🏞｜general-with-images

[smoky oak]

There. They should be much better now

[wispy nest]

noice.

[smoky oak]

Discords compression is impressively shit

https://imgur.io/a/jsi4sAM#0RDvTQq

Here, have a look for yourself

[oak osprey]

that internal testing version has been on the bot for a few days now at least

[dry cosmos]

2.048 x 16.384

[oak osprey]

it's been complained about how it has regressed in some things and does a lot better in others, mostly photography and text

[dry cosmos]

zoom

[oak osprey]

i have faith they will find a way to merge the checkpoints in a way that preserves the best of both

which of these has the better faces at a distance?

[dry cosmos]

Better

[tawny forge]

which model is that

[oak osprey]

Restore Faces sucks

[tawny forge]

I think you could use inpaint and choose the "only masked" option

[dry cosmos]

Use FAZOL to improve results

[smoky oak]

@dense tapir

[dense tapir]

LOL

[oak osprey]

what the hell is FAZOL

@smoky oak so i'm trying img2img for this face fixing and i don't think it works just that easily

https://images-ext-1.discordapp.net/external/aooLaT97s1m3ER0UcBWTdd-QMicSSWINBYYZfYthH_A/https/tripleback.net/public/discord/1686613741.596092d836c019cf6f795eb159b00dcd4a82a1.png?width=1440&height=960

[smoky oak]

Could be a difference between 1.5 and 2.1

[oak osprey]

wow, nightmare fuel

if ya squint, they mint

i forgot i force-disable VAE tiling

[smoky oak]

You can see how much high res fix helps here

[oak osprey]

with 1.5

[smoky oak]

Yeah ok, something is definitely different between your workflow and mine

[oak osprey]

hmm

[smoky oak]

But all high res fix is, is a pixel upscale, into an img2img with the same prompt. I don't understand

[oak osprey]

yeah, i'd tried that so many times before. so i was confused, and thought, maybe i just need a better model

or samplers have changed since then, etc

Here's a quick comparison between the old highres fix and the new one.

Before:
-Regular width and height sliders suddenly became the secondpass sliders when highres fix was enabled. The functionality of the width and height sliders was moved to the firstpass sliders in the highres menu, and thus the regular sliders then controlled the secondpass resolution. Their functionality was basically flipped.
-Setting any of the firstpass sliders to 0 was hardcoded to always default to defining the desired resolution as 512x512. There were TWO different operations done in the code to define what the actual firstpass width and height would be, more on this below because math*. Before, if you manually set your firstpass to 512x512, you got a different result than just leaving firstpass = 0 due to this.
-Steps for the secondpass were always hardcoded to be half the steps defined in the firstpass.
-Upscale latent and upscalers options were semi hidden in the Settings menu, needed to go back and forth to test them, unless added to the Quicklist settings.

tl;dr too confusing

Now:
-Regular width and height sliders always control the firstpass, they always control the base image and no longer flip back and forth between highres fix enabled and disabled.
-Every single option in the highres fix menu now directly controls (almost) everything in the secondpass.
-Can now define the steps for the secondpass independently of the steps for the firstpass. More steps = more details on the secondpass but slower upscale, less steps = less details on the secondpass but faster upscale
-Can now change upscalers directly from the highres fix menu, no longer need to go back and forth in the Settings.
-Can see the before and after resolution in the highres fix menu

tl;dr easier to understand

(this is from some random comment i found on github)

[oak osprey]

i tried DDIM at 100 steps even to see if more steps = more detail. it does not

[smoky oak]

It made them all look the same with high res fix lol

[oak osprey]

i noticed that

lmao

uncanny

https://images-ext-2.discordapp.net/external/Wq5_BV9cNtcmUiRCsMsRA8bmA3q9tRPfzLUCRTdrXkc/https/tripleback.net/public/discord/1686614268.9974217083b5d8b6de98cd7f1400486aeca3e57.png?width=1440&height=960

try and fix this one and see what happens

[smoky oak]

I can't from a base image

[oak osprey]

😦

[smoky oak]

Especially without the primot

*prompt

[oak osprey]

A family enjoying a picnic in a blooming spring meadow, high contrast, UHD, 8K

[sterile temple]

I was playing around with img2img with non ai generated images, and found it made things worse even at 0 denoising and no resizing, like it was still using the upscaler model before starting the denoising process

[smoky oak]

Here we go, much better example here

[oak osprey]

so it also can't fix faces at a distance?

[smoky oak]

It has been able to help significantly, but it can't fully fix them

@oak osprey what word did you say to use for cultural races?

[oak osprey]

ethnographic

[smoky oak]

I knew it was something ethno

Thanks

Localized inpainting will always be the best option for long range face fixing

[oak osprey]

i tried to do GPFGAN on the images, and it just doesn't fix these as well as it does normal portrait images

[sterile temple]

lol random 🫰

[oak osprey]

@smoky oak this stuff always makes me so suspicious and curious of the images i see on civitai. that Freedom model, the new 2.1 realism one, doesn't have prompts for its best demo images

the best results i see in posts for it are toasty looking

[sterile temple]

I was thinking civitai should have a bot that loads the model, uses the same few different prompts and posts them to every model's page

[oak osprey]

yeah just call it The Great Destroyer

have it use the diffusers default DDIM settings

watch all the models go stupidly baked

[sterile temple]

the person posting can set the recommended CFG, sampler and steps, but that's it

[oak osprey]

well i also don't remember seeing a single model that has a group shot on its cover

look at all those solo subject models

[sterile temple]

some people use so many loras and embeddings, they might as well be using a different model

[oak osprey]

absolutely

"look at what this model can do!" no it can't lol

[sterile temple]

ah, I think I figured out why img2img was ruining my images even at 0 noise. I had the upscaler in the settings set to Lanczos, so it must have been using it even without resize and noising at 0

[oak osprey]

yes

[sterile temple]

it's like the super tall images for character posts. they must be starting off with a wider image and then cropping it

[oak osprey]

dude the enormous portraits were all controlnet tile upscaler with ultimate upscaler

that alone is just giving their model a major advantage

it's ridiculous. that model is not making a 1080x1920 image at full quality with no repetitions, not without tiling and upscaling

[sterile temple]

and maybe controlnet from a img the model didn't even create

[oak osprey]

this is SDXL. i'm not saying this is all it can do, or all it can ever do. buuut..

it's like every time i go to use that bot it's just saying it's not available -_-

[smoky oak]

When people do that, they are likely either hiding a workflow/prompting process that makes them money, or they're hiding some form of doctoring the images

I know for me, I typically don't upload my prompts when I upload images for people's models, just because I use my very specific prompts for my clients when they buy LoRA's from me

[oak osprey]

idk, they claim they want it to be a base model for others and you'd want their validation prompts to ensure you're not, you know, breaking it

i don't usually care unless the author of the model makes such bold claims

what's with the specific nature of these negative words? does that indicate a model merge?

