🏞｜general-with-images

[smoky oak]

I might be able to

[oak osprey]

it's a hard one in my experience

[smoky oak]

Let's see what I can do

[oak osprey]

LMAO i love it but that's a hybrid

[smoky oak]

It's not perfect, but it's not terrible either lol

[oak osprey]

are they really the same height?

[smoky oak]

I have no idea

I do know you could easily do this with prompt zoning, but I have never done that before

[oak osprey]

img2img i love using as a base for composition too. seems to make it a lot easier for the output to work well

i used this

[bold cave]

use latent upscaler

but draw the blobs waaay far from each other

otherwise they end up fusing

[oak osprey]

draw?

[bold cave]

yh

[oak osprey]

blobs

wat

[bold cave]

thats how latent coupling works

u draw a black blob and a red blob and they become ur subjects

[oak osprey]

i need a way to make an app inside discord to generate masks

[bold cave]

I thought u were a master coder

[oak osprey]

i have no idea how discord works

[smoky oak]

there are several ways to do this sort of thing, some of them are even easier

I just need to find the videos on how to set it up

[bold cave]

latent coupler is the most effective yet its still ass. I say it has about a 10% success rate

but when it does it well, it does it well

[smoky oak]

There is the one that does the ratio boxes, and that one works extremely well

[bold cave]

its the same one

[smoky oak]

I just need to find it

[bold cave]

they just changed the interface

[smoky oak]

and you're saying is really bad now?

[bold cave]

you mean this one?

[smoky oak]

I think, I do not remember

[bold cave]

it was never good, at least a lot of ppl say it worked at random

[smoky oak]

all I know is it works really good, so I am not sure what this 10% success is

[bold cave]

hmmmm

cuz ive roamed reddit threads searching for something better

and couldnt find anything. something better than this would be revolutionary

[oak osprey]

https://github.com/JamelHammoud/paint-bot

whoa this is kinda neat

[smoky oak]

Alright, lets see if I can find the video

[bold cave]

I'll put it to test right away with my brand spankin' new dreambooths. Im doing about 10 subjects per model to save space, I'm just hoping it works

[rain gazelle]

evenin'

[smoky oak]

howdy doody

[smoky oak]

What is this extension called exactly?

It should be called Latent Couple, right?

[bold cave]

yeh

its the same one u thought of isnt it

[smoky oak]

it says I have it installed, yet its not in my UI

[bold cave]

I did have issues with it not displaying

I think I just updated the whole thing and it showed up, or I did something else that I can't remember

[smoky oak]

the downloader says its installed, but A1111 doesn't show that its installed

yeah, its not in my extensions folder

WTF

oh wait

its named "Two Shot" not Latent Couple

Yeah, cause that makes sense

alright, so the extension is bugged

fun

yeah alright, its a broken extension

great

I'd love to mess around with Latent Couple, but the extension seems to be broken at the moment

[bold cave]

umm no its not

it works on my machine at least

so is the good one supposed to be two shot?

fuuuuuuuuuckkkkkkkyeeeeessss

this upscaler is pure magic dog I'm lowkey tearing tf up

well it upscaled her so hard her waistband turned into hair but still. it's orgasmically good

[oak osprey]

OHHHH YOU HAVE TO USE A VAE

the mse one is pretty good eh

[sinful mirage]

🤔

[smoky oak]

https://tenor.com/view/fire-kill-fuego-matar-incendiar-gif-5307315

[smoky oak]

HMMM lol

what

[wispy ether]

Hes behind the beat

Bear

🗿

[smoky oak]

wow

literally no need for jokes about that, but anyways

[smoky oak]

Realism gang rise up

[normal spruce]

Nailed it 😎

[dry crow]

The 2.1 768 model needs a base resolution of 768x768
Also use some negative tags like:
blurry, deformed, low Quality, mutation,

[normal spruce]

Good catch.. I had that set before however was trying to resolve an issue with xformers dependencies being installed so had to restart.. That probably is my issue 😂

Oh hey, would you look at that.. Thank you

[dry crow]

Hey that looks really nice

You can also adjust the resolution to height: 1024 widht: 768
To get a portrait view.
Then say photo of instead of portrait

To get a similar image to the other one

[normal spruce]

So when you say that model needs a base resolution, I can change the resolution to generate an image of an aspect ratio of whatever I like, so long as they're what... Greater than, or multiples of the base resolution (768)?

[smoky oak]

generally speaking yeah

[dry crow]

Yes

[normal spruce]

That makes sense.. Thank you

[smoky oak]

you want the smallest side to be at least 768 for a 768 model

[dry crow]

But dont go to high over the base resolution as you will then get duplicates or artefacts

[smoky oak]

you can go a little lower, but best results will always be at the 768 sweetspot

[dry crow]

For 1.5 models its the same case but with 512x512

[smoky oak]

yuppers

[normal spruce]

Yeah, so stick with a scaled/multiple of the base resolution? Then use upscalers to generate a higher resolution, yeah?

[smoky oak]

no need for multiples, just try to stick as close to 768 as possible

[normal spruce]

Okay, I gotcha.. Thank you

[smoky oak]

no worries

[normal spruce]

Nice to finally put some work on the 4090 I spent all that money on.. Never seen 23.9GB/24GB of VRAM usage.. Finally got the tensor cores doing something other than DLSS for gaming

Does "restore faces" generally do much on txt2img versus img2img when rendering human faces? In fact, does it help with animals/whatever else too?

[smoky oak]

I recommend never using restore faces

[normal spruce]

Just asking based on the lopsided eyes from this being generated:

[dry crow]

Its only for humans. I would say stick to highres fix. That will fix your faces too

[smoky oak]

generally speaking, its worse than just generating or inpainting now

[dry crow]

Yea Inpainting is also very good

[normal spruce]

Is there any difference between running the high-res fix during the initial passes, versus putting that through the upscaler afterwards?

[smoky oak]

something like that can be fixed super easy with inpainting

[normal spruce]

Sorry, I could probably Google half these questions

[hearty karma]

o.O

[normal spruce]

Will Google, thank you for the pointer

[dry crow]

Hey no problem if you ask here or in #📝｜prompting-help

[smoky oak]

high res fix is basically an img2img process with a pixel upscaler inbetween

its a great tool, but you can go higher res with more control using an ultimate upscaler, specifically the new control net one

My workflow is usually generating a ton of bases, refining with high res fix, then ultimate upscaling from there

an example of the new ultimate upscaler process with controlnet is this

base

4096x4096 upscale

(this is a screenshot of it)

[normal spruce]

Wow.. Okay, that looks a tonne better.. That's awesome man

[smoky oak]

but the detail is for sure there

[hearty karma]

is the ultimate upscaler an extension?

[smoky oak]

It should also be noted that I have 100's of hours with upscaling, and 10,000's of upscale generations

[normal spruce]

Sheesh.. Yeah.. That looks like a 48MP DSLR image 😲

[dry crow]

@smoky oak is that done with the Controlnet tile + ultimate upscaler?

[smoky oak]

and I am also kinda one of the resident realism experts, so there are a lot of things playing in here

[dry crow]

I got worse results using both together

[smoky oak]

yeah

I have done 1000's of upscales with just ultimate, and the new controlnet works wonders at super high res

[normal spruce]

Do you have any guides on what you're using to generate that sort of stuff? That's kinda my goal, but I'm just scratching the surface and still overwhelmed with sliders and settings 😅

[smoky oak]

some of my show off gens

Realism and artistic realism are my forte

omg, a super react haha

allow me to share some more really fast

these are less impressive as they do not use the same process, and are from when i first started my venture into super realism

[normal spruce]

Could you do much if you trained on self-portraits, based on that?

