🏞｜general-with-images

[boreal falcon]

for the sake of the workflow

[smoky oak]

deepfloyd is just overbloated with parameters, but SD 3.0 will be way more reinforced high quality and lesser parameters, based off what the devs said

[hearty karma]

tbh i might as well find myself an m2 while im on this site since i need a new storage space for my os anyway

[boreal falcon]

an A100 runs games like sh*t

[smoky oak]

yeah, my friend has one lol

[boreal falcon]

those high numbers of VRAM etc. mean nothing for gaming with those cards

[smoky oak]

he has two GPU's, one for gaming, one for AI

[boreal falcon]

because they arent made for games

[boreal falcon]

he has one thing in the first place: Money

or he took loan from bank

which is also an option

[smoky oak]

nope, got it for free lol

[hearty karma]

or he stole the money from a bank :-)

[boreal falcon]

i dont know if i want to take a bank loan for that

[smoky oak]

He might be getting a H100 soon as well, but it doesn't seem to be very beneficial to AI right now strangely

But he got his 3090ti and A100 for free

Being the head of one of Unity's AI development teams has its perks lol

[hearty karma]

yo what..

[smoky oak]

He and I made a tool that unity bought from us, and they took him to be part of their team. I didn't know enough code to be taken

[smoky oak]

It was called Barium AI, and its now one of the core parts of their next level AI suite they are working on (of which he is overseeing)

[hearty karma]

damn lucky dude

[smoky oak]

Yeah, he got his own department and stuff, and hes mentoring me to be hired under him as a machine learning engineer

[oak osprey]

openjourney-v4 is interesting, i don't think it is useful to put in as a CTU base model

[hearty karma]

i love having to upscale my images by exactly 1.334 for high res fix :-)

[smoky oak]

1.33333333333333...

[hearty karma]

no not 1.3 recurring

just 1.334

because im upscaling to 1024

[oak osprey]

CTU

[bold cave]

bru be careful

this shit can become an addiction

[hearty karma]

what if he wants to be addicted?

[bold cave]

loool. then welcome to the club

dog this is lowkey awesome

you could generate 8bit splash screens for a game with a retro theme

you should try 16-bit next

[bold cave]

noone said it was for gaming m8

https://www.youtube.com/watch?v=zBAxiQi2nPc this video pretty much sums it up

it doesnt even support directX so you physically cannot use it for gaming

[hearty karma]

the fastest gpu that sucks ass for anything other than server or ai related stuff

[bold cave]

hey, by any chance does anyone know how much you can mine with a 3090 per hour in terms of USD

[hearty karma]

🤨

[bold cave]

in ethereum

once all the conversions are made

so maybe 50c per day on a 3090?

because I have a few business ideas for setting up a mining rig somewhere where electricity is cheap, like in third world countries

my life goal is to never have to work again

[hearty karma]

👀

[oak osprey]

@bold cave bam, CTU'd

[hearty karma]

what have i done to spiderman...

[oak osprey]

https://giphy.com/gifs/nounish-dao-nouns-goldypix-VqPheGMC8WLrkahF8Z

[bold cave]

I thought I was having a stroke there for a second

[bold cave]

tyrone biggums

[oak osprey]

i haven't tried the GPT3.5 prompter til now on the fixed pipeline

[hearty karma]

it looks like it gave spider-man toes

[oak osprey]

clip clop horsey

oh wait that's a T-Rex's feet

[hearty karma]

ok.. now this is where i want it to go

[oak osprey]

[hearty karma]

perfect 😈

[oak osprey]

dats tony stonks

[hearty karma]

yes but actually no

[oak osprey]

man even good images become better with CTU

[oak osprey]

@bold cave have you plugged DF into CTU yet?

[hearty karma]

deep floyd into ctu?

[oak osprey]

using that instead of the stage 3 upscaler

[wispy nest]

whats CTU im behind lol

[oak osprey]

controlnet tile upscaler

the two images i posted above are before/after

[hearty karma]

actually it should be CTR

if you think about it

[oak osprey]

dont think too hard lol

[hearty karma]

wym dont think too hard

if i want to think too hard i will

[bold cave]

its an upscaler that uses resampling as its MO

also whats DF

deez fnutz

?

[hearty karma]

deep floyd

[bold cave]

o

Im behind on that apparently

oh its for text

eh not interested although would be nice to learn

[hearty karma]

i love using high res fix :D

[bold cave]

can someone help me settle whether or not its better to do base gen > CTU or if it's better to do base gen > hires fix > CTU

im too lazy to try it and im about to finish my first dreambooth on 1.7

[hearty karma]

id say base gen + high res fix

and then do the CTU

[bold cave]

interesting

[hearty karma]

because highres fix seems to do a world of good for gens

[oak osprey]

and photorealism

it uses cascading diffusion

[oak osprey]

@bold cave upscale dis

dats DF

so cute

[hearty karma]

i doubt that its even going to work

[bold cave]

lmfao

u can

trust me I have a 150x150 that I want to upscale

[hearty karma]

really? huh..

[bold cave]

the issue is that the smaller the starting image, the less faithful the upscale will be to the original

[hearty karma]

i am trying to make samurai spider-man now :)

[oak osprey]

hmm deepfloyd's stage 2 seems like magic

[oak osprey]

there we go. this is CTU's version of upscaling the DeepFloyd bear.

[bold cave]

is it img2img

[oak osprey]

yes

[bold cave]

I really dont want to have to learn something new. I cant keep up with all these new tools

[acoustic moth]

any new best photorealistic models since 2 days ago? ai moves too fast

[bold cave]

ultra

called zegovya or zegoya or some shit like that

[acoustic moth]

i just got edge or real and cyber

[bold cave]

nah

zegoya or go home

[acoustic moth]

is that the 16gb one?

[bold cave]

jus my opinion

yh

[hearty karma]

is it on civit?

[bold cave]

yh m8

[oak osprey]

cranking it up lmao

[acoustic moth]

that one guy who ran all those tests said it was that good, you found a noticeable difference over edge and cyber?