[sterile temple]

badhand4, easynegative are text embeddings

by bad artist is a booru tag

i've gone off negative embeddings as they change the look too much

some even made the model NSFW

in this case the faces are true to life 😄 real housewives of new york

[oak osprey]

well i dont feel bad anymore lmao

here's the first 4 results for a prompt and i'm pretty stoked on that considering how little care i gave to the dataset of faces

it's interesting how few hand deformations i see now that i've fed it 10k photos from the Hands HD dataset

i trained on so many hands i thought they would be coming out of the ground etc

[sterile temple]

they are multiplying! 😱

[oak osprey]

@sterile temple try just the prompt alone, face-off and try it with other concepts

eg. celebrity face-off or wild animal face-off

mine had one with elephants that was interesting, but my friend's settings have it all baked-looking

for example, he prompted john trabolta 🤣

[sterile temple]

celebrity face-off with no negatives gives me these tiled celebrity pics

[oak osprey]

https://images-ext-1.discordapp.net/external/lIgbl5fwwo9XSE8JcyLRsKK7mZiQ4PeTSBkJPrkENm0/https/tripleback.net/public/discord/1686618411.562704101dbb32ab2a4c2a6adaaaa950ac6f6a.png?width=960&height=960

mine did this

[sterile temple]

my model finds the prompt john trabolta very confusing

[oak osprey]

what does it think of jews for jesus, 1985, ethnographic 😄

best group faces i've seen so far

oh you might regret it but taxidermy tends to make animals very high quality because they're sooooo still when they're photographed

man, i love this thing 😄

[grizzled olive]

That's actually a really great tip

[sterile temple]

Leonardo DiClaprio is a hunk. Arooga!

[sterile temple]

George Clooknee and Antonio Banberas

[oak osprey]

@sterile temple i put your Leonardo DiClaprio is a hunk. Arooga! verbatim

[sterile temple]

I think he went through a shredder to get shredded

[oak osprey]

Shia Laboeuf

[grizzled olive]

I did "Leonardo DiClaprio is a hunk. Arooga!" and my first image is extremely nsfw

[oak osprey]

Shia

[sterile temple]

The prompt was just Leonardo DiClaprio

Is that the correct spelling?

[smoky oak]

DiClaprio lmao

[oak osprey]

sexy shia!

Shia LaBeouf

[sterile temple]

looks like any shia will give you Shia LaBeouf

[tropic shell]

He looks as tired as I feel lol. That hair tho

[smoky oak]

Mom and I went on a candy spree lol

Don't need the parental supervision when you are the parent >:D

[sterile temple]

my teeth hurt looking at that

[smoky oak]

Ehehehehhe

We're flying to Mexico with our family, so it's for all of us haha

And not all of that is going either kok

*lol

[tropic shell]

That's right! All you need is...that sugary feeling, followed by some regret the next day.

It's like eating a tub of ice cream, which, I can't say I have experienced, but I'd imagine that it's followed by a similar feeling lol

[oak osprey]

all you need is glove

[tropic shell]

Those Harvest Snaps are so good

[oak osprey]

ok you can tell that's Shia but why he all hair hiar

[tropic shell]

Look, He's become a man of the Misty Mountains, okay?

His brilliant beard makes sure everyone is in awe!

[oak osprey]

sounds like something from Monty Python

[tropic shell]

He's a very tall man with the lineage of...miners.

[oak osprey]

monty python and the ethnographic photograph of shia laboeuf

@smoky oak this is what it does for wookie

this some 2024 remake vibes

"im not just a wookey, im also a wookey MAN UwU"

[smoky oak]

||||

[oak osprey]

lord of the rings monty python and the ethnographic photograph of the bee gees

bunch of derps

[smoky oak]

Johnny derp

[oak osprey]

ethnographic 1974 pirate jesus of nazareth, greece

[smoky oak]

Oops lmao

Wrong gif haha

Look at them teefers

[oak osprey]

all them teef and no toofbrush

[sterile temple]

wouldn't want him biting me on the ankle like my cat does when he's really hungry 😄

deep fried tom hanks

[oak osprey]

steampunk pirate jesus, 1974 ethnographic photo

you want deepfried, eh

so i guess i should put deep fried into my negatives lmfao

dear god

[smoky oak]

He could barely even open his jaws for his teeth to clear lmao

He got them /\ / ---------- / /\ teeth lmao

[oak osprey]

lmao how does deep fried produce such deformities

[smoky oak]

Do you not know what a deepfried image is? Lol

[oak osprey]

i definitely do but how does my model knooow

it overcomes all of my strong negative embeds lmfao

[oak osprey]

this was best one

[smoky oak]

Maybe try being more strongly positive next time, you pessimistic jerk >:C

Lol

[oak osprey]

[sterile temple]

i was also getting some weird shaped heads with deep fried

[oak osprey]

oh i get that with very smart

hmm i wanted more cat teeth

[sterile temple]

https://github.com/DominikDoom/a1111-sd-webui-tagcomplete

[oak osprey]

he is resisting the footie pyjamas

[sterile temple]

installed this neat extension for A1111 yesterday

[oak osprey]

i have a thing in my discord bot, it just directly gets prompts from GPT based on a theme i give it

it can make some great stuff

long descriptive prompts tend to do really well, so, it just rambles

[sterile temple]

try this real housewives of insert_place_here

[oak osprey]

Housewives, Nicaragua, Reality TV, Drama, Fashion, Luxury, Beauty, Lifestyle, Relationships

Luxury, fashion, drama, wealth, party, cocktails, beauty, fabulous, gossip, fame.

[sterile temple]

real housewives of disney

real housewives of tokyo

[oak osprey]

detroit rock city

seems like deep fried in negatives, weighted heavily, is helping with tiny faces

[smoky oak]

Mmmm, these shits are fire

[oak osprey]

use icy hot to cool it down

and then... heat it back up

oh

nvm i mis-understood

@sterile temple the real housewives of steampunk scrap metal

[sterile temple]

If you can find them in your local supermarket/Target international foods area, I urge you to try Tim Tams from Australia

[oak osprey]

oh, deep fried negative is so good

good one, mikey

thanks for the idea, indirectly

@smoky oak look, faces

[sterile temple]

real housewives of gotham city

[oak osprey]

that prompt is weird

[sterile temple]

silly 😛

[oak osprey]

looks like the grocer is fully stocked up on tasteless blocks of matter

[smoky oak]

Gonna have to try deepfried as a negative

[oak osprey]

yes, please do

i used it like ((deep fried))

[smoky oak]

Wonder if shit post works too lmao

[oak osprey]

ooo

((youtube poop))

lmao

((meme))

[sterile temple]

I found rough textures,grit helps get rid of that spotty black texture that makes me barf

[oak osprey]

yeah with deep fried in the negs i can get a good enough pic that GPFGAN can fix this

don't look at her arm lmfao

[sterile temple]

real housewives of lego land

[smoky oak]

Deepfried seems to provide no real benefit for my gens

[oak osprey]

real housewives of a dirty apartment, surrounded by garbage

@smoky oak its likely from my midjourney or LAION aesthetic datasets

i can check

it's in LAION aesthetics

that did not exist when 1.5 was made iirc

[smoky oak]

Without, vs 1.4 weighting

[oak osprey]

try it as a positive prompt first and see if it makes things go haywire imo

if it does, idk, it should help in the negs

the most improvement i saw is in faces at a distance

and it's not that it makes them perfect, it just helps them trigger the facial detection so that a face restore can actually help

[smoky oak]

It makes a difference, but not a worse one I'd say

[oak osprey]

just couldn't even do that many people in a group

[oak osprey]

might not look like much but i consider that a success

jeez i think it made yours look better

[oak osprey]

of metal gear solid

[oak osprey]

try it on a 2.1 model next, if you still have one around

[sterile temple]

real housewives of studio ghibli

[oak osprey]

me: "you just can't do faces at a distance "
ai: "generating: faces at a distance"

[sterile temple]

-_-

can it do baseball caps?

