🏞｜general-with-images

[weak sage]

The individual is the ultimate minority.

[oak osprey]

are you on the latest versions of everything?

[smoky oak]

Also, I may be able to collect Native American benefits, which I didn't realize before

[weak sage]

i ran updates yesterday

[oak osprey]

that might be the problem

[sterile temple]

updated the extensions, or just A1111?

[smoky oak]

I love your pronouns haha

[weak sage]

extensions

[weak sage]

no one even tries to use them, it's all a farce

[smoky oak]

I can try to use them for you lol

[weak sage]

thee*

[smoky oak]

Oh, I see, it is 3, Interesting

Discord has a new section just for pronouns BTW, just found that today

[oak osprey]

you mis-pronouned them earlier thy saying when ye said "your pronouns" and not "thee pronouns"

or rather, "the pronouns of yore"

[weak sage]

thy*

[oak osprey]

nay i am qyte sure it is thee

thine

[weak sage]

Thou needst to learn thy pronouns

[oak osprey]

thour

thour pronouns

[smoky oak]

I will make sure to try and use thy pronouns correctly, to the best of my ability

[oak osprey]

full name: themothy

[smoky oak]

*Thymothy

[weak sage]

Verily, I thank thee, Sytan

[smoky oak]

[oak osprey]

where do you get all these cursed fucking animoot

they're awful

[smoky oak]

We stan the gay cockroach

[smoky oak]

[oak osprey]

i don't care what thy orientation is thee is a roach

ergo BANNED

ipso facto

[smoky oak]

[sterile temple]

someone told me roaches are actually pretty clean, they just prefer to live in filth

[smoky oak]

Mood

Side note, cause you made me think of it

Opossums are like super harmless, and it sucks that they get such a bad rep cause they do fantastic things for the environment

[weak sage]

as long as we dont have to eat them

[oak osprey]

you don't like, "have" to

but my mother will be super embarassed if you refuse to

[weak sage]

gonna slingshot them onto soros' island

[oak osprey]

,_, is it nice there?

[weak sage]

if thou likest ze bugs

[oak osprey]

i don't get it but i'm not in the illuminati, so

i probably never will

[weak sage]

illuminarty

thats the anti-AI site

[oak osprey]

ಠ_ಠ

[weak sage]

basically if it looks detailed they mark it with the scarlet letter

[oak osprey]

how do i get it to not fuck up their faces

[weak sage]

inpaint

[oak osprey]

another damn AI cryptoid

[weak sage]

a whole face is probably easiest thing to inpaint

[oak osprey]

this is on Clipdrop. i don't think it has that.

[weak sage]

the fuck is clipdrop

[oak osprey]

i think that's the killer from Scream 2

[weak sage]

there's a faceless undertaker in witcher 3

hearts of stone DLC iirc

[smoky oak]

How are thou on this fine day, @weak sage ?

I will be home in like 5 minutes, and I am so excited

[weak sage]

im okay, just trying to generate some paraphernalia into an image with inpainting and it doesn't know how to draw it

[oak osprey]

bongers?

if it's not making tiny faces, my WIP model does alright

faces are rough though

i should really remove the "embarassing" term from this family photo prompt though

it throws it off so much lol

SDXL's answer to the same prompt

lol uncrop XL

i thought Uncrop would be better based on the demos they showed

lmao the Cinematic style

[sterile temple]

Like a thumbnail of a True crimes show

[smoky oak]

I am hommeee

I am at my PC for the first time in 8 days and it feels weird lmao

[oak osprey]

whoa

maybe you're outside of the matrix

did you take a red pill at any time from a mysterious man in sunglassae

[cyan snow]

i agree

nah, my handwriting is way worse than that

[oak osprey]

idk what that tool is

[cyan snow]

dude, the AI is better than me in writing, my handwriting is even more messed up than 1.5 text in images

[smoky oak]

No sir, but my mom did buy a Mystic Topaz ring lol

[oak osprey]

i used the bug's life font

[smoky oak]

It did lmfao

That is amazing

Like, actually amazing haha

[oak osprey]

no i edited it to do that in there

but thank you thank you

[cyan snow]

DUDE is the text in that image made by SDXL

[oak osprey]

[smoky oak]

Ah damn lmao

I was gonna say lmfao

[oak osprey]

i tried to get it to do that

even Bing, and that's where the jackhammer whisperer image came from

[smoky oak]

That explains why Life looks low res haha

[oak osprey]

oh shit you can tell? you have really good eyes

[smoky oak]

Have you not met me? I am a pixel peeping king lol

[oak osprey]

i hate the red tbh

yellow feels more 80s like i was aiming for

[cyan snow]

the image of the people in the background looks like it's missing the VAE

[oak osprey]

it's just how SDXL looks

it's either too much of one thing or too much of another

[cyan snow]

no, it's not i usually don't see that in SDXL

[oak osprey]

hm

you mean the arizona haze effect?

[steep jackal]

Coffee

[oak osprey]

looks like a very polluted area maybe

[sterile temple]

I ordered my new PC last night, exciting a couple of items on backorder though so have to wait

[cyan snow]

no, the no VAE effect, when you make an image without VAE it does this

[cyan snow]

your a pre-built boy?

or did you order parts

[sterile temple]

parts

[oak osprey]

all images have a VAE, the models (incl 1.4) all come with one "baked in"

they trained SDXL's VAE a lot and the previous ones were chronically undertrained

[cyan snow]

idk man, when setting the VAE in A1111 to automatic it does something simmilar to this

[sterile temple]

Would be the first computer i built in a long time

[oak osprey]

i don't know what automatic means but the 'none' means 'the one that comes with the model'

automatic might use the one named alongside the model in the dir

no idea

[sterile temple]

matching names is automatic

[cyan snow]

it's set to that by default, whatever it's using by default is messed up.