[smoky oak]

those first 2 are with my new workflow

[smoky oak]

potentially, yeah

[normal spruce]

Aside from the goatee on that kid with his dog, I'd have a hard time telling that apart from a professionally capture DSLR image with a bit of photoshop on the end

[smoky oak]

Thanks, I am getting better and better over time

[normal spruce]

Yeah man, you're smashing it ❤️.. That's awesome

[rich creek]

Is there anywhere here I can hire someone to make a Lora?

[smoky oak]

I would be happy to do it, depending on the subject

I have experience with a wide range of character and subject LoRA's

[rich creek]

Do you have a portfolio or anything I can look at?

[normal spruce]

@smoky oak , also crazy how nice that renders.. The Asian kid for example, you get proper bokeh but without the software-processing issues around the strands of hair versus the background seen in iPhones, etc.

[smoky oak]

I can share examples in DM's

[smoky oak]

Its a lot of clever prompting haha

[normal spruce]

@smoky oak , do you have any guides on what you're using to generate that sort of quality? Assuming you're willing to divulge your secrets 😂

[smoky oak]

A good model is one of the biggest things

[normal spruce]

.. I'm listening

[dry crow]

I would say if your new to SD try out the basics first. Then go into upscaling with highres fix or the ultimate sd upscaler extension.
Try other models from civitai.com

There is definitly a lot to learn 🙂 and at start its overwhelming

[hearty karma]

how do i use the upscaler i am confused on where it is

[dry crow]

In txt2img its the checkbox "Highres Fix"
In img2img its under scripts at the bottom called sd upscale script or by installing the extension sd ultimate upscale

[hearty karma]

oh wait really? thats what highres fix is? thats the ultimate upscaler? huh... ive been using it all this time without realizing

[dry crow]

No thats highres fix. The Ultimate upscaler is an extension for img2img

[hearty karma]

oh

how do i use the extension?

[static tusk]

hires fix is for txt2img, UU is img2img processing

[hearty karma]

oh...

[static tusk]

you can find it under the Script dropdown once you install the ext

[dense bridge]

does anyone know the keywords for hairstyle like this ?

[lone storm]

I would try “shoulder length brunette layered hair cut”

[dense bridge]

thanks,it sorta works but i guess there just isnt enough pics with that hairstyle in the dataset

[digital hornet]

anyone know why you don't get the VAE option in webui v2.2 (it's there in v2.1 next to model)?

[hearty karma]

maybe vaes only work on certain models?

¯_(ツ)_/¯

[static tusk]

settings > user interface > quick setting add
sd_vae,

[digital hornet]

that did it, thanks

[wispy nest]

what's vae?

[static tusk]

its a model that enhances the final colour and lighting of a generated image

like this https://huggingface.co/stabilityai/sd-vae-ft-mse-original/tree/main

some models have their own specific vae files. the above link is the general purpose one

[sterile temple]

where can I find other vaes to try?

[static tusk]

on the civitai description page it will mention what vae files are recommended. Depends on the model, but you wont need anything other that the one above unless its specified

[sterile temple]

I've been playing around with mixing models, and the pictures are coming out a little lack luster. Not as colorful

Thought maybe if I tried a different vae it might look better?

[static tusk]

faded images is vae related yeah. go to the page of each of the models on civitai and see what is required. If it has only started since your merge then its a bad merge

[fleet olive]

Hello, I want to make this 2D picture to a 3D picture, Can i do it with SD?

Any aces can help me with this?

I have been finding solutions for days

[sterile temple]

Have you tried controlnet?

[fleet olive]

No, I don't know that yet

[sterile temple]

I'd recommend finding a youtube video tutorial and giving it a go

[fleet olive]

I mean 3D picture just like this

I have found youtube about how to do it

You meant use contronet?

[hearty karma]

upscaling using ultimate upscale takes a long while

[sterile temple]

yeah, use a model like the canny with your 2d pic

[fleet olive]

@sterile temple ok, Thanks, I'll try it

[hearty karma]

4k spider-woman image

[neon crystal]

More Realism, Collage and general weirdness experiments

[smoky oak]

oh my goodness gracious, I made a wonky new sound

I like this sound wayyyy too much lol

In context in a small demo

@glossy herald Little synesthesia sound for you

[glossy herald]

Thanks mate !

[dense tapir]

That is the old Courtney Cox hairstyle so 1990s.

[hearty karma]

ah yes

perfect face :D

[static tusk]

@glossy herald not sure if you saw this from last week or so #🏞｜general-with-images message

[dense tapir]

Damn, that one came out hot, lol

Added more of my special sauce.

[hearty karma]

ah yes an old gen i found

[dense tapir]

Use it and revamp it. I love doing that as the tech advances.

[hearty karma]

hmm i will try and do such things

[dense tapir]

It is fun. Sadly I just lost all my old saves 😦

[hearty karma]

yeah i kept all my old gens in a folder

and that folder has 3000 items in it

[latent atlas]

https://www.instagram.com/p/Cr-0hvtIo98/?igshid=NTc4MTIwNjQ2YQ==

veiled.vision.studio

"Courage is not the absence of fear, but the triumph over it."

Instagram

[dense tapir]

Mine was just as many. The only thing I could save was the old prompts that were their filenames so some were too long to have.

[hearty karma]

i actually did another prompt

and redid it now

this is the result of redoing the prompt

i half expected the prompt to not work that well

[dense tapir]

That leg is a hard pose to get

[hearty karma]

i should probably switch my ui settings from med vram to low

[dense tapir]

I had to do that as the tech grew only my vram didn't

my ram grew though

[hearty karma]

my ram has grew as well

i used to have 16

[dense tapir]

same

now have 48gb

[hearty karma]

i doubled to 32 and if i desired i could double again because i have two more slots :)

[dense tapir]

same

I could remove the 16 and go for another 32 so 64 vs 48gb

I do know my new system will begin with 64gb

[hearty karma]

altho one downside

i have my os on a hard drive...

._.

[dense tapir]

OUCH

[hearty karma]

which is where my photoshop also is

which is why it crashes 5 times before launching properly

[hasty nova]

I have 32 on my PC, but I did win the computer for free. I desperately need a bigger case becauseit originally came with a 1070 and we upgraded it to a 3070 which is much bigger and restricts airflow (I just took a side panel off to make sure it stayed cool enough)

[dense tapir]

Oh, I saw a vid with crystal disk mark and 5 Seagate HDD (2tb each) in raid 0 made 10TB and was as fast as my NVME M.2 gen 3

[hearty karma]

tbh i might get an nvme or something

put my os on that

[dense tapir]

that is what I did years ago

My b450 is too funky to use the other slot so I am going to get 2 1tb SSD sata and raid 0 them.

[hearty karma]

you got any recommendations for what brand is good?

[dense tapir]

Samsung

if the price difference is more than 20 USD get WD black

I use WD Black

My WD Black was 3.3-3.5GB/s on reads and not much less than that on writes.