I've seen some models really struggle with them, look like broken hoodies

[oak osprey]

oops

[sterile temple]

🤔

[oak osprey]

this model can do everything but mend a broken heart

[smoky oak]

I don't, I deleted them all

[smoky oak]

Jack of all trades, master of none, that's for sure haha

[oak osprey]

it mastered hats, mothafucka

[oak osprey]

i didn't even make a ckpt for a1111 for this one

if all of the reception is so chilly and cold, i do not wish to release my work lol

[smoky oak]

It's not chilly, just realistic 😅

[oak osprey]

it isn't a specialised model

it's a minor correction to 2.1

[smoky oak]

||unlike your model ||

Sorry, had to lmao

Yeah, for just being a correction it's a step in the right direction for sure

[oak osprey]

i plan on making more experiments with it, like trying to make a high quality photo centric version where i don't care if i lose the more dreamlike qualities

trying to keep it able to dream as well as be more realistic when asked, man. it's hard

currently the model seems to have converged, it isn't learning much more from my dataset

probably need more data to train it on next

the CelebA-HQ dataset is made traditionally using Real-ESRGAN and that thing is trash, i think if i mix GPFGAN and CodeFormer, i can likely beat that quality and improve the dataset

im mostly glad i determined how to save a burnt model. or at least, one way to do so

"stopped" LMAO dude was tackled and looked like he had a blast

[languid tide]

Thor playing guitar in a bar, oktan render,, beautiful, aesthetic, professional photo, 4k, high resolution, high detail, neon lighting

[oak osprey]

that'll be $50

SDXL's deep fried tom hanks

[sterile temple]

extra crispy, can I get some aioli to go with that?

[oak osprey]

that's $5 extra

[sterile temple]

oktan render

[oak osprey]

$2.50 if you are cool with a finger having been dipped into it

oktan render

[oak osprey]

aioli?

[sterile temple]

yes much better

[oak osprey]

i asked a waiter for aioli once and he said "god bless you"

[sterile temple]

shoulda asked for raaaaaanch dressn

[oak osprey]

deep fried chris farley

@smoky oak admit it, this is amazing lmao

deep fried antonio banderas

[oak osprey]

hes so pretty

[oak osprey]

RDJ?

[sterile temple]

yep

look at those big beautiful eyes

[oak osprey]

i get why he was considered so handsome

[oak osprey]

LMAO

bob ross

ornaldo bloom is quality

it came out just like that

oh... same batch lmao

[sterile temple]

what's with the flowers? bloom?

[oak osprey]

i guess i didn't make the connection but yeah lmfao

orlando bloom, blooming

lmfao AI art on canvas

[sterile temple]

uh that face gives me the creeps, how can anyone listen to his advice? 😄

[oak osprey]

he's an electrician lmfao

deep fried home alone 2

why is it so good lmao

gary busey

[sterile temple]

mmm what is that? fried crab?

[oak osprey]

does look like that

[sterile temple]

she wants some 😄

[oak osprey]

yes sir

john is on the way

he has arrived

john lennon

[sterile temple]

All my friends know the low rider

[oak osprey]

george carlin

imagine john lennon riding a deep fried bicycle

[sterile temple]

Imagination. This image. They are the same thing.

[oak osprey]

it's like trying to run from an enemy in your dream

^

[sterile temple]

that's not how you dig trenches

[oak osprey]

it's how you live them

well i was curious what it would do. visiting the great deep hole tourist trap

looks like it's about to cave-in any minute

[sterile temple]

get elon's submarine, he will save them

[oak osprey]

it won't fit

dear god they look so happy

[smoky oak]

@oak osprey do you know how to remove all weights associated with a specific subject from a model?

[oak osprey]

train it on pure garbage lmao

[smoky oak]

Welp, nevermind then I guess

[oak osprey]

it's unlikely that's possible. when people think of that they think of 2.x with the missing nsfw but the present composited images that show nsfw

that's because SAI trained on frozen text encoder which knew that stuff but the unet has added filtration

[sterile temple]

is it possible to rename an embedding in a lora?

[oak osprey]

probably by using a token map

[smoky oak]

Was trying to merge 2 models together to get a cool hybridization, but one of the models is so bad at some things that it completely nukes how good the second model is at them

[oak osprey]

oh, that might mean you need to include only the later layers of the bad model

or merge just the unet and leave the text encoder alone

[smoky oak]

No idea how to do that

[oak osprey]

it's not easy to implement this stuff in a capable tool without making it way too complicated to code and use so it's unlikely it exists

it's easy to approximate but you waste compute by training a unet you'll be throwing away at the end or something like that

[sterile temple]

block weight merge

[oak osprey]

oh, that

yes. that's the good stuff but it's hard to master

you will spend as much time as i have learning to train 2.1

i hope you do it though, and teach us

[sterile temple]

I use supermerge extension, run all the preset weights, which merges using the weights, generates an image and then shows an XY chart at the end

That's how I've been mixing my model for a while

[mossy bolt]

Prompt Help. How to make image like this.

I am trying.

Prompt : 32 year old men, holding small white paper on his left hand, human hand writing on paper, boil point men, (Written Code : 12345, Name Albert, username albertx74), clean background, portrait selfie.

i am not getting success, trying over 3 hours. please help.

[oak osprey]

lmao

use DeepFloyd

try img2img

[mossy bolt]

Let me try

I am not finding signup in there websiter

[oak osprey]

deepfloyd is on huggingface hub

@smoky oak deepfloyd was trained on frozen text encoder too

that's wild

magnets man, how do they work

[sterile temple]

verification photo achieved, what's next?