[oak osprey]

it's possible that the stuff the model represents isn't known by the default VAE

but that's unlikely to be the case for SDXL

they said the settings are going to be tweaked for Clipdrop and i'm not sure what those settings are

[cyan snow]

then something went wrong in that image you generated, i never saw SDXL do that

[cyan snow]

makes sense

[oak osprey]

idk man it's what the model thinks "cinematic" means

give me a prompt, i will have it generate it

why does that one not have a watermark

[cyan snow]

cosmic phoenix, epic

[oak osprey]

it's spinning

[cyan snow]

this is what my model makes with that prompt

[oak osprey]

idk the image gen crashed. trying again

[smoky oak]

It only adds the watermark if you select the image to upscale, which from what I have seen, does ad dgreat fine details like body hairs and eyelashes even from far away, but reallyyy struggles with bokeh (it has a tendency to sharpen out of focus details and make them nasty)

[oak osprey]

this one did get upscaled

oh but i was upscaled when i genned it. i guess you can skip watermarking that way

[smoky oak]

god damn, my new mouse is so smooth lol

I got used to it before I left, but I forgot just how smooth and nice it is

[cyan snow]

EW WTF

what happened to that dog on the left

[cyan snow]

i think it did better than my model..?

what's your opinion

[oak osprey]

yours followed the prompt better

if you want super colourful versions of it, i guess SDXL is better. if you want sharp and high contrast, your model did better

[cyan snow]

=[

[oak osprey]

i switched to fantasy art style. sec

the 4th one crashed

[cyan snow]

i actually expected SDXL to do better =[

[oak osprey]

for a base model i am really happy with all of these, i guess the most i can hope for in fine-tuning is to improve the default responses to be a bit more like i'm expecting

like, it knows a lot, i'm just not sure how to prompt it to get what i want

[cyan snow]

yeah, i bet if i will do to base SDXL the same thing i did to 1.5 it will be a WAY better model than mine

[oak osprey]

i find it the opposite of SD 1.5 and 2.1 for photoreal stuff and fantasy. as in, it's very easy to get realistic looking people from this, but not easy to get fantasy stuff that wows me

[oak osprey]

2.1 is super happy to be trained on mysticism

[surreal thistle]

How good is this?

[oak osprey]

adopts elder magic syndrome like a child to love

[cyan snow]

hm....

[oak osprey]

it has the same smooth skin issue that a lot of the basic photoreal attempts do

[surreal thistle]

Yes, and how do I fix it?

[oak osprey]

try adding a year and a location and a camera like 1998, California, Nikon lens

[surreal thistle]

Here are some others made with SD 2.1 instead but the issue is that they all have the same face.

[cyan snow]

which one is better?

[oak osprey]

that looks like a taxidermy thing

very stiff

[oak osprey]

you might have faceblindness

[cyan snow]

did SDXL do better?

[surreal thistle]

I might have what again?

[oak osprey]

i changed your prompt a bit

i quite like this

dark cosmos vs sharp high-contrast phoenix

'vs' made it a face-off tho

[cyan snow]

better than this?

[oak osprey]

yeah actuallt because mine looks almost like you'd see it in deep space

like the eagle nebula

you see it too? how it could be a legit space dust formation?

it's just a crazy thing that exists in our universe lmao

nevermind

[cyan snow]

i don't think this is a fair comparison actually, because that was a 3 word prompt

[oak osprey]

it was the fantasy art style

same prompt without the style

what the fuck lol

[cyan snow]

oh no

what happened there

[oak osprey]

idk clipdrop is breaking down

[cyan snow]

well, in comparison to that it's not even a comparison to what my model does

[oak osprey]

i know

i'm just hoping it's the Clipdrop backend config because this is worse than the bot for sure

the bot's incredible

but it's so slow

[oak osprey]

yeah i don't know how to get something like that out of SDXL right now

[cyan snow]

you think if i will put SDXL through the same process i put 1.5 through it will make for a better model?

i hope so

[oak osprey]

yeah mostly while reducings its ability to do things you don't care about

[cyan snow]

SDXL finetunes should be the best diffusion models made so far

there is no way it's another 1.5 downgrade

[oak osprey]

seems like if you're really bad at prompting, SDXL will be frustrating 😅 i am frustrated lol

[cyan snow]

what do you mean? i usually write an entire paragraph for a prompt

[oak osprey]

too much prompt with sdxl, hurts it

so you have to be succinct and short

[surreal thistle]

Will the hands always be botched?

[oak osprey]

so far im seeing some 1.5 prompthero prompts look good out of te box on SDXL but they look better if you start deleting keywords

[oak osprey]

forever lmao

[surreal thistle]

https://tenor.com/view/hp-harry-potter-snape-always-stare-gif-17584635

[cyan snow]

also elder magic

[oak osprey]

SDXL 0.9 has none

[smoky oak]

clip drop is annoying me

2/4 images are failing, and I am not sure if they are using credits

also, how are people getting aspect ratio images?

[cyan snow]

this is what the SDXL bot makes

[oak osprey]

they are

[smoky oak]

lovely

[oak osprey]

#1100170365604483202

[smoky oak]

so people are using the bots to get the higher quality aspect ratio images?

[oak osprey]

use the arrow key to select the option

[cyan snow]

i just change my A1111 settings kek

[smoky oak]

But the discord bots aren't using the same models, I am confused lol

[oak osprey]

the discord bot runs that clipdrop model plus a newer one and two older ones

you can get any one of the four

[cyan snow]

also, will lora's work on SDXL?

[oak osprey]

yeah but not any that already exist

[smoky oak]

@oak ospreyOh my god, the differences between some of these models are for SDXL are insane lmfaooo

like, some of them look HORRIBLE

[oak osprey]

welcome to the party, sir

i'm pretty sure the horrible ones are without the refiner in the pipeline

[cyan snow]

what are you talking about

[oak osprey]

they're very smooth

[smoky oak]

look at the difference here lmfaooo

[oak osprey]

fuck man, the one on the left looks like Spud from A Christmas Story

[smoky oak]

the image on the right looks amazing BTW, very impressed with how that came out

[smoky oak]

the SDXL bots use 4 models, as Pseudo just said

[oak osprey]

mused?

[cyan snow]

ah i see, we did a comparison before, and my 1.5 finetuned model beat SDXL =[

[oak osprey]

for fantasy art.