Gen 5 drives are almost 14GB/s

[hearty karma]

i can get a 250gb samsung m2 for £41

surprisingly cheap

[dense tapir]

They really are now

I use the 512GB WD Black cause it was over 20 less now I would get a Samsung 1TB for the same cost

[kind quartz]

ssd with cache is important thing as well, but bit more expensive

[dense tapir]

No way would I buy any SSD without dram

I had one I used in my mom's system and boom it almost went to zero on a really large file. For what she does it was fine.

dram is a thing in M.2 as well

[hearty karma]

would having sd on a hard drive affect its loading performance in anyway?

[dense tapir]

yes, and why I am buying for the raid

[hearty karma]

ah...

[kind quartz]

I have 500GB samsung with cache and cant complain, it can spare some small grey cells 🙂

[dense tapir]

my HDD, as it loads, is at 100%

[hearty karma]

i really cant wait to get this hard drive replaced 😭

[dense tapir]

takes 2m to load

when I had it on my SSD took less than a minute. On my nvme it was 20s

[kind quartz]

m2 as well

[hearty karma]

another old gen

[dense tapir]

If I had the cash I would SSD 4 in a raid 0 as that would be about as fast as my NVME M.2

I'll do 2 and see how it goes, but I sure wish B450/550 wasn't so funky with the other m.2 slot

A 570 doesn't have this issue

[hearty karma]

which m.2 slot is the funky one?

on b450

[dense tapir]

2_2

2_1 (the top one nearest the cpu) is fine

[hearty karma]

so the one that would be under most gpus then

[dense tapir]

yes

[kind quartz]

i have as well b450 🙂

[wispy nest]

oh okay

[dense tapir]

love the board except for this

[hearty karma]

i think i can only get a gen 3 m2

[dense tapir]

You know even using the 2_1 it removes two SATA connectors? Yep, and why I went with six so I still had four left.

all b450 can do is gen 3

[kind quartz]

https://www.samsung.com/us/computing/memory-storage/solid-state-drives/ssd-970-evo-plus-nvme-m-2-500gb-mz-v7s500b-am/#specs
this is ssd i have.

[dense tapir]

that is the SSD I am going with but in SATA 1TB. Two in raid 0 to make 2TB and speeds half of my gen 3

nvme m.2

[hearty karma]

oh wow

only took 11 minutes for automatic 1111 to load up

:-)

[dense tapir]

damn, what are you loading?

Takes around 2-3m for my HDD

[hearty karma]

like do you mean what model?

[oak osprey]

ah I am spoiled with my 4090 system that has 16tb of nvme and 128g ddr4 and a 5800X3D

[dense tapir]

yeah

[hearty karma]

digital diffusion 2

[dense tapir]

that one loads really fast for me

[hearty karma]

yeah im on a pretty old hard drive..

[oak osprey]

maybe he's loading it in fp32

[dense tapir]

sounds like a 5400rpm

[oak osprey]

haha

[hearty karma]

whats a fp32?

[dense tapir]

DDV2 is fp16 only

[oak osprey]

32bit precious

oops precision

[dense tapir]

Snoop

you have 32GB, tell ya what I now tell it to load into vram 4 checkpoints and it helped. Especially when I switch back and forth.

the next time I load it the time is seconds

I mean the entire auto too

[hearty karma]

how do i do that?

[dense tapir]

In settings

[hearty karma]

is it the checkpoints to cache in ram setting?

[dense tapir]

yes

under stable diffusion

[hearty karma]

alrighty set

[dense tapir]

Switching models is so much nicer now as long as you switch to the same ones.

I was doing some funky stuff like generating in 2.1 then fixing the gen with 1.5 models then back to 2.1

Without them in ram it was torture

[hearty karma]

do you know if there are any fixes to get it to stop installing and reinstalling stuff everytime it loads

because my install seems to love installing and reinstalling stuff

[dense tapir]

Mine doesn't do that unless venv was removed

[hearty karma]

and oh crap... it was a hell of a lot faster

[kind quartz]

if you have git bash in webui user.bat, delete it

[hearty karma]

19 seconds

[dense tapir]

told ya

yeah, no git pull in webui-user.bat

[hearty karma]

this is the only stuff in my .bat

[dense tapir]

weird

set COMMANDLINE_ARGS=--theme dark --medvram --xformers --disable-safe-unpickle --port 9000 --api --opt-channelslast --always-batch-cond-uncond --skip-version-check

that last one, add it

might help

[hearty karma]

not for me ig ;-;

showed up anyway

ill probably have to live with that for a while

[dense tapir]

I wonder if you HDD is dying?

[static tusk]

why do you have --no-half?

[hearty karma]

1660

[dense tapir]

I was going to ask about that myself

no need with --xformers

[hearty karma]

oh

🤔

[dense tapir]

any memory attention but no mem attention you have to have it

[hearty karma]

so i can completely erase --no-half from my .bat?

[dense tapir]

yes

[static tusk]

yep

[hearty karma]

wow my models are actually loading faster

[dense tapir]

Of course

try a gen

ditching that and using ram you will feel so much more comfortable.

[hearty karma]

woah...

im only using 2/3 of my vram now for gens

[dense tapir]

yep

[oak osprey]

so you were loading it in fp32 then? that's funny

[dense tapir]

you were, artificially, taking a fp16 and doing fp32 in the card

[oak osprey]

I called it

[hearty karma]

whoops ¯_(ツ)_/¯

[oak osprey]

yeah that doubles your vram use

in general even an a100 80g struggles to do fp32 models

the 4090 doesnt

[hearty karma]

and i was accidentally forcing my little 1660 to do it for long durations of time

[oak osprey]

tragic

[dense tapir]

SMH, lol

welcome to a new world

[oak osprey]

use enable CPU offload to get even more out of it now, in addition to xformers efficient attention that will let you run even larger gens

[hearty karma]

👀

how do i enable cpu offload

[static tusk]

you could even technically move to medvram, instead of lowram setting

[oak osprey]

you might already

[hearty karma]

yeah i already did move back to medvram

[oak osprey]

when CPU offload is enabled at least in pytorch 2 and diffusers .16 it gives me some warning about how I've enabled offload and now am manually moving the pipe to cuda. saying memory savings could be lost. but if I don't move it to cuda it complains all tensors aren't on the same device

so despite the warning it actually does offload and saves me a bunch of vram on eg. 4k upscaling

[dense tapir]

I am not a fan of PyTorch2 on the older cards.

[oak osprey]

I saw greater memory savings in an earlier version that didn't print that warning but it was also slower. so something seems broken but not broken enough to make me care

[wooden jolt]

hey friends, a special kind of weekly post from me as this one comes with a published model. I basically attempted to create a general purpose model, in this case using MidJourney v5 data and documented every single detail what I did there in this post. The model is not perfect but can generate some cool images.
I'm super interested in hearing feedback from SF community as there is a ton of potential improvements if I were to continue working on these series of posts
https://followfoxai.substack.com/p/guide-to-fine-tune-your-own-general

[dense tapir]

3k, and especially 4k, cards it rocks

[wooden jolt]

or link to the model directly https://civitai.com/models/61086/vodka-by-followfoxai

[oak osprey]

I don't even think I have access to anything earlier than a 3070 from Nvidia

I might have a 1050ti or something in storage

[dense tapir]

My 1060 is barely able to keep up with this

SDXL is a forget it I can't run it

[oak osprey]

sdxl is released?