[oak osprey]

vaping contest, intimate gathering, 1983

the 1995 event

[sterile temple]

looks like they're getting tear gassed

[oak osprey]

to me it looks like punk kids taking it too far

[sterile temple]

lmao boil point men that produces some very buff dudes. no wonder he's been trying for 3 hours 😏

[oak osprey]

ironically when you said that i was vaping and i began coughing

and laughing

[sterile temple]

so small and portable, convenient to use

[oak osprey]

it's funny to get them vaping through spongebob

poor model gets so confused

hmm

like this baby that fell asleep while welding. i shall pass out now

https://images-ext-1.discordapp.net/external/9VDUy6OPi7C-LbCuBJeqPd_lQvKx_j4q6ZzGgNt7BKc/https/tripleback.net/public/discord/1686634447.7157307a9d1dcbdbd67bc32c915f74f239596b1.png?width=1003&height=1003

[sterile temple]

are these all deep fried images?

[cyan snow]

yeah, looks like bad cnet or overcooked model or even missing VAE.

[shut sinew]

back with the weird stuff

[sterile temple]

weird stuff is the best stuff

[shut sinew]

For sure!

[sterile temple]

get a room you two 😒

[shut sinew]

Nice wood

[sterile temple]

bubble head

baa

[shut sinew]

goated

[sterile temple]

mr flashy hat

yes, i think this flame feels heavier

[shut sinew]

@velvet wren Something like this? I made it a few days ago

[velvet wren]

yup!

[shut sinew]

It should have my settings in there, maybe it helps

[velvet wren]

in the metadata?

[shut sinew]

Yeah

[velvet wren]

I am unsure if its me but

metadata works only with my own pics, not when I download it

[shut sinew]

I just tried and it seems to work

I downloaded it through discord

All you need is your own QR code in the img2img tab and into the controlnet

[velvet wren]

hmm... gotcha, will try

[faint skiff]

A friend of mine asked me if I could use SD to create a line-art drawing (in approx. this style) of her and her coworker. I'm certain it's possible, and have been looking through different ControlNet models and LoRa's, but don't yet have a solid lead. I wanted to ask the community before I spend too much time down this rabbit hole, what would you recommend?

[sterile temple]

the qr code thing is hard

[sterile temple]

did you try the scribble controlnet model?

[faint skiff]

I haven't yet. that was one of the ones on my list to try, though. Scribble + a line art style Lora like 'Simple Line' (https://civitai.com/models/83244/simple-line?modelVersionId=88458)
Thank you for the suggestion, though

[faint skiff]

ha! that's awesome

[faint skiff]

may i ask what model/lora you used for that?

[sterile temple]

just a model i mixed, no loras

[faint skiff]

word. looks great!

[sterile temple]

sending it back to img2img

[faint skiff]

haha, whoa

[smoky oak]

ah yes, thank you FL studio lmao

[sterile temple]

🔭👀

[smoky oak]

yeah, its been an issue for a while now lol

it happens sometimes when waking from sleep

[sterile temple]

my work laptop does that with apps all the time, opens in the tiniest window at some random location

[sterile temple]

the smokeanywhere lora is pretty cool

[oak osprey]

i do vape contests

i trained vaping into 2.1

and smoking

[smoky oak]

Finally starting to cache my models in RAM

so much faster for large model comparisons

[wispy nest]

I just removed another 100 old checkpoints myself, 2TB SSD isn't enough for my ai hoarding! :P
There's like 4 models I use often, and one of them is the one you PM'ed. I've yet to unlock its features for other stuff though :(

[smoky oak]

hey, that model just got an update, I am checking it right now

they released 2.2

I also found a new alternative thats focused more on males, and does even better

[wispy nest]

dope

[wispy nest]

what model is this?

[wispy nest]

thanks, I'm downloading it now!
dreamshaper had an update a few days ago as well. the one you showed and dreamshaper is so close to being the best model I know of, or merge, etc :D

[sterile temple]

a mix i made

[wispy nest]

noice.

[native holly]

Hey everyone, I got a bunch of these sheets (hundreds upon hundreds) that I am trying to use for LoRa training

I'd need to crop them into pieces.. Is anyone aware of a script or software that does this automatically?

Doing them one by one in photoshop is a bit of a pain

[smoky oak]

I am not sure about automatic, but you will need to square them as well

[native holly]

Yeah, doing it manually always results in a rectangle of sort

I think the lora training script automatically scales everything to 512x512

So I dont need to really worry about resolution

[sterile temple]

ask chatgpt to write you a script 😛

[wispy nest]

there's a setting, it looks like it is auto cropping empty (masked/transparent) space

[native holly]

In photoshop or A1111?

[smoky oak]

yeah, it randomly crops them in to 512x512, which will cut out all other parts of the images

[wispy nest]

photoshop, or gimp, or most image editors. Probably not in auto1111, or none I know of anyway

[smoky oak]

@native hollyI do just wanna warn you again, that it will not work how you want if you put in non square images

like this image

it'll just randomly crop a square out, like this

[native holly]

Wait, what is the solution then?

[smoky oak]

you will kinda need to manually make them square

its a pain, I know

[native holly]

But uh, its going to be upscaling them to 512x512 afaik

None of these are larger than 100-200 px anyway

So it'd not crop anything out

[smoky oak]

then I don't think the training will work in general, unfortunately

you're gonna get a lot of blurry, messy, unreliable images

[native holly]

I disagree

[smoky oak]

Its not like... an opinion lol

[native holly]

https://civitai.com/models/45713/sxz-wow-icons-concept

[smoky oak]

I train LoRA's daily lmao

[native holly]

Which was trained purely on icons

[smoky oak]

Yeah, but they weren't super low res or misshapen, I guarantee it

[native holly]

Not exactly high res either

[oak osprey]

theyre square images

listen to people that are trying to help you

[smoky oak]

I mean, got for it if you want to, but I am saying as somebody who has made dozens of LoRA's, its not gonna work with so many bad aspects combined (low res images, messed up crops, overly simple subjects)

[wispy nest]

I'd say just to try it, if it works, then it's great. If it doesn't then you learned something new! :D

[oak osprey]

its not just lora. any training nedds a minimum resolution of 512x512

[smoky oak]

that would be a lotttttt of wasted time

[oak osprey]

no Walach

[smoky oak]

its just flat out not gonna work

[native holly]

If I understand that correctly, it does upscale stuff to 512x512 if its not there

[oak osprey]

it will upscale and make it blurry

[smoky oak]

no, it doesn't ._.

[oak osprey]

and.the unet will overfit on blurriness

[wispy nest]

I assumed they'd make them 512x512 before hand*

[smoky oak]

what it does, is it pixel stretches, not upscale

which will make everything super blurry

its like taking this

[oak osprey]

if it upscales to 512x it does so by conditioning it so the larger side is 512. and then pads the rest with pure white

it will be trash

[smoky oak]

sure, it may be 512x512, but like... lmao

[oak osprey]

but you seem to want to do it anyway and you know best

so just do it

[wispy nest]

just as I said as well :P

[smoky oak]

we're just trying to give you the facts on why it won't work before you spend dozens of hours for nothing, but we can only provide information, not change your mind

[native holly]

@smoky oak

[smoky oak]

and you think that low of an image quality will give you good results?

you are gonna get dirty, messy, blurry, artifact filled results, like I said ._.