[bold cave]

It was better than my own model which was already pretty damn good

[hearty karma]

its zovya

[bold cave]

sytan tested edge and cyber and I tested ultra, ultra is unbeatable atm

[acoustic moth]

i mean i do have the space, might as well try

[hearty karma]

https://civitai.com/models/57319/a-zovya-photoreal

if you mean this thing sen

[bold cave]

yh if its 16gb then its the one

[hearty karma]

yep its 16gb

16.3 💀

i aint even got 7gb free

i really need to clean my hard drive

[bold cave]

do my method

2x SSD m.2 x 2TB each

and 1x HDD 4tb

6tb I mean

that's 10tb space total for the best bang for ur buck

I moved 118gb of old models to my HDD

[oak osprey]

using your method of passing the same negative/positive prompt embeds from deepfloyd to CTU 👀

man this uses way less VRAM than the default DF setup

[bold cave]

brooo that is dope

[oak osprey]

and it's just a simple python script

not even tangled up in discord code

lemme run the DF version to compare against

[bold cave]

i still dont understand what makes DF better than good ol sd

[oak osprey]

the amount of money spent on it probably

aw man am i being cynical again

[bold cave]

nah. im just asking from a technical perspective

I havent tried it yet

[hearty karma]

i mean the fact its got really good upscalers ig

and the text thing

[oak osprey]

it has a text embedder called T5 which on its own is essentially a LLM comparable to GPT3

you can add a 'head' to T5 and use it that way, but DeepFloyd is apparently also a possible 'head' to use it with

this allows more easily tuning the prompt engine eg. to add instruction based prompting "make this ball red" rather than "red ball (red) ((red))"

[bold cave]

ohh I see

[cursive sun]

hi guys
i have problem
why my generates goes this?
in realtime render all looking good but in final...

[oak osprey]

ghosts

[bold cave]

pass the prompt data

[hearty karma]

what are your settings?

[cursive sun]

you didn't get one

[bold cave]

I fink u need to add -half or --no-half-tile

something like that

[cursive sun]

ah

okay

i will test it

[bold cave]

or --no-half-vae. I saw that problem a long while ago

[hearty karma]

ok well i just moved my sd 1.5 and 2.1 to my 2tb hard drive

now i have the space for that model

[oak osprey]

@bold cave i just realised CTU nuked the deepfloyd watermark ahaha

[bold cave]

thats the same principle behind patching images

if theres a spot missing for example, u color it in photoshop or even paint and run it thru img2img and tadah. it makes it look real

[hearty karma]

i just realized i have 14gb worth of recordings ._.

atleast that makes more space for the model

anyway sen how good is that model?

[bold cave]

so good that my new senblend is only 20% senblend 80% ultra ):

my model cant compete

[oak osprey]

https://gist.github.com/bghira/5cccb5220ec101fd48c4dd9329754945

here you go, if you'd like to try the DeepFloyd-to-CTU pipeline, this one will do both. you can set deepfloyd_stage2 = False to use CTU, etc

you'll have to do the login() jazz if you've never done that before - had to fix a line.

[cursive sun]

no-half calls to cpu render?

cuz i have runtime error

RuntimeError: [enforce fail at ..\c10\core\impl\alloc_cpu.cpp:72] data. DefaultCPUAllocator: not enough memory: you tried to allocate 58982400 bytes.

[bold cave]

no thats a ram error

have u already modified ur pagefile size

I cant guarantee this is the fix tbh. im just guessing

u can also ask #🤝｜tech-support

[cursive sun]

i have did nothing

i'm just add --no-half in start file

[hearty karma]

what kind of gpu you got?

[cursive sun]

hm

4x 3060

[hearty karma]

💀

[cursive sun]

but now work only one 3060

[hearty karma]

bloody hell

[oak osprey]

@bold cave so CTU is a superior upscaler to DF's stage 2/3

unless the bear was intended to go through battle, idk

[hearty karma]

ngl spider-man with like a wig on his head is a horrifying thought

[oak osprey]

here's deepfloyd stage 3 and CTU side by side

i imagine you could play with the noise strength to improve either side's result

time for the caustics prompt

[hearty karma]

💀 not caustics

i hate caustics

[oak osprey]

remember that part in ace ventura, when the helicopter pilot guy is like "I wouldn't do that if you were you!" and Ace is like "well if you were me, i'd be you, and i'd use YOUR body to get to the top!" ?

[hearty karma]

woah..

i think the logo is a bit buggered up

[bold cave]

does anyone know if this is outdated or makes a difference in db

[hearty karma]

no clue ¯_(ツ)_/¯

never properly tried training

[bold cave]

yh its probably outdated shit

[hearty karma]

well.. i tried to train a lora but it didnt work

the image directory wouldnt work 😭

[bold cave]

use a1111 broh

or even better do db

[hearty karma]

i was using a1111

[bold cave]

loras are overrated af

[hearty karma]

can i use db on low vram?

[bold cave]

how much vram u got

[hearty karma]

6gb

[dense tapir]

10 min 12 better 16 is great and Joe Penna DB 24

[wispy nest]

Yeah it is

I ran some batches and forgot to uncheck resize

[bold cave]

too low sorry

[hearty karma]

yeah which is why i wanted to try making a lora...