[dense tapir]

The XL in the name says it all

not yet but soon

still in beta

[oak osprey]

it generates 64x64 image

so you can likely run the first stage by itself

[dense tapir]

yeah, let's rock and roll back to a decade ago

[oak osprey]

just run the upscaler separately

[dense tapir]

I might can do 256x256

[oak osprey]

no need to chain them immediately

that just wastes vram

[dense tapir]

I don't believe in upscalers if they are in the pixel space. those are ugly

Sytan storms in the room.

LOL

[oak osprey]

deepfloyd uses the x4 upscaler and it looks great

[hearty karma]

it only looks great for certain things

i fat fingered on the args i put medram instead of vram...

[dense tapir]

I have someone with a 3090 who uses DF a lot and he has now uninstalled it. Called it junk and just not worth the effort. It used a shit ton of vram too.

[hearty karma]

isnt deepfloyd the biggest vram sucking model that is out right now?

[dense tapir]

Don't quote me on this but around 11-13gb

For SD yes

Until SDXL drops

I wonder if that new neural compression thing from Nvidia will help with this?

[oak osprey]

deepfloyd needs better prompting

I also called it junk tech

waste of an a100

[dense tapir]

Oh, he is very capable at prompting, but he was not in awe of it at all.

[hearty karma]

yeah uhh i cant use no-half

[oak osprey]

disable nan check

[dense tapir]

Do the upcast

[oak osprey]

😄

[dense tapir]

only after what psedo said

if disable works viola if not upcast

[oak osprey]

🎲

roll dice on your fix, baby

[dense tapir]

LOL

[oak osprey]

I rolled a damn 20

hot shit

[dense tapir]

1650/1660 is a 1060 minus balls.

[oak osprey]

look I ignore warnings all day long and I'm a professional

[hearty karma]

ok i have checked both neither work

[oak osprey]

shitty, that's super weird. I wonder what no half even does under the hood

[dense tapir]

See, I never could stand warnings because warnings should only ever mean a bug might happen. This is why I treat all warnings as bugs.

[hearty karma]

screw my graphics card to the moon 😠

[oak osprey]

buy a new one. if anyone gives you gruff, tell them "pseudo sent me"

[dense tapir]

1060/1650/1660 has no real fp16

we emulate it

[oak osprey]

ahh

use a PlayStation 3 instead of a 1060 then

[dense tapir]

when a 1060 sees a fp16 string it shoves it to fp32

[oak osprey]

that's backwards

[dense tapir]

it can't do floating point 16 bit

[oak osprey]

half precision was easier forever. why did Nvidia do that

[dense tapir]

no idea

[hearty karma]

i might remove xformers from my arg

[oak osprey]

no

[dense tapir]

oh, god you better not

[oak osprey]

you can if you want to experience the year 2020

it wasn't great

[dense tapir]

See, Nvidia has been screwing people for a long time now only people didn't notice, or care, until things like this AI hit. Now they can't hide it.

[oak osprey]

true

[dense tapir]

I mean what is up with less ram on higher cards etc...? Should be the most then gets less and less in everything unto x050

[oak osprey]

market segmentation is something they perfected

[dense tapir]

LOL, yeah

AMD has the ram side down at least

I am wondering about their rsn 7950XTX release.

[oak osprey]

well vram is pretty expensive especially when it's on die

a 2.5 million dollar wafer scale system from cerebras has 40gb on-die SRAM

so maybe Nvidia isn't so bad eh

putting just 40gb on-die would consume more than 56 percent of the available die space for the 4090

[dense tapir]

I can't wait to hear the screams of millions when the 5k series of cards comes out at 1.5-2x the price of 4k cards. Already Jensen said they will have to be more expensive. 6k cards, if they go to 2nm (if they can) will be 5-6 times the cost of a 4k card.

[oak osprey]

100k card then?

that's ridiculous LoL

[dense tapir]

I know but look at the H100 at 45k

[hearty karma]

im sure someone would buy a card at that price

people are idiots

[dense tapir]

They do

[oak osprey]

there are better cards at that price that don't involve Nvidia and are called TPUs

win win

most of us relying on Nvidia want to spend under 40k per accelerator

hahaha

[dense tapir]

TPUs I was told are funky and need all kinds of serialization and stuff

[oak osprey]

that's all being done under the hood

like AMD cards use CUDA interfaces in pytorch

but they don't support CUDA

[dense tapir]

Well, I still wonder why ROCm is so bad in comparison? Hell, they do not YET (even with 5.5.0 just released) work on 7k cards

[oak osprey]

7k cards aren't workstation grade other than w7900

[dense tapir]

Well, the PR was just released and all 7k cards (ones not yet released too) are mentioned in it so I suspect 5.6.0 will get support for 7k cards.

[hearty karma]

i hate having to use xformers and half precision

._.

[dense tapir]

NO half precision you mean

no half means full precision but go figure which nimrod thought --no-half was cool.

[hearty karma]

oh yeah... 🤦‍♂️

the fact 16xx series cant go half precision is stupid

[dense tapir]

on a 1060 on 2.1 I have to use memory attention or the --no-half or I get a black box. One or the other. I could use both but no reason as --no-half sucks all my ram and makes it slower.

[hearty karma]

tbh im happy that i have atleast dropped some vram usage

i went from 5.8 to 5.2-3

[dense tapir]

btw, I watched a video last year on the internal workings of the 4090 and they really knee capped it except for fp32. FP32 it shines so bright and is why, despite Nvidia's internal kneecapping, it is up to 56% faster than a 3090. They kneecapped it because it VERY EASILY could compete with their business lineup if they hadn't.

As the guy said he was amazed they gave us the FP32 stuff that only their business lineup has. I suspect they did that cause they hampered the card so hard everywhere else.

This is the problem with these companies because they are so worried about their business lineup while we get the crumbs. I wonder if Intel GPUs will be like this?

https://archive.ph/c2Eev

WOW, civitai takes 10-20s to come up for me lately

[oak osprey]

run your browser in half precision

😄

[dense tapir]

hehehe

[lofty yew]

General, what are these portraits from?

[dense tapir]

A prompt I was doing. I decided to attempt a batch size 4.

[lofty yew]

i thought it might be midjourney. They had a free trial over the weekend.

I used it all up in 1 day

[dense tapir]

Oh, nope, I don't touch MJ with a bilge pole.

[lofty yew]

that is a cool look. Were you using Digital Diffusion when you made it?

[dense tapir]

Yes

V2

[lofty yew]

nicwe

i mean nice

[dense tapir]

I was seeing how consistent they were

Her teeth are so good

[lofty yew]

yeah. And it doesnt have the normal face distortion. Good Proportions

[dense tapir]

yes

I was shocked at the teeth cause I always get some funky teeth in my gens

so bad I started to add teeth in my neg

[lofty yew]

was that just digital, or did you include positive embeddings for that type of artwork look?

its kind of like MJ v3

[dense tapir]

Just DD as I don't have any artwork embeddings for 2.1

What I wish I had was MJ5.1 model

[dense tapir]

nice one

Seems all the MJ models are for 1.5

[lofty yew]

I did these in Artius, but i was playing around with a contrast fix lora

[dense tapir]

I gave up on contrast fix as I found it to really tear stuff up, or makes it purple. Now on my death vibe artwork I use nfixer instead due to the purple tinge of the contrast fix.

cyberpunk I use the contrast fix lora for the purple

[lofty yew]

yeah, i also use nfixer. Its the only neg emb i use.

[dense tapir]

nfixer has one issue

I ran into it recently and that is it makes everything too contrasty and dark. I mean so dark I can't see the face.