😛 we didn't try photos

[smoky oak]

no idea how the hell that happened lmao

[oak osprey]

i'm pretty sure i could get deepfloyd to kick a fantasy art finetune's ass

[smoky oak]

I am doing photos right now, cause I am a photo snob lol

[oak osprey]

yes you are

try the Analogue Film style

[cyan snow]

my model isn't intended to be a fantasy art model

[oak osprey]

whoops 😛 well it is oneee

try this then:

gorgeous red-hair girl, 1977, ethnographic photography, sweden, best quality, highest detail, balanced contrast

negative: (bonnet), (hat), (beanie), cap, (((wide shot))), (cropped head), bad framing, out of frame, deformed, cripple, old, fat, ugly, poor, missing arm, additional arms, additional legs, additional head, additional face, multiple people, group of people, dyed hair, black and white, grayscale

[cyan snow]

and the SDXL model my 1.5 model outdone was the one on clip drop, which was fucked

[oak osprey]

^

we didn't try the bot

[cyan snow]

doing this now

[oak osprey]

this is 1992 and 1977 in deep floyd

[oak osprey]

yeah, fantasy art model 😉 it's a good fantasy art model though

i like it a lot

for example i bet if you prompt it simply space llamas you'll get straight fire

[cyan snow]

oh no, it just made porn

[oak osprey]

welp the sdxl bot lost at that prompt too

2nd result

[cyan snow]

THAT DOESN'T EVEN RELATE TO THE PROMPT

[oak osprey]

SDXL prompting is strange

it'd be cool to get made as one of those velvet posters though

[cyan snow]

i can't stand the fact that my 1.5 model has beaten SDXL on anything

[oak osprey]

you know where you can use neon colours to fill it in and then put a blacklight on it

it's ok man it's a research release, just calm 😛

they stated they left 'training gaps a mile wide' (mcmonkey's words) for fine-tuning to be easier

[cyan snow]

i was hoping that SDXL finetunes will be the best diffusion models made so far

[oak osprey]

so i think you'll make a really cool model with it. it's weird to see that it's so hard to get the stuff out of it though, honestly i've seen really good stuff and assumed it'd do better

it's still got the abilities, i just don't know how to get at them

could swear i asked for wide aspect lol

[oak osprey]

hmm cell phone backgrounds tho

[oak osprey]

do you have a negative prompt

i can use those on the bot but not in clipdrop

[cyan snow]

of course, hold up

[cyan snow]

stinky toes, twins, two people, lowres, ((bad anatomy)), bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, (jpeg artifacts), signature, watermark, username, blurry, deformed, (bad eyes), (misaligned eyes), squashed face, pixelated, people

[oak osprey]

oh

SDXL told me it truncated the prompt

it's too long

[cyan snow]

that's what she said

[cyan snow]

i know i'm a fucked up person, no need to tell me

[oak osprey]

it's still the same damn thing

different seed, same prompt, not too different at all

[cyan snow]

that's a weird aspect ratio

[oak osprey]

the two models even did the same image

SDXL can do pretty much any aspect ratio that is 1 megapixel

[cyan snow]

yeah, it's not that fair because im using hires. fix

[oak osprey]

sdxl doesn't need hires fix

[cyan snow]

i think it's better with detailing than SDXL

[worldly coral]

These are so sick

Hello by the way!!

[cyan snow]

still, i hope SDXL can beat this, MJ can't

[oak osprey]

thank you but i'm frustrated because the same prompt cannot get much variation out of the new model

[worldly coral]

WHere can I access the new model?

[cyan snow]

you can't, only the bot works atm, but i can't get it to outdo my 1.5 model

[oak osprey]

clipdrop.co

[oak osprey]

@cyan snow WatermarkXL

sucks that they still clearly don't own the dataset

[cyan snow]

oh no

[oak osprey]

putting watermark in the prompt doesn't give you one, so, putting it into the negative does nothing

if they trained on images with text in them, it might explain the watermarking issue. the watermark scoring happens to overlap with just general text existing in an image

[cyan snow]

https://tenor.com/view/pepe-lonely-sadge-sad-pepe-lonely-pepe-gif-24375295

[oak osprey]

so you can't put the watermark score threshold very high or you're eliminating all of your valid text targets and you're left with low-value text training data

i noticed this when looking at the laion data last week

it's almost like the watermark detector needs to be trained more 😛

the good news is SDXL happily generates watermarked jungles so they can use that to train the detector

[cyan snow]

i gotta go, if you can outdo this using SDXL please tell me, id love to know. @oak osprey

[oak osprey]

i would have to be very descriptive

[cyan snow]

i wasn't the prompt i used here had 3 words in it

[oak osprey]

i kno lol

[cyan snow]

well, good luck, i gotta go

[oak osprey]

too much nebula

[oak osprey]

@smoky oak turns out the bot doesn't run 0.9, at least not according to fruit's announcement

that's on clipdrop

the bot is the "previous research model"

[smoky oak]

interesting, well some of them look amazing

I really like this one

[oak osprey]

i want them to release the bot's model

can you like, vary him around much at all?

i'm struggling to get variation out of it without going 'too far'

it's either very similar or too far away from my expectations but we can't set the seed statically yet

interesting the forest he's in doesn't have him standing on a trail / path

that was something that was difficult to stop from appearing when you prompt for just a forest

[smoky oak]

I am just getting single gens for now to see how well they work, so far it seems like when the good one hits, they look insanely good

[oak osprey]

< when it hits good

[smoky oak]

I am gonna keep messing around, cause man, it does hit sometimes

[oak osprey]

meanwhile, my robot

[smoky oak]

the difference is insane lmao

I wonder which is the new one

[oak osprey]

a ways to go until we're at 1920x1080 alleyways not splitting

rainbows too

1920x1080 horses are doin better than expected

[sterile temple]

can it do a circle shaped moon behind someone?

[oak osprey]

at this resolution?

i'm sure it'll do several

somehow the vape shop is ok

everything's on fire but this is fine

and now, for my next song: Phase Shift

[smoky oak]

uh oh

I think the bot just died

I am getting "Command is outdated, please try again later"

now they have changed to "The application did not respond"

[smoky oak]

not sure whats happening

restarting discord

That fixed it

weird

my discord did update, so I wonder if thats what it was

[oak osprey]

you would want to run

[wispy nest]

can we add bot to our own server 🤔

[oak osprey]

you can in theory use the API to make a simple discord bot but it'll charge you for each gen, some fraction of a dollar

what a rip-off

[oak osprey]

@split rover so i'm at 1920x1080 training still and i've almost got her kinda, uh, like, not all duplicated and stretchy lmao

im hoping it turns that post on the right side, into a cat

crazy architect

meanwhile, inside the pub

[split rover]

haha nice, the high res certain looks fun. At a certain point though the transformer layers on the older models probably wont be able to keep up with the resolution for keeping structural integrity on the entire scene without some additional help

1920x1080 is a whole heck of a lot of pixels hah

[oak osprey]

well i haven't hit a ceiling yet but so far i do see some things will struggle more than others. for me teaching it this resolution isn't so much about making this one work but stopping bullshit from happening in smaller pixel space

like at 1024x1024 it does pretty well now with less training than when i was doing 1024x1024 trying to fix that by itself

[oak osprey]

huh, that last prompt without negatives

[neon crystal]

Just wanted to congratulate the SDXL team! Like WOW! Only a couple of days ago I was lamenting how behind SD was compared to Bing ( for out of the box fun / art stuff and not professional pipelines etc ). SDXL 0.9 is 🔥 and so is Reimagine. Incredible Stuff! This will finally help SD reach the masses and popular consciousness in general I think.