[wispy nest]

I don't think the right one is of low quality, it looks more like an artstyle :D

[smoky oak]

.>

I don't consider compressed and muddy an art style personally lol

anyways, whatever, I am not gonna waste any more time trying to warn you, do whatever you want

[native holly]

The art assets were made in 2004

So I'm not exactly expecting HD here

[wispy nest]

it reminds me of one of the remakes of the monkey island serie with the new artstyle :D

[smoky oak]

I'm just a guy who has made dozens of LoRA's for months now, what do I know? 🤪

[native holly]

For what purpose?

Textures?

[oak osprey]

you can pre process them with a gan

[smoky oak]

@wispy nestLet me know how you liuke the new update to the model, I am finding some interesting differences

[native holly]

Huge difference between actual textures and loras that draw characters

[oak osprey]

GPFGAN, RealESRGAN, codeformer

[wispy nest]

I can PM some results? :)

[smoky oak]

Styles, subjects, characters, concepts

And also, it should be noted that I worked on Barium Ai, which was so good for texture generation that we got bought out by Unity... lmao

[oak osprey]

no they arent

[smoky oak]

go for it

[smoky oak]

indeed

[oak osprey]

textures are just the unet

its the same as improving skin or image clarity

[native holly]

Yeah, barium is excellent for anything that tries to be actual realistic in terms of textures, this is way more hand-drawn in style which makes it look blurry

[native holly]

I don't disagree on that, I've no idea how the actual trainer upscales the images

But if its just flat up in terms of pixels

[oak osprey]

it looks blurry because its blurry

[native holly]

then I better play it safe and upscale it myself

[smoky oak]

good luck then I guess

Hope it isn't a complete and utter waste of time

[oak osprey]

you will end up having to downscale all of sd outputs

if it doesnt destroy the model that is

like training it on upsampled 512x makes its outputs all look upsampled

the unet overfits on image artifacts before it overfits on textures

then, all of the outputs have those distortions

i have destroyed more models than you will ever use

[native holly]

I'm just having a really hard time believing that unless its 512x512 that the dataset becomes unusable

I trained a LoRa with nothing but 256x256

[oak osprey]

use controlnet tiles to upscale all of your training images and clean up the noise

[native holly]

And it provides good results

[oak osprey]

it upscaled then

i doubt the images were as bad as what youve shown here

[native holly]

Nah, they were larger tiled textures

Something closer to this

[oak osprey]

yo thats blurry.

and artifact filled

[smoky oak]

yeah

[native holly]

And yet the output is good

[smoky oak]

it looks like its supposed ot be aliased pixel arl, but the compression is so bad it looks like a downsampled and blurred image

[smoky oak]

the output being?

[native holly]

https://cdn.discordapp.com/attachments/268087129131843585/1117875749425254451/image.png <- Original

https://cdn.discordapp.com/attachments/556851372788547596/1117880733982265404/image.png <- AI

[smoky oak]

[oak osprey]

that thing is so janky the textures are the least of its problems

[smoky oak]

I am not gonna waste any more time on this lol

if you really think that 2 people who train daily are just making this up, then go for it, prove us wrong I guess

[smoky oak]

right 💀

[native holly]

I think you've already made your point that unless its UE, its not good enough

[smoky oak]

unless its UE? what?

[oak osprey]

well if the results are satisfactory for your needs who can argue but dont go telling others thats how it works, because you just dont notice what others will

[smoky oak]

^

[oak osprey]

your stuff being upsampled.doesnt matter because your assets can be blurry and no one will notice

[native holly]

So far your advice has been "don't try it, it will not work" even though I've disproven you that it can be used to create useable textures, they might not be perfect but they are by far more superior to the source

Sub optimal results is not same as unusable

[oak osprey]

in fact the game engine even might help you out there with infinityfx etc

[smoky oak]

Even the LoRA you sent is full of issues unfortunately

[oak osprey]

context is super important Azungar

[wispy nest]

I'd say you should try it. The worst thing that can happen is that you learn something new...wait...that's not a bad thing...or is it 🤔

[smoky oak]

we are just trying to tell you that you could spend your time and get wayyy better results with some proper training, rather than trying to brake almost every single rule in training

[oak osprey]

you also dont need to train a lora to upgrade textures

[smoky oak]

That as well

[oak osprey]

you just need cobtrolnet img2img

[native holly]

No I appreciate that and I am not trying to come off as rude or claiming that you guys don't know better, I'd be full of crap if I told you that I know more about LoRas

[oak osprey]

controlnet stupid phone

[native holly]

But you are also telling me to generate 512x512 textures out of my pathetic 50x62 that are blurry to begin with

[smoky oak]

I think that would actually produce even better results than most LoRA's for this

[oak osprey]

well Sytan and i dont mean to sound rude, but we are rude so its a challenge

[smoky oak]

hey, I am being a lot nicer here ._.

[oak osprey]

controlnet tile is out if this world as an upscaler

it pulls off miracles

[native holly]

I gotta be honest, I've never used controlnet

literally don't know what it is

[vapid bronze]

its magic

[smoky oak]

In that case, I suppose the jankier pixel upscaling is better than nothing in the end. The big problem is just that the LoRA will make results that are unusable above anything over that like 60x60 res, and will probably lose all forms of legibility

[oak osprey]

its just not an option for auto upscaling during training yet

[smoky oak]

controlnet reference and tile upscale would be phenomenal for this

[oak osprey]

controlnet tile example takes a 64x64 puppy and makes it 1024x1024 flawlessly

[native holly]

The sheet that I sent you for example, all of those flowers/leaves/assets are just small 2d sheets on the ground, nothing more, not even 3D. The 2D icon Lora I linked you earlier is capable of doing stuff like this:

[oak osprey]

you can take the short controlnet tile example code from their github and have gpt write a loop that uses it to process all your assets automatically

[native holly]

So its the right 2D angle and works in the same fashion

But the style is off

Hence why my idea was to try to train a lora with all of the plant assets from WoW

[native holly]

I'll have to research this a bit more

[oak osprey]

you should upscale them first, still. i promise your results will be EVEN BETTER

USING CONTROLNET ugh caps

[native holly]

in-game render

texture sheet

So I am not exactly working with good assets to begin with

[smoky oak]

honestly, your better bet would be to use non smoothed assets like the ones you just sent

where you can see each pixel, and then utilize them like a pixel art sort of thing scaled to 512x512

[native holly]

What about just culling older stuff and using newer assets that are more "HD"?

to get a model going

[smoky oak]

I think the pixel art aesthetic ones would be the best bet still

like this

but without pixel blurring, so just sharp pixels

like this

[native holly]

Okay so the way to achieve this would be --> Cut the sheets into invidual images --> Use controlnet to loop through all assets to upscale them?

[smoky oak]

A LoRA would be able to train on that

[smoky oak]

if you can get them all pixel sampled like I just showed above, you would have no need to upscale

if you can get all of them like the last one I sent, you would be golden to train a LoRA

[native holly]

Well thats the thing, the older stuff is not like that, newer is

So there's like..