[wispy nest]

I still don't understand lora much

Tried different setups and seem not much difference 🤔

[dense tapir]

now on illuminati it was made for it is fine

[wispy nest]

Dog pics look nice

How many pics of you do you think need to train for photorealism?

[dense tapir]

which method?

Dream Booth is 30 and 300 reg images

Lycoris is about the same.

Finetune could be 100-1k

[wispy nest]

I was reading you need images from different environments to be most effective. Not like same pic setup different poses

Have you tried any?

[dense tapir]

Not since 1.5

[hearty karma]

have you ever booted up 1111 on a 2.1 model and forgot to put the resolution to the correct thing?

[wispy nest]

What gpu are you using?

I think this server needs roles or tags to show which GPU each person is using lol. 😅

[hearty karma]

that would be a very very very long list of roles

[dense tapir]

All the time so I set the default in the json file to be 768x768 default instead of 512x512

[lofty yew]

Snoop, that is the most common thing 1.5 users do, and it destroys the image it at least one side isnt 768.

[wispy nest]

Not if you keep generic like 2000 series etc

[hearty karma]

time to just delete the past 9 generations

._.

as they were done at 512x512 on digital diffusion

[lofty yew]

it will give you a picture, but its all blotchy and patchy and the colors go bonkers

[hearty karma]

yeah i know it gave me pictures

and im deleting them because of the fact they looked melty

[lofty yew]

im hoping that in SDXL and 3.0 you can go lower res without destroying pics, because i really think a lot of people think 2.1 models suck because they are using 512 by default without knowing it

[hearty karma]

SDXL

[dense tapir]

Probably

I figure 1.5 is where everyone will stay forever.

[hearty karma]

nah not me

im staying on 2.1 now

[dense tapir]

Thank you waifus

[lofty yew]

just the people in this 2.1 chat. I sometimes think there is a group of about 30 people using 2.1 trained models

[dense tapir]

Stats show it isn't much

Training is almost exclusively for 1.5 too

way easier than 2.1 hell. I gave up style training on 2.1

[lofty yew]

its too bad, because people are missing out on a lot of extra detail, but I guess it doesnt matter because it doesnt have boobs or celebrities

[hearty karma]

wym it doesnt have boobs?

🤨

[lofty yew]

well, 2.1 base. trained models, yes

[dense tapir]

Oh, the porn on 1.5 is through the roof

[hearty karma]

oh...

so glad im not using base 2.1

[dense tapir]

Well, you are

[lofty yew]

only a select few 2.1 models have nudity

most of the best ones do not

[dense tapir]

Not to the level of 1.5 models

[lofty yew]

yeah, some of the 1.5 models are so heavily trained on porn it just throws it at you

[dense tapir]

First image I did on a 1.5 I downloaded was full on, in my face, lunch time. I was like wth? I had to add stuff to the neg NOT to get that

This was on my recent retrying 1.5 with new models.

[lofty yew]

i try a lot of them, but they all seem to look the same to me now. 1.5 just makes skin too fake and I swear i see Anime eyes in EVERYTHING

my negative prompt is large eyes, cartoon, anime, big eyes.....but yet i get a human with giant spider eyes

[tender thunder]

HI im stuck in queue while generating images. how to fix this? @lofty yew

[lofty yew]

try different model from dropdown

[dense tapir]

Yeah, but I am getting that plastic look in 2.1 lately

[tender thunder]

yeah i fixed it while closing command prompt

lol

[wispy nest]

Hm, I took a couple month long break, and it's saddens me not much seems to have happened, in my eyes, with the AI. :(

[tender thunder]

a good start

is it fun to learn stable diffusion?

[hearty karma]

ah

sometimes :-)

[kind quartz]

first picture take longer i believe. @tender thunder not learning just playing 🙂

[lofty yew]

it does get plastic face especially at a distance, but i feel like close ups on good models just make skin a little better

[dense tapir]

That is the 768 showing. 1024 base model should be even better

[tender thunder]

Great thanks

[lofty yew]

try typing a photo of a cat, detailed, sharp focus

then download some better checkpoints. base 1.5 is just ehhh

[hearty karma]

base 1.5 can be good sometimes

[digital hornet]

(definitely dont juist go for base 2.1 as your checkpoint upgrade)

[tender thunder]

how to upgrade?

[hearty karma]

download checkpoint from internet

[kind quartz]

what to do with eyes, each different size, and this is quite good actualy, not so bad (iris sizes)

[oak osprey]

base 2.1 gives me weird results in my current pipeline flow. i'm honestly not sure if i've done something wrong, or if it's just that bad

https://images-ext-2.discordapp.net/external/y1y1Ws7gmCJ4RhkA0NP2xl_dwwg6tB-JgW5zaSUbYhI/https/tripleback.net/public/discord/1683497219.3260243352ec4cad61f3a892210b1d1ec1c6c81.png?width=1600&height=900

the weird tiling it does around the edges, i am not a fan of

i forcibly enable VAE tiling and i've not experimented with that

[digital hornet]

it's kind of bad but there are 2.1-derived models that are pretty good

[oak osprey]

i tried to find one and haven't succeeded yet

[wispy nest]

using my pre-lora (have no clue what it is even) skills. Seems there are new stuff that I cannot name at least :D

[lofty yew]

Try Artius or Digital Diffusion and set your res at 768x768....best 2.1 based general use models

[oak osprey]

but literally on the model card for 2.1 it promotes the use of nonstandard resolutions

did they lie?

[lofty yew]

no, just need at least 768 on one side.

bigger is better

[dense tapir]

Digital Diffusion V2 vs SD 2.1 non ema base.

[lofty yew]

wow

that is a demon

[dense tapir]

Glad I wasn't the only one saying that

[oak osprey]

hmm so my 1280x720 needs to be 1366x768 ?

it also just never looks good even when it's not doing weird tiling stuff

like something I've done or haven't done is causing this I'm pretty sure

like what scheduler do I use? just the default? and then like 150 steps?

does the vae matter?

[lofty yew]

no no no 150 is way too much, and set your cfg scale down to about 4

[oak osprey]

the guidance scale?

[lofty yew]

yep.

put steps at about 40....it will be a lot faster too

[hearty karma]

doesnt lower cfg make things uglier

[oak osprey]

it's pretty much instant on my 4090 no matter how many steps I use

the a6000 too

[lofty yew]

you are overcooking it. 2.1 needs lower scale

[hearty karma]

so 7 isnt good for 2.1?

[lofty yew]

are you using negative prompts? very necessary

[oak osprey]

yep

[lofty yew]

even 7 will give it that blotchy look sometimes

[oak osprey]

reducing guidance to 4 kind of helped

[lofty yew]

you just need to use a better model also

[oak osprey]

I mean I've seen some incredible stuff from base 2.1 I can't even reproduce

just curious why

[wispy nest]

for me, most of my info is very old, but I noticed that using a upscaler really makes the images I create better. Not as symmetrical, but overall better

[lofty yew]

do you know how to download model and yaml file and place it into auto?

[oak osprey]

20 steps not enough lmaoo

[lofty yew]

use one of the karras samplers

[hearty karma]

why karras samplers

[oak osprey]

how do I even do that

just a simple diffusers example or docs page will do

[lofty yew]

These are from Artius model.....using karras at 20 steps

[oak osprey]

that's incredible

I would be super happy with that

[lofty yew]

those samplers need less steps, and give decent results, i use DDIM for detailed stuff

[oak osprey]

sampler is scheduler?