[smoky oak]

Why is nobody talking about all of the false claims of clipdrop? lmao

I am on their site trying out some of their tools, and man do they just bold face lie lmao

Whats up with that, SAI?

[dim crane]

I see Richard, I like

[novel ravine]

SDXL is ||fant||ass||tic||

[oak osprey]

uncrop?

[smoky oak]

their big lie one is their upscaler

they just straight up bold face lie about what it can do lol

[oak osprey]

yes

and uncrop is terrible

the demos are way better than the real thing

and i put a family photo in and it assumed the boy was pregnant

since i had only seen the demos i assumed it was just like that. i was so disappointed lmao

[smoky oak]

I did get some pretty solid uncrop at times, but the upscale is literally just lying

and not even trying to hide it

all their upscale images are high res images that were downsampled for before and after lmao

[wispy ether]

Dam

[oak osprey]

ah

[smoky oak]

they make 0 attempts to hide it

[smoky oak]

you're gonna tell me it was able to tell that the gray blob on the right was eyelashes from this?

yeah, sure lmao

[oak osprey]

clipdrops own watermark screws it up

[smoky oak]

this one gets me the most tho

[oak osprey]

wow

magic

[smoky oak]

oh, its worse, just a sec

they are just bold face lying lmfao

Come on Stability, shame on you

[oak osprey]

i got my model out of the minima it was stuck in last night

had a burst of improvements now

never seen this prompt work so well

one earth horizon

[wispy ether]

[smoky oak]

https://tenor.com/view/mid-gif-24456870

[wispy ether]

[oak osprey]

fingies.

[wispy ether]

Is it just me or did sd get better

[oak osprey]

anime is easy, so

[oak osprey]

2.1

[cyan snow]

is that the one world trade center?

[oak osprey]

i havent been therr

dont know lmao

[cyan snow]

look it up, it looks identical to it

[oak osprey]

i love how good the buildings look til you zoom in

then its like a photoshop disaster

sdxl gets around that by blurring the shit out of the background

my problem seems to be ive been aiming for sharp images

people love blurry images

[cyan snow]

the building in this look identical to the one in the left of your image

that building is the one world trade center

[oak osprey]

yuppp

thats cool

2.1 doesnt start out doing that

guess i added that

oh the one in my imh has an inverse design

the facade is upside down

[cyan snow]

still, it's the same building

[stoic pivot]

Banana Pizza.

[supple jolt]

What is this website?

[tulip agate]

image source https://handicraftshome.com/products/11x14-picture-frames-handmade-art-wall-hanging

they are all stock photos

[ancient hound]

hi guys, when I put midj stuff into img2img I get result that is completely smooth I mean way too smooth, like loosing so many details from midj, what am I doing wrong

[oak osprey]

you need a lower denoising strength then

fewer step count maybe, you have to play with the sliders

[ancient hound]

but I want stuff to change a bit

so I don't get it... how can you have more details and stuff changing

[oak osprey]

you might have to still mess with the sliders and your img2img prompt

[ancient hound]

ok lets try more thanks

[oak osprey]

try like, intricate details

if you don't get the results you want, try the Controlnet v1.1 extension for a1111, if you're using that. it will give you a new Tiles model that will let you do cool stuff with more control

[ancient hound]

yeah I have controlnet but not sure how to use it well, I have a potato PC so it's a tough business

[oak osprey]

you can pay a few bucks for about 1000 image gens on Clipdrop but i don't know how flexible the img2img tool is there

[ancient hound]

could you show an example of controlnet settings to get good output with more details and slight difference from original

[kind quartz]

is this possible 1.4 ?

[oak osprey]

https://github.com/lllyasviel/ControlNet-v1-1-nightly

[wispy nest]

maybe I should learn what controlnet is, heard it a lot but never looked into it even when I got tabs and settings for it on the webui :P

[oak osprey]

visit one of the moon colony's, uh, many lakes

@wispy nest i'm going to be very honest with you. i don't think we can improve beyond this point

[wispy nest]

"we," what did I forget yesterday. I know I was drunk but I can't have been that drunk…could I?

[oak osprey]

this is our family now

🥹

[oak osprey]

@smoky oak sdxl still kicks my model's ass at people lmfao

this isn't duelling banjos, homie

[wispy nest]

I wonder if it'll be better than my fine-tuned one, or rather, if it's better in my eyes than my merged checkpoints 🤔
Only time will say, hey, what if I add it to my… 🤔

[oak osprey]

you can merge this with those too

and here's this checkpoint's family photo

i merged in some of my pseudo-real-beta model. threw away the unet

kept the newly trained wide-aspect unet with the beta text encoder

[stoic lotus]

hope this is realistic enough

[oak osprey]

i'm still training it in the hopes that it keeps learning the wide aspect because some prompts still end up with duplicates, though re-rolling the seed resolves it

[oak osprey]

anime eyes, homie 😛

and glowy skin

@wispy nest that's one powerful street lamp

[wispy nest]

just add solar eclipse to it and have it even worse! :P

[oak osprey]

i prefer sonar eek lips

[stoic lotus]

@oak osprey

[oak osprey]

those are all very good, i like them! i have no issues but i know pixel people will say that purple hair girl has smooth skin, try giving her some more natural wrinkles and what video games would call subsurface scattering

are those all from DreamShaper?

[wispy nest]

I would just say one thing:
Where is the restaurant where I can order that much bacon??!! 🍽️

[stoic lotus]

no, this is not dreamshaper

i made it

[livid ermine]

Self-growing mushroom buildings:

[stoic lotus]

These are images from Crystal-Clear-Photorealistic and RunDiffusion-FX

models me and my team made

[oak osprey]

ah, cool

what was your training schedule, eg. how many images / batch size / gradient accumulations, which i assume you adjusted here and there?

layer freezing, backmerging of checkpoints, offset noise, zero SNR, scaling v-prediction loss?

[stoic lotus]

out of all these, nothing will help you achieve a big enough difference as regularization images

[oak osprey]

for 2.1?