[dense tapir]

even the 3060ti is too low

[bold cave]

model doesnt matter, vram my friend

[wispy nest]

What about the p40 or m40

[bold cave]

thts all that matters in the world of ai. vram vram vram

[hearty karma]

loras use less vram

[bold cave]

and thus are less powerful than db

[dense tapir]

yep, and the 3060ti is 8gb vs the 3060 of 12

12 can do DB training while the 8 is doa

[bold cave]

this model im training has kyrsten sinema in it, and ima challenge the so called lora master sytan to do one better

[acoustic moth]

the 16gb model seems terrible so far, but its probably a prompting style issue

[bold cave]

ima guarantee you, with this model and an expert dreamboother, he's not gonna beat my sinema

[bold cave]

yes it is

[hearty karma]

this thing takes ages

yeah thats likely going to be on you not the model

[wispy nest]

I really need to just buy a 4090 24gb

[bold cave]

this is what the model produced

[wispy nest]

T swift

[bold cave]

ye, upscaled, in detail, and thats without using a finetuned model

[hearty karma]

tbh i kinda just hope it can do good photorealistic spider-man characters

:-)

[bold cave]

wait until I train my 325 image dreambooth model

I will become the master prompter

[wispy nest]

I just want to do portraits of myself smh

[bold cave]

I already do

[hearty karma]

🤨

[bold cave]

sadly they will give you insecurity issues bc of unrealistic beauty standards

[wispy nest]

Do you look like tailor swift

[bold cave]

nope

[wispy nest]

I hair a thin beard and the model keeps giving me a butt chin

[wispy nest]

But I want to see smh

[hearty karma]

i kinda just realized its probably going to be a looong while before the model even ends up loading

[bold cave]

nah only if u got a puny processor

for me it takes like 10 sec

[hearty karma]

bloody 16gb 💀

[acoustic moth]

around 10s for me also

[wispy nest]

Spotted the European

[acoustic moth]

but the results were terrible lol, id need to figure out the prompting for it

[hearty karma]

yeah well my sd install is on a crappy old hard drive

[bold cave]

did u also download the vae

u gotta make sure ur using the right vae too

[hearty karma]

you need to download a vae too?

[bold cave]

yh m8. its in the description

I just googled it and got it from huggingface

[wispy nest]

https://www.reddit.com/r/StableDiffusion/comments/13bcj09/what_the_hell_is_even_that

Lmao

[acoustic moth]

its not in the description, but ill try with the vae

i need to see what a vae even is

[hearty karma]

i dont see anything about a vae in the description 💀

[acoustic moth]

yea its not there

[bold cave]

vae-ft-mse-840000-ema-prune

pruned

jus google that joint n put it in the same folder as ultra

[acoustic moth]

ok i see that

ty

right in models folder?

[hearty karma]

isnt there meant to be a vae thing next to the model thing?

[bold cave]

nah

as long as its in the same folder as the model it works

i fink

[acoustic moth]

anything to enable it in the ui or once its in the folder it works its magic nnnnnnnnnnnb

[oak osprey]

LMFAO

[hearty karma]

ew

[oak osprey]

i wanted the text pickle party in a movie poster

friggen ghost miley appears

[bold cave]

yo thts legit dope

if miley didnt look like a wax figure

[oak osprey]

you guise like candy?

[hearty karma]

looks like jawbreakers

[oak osprey]

looks like a wiiild plate

[hearty karma]

5gb away from finishing the download

id be pissed if my internet went buggered rn

[oak osprey]

right, yea i mean hollywood people do photoshop themselves beyond recognition

[hearty karma]

just used that VAE on digital diffusion

and it completely changed how my prompt worked

[oak osprey]

yea

[bold cave]

for better?

[oak osprey]

it should do way better stuff

[hearty karma]

yeah

[oak osprey]

@bold cave sorry babe i didn't know you'd like it or i'd have CTU'd it in advance lol

[bold cave]

oh god

[hearty karma]

ew...

[bold cave]

i mean its a better looking person but looks nothing like miley

[oak osprey]

she looks like the lady from Weeds

boob doctor lady

[oak osprey]

i didn't actually put miley into the prompt

that's just who it looks like to me 😄

[hearty karma]

that image and this have the same prompt

lemme try with the same seed

[oak osprey]

holy shit snoop, way better

[bold cave]

just finished filling out the concepts list for db lets hope this works

last time I used a concepts list it was bugged

please db gods just let me have this one

you also need to clone your reg image folder which is 2000 images each lmfao

just because of a db bug

[bold cave]

looking good so far. 16000 reg images

[dense tapir]

Why so many reg images?

unless you have several thousand images for training

[oak osprey]

awww crap it went poorly this time, don't look sen

i just think it's really interesting what happened to his mouth/face and how it hurts to look at

[bold cave]

its 2000 images, but I have to clone the folder for each subject due to an a1111 db bug

[bold cave]

at least he looks realistic

[dense tapir]

Ahhh

the old rule of thumb was 10:1

[bold cave]

I do 20:1, and I was doing about 15:1 prior, Im confident they'll work fine

[dense tapir]

8:1 with that many images should be fine.

[bold cave]

yaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaas it works

ok im gonna have to close discord to not risk having this shit crash

gonna take about 3 hours, will post results later. cya guys

[dense tapir]

Use teeth as a neg and you will see less of this but will still see teeth

cya

[bold cave]

3 hrs for 8 subjects is pretty good tho

[dense tapir]

You know auto done fucked up

I couldn't get prompt editing to work then remembered the setting I had to turn off which means now all my () no longer works.

prompt editing, and dephasis use the same bracketing scheme [] which is bad

soon as I unchecked it my prompt editing works again

[oak osprey]

switch one to {}s

[dense tapir]

I tried that but no dice.

[oak osprey]

i mean in the sauce code

[dense tapir]

That is why I went back to his wiki to make sure nothing had changed

[smoky oak]

Cyber and edge of real are 98% as good as Zegoya, but Zegoya does win sometimes

[oak osprey]

there we gooo

[kind quartz]

reminds me my picture from dailies. Why you dont take a part?

[smoky oak]

MMMMM I am proud of this new demo

Making me feel all Synesthetic and crap

[oak osprey]

thanks

[smoky oak]

hate

big

hate

[oak osprey]

yeah i wish i could unsee that

[kind quartz]

😦

[oak osprey]

pwned 😄

[kind quartz]

🙂

[wispy nest]

Kinda looks like a kiwi

[oak osprey]

me: "what does this number change?"
SD: "this is what hell looks like"

[crimson dawn]

i have 64gb of system memory and 16gb of vram. will the 64 benefit me for any kind of ml stuff?

[bold cave]

no

[oak osprey]

for carrying a large model with sequential model offload

it will

[bold cave]

anything more than 32gb of ram is a waste of money although that's just my (generally agreed on) opinion.

unless ur doing what shrek is doing

which is rare

[oak osprey]

less rare now

[bold cave]

my training is taking a whole 22gb of vram. hoooly fook

I can see why so few ppl like dreambooth now

also, the area where my pc is feels warm af

it just heated up its entire surrounding lool

now I know why my cat and dog like to lay down next to the pc.