Two generations of assets

But I guess you're telling me to cut off the older stuff in that case

[oak osprey]

anti-aliasing

[oak osprey]

make two

@smoky oak apparently dropping 10% of the captions randomly to make them empty strings helps with CFG training

DeepFloyd: The output from the text encoder is used in two ways. Firstly, it is fed into a projection layer (a linear layer without activation) and used in the UNet backbone of the diffusion model. UNet is a type of neural network architecture commonly used for image segmentation tasks. It uses a series of convolutions and deconvolutions to analyze the image at different scales. The model also uses a technique called controlled hybrid self- and cross- attention, which allows it to focus on different parts of the image and the text prompt when generating the image.

The output of the text encoder is also pooled (i.e., reduced in dimensionality) using attention-pooling and used as additional features in the time embed, another part of the model.

[smoky oak]

Anybody elses discord just hard die for like 5 minutes?-

ok yeah, discord is having regional outages right now

[native holly]

@oak osprey Original

Below, one that I ran through control net tile

Did I do it right?

[oak osprey]

can you set prompt? use."best quality" and strength 1.0

you might also try cutting the sheet and parsing just one item at a time

[native holly]

I used "high quality" but I can use best quality instead

[oak osprey]

sytan is an expert on that stuff

the base model you use also has an impact

digital art centric ones might work better than realism

my realism models add leaves and flowers everywhere

you can also install Ultimate Upscaler extension

[native holly]

Already using it

https://stable-diffusion-art.com/controlnet-upscale/

Followed this guide pretty much

I'll have to try different models

[oak osprey]

ah

nice

yeah for those sheets an anime-centric VAE might be best

[native holly]

I'm running sxzLuma atm

Also thats really nice, by ROI region you mean like

UV's or?

[oak osprey]

https://giphy.com/gifs/netflix-love-happy-ukCFEU6Cg5MCCDoaVN

[native holly]

Ah, interesting. Is your script public by any chance?

[oak osprey]

i made you an asset sheet for Noah's Ark Adventure

that'll be $2500

[native holly]

I'll have to figure out a good way to extract stuff out of these sheets, it'll take way too long to do it manually. rembg was kinda my hope but its not that great with multiple things in one image

I'll Google around, thanks for the tip

[oak osprey]

stable diffusion's text is so bad because the training process randomly flips the images 🤣

thanks for helping them so much @thick anchor that's really cool

if you start burning a model that you're fine-tuning, try rolling the unet back to an earlier checkpoint, and keep the text encoder from the newer one

the textures will revert to a clearer state and the composition will mostly remain in tact

left = 4200 steps at unet and text encoder
right = 11,100 steps at text encoder, 4200 at unet

[native holly]

I have to say, I am not really blown away by the control net tiles in my application, the real world difference is marginal imo

Original

https://cdn.discordapp.com/attachments/1106168245569986581/1118196829331390485/image.png

Upres one

https://cdn.discordapp.com/attachments/1106168245569986581/1118196820070383616/image.png

[oak osprey]

oh dude that looks way better to me

@smoky oak look at that

it's like you remastered the game @native holly

[native holly]

It is

[oak osprey]

you can tell the textures are way better, does it use a lot more vram?

[native holly]

shitload

[oak osprey]

perfect

then you HAVE remastered it 😄

the 3D models need more tesselation but you can't really do anything about that. maybe bump mapping can help

[native holly]

Yeah, I kinda need a roi script or something to extract hundreds upon hundreds of sheets

[native holly]

I mean I didnt change the texel density

[oak osprey]

oh i know

[native holly]

and lol, I wish I had access to bump map or height map

WoW doesnt even have emissive textures

[oak osprey]

depth map exists in controlnet

😄

[native holly]

oh simulated one

[oak osprey]

yes

[native holly]

i'm not cutting all of this

by hand haha

[oak osprey]

python-opencv will do that for you

[native holly]

Oh? I was told that it'd be better to train a LoRa by invidiual assets

instead of sheets

[oak osprey]

i think they were assuming you're just upscaling each sheet rather than training a LoRA

[native holly]

Ah yeah, just for pure upscaling you dont need to cut them up

Its better not to

[oak osprey]

it is correct

[native holly]

for sake of UV's

I know, wow-tools, wow-export

cascview

@oak ospreyany idea on how to reduce some of the noise on control net?

Input

Output

That sounds really awesome but I don't have the faintest on how to achieve that

[wispy nest]

you don't learn programming, you survive it ;P

[oak osprey]

prompt, denoising strength, base model, step count resolution, tile size, etc

[native holly]

Once someone teaches stable how to deal with transparency and not to fuck up on texture sheets with padding, I'll consider it :P

Beyond that, unless it's a square texture, its still in too early stages to do much

[oak osprey]

use GPT4 without the Browsing capability (that thing sucks) and it can absolutely teach you Python

[native holly]

Yeah but it doesn't grasp UV maps at all yet

Oh right now I am just doing 2d plant sheets and maybe a LoRa for those, I had a good success with a texture LoRa already

[oak osprey]

https://huggingface.co/spaces/ronvolutional/sd-spritesheets/blob/main/app.py

uhhhh that's cool

[native holly]

https://cdn.discordapp.com/attachments/268087129131843585/1117875749425254451/image.png

https://cdn.discordapp.com/attachments/556851372788547596/1117880733982265404/image.png

I'm an admin on a private server

So new assets

I am crap with texture assets so using stable to supplement that

[oak osprey]

didn't know we consorting with a hardened criminal

[native holly]

Actually, running a private server is illegal, I'm not running it

just helping develop

[oak osprey]

conspiracy cough

[native holly]

Not a clue what BONS is

ah, my condolances

must be a stressful job

[oak osprey]

remember "Bowser"? that guy that made mod chips for the nintendo switch? well he didn't make the mod chips, he merely made their website. he was the only one that had jail time

his marketing approach was what led to the success of their company so they said it wouldn't have been possible without him

also it's hilarious that a guy named Bowser was taken down by Nintendo

and then the gymnasium floor opened up and everyone fell into the pool during your school dance?

(reference) https://www.youtube.com/watch?v=2rha-6qG4OQ

fuck, i'm old

you had a modem at home?

[native holly]

I trained the LoRa on various assets, depending on the race

[oak osprey]

i had to go into the university lab to use an academic network and have a reason for doing so!

[native holly]

Bloo elf on blood elf assets, human stuff on human

So style I guess

[oak osprey]

we walked uphill to get to the internet. both ways!

[native holly]

I was told that I should get the plant stuff into high resolution before I even attempt training with them but

I do feel like you lose the style if you do that, so its a bit problematic

[oak osprey]

no one really knew what you were trying to do

usually, loras are used to add stuff into, you know, traditional 2D images

everything we said is true for that use case

it changes when your assets end up downsampled and texture a 3D item with

[native holly]

@thick anchorhttps://civitai.com/models/53858/sxz-texture-bringer-concept

So this was the inspiration for me

[oak osprey]

poor kitty cat looks toasty

[native holly]

But it was trained with a very small dataset

(50 images or so)

You mean to generate new things?

