prompts, comparisons on most params that exist, disclosure on bias, ...

impressive work

(the samples)

[steep cloud]

how do I use the AI art bot here?

[crystal delta]

@steep cloud

FAQ: I'm new here! How do I get started and is there a bot?

Welcome! There's currently no bot on the server to generate your images. Start by heading over to #1072220168534642768 to get yourself situated and help find the channels you are looking for! Please make sure you review our #✍🏼｜rules-and-tos and feel free to assign yourself some #👥｜roles as well! Answer any questions your may have at our #1072229020520947753. There are many ways of accessing Stable Diffusion, take a look at #1080946152318443610 to start your journey!

[sonic vessel]

New model soon!

[smoky oak]

Looks awesome!

[wispy nest]

Yea their guide is awesome, creator doing alot of work just to explain us how to work with it

[dense tapir]

All removed now.

[wispy nest]

The only issue I had is that rpg4 often makes clothes too "clean" and perfect , if it makes any sense , but maybe I just wasn't doing good prompting job when I tested , idk

well, and onother issue was - if you are trying to use words model don't know , it might give you really lame results lol
Looks like it's trained on very specific things and when you go out to try something new - you might get cartoon-y result , probably just from base 1.5

Magic related things - fireballs, fire orbs , or whatever , which kinda makes sense for rpg - it can't do , unfortunately 😦

just clothes

[fleet goblet]

I kinda wish OneAPI had better stability

getting failed images sucks lmao

[smoky oak]

Man collecting this data set is tiring

[dense tapir]

Yeah, then the captioning beats the rest out of you

[smoky oak]

For me, i am collecting a massive data set

I am at 76 images already, and I have a feeling I am gonna have closer to 200 at the end

and then I will have to pick the ones that are the best for my LoRA. Hopefully a total of 40 for this half of the data set, then I will need to wait until early April when I can get access to the new release on bluray

[grizzled sage]

is that when avatar 2 comes out?

can't wait

[smoky oak]

Yeah l. Then I will be able to make the ultimate Na'vi LoRA

[grizzled sage]

avatar 3 is gonna be lit. the fire nation attacks and changes everything

[smoky oak]

That's one way to put it lol

[stoic pivot]

Just a random image I made... well, I typed some words and the Stable Diffusion AI did the rest, but it is not bad

[smoky oak]

With how good the results are out of my beta model, I have full faith that I'll be able to make an insanely good Navi LoRA

These are just examples from my micro set beta version

My total data set is going to be 8 times bigger, with significantly more variation

[grizzled sage]

i'd really like to see some good usage of the bioluminescent word if i can offer a suggestion.

i have no idea how to make a model do that kind of stuff yet though

i may buy a bluray player just for avatar 2 bluray. i haven't had one for a few years now lol

[stoic pivot]

Can you guess one one of the tags/word in this prompt?

[smoky oak]

I have one data image that includes the bioluminescent night, even with just one, it's capable of doing good bioluminescence

Tada

[grizzled sage]

it's good to know you're testing towards that already!

[smoky oak]

This LoRA is only trained on men, but it does on fact work on women

It can also hybridize with other lauras and people trained into the model

[grizzled sage]

woah woah did anyone ask laura firsT?

j/k i'm excited to toy with this

[smoky oak]

I'll keep you updated. I'm hoping to make it the Avatar LoRA to end all others haha

Lotttsss of work

Full quality Blu-ray dataset

[grizzled sage]

My lore theory crafting for the fire tribe of na'vi is going wild. In my head, i've already got an idea of it. When I heard of way of the water first, i figured they'd be the Zora from zelda stylized and well, here we are. Then comes fire tribe? well probably mechanically inclined, living near volcanos, and forging like some sort of na'vi dwarves. My theory craft is that they obtain tech the sky people left and develop it, turning those mecha exo suits that are so awesome into na'vi sized exo suits. hell yeh. Fire tribe attaaaack

I've always found James Cameron to be very predictable. Like, in Titanic, I knew that the ship ||would split||. saw it coming way before it even came out /s

[smoky oak]

I love the little titanic anti spoiler haha

[grizzled sage]

I'm honestly probably so wrong but it's such a fun lore to theory craft around

[smoky oak]

I just find the race very attractive, and their culture very pretty

[grizzled sage]

i like parts like when the tulkun goes ||WHUUHUHHUUUURRURUHHH and free willies onto the giant catamaran.||

||and then the next hour is the ship sinking because James Cameron is predictable||

[smoky oak]

I mean, to be fair... That's how it historically went down, so I mean

The ||ship split I mean||

[grizzled sage]

LOL yeah was the joke. we found that out cause of ballard

i got stuff to go do. fun shooting the avatar shit here. good luck with your data setting

[smoky oak]

Thank you. It's gonna be hard, but I hope to have good multi angle data on males, females, and adolescent characters

[dense tapir]

I gotta break it to you but that isn't massive. Trust me, I was in the same boat and massive would be 350k+

I have a life outside this so spending 3+ months to hand caption them, and fix them, etc... is not for me

100-200 images is considered micro

to you, and I, that is massive

[smoky oak]

I didn't say it's big

I said it's time consuming, especially for a free LoRA

I wish I had a way to monetize this all, as I am struggling for money

[fleet goblet]

A bottle atop a wooden table, made of plastic.

I wish there were better methods to use Arc.

🤷‍♂️

But DirectML is at least able to generate images without either needing stupidly large prompts like OpenAPI, but it's slower.

[celest geyser]

I was making some starcraft ghost stuff... and when I switched gears I forgot to turn off controlnet. No regrets.