[lofty yew]

yes

[oak osprey]

https://huggingface.co/docs/diffusers/using-diffusers/kerascv

so KerasCV is what you're referring to? this is a bit confusing to me. it sounds like they're pretrained models based on SD

[lofty yew]

you are using auto1111 correct? You need to locate the folder Stable Diffusion models. It will say place checkpoints here in the foler. put both the checkpoint and the YAML configuration file there

[oak osprey]

nope. i'm writing a custom Discord bot that uses a master-worker architecture to distribute jobs across workers

i have never used Automatic1111 before

[hearty karma]

why

[oak osprey]

why make my own Discord bot? or why never use a1111?

[hearty karma]

a1111

[oak osprey]

idk, i don't like its code

i hope some of the neat stuff gets upstreamed to Diffusers but they don't seem interested in that kind of work

[wispy nest]

hm, seems like anime style is slightly stronger now than before. Or that's how it feels like :P

[oak osprey]

[more context needed]

😛

[wispy nest]

I haven't made an ai image for something like 4 months and are just starting to "feel the waters again" can't remember most what I learned, but I noticed a certain anime focus toward most images. But that might also be because my 4 month old models etc. Downloaded some new ones but the rest is old in tech age :P

[oak osprey]

ohhh yea yea yea. if you don't pin your checkpoints to a certain LFS revision they'll be auto updated on run

[lofty yew]

i have only ever used custom models on auto, so im afraid i cant help much outside of that

[wispy nest]

yeah, I noticed when auto1111's upscaler changed how the numbers worked

[oak osprey]

well, no worries, it doesn't seem like any singular individual has all of the pieces to the puzzle

just have to get pieces from many people as i go

my images are remarkably better than when i started this project a month ago tho 😄

[hearty karma]

30 samples lms karras

[lofty yew]

yeah...it is a learning curve, especially with 2.1 models, but it worth it. The best stuff i have ever made is from 2.1 trained models with a mix of embeddings and negative embeddings

its the details

[oak osprey]

the first image my bot ever made

[hearty karma]

woah...

[wispy nest]

I felt 2.1 had too much work of embeddings than I wanted. Mostly because I felt it was too much busy work like loading this, testing that than I had to do with 1.5

But I also used 1.5 for most of my time so I'm probably more than a little biased for it :P

[hearty karma]

thats a mess

[oak osprey]

welcome to waifu-diffusion-1-4 or whatever

it doesn't tell me which model i was using back then

[wispy nest]

my first AI image wasn't even stable diffusion, but that other one I can't remember the name of, the one that slowly painted more and more abstract stuff into the image. And good luck getting it to paint anything quickly, and also having people in it :P

then I tried 1.4 and it was amazing :P

[hearty karma]

1.4 is crap now :)

[oak osprey]

ah i was wrong, that's the first image in the proper channel but there were 5 others it made before that:

[lofty yew]

im a portrait guy.....these are some older pics i did with 2.1 models.

[wispy nest]

at least people can see what it's supposed to be! :P

[oak osprey]

i really wonder what prompt i gave to achieve such awfulness

lol it was ||big ol tiddies||

with zero negatives

[hearty karma]

🤨

[wispy nest]

and I'm on the very other end, I love "none realistic" images a lot more for some weird reason :P

[oak osprey]

for context at that point the bot was merely creating a model from pretrained checkpoints and then stuffing prompts into it and returning whatever garbage came back, i didn't mess with schedulers or VAEs or even make use of Xformers

oh and i didn't use Compel back then for prompt embeds so it was just really hard to improve the output, as the prompts would only go to 75 tokens before being cut off and padded to 77

eg. the famous prompt from "Deliberate_v2" for the maid, produced this output

expectations vs reality

i wonder how close it can get to the proper one now

https://huggingface.co/docs/diffusers/api/pipelines/stochastic_karras_ve

@lofty yew so i found this pipeline option and it seems pretty obscure, but maybe that's what Auto does when you select Karras

but the call() method oddly doesn't have any kind of prompt or inputs arguments so i'm not sure how to actually use it

[wispy nest]

I think this was one of the first ai images I made...good luck guessing what it's supposed to be because I can't even remember :P

[oak osprey]

https://huggingface.co/docs/diffusers/api/schedulers/stochastic_karras_ve

ohhh there's this

is that what this one does?

well, yeah

crosses fingies

2023-05-08 15:56:52,240 [ERROR] Error while generating image: step() missing 1 required positional argument: 'sample_hat'

f

[oak osprey]

@stone cipher are you using self-attention guidance?

[latent atlas]

https://www.instagram.com/p/Cr_OJdIosAD/

veiled.vision.studio

Anastasia sat on the park bench, lost in thought as she watched the leaves dance in the breeze, wondering if she would ever come back 👑

Instagram

[static raven]

Hi All, Can you point me into some resource where Ian read about how to use generation steps? I have generated some images with 10 generation steps to 'save on credits' but they turned out far from ideal so I tried 100 generation steps and they look very good - what is the sweet spot for generation steps?

these are at 100 generation steps - same prompt

[oak osprey]

the steps depends on your chosen scheduler and model combination

@stone cipher look up how to enable SAG

[static raven]

ok - ill have a look - thanks

[oak osprey]

self-attention guidance

jesus christ

it's oppositional to CFG

[mortal relic]

One walking down the street

[oak osprey]

i'm probably wrong how it works but the CFG scale looks at the prompt embeds and the SAG scale looks at the image it has already generated

well not knowing is one thing but the circular questioning was what led to that outburst 😛

to be fair, i did that to people when i was young, so, it's likely some kind of universal payback

[wispy nest]

I mostly see CFG as a "the lower the value, the more freedom the ai has to interpret what you wrote" :P

[oak osprey]

yeah "CFG looks at the prompt embeddings" is how i wrote that 😛

[wispy nest]

never heard of SAG, but like I said, my general ai knowledge is like 4 months old hehe

*out of date, I mean

[oak osprey]

SAG scaling, aiui, requires its own whole-ass pipeline and it is kind of rigid which models it will/won't work with, or at what resolutions outside of a perfect 512x512 square

i had it supported in my Discord bot's pipeline code until last week when it upset me so much i just removed all of it. i saw dubious benefits for high maintenance cost

maybe i did it wrong 😄 but it doesn't seem to be a very popular piece of code and so it goes unmaintained. i opened an issue about the resolution problem and have received no response from the author

[hearty karma]

ah yes the roads of lava

[wispy nest]

like how I see 2.1 when it comes to "none photo" images 🤣

[oak osprey]

2.1 is really interestingly broken code.

[hearty karma]

ok im so damn confused

my vram has shot back up to 5.8 for 1.5

it was at 4 before 😭

[oak osprey]

your prompt also impacts vram use

using same ones?

[wispy nest]

yeah, by the by, I noticed my graphic card don't get as hot as easy as months ago, and that's very, very nice :D

[oak osprey]

@stone cipher no need to delete all that, i was just wondering if you were using it because your images are super high quality and i can only dream of producing stuff like that

but honestly if it needed SAG to achieve that i'd just be okay with producing terrible images forever

https://www.reddit.com/r/StableDiffusion/comments/12yxle3/sag_is_here_sag_and_cfg_comparison/

here's the post i grabbed that comparison grid from

[dense tapir]

SAG ceased to work for me

[oak osprey]

@dense tapir glad i'm not the only one

[dense tapir]

Oh, snap it is a known issue?