[oak osprey]

its because you never feed them

[bold cave]

that was true until I splurged on an autofeeder

so they will let me sleep in peace :kek:

[oak osprey]

im scared, i dont know if i like this anymore high-guidance CTU is weird

[acoustic moth]

with my current prompts, edge seems to work best, training a model on it now to compare it with the rv2.0 model with the same dataset

anyone use xformers when training in dreambooth? I've installed it a couple ways according to different guides but it never shows up as an option for me in the memory attention setting

[oak osprey]

man, MSE is just so good on its own. this has had zero CTU applied

[calm canopy]

wtf is this O_O

[oak osprey]

progress bars

[calm canopy]

ive been "generate forever"ing for about an hour

it randomly downlaods a json file

i didnt change anything

[oak osprey]

well it will go on forever

it does have an overall progress bar at the very bottom fwiw

[calm canopy]

no, my point is the json that it just randomly downloads

that shouldnt have occured to my knowledge

[oak osprey]

no idea

[calm canopy]

oh i take it back, i changed themes about an hour ago and completely forgot about it hahah. this was just the theme being downloaded. whewhh all clear.

[oak osprey]

what you guys think they're doing for their new model for this?

it has really good text whatever it is. seems like finetuned DF maybe

[calm canopy]

ive been wondering why tile wasnt working as great for me as everyone else
turns out there was newer updated version that came out after the initial 1.1 models were released. i still have the old one 😅

[bold cave]

reddit SD threads are nothing but CTU praise now

this thing really took SD by storm

ughh my models ended up a tad undertrained. this is the plight of training. every model trains differently. looks like I wont see the good shit until tomorrow

you will see the massive difference tomorrow. I fink instead of increasing the learning rate I'm just gonna let it do twice as many steps

[sterile temple]

what is CTU?

[bold cave]

controlnet tile upscaler

[sterile temple]

oh okay, is that an extension for A111?

[bold cave]

yeh

[grand egret]

Manga girl

[bold cave]

I just thought of something genius. What if you deliberately overtrain your model so that when you generate at higher resolutions, it doesn't deform it as much yet you get much better base definition?

Ima try that shit o:

has anyone tried this extension?

[oak osprey]

original, 0.3, 1.0

it really wants to make the groundskeepers ashamed of their lawns

[sterile temple]

sigh why do I keep forgetting to check the enable box on controlnet

[oak osprey]

i think the original pic works just fine for basic needs, the .3 is probably the most appealing detail-wise and the 1.0 almost looks fucked up / dirty but it looks very HDR

the one above it is 1.0 and it somewhat looks like it fixed a major terrain issue behind the upper pup's head. it's still there, but more sensibly

@bold cave

[bold cave]

r0fl

the 1.0 turned the grass incredibly aggressive

it hurts to look at it

the second one is just perfect tho. excellent balance

[oak osprey]

agreed but if you're trying to shame the local town council you use 1.0

"Papa, where's Riley? And why do these other dogs look and smell funny?"

the dreamlike models are resistive to making sensible and coherent results. it's fun

[half veldt]

genda

[oak osprey]

this lmao "golden grizzly retriever"

[lime lotus]

I'm trying to make some desktop backgrounds, anyone got any checkpoints that are for a more "fantasy stylistic" feel? I'm currently using rev animated but Im open to anything

[oak osprey]

@bold cave well it's like discovering AI image gen again in a way. i am going to be connecting DeepFloyd up to CTU in a more solidified pipeline in my bot later this week. it's going to be tricky because it'll make me implement some other architectural stuff that'll make it easier to eg. unload models that are no longer needed, when a new one is going to be loaded. currently it's super easy to do that for my image gen models but i need to do that for TTS and LLMs now too

@lime lotus isometric dreams maybe

same model

prompts:

• space llamas
• space-bound traveller wandering through mossy PNW forest, looking up at spaceship looming overhead, misty foggy

[bold cave]

can you run deepfloyd locally

cuz if u cant its garbage

[oak osprey]

yes

igave you my script for it earlier

https://gist.github.com/bghira/5cccb5220ec101fd48c4dd9329754945

it works on Apple Silicon now thanks to someone who reported the required changes

used about 24GB of VRAM on MPS

good thing he has unified memory architecture and a 32gb laptop

this isn't a script you use in a1111 btw it's just a terminal script and you can set it up to go through CTU or DeepFloyd's later stages, it might use less vmem to do it through CTU.

[bold cave]

I made my SD icon using SD... it essentially created itself.

there's something so sci-fi about that I can't describe the feeling loool

[bold cave]

so no gui yet

eh... I'll put it on my to-do list

by the time I get to it tho, something better will have probably come out. Im just not a big fan of sending shit thru a cmd prompt

I swear torch2.0 somehow made my gens better...

I also need to stop simping swift and find a new muse

I just think this jumbo-tittied hoe is way overhyped

[kind quartz]

aliens are staring at jumbo.

[kind quartz]

been trying James Bond's main enemy cat but cant get close with necklace and fur.

[smoky oak]

Yup, civit is down again

shocker

[sinful mirage]

face is super cooked but i like how the hat says HD instead of FD lol happy accident

[smoky oak]

jesus christ that is overprocessed lmao

[sinful mirage]

ye lol

[smoky oak]

its giving extreme HDR lmao

these make me feel physically violent lmao

[sinful mirage]

gotta mess with the controlnet tile/upscale loop thing

[smoky oak]

ah yeah, i already figured out how to stop it from doing that :p

[sinful mirage]

ye those are crisp lol

mind sharing 🤔 ig trial and error is fun lol trying it with just clothing rn, gonna start easy

[smoky oak]

I am still optimizing its use, so for now I am gonna keep it to myself

but I do have goals to share

[digital hornet]

what specifically do you usually change to make it seem less overprocessed

I only played with it a bit so far but in retrospect it all got pretty overprocessed by the 3rd upscale

[hearty karma]

@bold cave thank you for this model

it is very damn good

i didnt even have to use the face restoration thing

[bold cave]

thts a nice spideygirl

anyone here who can explain noise offset?