[stoic lotus]

all this is 1.5

[oak osprey]

oh, ok

why regularization images, are your dataset very small?

[stoic lotus]

'cause you can have both

regularization images help improve accuracy

they are not to replace your dataset, they are to help

[smoky oak]

It's unrealistic to expect that it won't 😅

[oak osprey]

this is why i laugh

[stoic lotus]

right now, i'm patching up the bad_hands in a new model that will be released in a few days
side by side, left one without the fix, right one with:

[oak osprey]

me, researching AI

[stoic lotus]

this is just the 1st version after training, have not fine-tuned yet

[wispy nest]

I recognize that throne, why am I recognizing it all of a sudden, I know I've seen that before 🤔

[oak osprey]

my darkness is fixed now at least

i still haven't added leopard geckos to the training data i guess

[smoky oak]

Fucken niceeeee

[stoic lotus]

lol

i do have something funny for you guys

[oak osprey]

the darkness is so strong like senator palpatine would love it, you can't just throw 'high contrast' into a prompt all YOLO anymore

[smoky oak]

Bro got that geopard lepcko

[smoky oak]

CRUNCH!

honestly, being able to easily crunch is kinda a mood

[oak osprey]

totally

[smoky oak]

Just like how SDXL can easily wash out if you prompt it

[oak osprey]

it's nice that i can do it on command lmfao

it means i can exclude it too

[smoky oak]

SDXL is basically gonna remove the need for VaE's cause it has such a good understanding of styles, and color processing keywords

[oak osprey]

you can ask it for adobe raw photo style

SDXL has a new VAE

you can fine-tune it still

well

[smoky oak]

A very adaptive one I am assuming

[oak osprey]

it's the same VAE. they fine-tuned it a lot

no, it has the same issues tbh

but i want to try VAE tiling on it

i will withhold ya know, any final statements on that lol but mcmonkey curses the VAE

just don't do small faces on any AI model and you're good to go

you can do like area generation to try and inpaint a crowd in an audience at a concert

[stoic lotus]

found it, lol

[oak osprey]

essentially making large faces and shrinking them down

[wispy nest]

won't be a problem as long as you upscale it and then downscale it 🤣🤪

[oak osprey]

that sounds like hell

[wispy nest]

it's worse when you can't decide on which upscaler to use :P

[oak osprey]

the fuck is going on with his arm lmao

[wispy nest]

looks like the "when two images turn into one" :P

[oak osprey]

so merely fixing the aspect proliferation isn't enough to fix all of the small details too

i like it though, still. if i could use img2img to fix this, i'd be happy

at least it is starting to reliably create the scenes i want in widescreen, with the proper spacing, and everything

never had that before

[wispy nest]

I mostly fix it by having more/less steps on the same settings, but I'm also too lazy and just lets it create a new image instead :P

[oak osprey]

yeah good point, my validation pipeline is a fairly fixed set of settings

[stoic lotus]

this was done with a low cfg, it helps give things a more realistic aspect

[oak osprey]

i train for higher CFG

[stoic lotus]

it does not really work like that

[oak osprey]

it does for v-prediction

i have a very wide range of CFG i can use in my base model, up to 9.2 is very good, and once you rescale CFG, it opens that level of guidance up to more options

[wispy nest]

can't see the CFG, I use a lot lower now as well because I'm testing some stuff people showed earlier in another sub-channel here. I'm at around 3-4 CFG, and also at clip skip 4-5

[stoic lotus]

@wispy nest thats exactly it

it's one of my testers that just posted that

he must;ve used 4 or 5

[oak osprey]

https://tripleback.net/public/grid_vanilla/woman_grid.png

[wispy nest]

I read about it from the 2.1 sub channel about a week ago

so if it's still the same, then say hi to them from me :D

[oak osprey]

trailing vs leading timestep and 0 or 0.7 cfg rescaling

[stoic lotus]

i can show you up to 30cfg if you'd like

[oak osprey]

your model is epsilon prediction based. it will not work the same.

[stoic lotus]

did a model where i used cfg as a way to change the art style of the render

[oak osprey]

one of my validation prompt grids that i make on each ckpt

i know how the CFG works

if higher CFGs stop working, i revise my training schedule

[stoic lotus]

@oak osprey i can show you one, where cfg ages people

lol

[oak osprey]

probably not across every prompt

cfg rescaling takes a normal sized woman in normal chair and makes it a novelty chair

[stoic lotus]

you train on standard model ?

[oak osprey]

2.1-v

i started with 148k images in the training pool and the filters have deleted about 48k as it has been running

[stoic lotus]

👍

we decided to skip version 2

[oak osprey]

i got pretty good results i was happy with for a 2.1 adventure model and a 2.1 photo model but i then realised i can't really do 1024x1024 without mutations and set out to fix that

in fact this morning i got some unexpected improvements in output by mixing some of the photo model with the work-in-progress wide-aspect unet

i haven't messed with textual inversions, LoRAs, or any of that. i just want a good zero-shot model that does widescreen adventure AND photo well

[stoic lotus]

do your thing man

[wispy nest]

I want more of a "half real, half painting" styled ai art. photo just by itself isn't doing it for me...I...I should NOT have written it like that, and I shouldn't have then continued to write this and then post it 🤣

[oak osprey]

it'll do that

do you want the prompt to try? maybe you already have it 😛

[wispy nest]

do people need to buy stock photos anymore

[oak osprey]

especially not when we can make it like, butt cheeks or whatever we want, without paying some barista

[wispy nest]

righto

[wispy nest]

I already got some idea but if there's some super prompt word in there then sure :D

[oak osprey]

digital art, fantasy, portrait of an old wizard warlock man, detailed, best quality, UHD, 8K

i think it's the digital art and fantasy. UHD definitely makes stuff look not-photorealistic

[wispy nest]

yeah, I've seen you type it before. I don't think a lot of those words do what might think they do, but that's more of a personal way to write the prompt itself than what they do. I try and "trick and lie" at the ai to create the thing I want more than just telling it straight out :P

[oak osprey]

illegible

[stoic lotus]

wait for it to load

[wispy nest]

yeah, and then click "open in browser" as well

[oak osprey]

i did

[wispy nest]

not the discord's smaller display

[oak osprey]

it says it's 168x1004 lol

[wispy nest]

I got some super large, super crisp image grid when opening it

[oak osprey]

strange

[fleet goblet]

Well I got vladmantic's repo working in WSL fine.