[wispy nest]

seems like protogen handle swords better then other models I tried so far

But it keeps making clothes leather for some reason

and results are brighter for some reason, I kinda liked dark atmosphere with other models

guess it needs completely different prompting

[static tusk]

try the epi-noiseoffset lora for 1.5

[deft bear]

#1072220168534642768

[wispy nest]

nah, I have darker model, no reason to throw lora in the mix

[ripe cedar]

Blender + SD can render images based off the viewport, or the depth. But is there something that integrates both?

Let's say I have a 3d castle with stone texture. I want to retain the structure of the castle using depth, but also want to use the visual stone information and not something that could stray into a variety of things like wood or sandcastle.

[sterile kiln]

Clip_Vision

[smoky oak]

This invoked a fight or flight response in me lol

[smoky oak]

You could train a LoRA on it. Would be really easy, assuming you have 8GB VRAM and a few good photos of jt

[ripe cedar]

I'm more like trying to find an efficient workflow for comics, where a variety of backgrounds from many angles would be required.

I have kitbashes ie lots of different buildings that can be quickly placed in Blender. But the problem is, how to make that realistic castle both stylized with consistent structure from different angles?

[smoky oak]

@ripe cedar if you have some images of the castle, I would be willing to train a LoRA of it for you really fast

[smoky oak]

Ah, not sure in that case

I specialize in specific object/subject LoRA's, not styles

[ripe cedar]

Appreciate it, but I can train. I suppose it's more like trying to find a workflow or something

Maybe the technology needs to improve

But I think multi control net is the best way to go about it so far

[smoky oak]

I'm sure it's possible, but you would have to learn how to do it. It's not something I know how to do

[smoky oak]

Yeah, that's a really good bet

[sterile kiln]

Why use LoRA to do a castle? Use COntrolNet + Clip_Vision

[sterile kiln]

Wow

[smoky oak]

They want a consistent castle with a consistent texture across different styles. Controlnet isn't gonna be able ot make multiple different styles/views of the same thing

[sterile kiln]

That's why personally, I often mix Prompt+Depth+Clip_Vision, a lot of variation

[smoky oak]

again, not gonna give you multiple different angles from one image. Which is what they want

It'll give you the same angle over and over

[sterile kiln]

Yes and no. If you use differents images

But i understand what you mean

[smoky oak]

Well yeah, if they had different images, it wouln't be a problem lol

[dense tapir]

It is funny but we can train loha as a friend did then found out there was no way to see the loha. :/

keys not found and mismatches. soon I bet though

[smoky oak]

This car looks like it came out of SD 1.4

lmao

[sterile kiln]

What did you used to do this mario?

ControlNet?

[dense tapir]

No

Just my lora and I don't use that controlnet stuff just straight prompting

[sterile kiln]

He represented the mario style of the 90s well, at the level of the face

[dense tapir]

His hands and stuff is whacked though

[sterile kiln]

It is in this kind of specific case that OpenPose is really great

[dense tapir]

I misspelled a word and it still got it right

dyslexia but still managed it

[smoky oak]

Oh wow, turns out Topaz Gigapixel is actually wayyy better at suppressing noise than Topaz DeNoise

crazy lol

[dense tapir]

Yes

[smoky oak]

who would have thought the upscaling AI is better at denoise than the denoise AI lol

and its by a lot too

DeNoise

Gigapixel (no upscale mode)

[dense tapir]

Considering they are from the same company it boggles my mind AND Gigapixel has not updated their models in a long while now

[smoky oak]

All I know is gigapixel is saving my ass with these data set images lol

[dense tapir]

yep

I used it for my stuf back in 1.5 days

I wish they would update their models though as free now beats them

[smoky oak]

what are some good free options?

[smoky oak]

specifically for denoise, rather than upscale

[dense tapir]

I used to have the BM to all of them but lost thanks to linux

[smoky oak]

I don't need more detail, just less artifacts/noise

[dense tapir]

*them BM

there are some rather good ones

[smoky oak]

The improvement is monumental

[dense tapir]

https://upscale.wiki/wiki/Model_Database#Noise

[smoky oak]

oh my goddd

[dense tapir]

A lot of good shit on there as I found the bm

[smoky oak]

a banding smootherrrr

I asked for weeks if people knew of good banding smoothers back in the day, OMG

[smoky oak]

oh my god that difference

[dense tapir]

yep

told ya there were good ones on there even for video

[smoky oak]

How does one run these?

they are .pth

[dense tapir]

well, I forgot the program I used

any esrgan can be ran in auto

[dry crow]

They go into models/esrgan folder

[smoky oak]

so they are something I could run in A1111 then?

[dense tapir]

yes

[smoky oak]

amazing

So I would drop it there, launch A1111, and I could access them from the extras tab?

[dry crow]

Yes and from Highres Fix and sd upscale script

[smoky oak]

amazing

I wonder how its gonna handle a non upscale AI

[dense tapir]

just set it to 1x

[fleet goblet]

A pencilsketch of a detective wearing a trench coat and fedora, walking down the streets of Manhattan.

[smoky oak]

ITS THERE

[fleet goblet]

Heun, 10steps, 2.1 x768, Intel Arc A770.

25 seconds.

DirectML is the only reason why it's this slow.

The A770 is possible of half this time, but the sad part is DirectML is unoptimized.

[dense tapir]

Arc770?

[fleet goblet]

Yes.

Intel Arc A770 LE.

[dense tapir]

Now if we can get AMD, or Intel to have cards we can train on

[smoky oak]

Brooo

it worked

that is INSANE

[dense tapir]

🙂

[smoky oak]

Original

[fleet goblet]

If ANYONE ELSE is using Intel Arc, please direct me to a better way to inference images.

Thank you.

[smoky oak]

New

I cannot believe that

that is insane