[tight mural]

what model did you use to generate this

[bold cave]

my own custom model which is based on Zegoya Ultra

[hearty karma]

Zovya

[bold cave]

^

[hearty karma]

i think the model has been reduced in size

[bold cave]

yeh, but I dont bother since bigger is always better

[hearty karma]

woah

spider pimp

[sterile temple]

spidey got money now

[bold cave]

@smoky oak should I add positive offset noise or negative if my gens are looking too shiny

[smoky oak]

peeng

I am not sure, I have not had that issue

[bold cave]

this new option got me stumped

[bold cave]

U were talking about noise offset earlier and all the explanations out there suck ass

[hearty karma]

the spider logo is a bit lopsided

but the face is amazing

[smoky oak]

I never talked about noise offset?

[smoky oak]

oh, several days ago lmao

[bold cave]

yeh I hadnt seen this new feature until I updated torch

[smoky oak]

I would assume Noise offset for a model and for a CN are two different things

[bold cave]

my only question is, more offset = brighter?

or more offest = darker?

[smoky oak]

I am not sure how to use it or train it, I just know how to generate with it

[tight mural]

mind telling me how you trained ur own model?

[smoky oak]

I have no idea, cause I have models that can go pure white or black

[bold cave]

I merged checkpoints. I merged 80% Zovya with 20% Senblend

[hearty karma]

damn the full fp16 zovya is only 5.2 gb now

💀

[bold cave]

shit. I hate having to do things via trial and error.

[smoky oak]

[bold cave]

alrigt... fingers crossed this doesnt ruin the model...

[smoky oak]

I am sold

[bold cave]

nah G are u sure its the same model

[hearty karma]

yes its the same

[smoky oak]

but for now, I really need to sleep

[bold cave]

as longs as its the ultra

[hearty karma]

https://civitai.com/models/57319/a-zovya-photoreal

[smoky oak]

oh, the website works now

[hearty karma]

or i think it is

[smoky oak]

thats non ultra

[hearty karma]

oh i didnt see that

[smoky oak]

I will test the difference tomorrow

[hearty karma]

oh...

[smoky oak]

See if its worth 3x the size

[hearty karma]

its still 16gb 💀

[bold cave]

TOLD YOU

[dry crow]

Offset noise is to get darker images and more Contrast

[hearty karma]

anyways how do i make it do the ultra thing?

[bold cave]

so positive offset = darker images and viceversa?

cuz the issue im having is that at cfg7 they already look overcooked and oversaturated

[smoky oak]

but it also works in the opposite direction, so I am not sure how people make models that can go pure white or pure black

[bold cave]

I need to go down to cfg4 for them to look normal

[smoky oak]

sounds like an overdone VAE

[bold cave]

its the same vae that im using for zegova

[dry crow]

Yes the over8comer has also a lora for making image brighter

[smoky oak]

No I mean like

Noise offset models can go pure white or black

I am not sure how you train that

[hearty karma]

you really forget the name of that model quick huh sen?

[bold cave]

lmaoo let's just call it the zoidberg model

[smoky oak]

the malldall

the mod-doll

[dry crow]

Here is the blog from the guys who made the offset noise:
https://crosslabs.org/blog/diffusion-with-offset-noise

[bold cave]

I saw it, but its all stuff unrelated to dreambooth

[smoky oak]

I will have to read sometime, cause I am not sure how to use that slider to go brighter AND darker, like in models I use

[bold cave]

this is why im such a square and never update, more stuff for me to learn with 0 documentation to back it up

[smoky oak]

cause it seems like negative would be darker, and positive would be brighter but like... how are there models that do both?

[hearty karma]

what the heck is the noise offset thing??

i dont have one of those

[bold cave]

Its for dreambooth training

[hearty karma]

anyway another gen

[smoky oak]

Noise offset is a huge deal, it really makes a difference

[hearty karma]

ah right

that thing i cant use

[bold cave]

yeh thats what I picked up from last time so I decided not to ignore it

[smoky oak]

it allows you to go way darker or brighter with generations

[bold cave]

im gonna train at 0.1 and see what happens

[smoky oak]

Its model dependent, you need a noise offset model

I have a custom mix noise offset model, and man is it magical

[bold cave]

hopefully no more pictures taken at the surface of the sun

[smoky oak]

if we can get a proper full quality Zozya photoreal with noise offset, then I would go as far as to say thats as good as it gets

[bold cave]

should I also mess with clip skip 😮

[smoky oak]

0 clue

[bold cave]

If I set it to 2, it will instead throw the second to last image as the result, right?

[smoky oak]

Have you guys ever actually seen just how amazing noise offset is?

[bold cave]

it might also help with overcooked images?

ive seen examples

but if it works in my model you will see it first hand 😏

[hearty karma]

I LOVE ZOVYA

[smoky oak]

I don't have much faith, as I have never met somebody that knows how to train noise offset

[hearty karma]

https://tenor.com/view/spiderman-woo-tibet-tobeymaguire-gif-21044414

[bold cave]

yea looks like I have a ton of experimenting to do

[smoky oak]

yeah, but the problem with models trained that way is they will all be darker

[bold cave]

gonna leave this training overnight and see what happens

[smoky oak]

I have used models trained at 0.15, and they are all naturally darker images

[bold cave]

I hope so. at cfg7 they all look blindingly bright

[smoky oak]

which is not how its supposed to work

[bold cave]

aight ima do 0.15

[smoky oak]

you're probably gonna end up with half assed noise offset unfortunately

[hearty karma]

zovya is making me extremely damn good gens

[bold cave]

its what we've been telling u broh

[hearty karma]

i had to load the model overnight 💀

[smoky oak]

I am still using Cyber and edge of real over Zovya ATM

[hearty karma]

and then i learnt this morning that i needed the yaml file

[smoky oak]

I wish we had proper documentation on good noise offset

[bold cave]

agreed^

well you have me.