3.39IT's per second on the 768 model, 20 steps.

[wispy nest]

I think the oil painting and portrait is the ones closest to the style I like :D

[stoic lotus]

thats why i posted that

[fleet goblet]

Ah, coffee.

[wispy nest]

but I like them all, except smelly photo >:P

[fleet goblet]

Is 3-4IT/s bad for the 2.1 768 model?

Hmm.

[oak osprey]

ohh i see

[fleet goblet]

My man's drinking coffee.

Face is great. Hand, uh.

Hm.

[oak osprey]

i tried to get bojack horsemen and it made centaurs

[oak osprey]

look, you're doing great. don't try and fix the hands

that's where madness lies

[wispy nest]

this one was the one I got that I liked the most :D
even if their shoulders can not only cut diamonds, but also turn them into dust! But that happens at high clip skip so nothing new really :P

[wispy nest]

wrinkled skin, old wizard, wizard hat, (detailed face:1.1), high quality,

[fleet goblet]

Hell yeah.

I made a dark souls character look like a warhammer character.

Somehow.

Thornhead.

[oak osprey]

try to make ben stiller

[fleet goblet]

Even the base 2.1 model is very good at generations given the right prompts.

[oak osprey]

he and john cusack seem resistent to being absorbed into the AI

[fleet goblet]

Yeah?

Well he's missing hands.

[wispy nest]

but I dislike wizards because of what they so often want to wear >:(

[oak osprey]

2.1's ben stiller

[fleet goblet]

Nothing like sitting on the back of the bench.

I wonder.

How does one like me upscale with Controlnet Without ultimate upscale?

ultimate upscale has precision problems for me.

Lol

[oak osprey]

clench your cheeks and pray, Dan

[wispy nest]

I never understood how ultimate upscaler worked, I wanted to but it was beyond me :(

[oak osprey]

i thought you are german, and ergo good with advanced technology

[wispy nest]

no, I'm a swede, if that was for me :P

[oak osprey]

ohhhhh do you know Felix? i met him when he was here on vacation

he is from sweden too

[wispy nest]

I knew a felix from school...way back in the forgotten past :P

[oak osprey]

well i suppose it could be a different one, but we think of sweden as just one small village where everyone knows each other and it is always winter

is that not how it is?

[wispy nest]

I moved from a city I thought was big with a population of around 10,000 to one with maybe 30,000 or so and I get stressed out when there's more than 5 people on the same street as me!

bah! when I get the wizard to take off their hats they still want their photo to include the empty space above 😵‍💫 🤣

[fleet goblet]

Weezard.

He's coming for you.

A cool character design.

[wispy nest]

purple/pink, my greatest ai enemy!

[fleet goblet]

Fairies.

[wispy nest]

I only see one though ;P

[fleet goblet]

Realized why my thing was going slow

wslconfig had it set up so WSL was limited in some level

Reset it, back to 8IT/S

Why he lookin at you like that

[oak osprey]

("a digital art piece of hauntingly pale skulls with black craft cult aesthetics", "opulent and gothic", "high contrast", "bold and macabre", "with a touch of mystic.").and()

one of GPT's prompts it came up with. a real banger

[oak osprey]

i like the edges on this one

that model without negatives 😄

[fleet goblet]

This is a cool one.

Negative prompt: Disfigured, Mutated, Multiple Limbs, fingers, toes and/or eyes.
Steps: 50 | Sampler: DPM++ 2M SDE Karras | CFG scale: 6 | Seed: 64272743 | Size: 768x768 | Model hash: ff144a4984 | Model: v2-1_768-nonema-pruned | Clip skip: 1 | Version: 4867daf | Parser: Full parser```

[oak osprey]

what the hell

cursed but better now i guess than before

[fleet goblet]

He's staring.

Lol

Prompt: Humans discover new life in the shape of a sentient frog.
Negative prompt: Disfigured, Mutated, Multiple Limbs, fingers, toes and/or eyes.
Steps: 20 | Sampler: Euler a | CFG scale: 6 | Seed: 1385135943 | Size: 768x768 | Model hash: ff144a4984 | Model: v2-1_768-nonema-pruned | Clip skip: 1 | Version: 4867daf | Parser: Full parser

[oak osprey]

no negative prompt for these, just the positive prompt

[fleet goblet]

The negative prompt does do something.

Lol

[stoic pivot]

I am trying to use SD2.1 and can not get any good image, I try "Female walking in a Snowy forest" and the images are horrible, not only in the rendering troll faces but that it is trained on watermarked images that was made to prevent images to be used. But some swear by this model and say that I just need "good negative prompts", any negative prompt suggestion to fix this?

[fleet goblet]

Hmm.

Eh, that one's iffy.

[wispy nest]

I gave up on 2.1 in general, but I noticed that using any sort of "environmental" prompt word could create those watermarks

There aren't really any real fixes in my opinion, that's why so few people went to use 2.1 over 1.5 :(

[fleet goblet]

I'm using 2.1 768.

Lol

[wispy nest]

I'm sorry to hear that 😉🤣

[fleet goblet]

I'm not getting bad outputs, though.

[wispy nest]

yeah, quality is up to each person, that's why I added "in my opinion" because I know that I don't have an objective standard, even though I wish I had the one true opinion! :P

[stoic pivot]

I just get images like this, but I am intrested in how so many say this model is good and I want to find the trick to tame it.

[wispy nest]

anyone done any animations?

[fleet goblet]

Add "Out of frame" to the negatives.

[oak osprey]

i will try on pseudo journey v2 (which is on civitai)

[fleet goblet]

Prompt: A man wearing a white tuxedo. Full body image.
Negative prompt: Disfigured, Mutated, Multiple Limbs, fingers, toes and/or eyes. (((Out of Frame.)))
Steps: 50 | Sampler: DPM++ 2M Karras | CFG scale: 6 | Seed: 2956969147 | Size: 768x768 | Model hash: ff144a4984 | Model: v2-1_768-nonema-pruned | Clip skip: 1 | Version: 4867daf | Parser: Full parser

[stoic pivot]

Negative "Out of frame".

[fleet goblet]

Nice face.

[oak osprey]

that's without negative prompt though

[stoic pivot]

... eeh. What kind of images was this model trained on?