Aye and I've had good success with it

Especially img2img

Just need to give it more stuff to work with

But manual tagging is very tedious and the dataset is large

texture of (short description of what its seeing, rock wall, wooden floor etc.)

I also got an activation keyword on all of them

Something I have to mess with

https://cdn.discordapp.com/attachments/1015497640659664936/1117533771244257310/image.png

Quite happy with how it blends in with older stuff as well

Its more of the "new" WoW style

Aye, so far I havent even fed it anything old

And I dont think I should either

I refuse to believe that someone actually hand-drew this

I dont even know what those do haha, I am still very new to all of this

In fact I think AI as a tool for textures in general

is very new uncharted territory

theres barely any information anywhere

Ah I think I used uh

BLIP

But it was honestly pretty bad with textures

Very inaccurate

https://cdn.discordapp.com/attachments/556851372788547596/1117068068049395752/image.png

[oak osprey]

well i use BLIP prompts exactly as they come out because i'm trying to fine-tune the model's current understandings with new styles

😄

but i use ViT-H-14g because, SD 2.1 training 4 lyfe

[native holly]

all of this sounds like a lot of python

which I am allergic to

[oak osprey]

this is something you will likely have to overcome, one way or another, or be continually subjected to the limits of reality and waste more time trying to bodge together workarounds and "workflows", most of those only exist in so many steps because the people creating them aren't developers

Deforum is great because it's created by developers. before that existed, users were using several things for tasks they weren't designed for

[versed blade]

画一只小黄狗

[oak osprey]

yes, but it's much better than when people were simply batching folders of img2img or using ffmpeg

[versed blade]

draw a little yellow dog

[oak osprey]

i think you're missing my point

learning how to code in python makes using these tools (and fixing/improving them) easier without resorting to hackish workflows

[rocky quail]

looks like i need a new ssd

[wispy nest]

https://www.youtube.com/watch?v=BgxklT94W0I
I swear I'm not posting this video because I love the humor. I love the humor AND the situation when to show it :P

[rocky quail]

they got 20tb ssd?

so you just copy the whole thing over?

how much have you filled up now?

[wispy nest]

it's not how big it is but what's on the inside...wait that didn't sound how I wanted it to!

me too! well, the horder thing anyway. I got stuff still from 2002! 😵‍💫

but I thought my folders holding about 2TB of checkpoints were a lot...well it might still be, but at least I now know I'm not the one with the most space wasted saving them for a never going to happen situation :P

I have moved, not deleted, about 400 checkpoints in total to another folder
I'm 99.9% that I will NEVER use them ever

but I still didn't delete them >:I

[oak osprey]

i used to make a storage appliance that was in use by companies like earnst and young and i've seen some shit

site reliability engineer

customers are basically my enemy because it's like none of them want their sites to stay stable

[wispy nest]

[oak osprey]

i'll never see steve buscemi the same way again

[smoky oak]

Sent this before, but wanted to send again lol

[oak osprey]

it's a bad format

lmao

are you pro user? pro Photoshop? anti Photoshop?

[wispy nest]

lol perfect. and fits one irl friend like a glove

[smoky oak]

What? Lol

[cyan snow]

i merged a new model =]

[wispy nest]

don't you mean captured them? ;P

[cyan snow]

dude, it's simmilar to what i used before, except here almost 100% of generations are the quality of the best generations from the last model i used

[wispy nest]

it was a joke (play on words) on the image where you "captured" the galaxies in a jar. And I joked that you didn't merge them, but captured them in those jars :)

[oak osprey]

idk man reading that back now, new hour new me, none of it makes sense anymore

[smoky oak]

You are clinically insane

[wispy nest]

this is the difference of using 'highly detailed face'

[oak osprey]

[wispy nest]

otherwise its all the same settings and prompt

[oak osprey]

highly detailed doesn't work the same in every model

for some reason, intricate detail could result in crazy patterns

[wispy nest]

im using revAnimated, pretty good model and those two images are from that

i have intricate in too

[oak osprey]

best quality becomes animated for my models

if you want the best real people photos, try adding a particular year, a real world location, and the terms ethnographic photography

[wispy nest]

there is another mod i've found that's pretty good, by the same author of dreamshaper, called AbsoluteReality

[oak osprey]

add (deep fried) to your negatives and see if that helps with faces at a distance. it does for me. it fixed faces in groups, too.

[wispy nest]

lol interesting

and oh btw, i used DDIM 15 steps

[oak osprey]

you can quickly add it to the positive prompt to check if it'll help, first

if it destroys the image, it'll help

[wispy nest]

would you suppose this is the result of ((deep fried)) in the positive?

or this ?

I'd say that if you don't spot what the negative prompt did, then it's not needed in the first part :P

not sure i see it, what is it?

[oak osprey]

yeah, exactly

1.5 models don't benefit from that negative as much, they do not have it in their dataset usually

the Deepfloyd team created that dataset

[wispy nest]

there's a lot of placebo prompts out there, both positive and negatives. I'd say it's better to add stuff you know does something, than adding more and more unkown prompts because that might cause one to miss a lot of good results because they use too many prompt words they aren't sure what they do

[oak osprey]

placebo might actually work on one model but not another

[wispy nest]

yes, just like placebo would function :P

think there are some common words often used for negative prompts for specific genre, i try to use textual inversion for that, but i'd hope SD would include those as built in feature

yeah, just know that adding a prompt word always does something but the, in my opinion, the important thing is to know what that something is :D

[oak osprey]

it reduces the probability of that block's weights impacting the result

[wispy nest]

some use a default template for negative prompts, i used to do that too lol

like putting in bunch of words as a fixed recipe for good output

the important part is that one understand and enjoy making art. We all have different styles of prompt magic :P

yeah agree, im better with using prompts now that i think what i want the output to be

[oak osprey]

a lot of abuse of negatives is because people are using others' models that they have no idea waht was trained on

with my own model was the first time i was like, okay. let's go without negatives at all. but you have to put too much into the positive prompt all the time then, like, 'best quality 8K UHD sharp'

so having a good negative prompt that gets consistently clean results on a single model is a pretty incredible thing. in 2.1 the negatives are far more powerful than 1.5, where you can try adding "child" and "nsfw" to the damn thing as much as you want and it just won't help

[wispy nest]

i'm suspective of 'best quality' and 'masterpiece' i mean that's like saying there couldn't be anything better

[oak osprey]

and best quality isn't tagged on the best images

ethnographic i have found, is

idk why people think cropped is going to help either. lmao

[wispy nest]

I think best quality masterwork/piece started with the leaked NAI (can't remember what it stood for currently :/ ) but I'm not 100% sure of it as I mostly use them for models merged or comes from that model or something similar

[oak osprey]

if a tag says it's cropped, it was done in advance. not randomly

and not-in-frame. what a pipe dream

[wispy nest]

yes also there are so many variable factors from mods to mods, it'd be so good to have a generic inbuilt algorithm with SD

[oak osprey]

yeah making a tag cloud from the text encoder would be sick

[wispy nest]

fingers crossed

[oak osprey]

?