im about to be the guinea pig

[smoky oak]

cause noise offset .15 is just gonna make all of the images a little darker, not actual noise offset

[bold cave]

ima do 0.15 to see what the difference is

[smoky oak]

I have a feeling its a 3 step model merge process

main model, then merged with a second model with high noise offset (like .3) and images that are all very dark in the data set, tagged appropriately, then the same at -.3 with all bright images

then merge at 50% main, 25% dark and 25% light

[bold cave]

any idea or is this chinese to you all too

[smoky oak]

0 clue

[bold cave]

I wish I knew what every single contraption did

[smoky oak]

I really want a properly trained noise offset realism model

that will be end game for 1.5

[bold cave]

ur about to get one 😏

[bold cave]

its based on zoyva

[smoky oak]

noise offset .15 will not work

[bold cave]

thats the best spidey chick ive ever seen

[smoky oak]

I keep saying that

[bold cave]

so then what amount

[smoky oak]

its not a one number solution

[bold cave]

but its a start right

[smoky oak]

either you have nutral images, dark images, or bright images with that slider

[bold cave]

aight ima do 0.1

[smoky oak]

proper noise offset can do all 3. I have a model that can do all 3

and the difference is monumental

without

with

[bold cave]

oh man

I certainly dont want them so dark

[smoky oak]

I wanted a dark and ominous rabbit at night

[bold cave]

ima do 0.1 and see how that goes

[smoky oak]

thats the thing, with .15, they will all be darker, but the model I use only gets dark when using words like "underexposed" and "dark" which makes me think its 3 models in 1, at different noise offsets

cause it does this too

without

with

[hearty karma]

i want noise offset now 🥹

[bold cave]

chill snoop

I have my previous model without

[hearty karma]

i want evil ominous bunnies

[smoky oak]

It is actually incredible

[bold cave]

and tomorrow we will see what it looks like with

0.1

[smoky oak]

night city without

with

[hearty karma]

aw man that would make things way better

[bold cave]

if my images are gonna be looking like that, ima have a heart attack

[hearty karma]

i was doing night time car things earlier

[bold cave]

especially since I suspect torch 2.0 also is helping with the quality

[hearty karma]

and it wouldnt go dark 🥹

[smoky oak]

they won't ._.

[hearty karma]

this thing is meant to be night time

[bold cave]

well we have an option to set the offset now

so its just a matter of tinkering

[smoky oak]

we have had that for well over a month

my noise offset model I have is well over a month old

I wish I could train dreambooths just cause I am pretty sure I know how it would work

Noise offset is also why MJ gens look so much more moody than SD gens

[bold cave]

I am also finally training with xformers now instead of flashattention o:

[smoky oak]

MJ has had noise offset for like a year

[bold cave]

aight tomorrow's gonna be looking like a good day to run tests

training's gonna take 6hrs so perfect to run while I sleep

[smoky oak]

Don't be discouraged when it does create full noise offset, but it may help with your issue

Also, I have my own hacky noise offset Cyber Realistic model, and while the colors and tones look better, the realism is hurt

[bold cave]

oi, does the scheduler u train with have any impact

cuz ddim has always been the scheduler of choice and I have no idea why

and when I tried euler a the training went 💩

the reddit post does say that it hurts the training

[hearty karma]

yo sen do you know if zovya can do you know images

[bold cave]

but I got a strategy to ensure I dont undertrain nor overtrain

im saving 8 checkpoints x 17gb = over 100gb of models lamo

[smoky oak]

It can

[hearty karma]

[smoky oak]

well, it can do women, no realism model can do men

[bold cave]

but once I narrow down the good ones I will trash em and keep the goldilock ones

[hearty karma]

ah yes realistic face

[smoky oak]

This is the difference between my non noise offset and noise offset versions of cyber realistic

[bold cave]

whats the offset value

0.15?

[smoky oak]

no idea, and its not fixed

[hearty karma]

oh yeah it can definitely do the you know what gens

[smoky oak]

yeah, if its women

[bold cave]

it does everything broh I have no idea how that bastard managed it

[hearty karma]

magic and a 16gb model

[bold cave]

he destroyed all models on civitai singlehandedly

I added 20% senblend out of pity

just hoping I can at least beat him in flexibility looolol

[hearty karma]

btw asked him if he could possibly add noise offset to his model

:-)

[bold cave]

well apparently its what Im doing

or not according to sy

we will find out tomorrow ig

[smoky oak]

what you are doing is making images darker on average, which is a form of noise offset, but not the holy grail noise offset

[bold cave]

but thts all I need. Imagine the extra contrast without having to lower cfg

[hearty karma]

anyways why the hell are the hands so damn good

[bold cave]

the skin details will look real

[hearty karma]

this is a 1.5 model

they shouldnt be allowed to be this good

[bold cave]

ur a fool to believe 768 was any good

[hearty karma]

in certain cases it can be

[dry crow]

I just merge the noise offset model into my merged models to get the nice Effect

Some models need 1.0 some need 0.5

[smoky oak]

yeah, only problem is they also shift in the direction of the noise offset model

so mixing a noise offset model with photoreal will degrade its realism

[hearty karma]

i only just realized that zovya is on front page of civit

[bold cave]

bru it had like 20 5-star votes on the same day it was uploaded

I knew from when I saw it that it was gonna blow up

[hearty karma]

96% 5 star

and the average is at 4.92

[dry crow]

I didnt tried it with Photoreal models yet

That would be a good test

[smoky oak]

Yeah, it works good for stylized ones, but I have tried with several photo real models, and it suffers a lot

[bold cave]

dude. torch2.0 + xformers is literally twice as fast in dreambooth

my last training took the same amount of time but with half the steps lol...

[hearty karma]

i have just done my first time using img2img properly

[smoky oak]

neutral exposure (SD signature)

Moody noise offset exposure (Midjourney signature)

That noise offset look is just so much more appealing IMO

I have an idea...

[hearty karma]

guys is it possible to change the brush size for img2img?

[smoky oak]

yes

[hearty karma]

how?