[oak osprey]

you need a negative prompt with base 2.1

no idea what the heck model that is lmfao

looks roasted

[fleet goblet]

Negative prompt: Disfigured, Mutated, Multiple Limbs, fingers, toes and/or eyes. Disfigured and/or out of place face.(((Out of Frame.)))
Steps: 50 | Sampler: DPM++ 2M Karras | CFG scale: 6 | Seed: 3027338645 | Size: 768x768 | Model hash: ff144a4984 | Model: v2-1_768-nonema-pruned | Clip skip: 1 | Version: 4867daf | Parser: Full parser```

[stoic pivot]

I try "(out of frame:1.2), crappy image, 1972"

[oak osprey]

put (deep fried)

[fleet goblet]

Question Kenny

Are you using the 768 model?

[stoic pivot]

I start to think I may have wrong SD models? Just now 1.5 and is not okay

[oak osprey]

same prompt no neg 😄

[fleet goblet]

Prompt: A bartender that is wearing a white tuxedo. Serves you a drink. Bar Counter. Monochromatic Image, Analog 1972 photo.
Negative prompt: Disfigured, Mutated, Multiple Limbs, fingers, toes and/or eyes. Disfigured and/or out of place face.(((Out of Frame.)))
Steps: 50 | Sampler: DPM++ 2M Karras | CFG scale: 6 | Seed: 2436356879 | Size: 768x768 | Model hash: ff144a4984 | Model: v2-1_768-nonema-pruned | Clip skip: 1 | Version: 4867daf | Parser: Full parser

Beat that one.

[stoic pivot]

Ooooooh, SD2.1 Knows who Leisure Suit Larry was... Yes, I am that old 😄

[oak osprey]

(deep fried) watermark, cropped, out-of-frame, low quality, low res, poorly drawn, bad anatomy, wrong anatomy, extra limb, missing limb, floating limbs, (mutated hands and fingers:1.4), disconnected limbs, mutation, mutated, ugly, disgusting, blurry, amputation @stoic pivot

well i know leisure suit larry, and bucky o'hare too 😛

[fleet goblet]

Let me take that, because that's just a better overall negative than I use.

[oak osprey]

no >:| its kennys

[fleet goblet]

:)

[oak osprey]

[fleet goblet]

Actually, that negative is giving me even worse images.

Lol

[stoic pivot]

I start to like the random crazy SD1.5 and SD2.1 can spit out.

[wispy nest]

my bartending skills—I mean prompts—aren't that good :P

[oak osprey]

this is kind of how the model starts out thinking of a time machine

strips it for parts

[fleet goblet]

Beautiful hands.

[oak osprey]

it's building some futuristic shit now

[wispy nest]

only broken a couple of times by the mafia! :D

[fleet goblet]

Prompt?

_>

A cherrypicked image, but a nice one.

[wispy nest]

a 1960s gruff male bartender in a bar

• a billion loras, and my own "fine-"tuned model that I'm starting to think isn't needed as I also tested the newest "dreamshaper" model which is…sadly…better than the one I merged >:I

[stoic pivot]

Prompt: "1975 photo of a Italian person looking like Leisure Suit Larry sitting in a bar wearing a white suit, creepy smile that even would scare a dog away".

[wispy nest]

this is a un-cherry picked version without all the loras

[stoic pivot]

That is a dream bartender.

[oak osprey]

this is my photo model with your prompt unmodified

[fleet goblet]

handy man.

[oak osprey]

this is my take on it

[fleet goblet]

Cartoonized moment.

[stoic pivot]

Fun.

[oak osprey]

if you all shut up i can balance a third one

ethnographic photography of a bartender that is wearing a white tuxedo serving you a drink. Bar Counter. Monochromatic Image, technicolor kodachrome 1972 photo nikon lens f1.4

[wispy nest]

cheers

[oak osprey]

the monochrome and technicolor fight each other nicely

[stoic pivot]

I have dyslexia and are afraid to use the word "cocktail" in A1111, extra look so I not use any furry model.

[fleet goblet]

Now I gotta question for you

Would you be able to dish that out with a base model of any kind?

[wispy nest]

I'd say you don't need to state that sort of detail, but then again, you may want very specific drinks :D

[wispy nest]

you mean only using the "vanilla" 1.4, 1.5, 2.0, and 2.1 versions?

[fleet goblet]

Yes.

[wispy nest]

no, never

[fleet goblet]

Heh

[fleet goblet]

What is hands down the best overall model?

That we have our hands on

A general-purpose model.

[wispy nest]

dreamshaper
You can get it on civitai

[oak osprey]

dreamshaper makes CP

i wouldn't touch it

[fleet goblet]

Oof.

[wispy nest]

I would just not write such prompts that makes those things, but that's just me :P

[stoic pivot]

I give up on SD2.1 for now, so here is a...

[oak osprey]

i prefer SFW models, mostly 2.1 based

try Artius v2.1

[oak osprey]

ManglerFTW's V3 mix is really good too

[smoky oak]

@wispy nest I would love to Collab sometime and put MJ and SDXL beta against each other for realism

I am curious to see how well they compare at the moment, as SDXL has been giving the most realistic images I have ever seen

[oak osprey]

including actual real images?

[oak osprey]

trust me this is realistic, tons of people eat at diners

[stoic pivot]

That look like a good suggestion, I will try that one. Ty.

[oak osprey]

i'm biased toward my models but i never recommend them to others anymore

[wispy nest]

jay leno might be too young for 1960's bars :P

[oak osprey]

well i guess in a way i did, ManglerFTW include mine as a model base he merged with

[fleet goblet]

@oak osprey

I assume I download CharhelperV4?

From manglers?

[oak osprey]

oh i've never tried that

https://civitai.com/models/5395/mangled-merge

here's the link to the V3 merge

[smoky oak]

I really wish I had access to MJ to really see just how much better SDXL is in direct comparison

[wispy nest]

one day the ai will know where the wings should be placed on people, one day, any day, soon'ish...maybe :P

[stoic pivot]

Coattail snorting, a typical Swedish proplem?

[wispy nest]

problem? no, that's just a normal monday

[earnest solar]

best harpy so far

[stoic pivot]

That Artius v2.1 was flexible, but it could use the Loras I had and like nor my own textual invesion I made for my old art style.

[cyan snow]

@oak osprey did you manage to do it?