[oak osprey]

probably yes because i reported it

[dense tapir]

I am reporting a lot of extensions lately with tickets

[wispy nest]

wait... how can the scale be in the negative? -1? :O

[oak osprey]

try to use SD 1.4 based models with SAG instead @dense tapir

works better

eg. widescreen works on 1.4 but not 1.5

[dense tapir]

WOW, my 1.4 I think I still have

[wispy nest]

the power of 2.1 :P

[dense tapir]

Cute

[frigid dew]

Can someone teach me how to create ai arts? And how to prompt?

[hearty karma]

guys what are the best settings for auto 1111?

[wispy nest]

woah hogwarts looks sublime

[tawdry mulch]

a man and a woman sitting next to each other, a stock photo by Mór Adler, pinterest contest winner, art & language, contest winner, sabattier effect, epic

[crystal delta]

FAQ: How do I generate images? Is there a bot on the server?

Currently, there is no public bot on the server that generates images. There is an experimental bot available for early server members & Stable Society. You can obtain the Stable Society role by winning our weekly events #⭐｜pow-info #1087493421209485393! However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610⁠ for more details! You can also stop by ⁠#1025467151206854736 for any issues you experience while using DreamStudio or ⁠#🤝｜tech-support for any problems you encounter while installing it locally!

[tawdry mulch]

a man and a woman sitting next to each other, a stock photo by Mór Adler, pinterest contest winner, art & language, contest winner, sabattier effect, epic @dense tapir @

[dense tapir]

@tawdry mulch

[hearty karma]

this is looking weird rn

[dense tapir]

Oh, I rather like this one

The brain in the jar sealed it for me.

[hearty karma]

hybrid car

meant to be a mix of a lada, delorean and mustang

[wispy nest]

lada? 🤣 ❤️

[crimson dawn]

russian pos i think

soviet engineering

[hearty karma]

yeah its russian

[wispy nest]

hehe, yeah I know of it, I just wasn't ready to see it beside the other two :P

[hearty karma]

same prompt ddim sampler this time

[crimson dawn]

state owned russian car company

[hearty karma]

now have yall heard of the trabant?

[wispy nest]

thanks to that one youtuber, yes, yes I have :P

1-60 in whooping 36 seconds :P

[hearty karma]

the worst german car to ever exist :-)

btw guys do yall know how to weight prompts in auto 1111?

[kind quartz]

if you have -1 in seed, it doesnt matter, each run different seed.

[crimson dawn]

the shittier knockoff of the lada made when the wall was a thing

[kind quartz]

(arm) or ((arm)) or (arm:1.4)

[crimson dawn]

"The man from Uncle" i think has oen in it

[wispy nest]

you can also use CTRL+up/down arrow to increase/decrease weights if you've marked the word

[hearty karma]

oh cool thanks guys

the help is appreciated 👍

[kind quartz]

and lower is [] or [[]]

[crimson dawn]

i noticed an update happen. when you ctrl down on the wights now, it'll actually remove the ()

lower can also be (arm:0.5)

[wispy nest]

yeah, having less than 1 in () is the same as using greater than 1 in []

[hearty karma]

got a bit of a delorean with a mustang shaped roof

[crimson dawn]

well no cause :1.2 doesn't work in [] ? unless theyv'e changed that like removing the brackets. mmm i love that nice clean update

[kind quartz]

also probably better keep rounded () because [] can be used for dynamic prompt

[crimson dawn]

hated having :1.0) all over my prompts

yeah i love the squares for the prompt engineering stuff. [Arm|Leg|Head|Shoulders] (don't try that prompt)

[hearty karma]

that is what i am doing for my hybrid car

[crimson dawn]

DANGER PROMPT

[hearty karma]

ew what am i looking at

[crimson dawn]

[Arm|Leg|Head|Shoulders]

[hearty karma]

why did you even try that prompt

[crimson dawn]

science

[hearty karma]

some science should be left undiscovered...

[crimson dawn]

but what if the enemy discovers that first??!

[wispy nest]

enemy? :P

[hearty karma]

ah yes i love having my cars have cropped fronts

[crimson dawn]

You want King Charles getting his hands on that?!

[oak osprey]

bonnet

[hearty karma]

time to mix the delorean with the countach and robin reliant

https://tenor.com/view/grin-evil-evil-smile-smile-gif-15036311

[crimson dawn]

the amazing tip over delorean

[kind quartz]

you can try subject_center_image or so

[hearty karma]

i just found a very weird transitional phase

that is literally made up out of none of those vehicles

[crimson dawn]

isn't it reliant robin? that one cgot me tilted

[hearty karma]

it might be

i forgot

[crimson dawn]

lol i loaugh whenever i see one LOL

[hearty karma]

beautiful car :-)

[crimson dawn]

i remember the top gear episode. they dump it on every corner in london i'm pretty sure

[hearty karma]

it showed up in a news broadcast too

jeremy clarkson going to hard on a corner

[crimson dawn]

https://www.facebook.com/CHandMTopGearVideos/videos/rolling-a-reliant-robin-top-gear/1392013524227576/ soirry for the facebook link. only good clip of it i could get

[oak osprey]

thats a Reliant Robin

[kind quartz]

they got it in only fools and horses i think

[oak osprey]

oh holy cow

that's SD 2.1

guess it really needs the Karras sampler

[hearty karma]

really?

worked pretty fine on other samplers for me

[calm canopy]

isnt that a bear

[oak osprey]

that's 1280x720 native output without tiling issues etc

i get tiling errors on other samplers/pipelines

[kind quartz]

he is just missing one leg 🙂

[hearty karma]

well you arent exactly meant to be generating 1270x720 images tbf so...

[oak osprey]

in the SD2.1 model card, it says it supports non-standard resolutions

[kind quartz]

yes for sure 768x768 minimum

[oak osprey]

@kind quartz the file is called cat.png, too

[hearty karma]

ok but im pretty sure the minimum a resolution can be is 768

[oak osprey]

@hearty karma not once you switch to Karras, is my point

you okay peter? 😛

[hearty karma]

why was that post delet?

[kind quartz]

problem is if AI, as her head breaks horizon. It is sign of AI. I mean first picture. But can be real with bit weird horizon 🙂

[hearty karma]

yeah he knows

hes just saying the horizon is screwed

i think

[oak osprey]

could be a real problem in the terrain tho lol

overly perfect ground would be declared a sign of AI, too.

[kind quartz]

yes i said it

but usualy it is sd issue

[oak osprey]

wasn't clear what you meant at the end

[hearty karma]

does the weird [ | ] mixing thing not work on 2.1?

[kind quartz]

i am bad in english, too old that i have not english in school 🙂

[hearty karma]

the horizon dips down towards her head

[kind quartz]

i would say real, but who knows 🙂

[oak osprey]

gravitational lensing

[hearty karma]

guys on automatic 1111 does the [ | ] thing only work on 1.5?

[kind quartz]

have you posted real ? 😄 @stone cipher

[hearty karma]

anyone?

[oak osprey]

the [] stuff is depedent on the prompt embedding engine and the pipeline in use. with Compel weightings, [] shouldn't do anything. you have to use -- nomenclature there.