[hearty karma]

ah thanks

👍

[smoky oak]

no worries, glad to help

[hearty karma]

[tight mural]

how do you guys even generate images like this? mine gives really bad images

[smoky oak]

Good prompting, models, and 1000's of hours of experience

@bold cave I have made a monster...

43 seconds to load model

💀

[hearty karma]

ok at the start of this

i was very confused because it looked like a large mess

and now its going not messy

[dry crow]

You need Community models for better Images, also good prompting, using negative prompts too

[tight mural]

al alright, ima try different models, tysm!

[hearty karma]

through 16gb of pain :-)

[tight mural]

._.

[hearty karma]

what?

[tight mural]

my drive is gonna be full soon

[hearty karma]

theres a 5gb version of the model

[tight mural]

of wat model

[hearty karma]

zovya

[tight mural]

where do I find models? huggingface?

[hearty karma]

civitai

[kind quartz]

looks like sfinx 🙂

[hearty karma]

i can agree on the pose

but say hello to the cat tiger lion cross breed :-)

[kind quartz]

@tight mural i like deliberate model and it is very small 2GB!
@hearty karma Hello cross breed 😄

[tight mural]

my drive size isnt the main issue, the issue is cf storage is really slow

[hearty karma]

i am now mixing an ant lion and wolf

im sure thats going to be good :-)

[kind quartz]

suggesting saving in jpg png disk is full very soon

[hearty karma]

i had to remove ant

it buggered up the generation

[kind quartz]

🙂 yes it is challenging!
@tight mural looking forward to see your first image! 🙂

[gleaming flame]

@dry crow @shut sinew

I want something like this

Its a video btw

https://media.discordapp.net/stickers/907134046419689502.png?size=160

[hearty karma]

i think i did a [ | ] that i shouldnt have

[kind quartz]

to your album 🙂 @hearty karma

[modest marsh]

how can i upscale or make something like this https://media.discordapp.net/attachments/1046079014156107896/1102781507904479322/nicequality.mp4

[oak osprey]

Mia Sara 1985

[wispy nest]

Not bad

[oak osprey]

crystal caves

@kind quartz original, CTU 1.0 strength to 1080p, and then 1080p re-CTU'd to 1.0 a 2nd time

[oak osprey]

.3 strength to 1080p adds more depth to her

[dense tapir]

I don't upscale

though I do have it at half FHD

[oak osprey]

whats that

[dense tapir]

FHD is 1920x1080

[oak osprey]

oh

[dense tapir]

Upscale that one

[oak osprey]

would love to but huggingface started 404'ing for all the models

[dense tapir]

damn

earlier github was down

[oak osprey]

404 Client Error: Not Found for url: https://huggingface.co/api/models/Duskfallcrew/isometric-dreams-sd-1-5

[dense tapir]

{"error":"Sorry, we can't find the page you are looking for."}

[kind quartz]

@oak osprey thanks, must learn a lot about ctu!

[oak osprey]

np

[acoustic moth]

https://www.reddit.com/r/StableDiffusion/comments/13bp1pb/here_are_some_stunning_highresolution_images/ thoughts on these? besides some flaws I think they look really good overall, any idea how these were created? model/workflow etc, Id like to recreate this but also train the face to be a consistent person

[oak osprey]

"loopback upscaler"

likely using MSE as the VAE and/or Karras sigmas as a sampler

[acoustic moth]

yea was looking into that, but even the base render before any scaling, wonder how they got that

[oak osprey]

MSE/Karras

VAE tiling off, Xformers enabled, lots of VRAM

lol

[acoustic moth]

unless its a selfie, my renders come out very unclean for full body, would i need to upscale then inpaint the specific parts?

[oak osprey]

you're likely using VAE tiling if things start to go incoherent like that

i couldn't get the faces correct when doing full body shots but portraits went well

disabling VAE tiling eats more memory but it allows the MSE to do its thing better

i don't think i ever saw any tiling artifacts with the EMA version of the VAE, but it had the same garbage output when there was more going on in the shot. almost like it couldn't focus on the mountains/trees and the subject's face

[acoustic moth]

some of these terms going over my head, with base a1111 i had this issue, not so much with models straight from civitai but when I train my own

I did get this vae recently https://huggingface.co/stabilityai/sd-vae-ft-mse-original/tree/main

ill have to explore the ui to see what this vae tiling is and how to disable it

it might be im running on 12gb vram so I do turn off ema settings for training otherwise it doesnt work

[oak osprey]

using ema versions of models produces better outputs for inference

[acoustic moth]

ill have to retrain then, ill probably have to lower training steps per image to make room for ema

ill try to get an output I want from a base model before I worry about training a face

[oak osprey]

https://civitai.com/models/23188/loopback-scaler

[wispy nest]

Did you put her in a swamp since you are Shrek

[tiny wharf]

Im trying to generate variations of this mascot logo. E.g:It wearing different clothes, different poses etc. Is there a way I can do that?

[acoustic moth]

inpainting for different clothes for sure, different poses tho? i guess it could but not sure how well

id be interested in results if youre testing it

[hearty karma]

i think my stable diffusion ui is buggered

i cant stop the generating

[tiny wharf]

Is there a way to train my own model using multiple images. Lora maybe?

[hearty karma]

how much vram you got?

[acoustic moth]

i personally use dreambooth, but lora for low vram yes

[tiny wharf]

12gb

[hearty karma]

oh you should be fine then

[oak osprey]

happens when the cuda kernel hangs. watch your cpu use in resource mon. it will skyrocket

[acoustic moth]

lmk how your training goes, im also on 12gb but i often have to compromise some settings so it doesnt run out of space

[tiny wharf]

I'll try to learn dreambooth

[hearty karma]

i might have to close my ui down

and restart it

🥹

[bold cave]

now thats a hot lady

mixed feelings about my new model

I honestly think the noise offset thing is marring the quality of the model. the extra darkness looks fake af. but I havent tested it thoroughly yet

[oak osprey]

amen

[hearty karma]

:D

[oak osprey]

I don't consider myself a religious person but she makes me feel like I could become one

[hearty karma]

for some reason i cant seem to get good skin on this model ;-;

[bold cave]

which model G

[oak osprey]

because looks like it's pulling unreal engine 5 training data

[hearty karma]

the photoreal one

[bold cave]

good lord something is getting fucked up with my training. now it cant even get tay right

yeah fellas please dont fuck with this noise offset shit

[oak osprey]

imagine gatekeeping noise offsets

you monster

[bold cave]

ima show you the atrocities it is creating

[oak osprey]

no

NO

ok fine

I can't help it I need to look

[acoustic moth]

👀

[bold cave]

wait it got fixed by loweing cfg

let me do some more tests

cuz now she looks just right

[hearty karma]

[bold cave]

do you see the extra darkness

it's there, but it doesnt look good

I might redoit but with something like 0.02

[hearty karma]

oh just learnt zovya has noise offsets in it

[bold cave]

it does?

maybe thats why.