[oak osprey]

no and clipdrop crashed

[cyan snow]

=[

[oak osprey]

yeah it's a weird one to prompt, you can't do long prompts with it and the short ones don't have much variation between seeds for certain subjects

i get bored of it

666/54650

[weak sage]

bug error log

[oak iris]

https://twitter.com/nftaleth/status/1672228905665568769?s=46&t=mgJhjYjx7J8E48-ILP3Bxg

Matkot (@Nftaleth)

AI's dance with art 💃
AI can mimic, but does it create? 🤔 We need more blending of human creativity & AI power, embracing #AI as a tool, not a replacement.
Let's sculpt a future where tech amplifies our collective imagination! 💡🚀 #AIinArt #FutureTech #AIForArt #ArtisArt

Twitter · Jun 23, 2023, 1:03 PM

SD with PS new generative fill

[oak osprey]

me , my wife, and my wife's boyfriend at our last family photo

don't know what she sees in him

[weak sage]

He kinda looks like you, so there's that

[oak osprey]

i don't see it

when you see it 😁

[languid pagoda]

he looks so proud of that trash

[wispy nest]

so no more clipdrop 😭 until monday at least

[oak osprey]

fuck yeah the time machine has WHEELS again

[oak osprey]

they all do

i love how the cooking pot is clean

gotta be safe

the prompt came from the CEO of Stability

Check out my cousin larry in his dirty room, he is such a damn mess

[wispy nest]

lol @ the prompt

[oak osprey]

my model is just like "dirty? i gotchu fam"

the alleyway prompts now work with widescreen 😄

[weak sage]

the effects of tags is pretty unpredictable, like all those "highest quality, best quality, trending on artstation" tags dont actually did jack shit in my imgage once i actual test it

[stoic pivot]

For me that have a cleaning schedule that include dusting of inside of my computers once a month and have 3 vacuum cleaners for different room and usage and like relaxed minimalism, those images give me nightmares.

[oak osprey]

some models do

[weak sage]

using senzus rn

[oak osprey]

c'mon old man, get that chip off yer shoulder >:O

[oak osprey]

billionares stranded in a van sized submarine at the titanic wreck 13,000 feet under the sea celebrate their rescue with pizza rolls

i love how those all 4 came back in 1 batch

[umbral fern]

do I see pizza rolls?

[oak osprey]

yes

[wispy nest]

making me hungry lol

[stoic pivot]

But it is not banana pizza.

[umbral fern]

Cursed

in a good way

[oak osprey]

2.1 at 1920x1080

damn

i really like this ckpt

the starting ckpt

[umbral fern]

Her glasses fused with her eye, sensational

[cyan snow]

"the mystery box is so happy to swim"

i love prompt gen

made a photo of a lucky block, then BLIP interrogated it, then i threw the beginning of the interrogated prompt to prompt gen, then it said this shit

[umbral fern]

Wow

That...that is beautiful :' )

[cyan snow]

actually Jesse, it's just basic chemistry. =]

[oak osprey]

trying an essential oil diffuser for the first time @smoky oak

[smoky oak]

peeeeeeeeeeeeeeeeeeng

howdy hello

[oak osprey]

didju just wake

[smoky oak]

your cloud machine looks mighty moist lol

[oak osprey]

from the jet lag

[smoky oak]

no lmao

I didn't have any jetlag lol

[oak osprey]

whoa, young people ftw

[smoky oak]

It was a short flight lmao

4 hours, and the time change was only 2 hours

[oak osprey]

i do'nt think that matters with jet lag i think that makes it worse lmfao

ohh

wait, you're not in florida?

[smoky oak]

I have been in car rides longerthan 4 hours

no bro, I have said it several times lmao

[cyan snow]

this is from Mario, right?

[smoky oak]

I am in California lmao

[oak osprey]

are you sure

[umbral fern]

PFFFT

[cyan snow]

i swear i saw those in a nintendo game

[umbral fern]

real life mario blocks

[cyan snow]

"super mario in real life"

https://tenor.com/view/mario-gif-20449358

[umbral fern]

Supa mawio in real loyfe

[cyan snow]

EW WTF, i didn't mean to click on that gif

[oak osprey]

@smoky oak hey with your young eyes can you ruin my ego again and tell me there's stuff wrong with this, i only see her sweater texture and maybe the glasses shape

[umbral fern]

lmaoooo

[umbral fern]

Yeeeah that's "Super mario in real life" for ya

[oak osprey]

she looks hella white too but that's another issue

[umbral fern]

looks like the roblox winning smile in the first pic

[oak osprey]

same prompt and seed from earlier but now after training, the two dudes are now women

some wild shit man

[cyan snow]

classic 2.1

[oak osprey]

2.1 is great

imagine using text embeds on this model

[cyan snow]

still, it fucks up WAY more than 1.5 and SDXL, we already discussed that

[oak osprey]

nope, i've tried training 1.5 and it can't do faces as small as 2.1 can

[cyan snow]

=/

[oak osprey]

others in #🔧｜finetune say 2.1 is easier to train too

[cyan snow]

idk man, i tried making a 2.1 version of the model i made, it's very weird

i think SDXL will be way easier to finetune because SAI said they will release tools made specifically for finetuning SDXL

[cyan snow]

not even close to the one i normally use, i did the same process i did to 1.5 to 2.1, it wasn't even close to the 1.5 model.

but yeah i don't blame the model. i think the reason 2.1 usually comes out weird after finetuning is that SAI didn't release finetuning tools for it. because 1.5 was a hit, the community made the tools, but that isn't the case with SDXL, even it's unfinished BASE model (SDXL0.9) seems to outperform almost all 1.5 models and MJ, so imagine what it would be like after finetuning with suitable tools

[smoky oak]

Oh yeah, I definitely can

want me to really go in? lmao

Cause I see a lot wrong with it

[cyan snow]

dude, i hate these fucking things

[smoky oak]

Also, fuck I love SDXL

[cyan snow]

wait, how no watermark?

you are running it locally?

[smoky oak]

SDXL bot in this server, it's even better than clip drop, and no watermark

[smoky oak]

All made with the bot in this server

[cyan snow]

and that's not even the finished beta model, and it already outdoes most current diffusion models

[smoky oak]

MJ should be scared lmao

Yeah lmao

and it will run on consumer PC's

[cyan snow]

i can't wait to fine tune it, we will make the best diffusion model made to date man

[oak osprey]

apparently it doesn't really use terminal SNR

[smoky oak]

IDK, all I know is it has proper non biased lighting control, so really dark or really bright images

IDC why or how, just glad it has it lmao

[oak osprey]

it's offset noise and it keeps changing the balance of dark / light images during the various beta releases we've had access to

in one of the ckpt we had zero contrast lol

that was like last week

[smoky oak]

whatever they have right now, I am in love