🫰🏽

[wispy nest]

that'll probably happen, one day, any time now...probably :P

maybe in 6 -12 more months SD will have features that are lacking now

[oak osprey]

yeah well SDXL still sucks at faces at a distance 😄

[wispy nest]

I believe that's because, well a billion things, but the main ones are probably the "small" sizes of the trained images and then the compression or what it's called when it gets turned to 64x64 noise pattern or what that also is called :P

im optimistic, this tech is still in its infancy .. so given a year or so things could change big

[oak osprey]

they've added stuff to its inference loops to optimize performance and as far as i know, shouldn't have the same issues as SD 2.x and 1.x

1216x832 is an odd resolution they selected

[wispy nest]

that sounds good, but im still using 512 base version

[oak osprey]

don't think i've ever tried the 512x model for 2.x

unless one of the 2.1 models is based on that.

[wispy nest]

and from how good quality 512 is possible of, when we get even larger default sizes, then life will be wonderful!
Maybe they will release that "two step sampling method" or what it was called at the same time? ;P

two steps sampling method?

[oak osprey]

i think you're referring to a distilled model

can effectively create an image in 5 steps or so

[wispy nest]

it was something the SD team talked about last year I believe. It was supposed to be around the corner, or maybe it was "soonTM" but from my knowledge it's still not released. Not that I've looked so I can, and probably is, wrong :P

oh wow

that sounds too good

[oak osprey]

it is too good to be true

the process of distilling the model results in a loss of generalized capabilities

lower prompt coherence, and so on

[wispy nest]

good logical points

but evne if the steps are reduced to 10 that'd be great

https://arxiv.org/pdf/2210.03142.pdf
I believe this was their documentation on it at the time. The only reason I remember it is because I bookmarked it and forgot to remove it :P

im using 15 steps in DDIM when i try to sketch out a concept art, and sometimes that 15 steps can produce pretty good images

[oak osprey]

a perfectly trained model is likely efficient enough to do that. but it depends on whether they've used traditional sequential freeze, block weighted merging, and other fancy tricks. honestly i assume they've used all of the tricks in the book, but when i mentioned SmartFRZ to TwoDukes, he hadn't heard of it yet

i can get really good images from my model out of 10 steps but when the prompts aren't as well learned (eg. plenty of unseen data in the training set) it will tear the image to hell

[wispy nest]

well sd is already capable of decent outputs with like 10-15 steps

just not high enough for quality standard

[oak osprey]

but the models that do that have reduced generalisation

they don't "Listen"

still it works for some needs

if i had an API that needed to fire out consistent image style, using a distilled model is a no-brainer

[wispy nest]

hugging face recommends 50 steps max but the changes are hardly noticeable over 30 steps, so maybe those 30 steps could be reduced to 15 with new features?

[oak osprey]

those are using PNDMScheduler or DDIM. and a lot of noise schedulers end up converging before or around 50 steps. so the number make sense. but Euler A never converges

[wispy nest]

it's another one of those personal style things, some people want it at X, and other feel it's better at X+Y, etc :P

yeah none of the ancestrals do

[oak osprey]

it can kind of repeat concepts every X steps but it'll generally keep changing

[wispy nest]

i think i also heard somone say the app could reduce vram usage too ...

[oak osprey]

what app

[wispy nest]

i mean SD

like some beta version ran on 2gb

btw im surprised with DDIM 15 steps when i always used DPM++ 2m karras with 30 steps most of the time

[smoky oak]

I know several people who use SD on 4GB

[wispy nest]

going to see the difference between DDIM 15s and DPM++ 2m karras with 30s on same seed

im using it on 4gb too now

15 steps ddim

ah, the headphones prompt :D

it's a prompt i made up girl sitting on a cloud of swirling psychedelic colors. she is wearing a large headphone, rocking to the beat of music, and her long lush golden yellow hair flowing in the air.

haven't used any fancy mod for the effects yet

[oak osprey]

DDIM i couldn't even use on 2.x til recently

it would sample from the leading timestep instead of the trailing one. and it wouldnt work with uneven timestep distribution. now, both are fixed

but yeah now that it works it's incredible

[wispy nest]

so this is the dmp++ 2m karras 30 steps in comparison

im not going to hate on ddim 15 steps for what it can generate

[oak osprey]

me n sytan lookin all cool n shit, with our trained models, but me lookin a lil cooler because im training 2.1 and thus, taller

@smoky oak it is you

[smoky oak]

Bruh

[oak osprey]

did i mess that up

[strong sonnet]

I am training the model with around 1k images, I saved a ckpt when 20% of the training was complete, can someone test?

I will share gdrive link to download the checkpoint

[dense tapir]

@smoky oak dig this nonsense.

[smoky oak]

Bruhhhh

[dense tapir]

I am SMH at that even if you paid for it. wtf?!?

[dense tapir]

DDIM 20 steps.

with my lycoris

[slim lark]

Is that doctor strange

[oak osprey]

Toronto

[slim lark]

How do i describe this

[oak osprey]

clip interrogator

[oak osprey]

https://huggingface.co/nousr/conditioned-prior @smoky oak well this is slick

[smoky oak]

Anyway to like summarize it super fast? I am getting ready to leave for mexico

[oak osprey]

LAION is reproducing DALL-E2 as open source

that's about as quick as it gets

[smoky oak]

Is that the whole point of that link?

Just to basically say that they're making dolly too but open source?

[oak osprey]

yeah and how it works and why it works

[smoky oak]

Very interesting, thank you for the summary

[sterile temple]

morning peeps

[oak osprey]

robert downey backstreet boys junior

i got a real place in Toronto

the subway station on the left has been replaced now and there's no wonky street posts but the rest looks right

[oak osprey]

portrait of person with bed-head wearing bathrobe, downtown Toronto, 1984, Kodachrome, ethnographic photography, highest quality, UHD

they look so well-posed

[oak osprey]

i dont think it understands bed-head other than "put a bed on their head"

[sterile temple]

more of a fur rug than a bed 😄

[oak osprey]

like dog bed

[sterile temple]

does the keyword 4k make the image look pixelated? I think I found that when I tried 2.1 models in the past

[smoky oak]

No they are not, you are not welcome here

Please, keep your discrimination out of the chat

[wispy spindle]

Banned

[sterile temple]

loving these Chrysanthemum portraits

[oak osprey]

on my model?

my unet is different

[sterile temple]

yes

[oak osprey]

it does not do pixelated images

i trained the unet on a super high res subset of data and then froze it and focused on the text encoder

so it can do film grain or smooth images

base 2.1 unet with the 4k keyword for a coral reef, vs the tuned/frozen unet

best example of the elimination of artifacts

thanks to @smoky oak for pointing that noise out so i couldn't unsee it and had to find a way to get rid of it