[hearty karma]

yeah confirmed by the creator

so you are probably doubling it and screwing it up

[bold cave]

also I hate how this model gives everyone chickenpox scars

yeah I probably shouldnt fuck with it then

sytan is a lying sob

[hearty karma]

im confused then

where is offset noise

[acoustic moth]

yea and thats hard to get consistent with, identifying features gonna be different in every render

[hearty karma]

how do i mess with it for my generations 😢

[bold cave]

you cant

its for training a model only

[hearty karma]

really?

[bold cave]

yh

[hearty karma]

🥹

[bold cave]

the shadows are more dramatic here for example

but it looks like a shittier tay

[hearty karma]

taylor swift after having a botched face surgery

[bold cave]

^

[hearty karma]

btw sen this is a thing i did before

[weary ember]

i fused some actresses

[bold cave]

is that Emma Stwatson

[weary ember]

nope, guess again

[bold cave]

one of my dreamboothed girls. Im getting closer to making this shit look on point

[hearty karma]

sort of looks like a female mark hamil

[bold cave]

I love the detail on her nose bridge

[acoustic moth]

cara and idk who else

[weary ember]

cara is first in row, correct ^^

[bold cave]

ben shapiro is the second one?

[acoustic moth]

i kind of see it

[hearty karma]

sen how did you get super detailed faces?

[bold cave]

DDIM 90 steps, CFG 4

[hearty karma]

💀

ah

[bold cave]

but these are just tests

I also think training multiple subjects in 1 model is somehow diminishing their quality, but it may be because Im naming them too similar to each other

[acoustic moth]

naming?

[weary ember]

i solve, its: cara delevingne, angelina jolie, christina ricci

[hearty karma]

coincidence

i think not

[acoustic moth]

yea cara is the overwhelming face here

howd you get it so consistently though damn

even with single subject training mines not as consistent

[weary ember]

deliberate model

[acoustic moth]

hmm, i always thought deliberate wasnt as realistic as others ive used

can you get full body renders out of this or is it just portraits

[hearty karma]

personally think the best model rn is Zovya

[bold cave]

deliberate was among the best until Z came out

[hearty karma]

Zovya is godly

[weary ember]

full body too 😉

[wispy ether]

i mean...

[hearty karma]

the thing is i havent even used restore faces on my Zovya gens

[bold cave]

dont

[hearty karma]

and the faces are pretty damn perfect

[bold cave]

restore faces destroys details

[hearty karma]

this is just straight from the model

[acoustic moth]

used dreambooth?

[wispy ether]

emo gwen

[hearty karma]

i do not plan to use it

[weary ember]

nah, local. bodies not as easy and good though

[hearty karma]

dreambooh is training not the website

[bold cave]

yes it can noob

[hearty karma]

do what specifically?

[bold cave]

peter ur outdated af

peter still thinks realistic vision 2.0 is the bomb

dont even bother. it doesnt matter what you show him, he'll find a way to naysay it

[acoustic moth]

tbh realv has given me better results than most thatve been recommended over it, but its all prompts and setting adjustments

[bold cave]

we've done all the testing here G

between me and sytan

there is only one clear winner

[hearty karma]

i think euler a might be the best sampler for zovya

[bold cave]

we even did some blind tests here

[hearty karma]

but im only saying that without trying any others

[bold cave]

Ok time to retrain my model without the noise offset and with even higher learning rate. Catch u guys in a bit

[hearty karma]

see you in a few hours

[bold cave]

nah G torch 2.0 and xformers made it twice as fast

[hearty karma]

oh 🤔

[acoustic moth]

what gpu you using

[bold cave]

3090

paying for itself. the training uses up 22gb vram lolol

I might try doing a higher batch count too

[acoustic moth]

for whatever reason i can never get the xformers option to show no matter what i install or what commands i put in batch file

[acoustic moth]

whats the vram on that

[bold cave]

24gb

[acoustic moth]

wtf

whyd i get a 4070ti lol im dumb

oh nvm they expensive as shit no wonder

my trainings pretty limited with only 12gb i guess

[wispy ether]

wth

[hearty karma]

what the hell am i looking at...

[wispy ether]

[hearty karma]

the bottom right character looks like ash ketchum mixed with mario

[bold cave]

Yo I just noticed something

the final dreambooth models weigh only 5gb 😮

so I must be doing something thats causing it to lose data? it's gonna be a nightmare figuring this out

I think it may be cuz i've been extracting the ema weights

[wispy ether]

[acoustic moth]

is there a way to train in pieces so my 12gb can handle it

[wispy ether]

12gb can handle training

[acoustic moth]

i often run into batch size error

[hearty karma]

could it be that you set the batch size too high?

and its eating away at your vram

[hearty karma]

doctor strange?

[wispy ether]

run batch size 8

i do

did

i stopped training

[bold cave]

8 is too high

[wispy ether]

still worked for me

with fp16

[bold cave]

hmm. thats what I use. perhaps I should start experimenting with much higher values

[wispy ether]

ew

ok?

[bold cave]

I told u not to argue with peter

[wispy ether]

who is peter

[bold cave]

theres so many variables to experiment with

[wispy ether]

https://tenor.com/view/bh187-family-guy-peter-griffin-laugh-laughing-gif-21485227