[dense tapir]

Still works

[hearty karma]

oh yeah just realized it works

looked back at my gen and its got delorean traits now

[oak osprey]

if the pipeline is the community LPW pipeline, there you can use [] hints

[dense tapir]

It does in auto111 [ and ] so does the pipe character

[hearty karma]

it was staying looking like a mustang for a long time and it confused me

[crimson dawn]

i use sd 2.1 with a variety of samplers often. i use unipc and ddim the most

[dense tapir]

for the [] you need to make sure it is on in settings

[oak osprey]

the community LPW pipeline drastically changes output even without using any embedding hints

[hearty karma]

id assume its on by default since it was working for me anyway

on 1.5

[dense tapir]

You do realize what [] is if on, right?

[oak osprey]

de-emphasis

[dense tapir]

yep

| is used to mix images so (dog|cat) every other step is either the dog or the cat.

[oak osprey]

so as i understand it, auto1111 when you have "long prompts" option enabled, uses Compel to generate prompt embeddings and that uses the syntax keyword+ and keyword- to emphasise and de-emphasise

[crimson dawn]

a lot of people have confirmation bias with SD research. have it so hard.

2.1 can even do boobies if you prompt it right!! mind you, need specific oil painting artists who have done natural life drawings to get them though

[oak osprey]

when i prompt naively to make it generate them, it makes Russian gothic buildings

[dense tapir]

We are forever stuck with 1.5 as the community refuses to move on due to "I don't consider that an upgrade" mentality.

[crimson dawn]

you can't prompt 2.1 the same way you prompt 1.x. yuou can but you'll not have great results

[hearty karma]

disabled the emphasis thing

[crimson dawn]

Those aren't gothic architecture

[oak osprey]

yeah it's similarly different in prompt style to DeepFloyd imo

[dense tapir]

yep, different clip different prompting required and a HUMONGEOUS need for proper negative prompting.

[wispy nest]

for me, 2.1 just don't have the same reach. It got slightly better area definition, and photo-realism, but that's about it in my mind. The rest 1.5 does better :/

[crimson dawn]

deepfloyd uses the t5 thing instead of a clip thing. i have no idea about that stuff

[dense tapir]

I hate the look of people in 1.5

Porcelain doll syndrome

[wispy nest]

new here, which label is image creation?

[dense tapir]

no bots on this server for image creation

[crystal delta]

FAQ: How do I generate images? Is there a bot on the server?

Currently, there is no public bot on the server that generates images. There is an experimental bot available for early server members & Stable Society. You can obtain the Stable Society role by winning our weekly events #⭐｜pow-info #1087493421209485393! However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610⁠ for more details! You can also stop by ⁠#1025467151206854736 for any issues you experience while using DreamStudio or ⁠#🤝｜tech-support for any problems you encounter while installing it locally!

[wispy nest]

I mostly just want paintings in one, or another way. I dislike the wrinkles, got enough of them in real life :P
I also noticed how 2.1 faces look "weird" somehow when being away from ai art for such a long time

[hearty karma]

👀

[crimson dawn]

most of 1.5 land are all merges which disintegrate the knowledge into one mushy hard to recall model. you know like whhen you're trying to remember soemthing but all the days blur together in your mind and you can't pull the details out? I figure it's like that in the model. Everything is blurred together and you get a lot of generic same face situations in 1.5

[oak osprey]

well in my discord bot's code which uses roughly the same pipeline process for both (which is likely why it's so bad) the 2.1 models are just garbage (768x512)

[hearty karma]

ok but the skin looks dead smooth

[dense tapir]

I'll stick with 2.x

[hearty karma]

on 2.1 models the minimum resolution is 768

afaik

[wispy nest]

yeah, that's the effect I'm thinking about, not as defined in most images but it's some weird "copy paste different focus bits into each other" that just seems weird whenever I look at it

[crimson dawn]

this feels like seeing what you want to see. when i looked at it, first thing i noticed was realistic skin textures

[oak osprey]

same. the people in here have really destroyed their perception of reality after staring at AI images for too long 😛

[hearty karma]

._.

[oak osprey]

constantly thinking "there's something wrong with those eyes" when i share real photos in here

[crimson dawn]

what is real?

[hearty karma]

ok well with this image has buggered up eyes

well the left one is

[crimson dawn]

before digital cameras had all that fancy post processing, most images had buggered eyes because people would blink during the shutter action xD

[kind quartz]

real

[crimson dawn]

eyes being all buggered in photos is my whole life

[kind quartz]

ai cant do line like this

[dense tapir]

Meh, I prefer 2.1. Now training in 2.1? It can go eff off I refuse to ever train on that turd again, at least for styles. 1.5 with the same data BOOM, and done. 2.1 3 months and still nothing. When that happened I tossed in the towel and said no more.

[oak osprey]

weird

have you tried training 2.1 in CV?

[wispy nest]

what I do like though, is that most people starting position around ai art was that "it isn't a perfect masterwork"
just try and think how people would react if someone was told that when they started learning how to paint :P

[hearty karma]

guys is it possible to change when the progress view thing updates?

[crimson dawn]

i haven't tried training 2.1 yet. i dont understand the clip language as much as 1.x and i use clip to caption my images so that it aligns with what the model's knowledge better

[dense tapir]

It is a known issue due to the TE and SAI not releasing tools to match the needs of the new CLIP for training. It literally fights you.

[crimson dawn]

yeah i need an openclip interrogator

[kind quartz]

you can skip or interupt @hearty karma

[hearty karma]

nah i mean i want to see the progress updating in the thing i dont want to interupt it or skip it

[kind quartz]

but you cant change prompt or settings

[oak osprey]

https://huggingface.co/spaces/fffiloni/CLIP-Interrogator-2

[crimson dawn]

live previews in settings

[kind quartz]

yes and set 10 value to 0

[crimson dawn]

ooooo. a sequel

[oak osprey]

and in this one, all te kids make it out alive

[crimson dawn]

tyvm i appreciate the link. will explore this today maybe

[oak osprey]

i haven't gotten into the training side of things yet so let me know what you find

[dense tapir]

Take the same dataset, right? Do nothing. Train 768x768 in 2.1 and 1.5. Change no settings for style training. 1.5 done with losses 0.05ish and 2.1 0.4-0.5 and the training is rubbish. I did manage a couple but nothing like 1.5 with the same dataset.

even the brand spanking new schedulers the best can't go below 0.29ish for people.

[crimson dawn]

while i don't think this interrogator will give me a streamlined process for training like the captioning with clip preprocessor extension i use, but this is a great start to my journey here

[hearty karma]

the cat bear

[oak osprey]

https://www.kaggle.com/code/leonidkulyk/lb-0-45836-blip-clip-clip-interrogator

have you ever seen/read this?

[dense tapir]

1.5 having losses of 0.004 is not uncommon

Yes

[kind quartz]

i made elephant giraffe

[wispy nest]

I don't actually mind having more details than this as well. It more than enough for me :)

[hearty karma]

ai i think

[oak osprey]

its definitely weird whatever it is

[hearty karma]

because of the sticks

[dense tapir]

What started me to realize just how bad the TE is for training is when the controlnet devs wrote an open letter to SAI detailing how much of a turd it is.

[hearty karma]

woah

[dense tapir]

It rubbed my nips seeing that as I had already about 2 months of constant daily pain with it.

[oak osprey]

TE = textual embedder or whatever?

[hearty karma]

those animals are a mixture that should have never been made

[dense tapir]

yes, it is CLIP

[oak osprey]

@dense tapir for what it's worth your inside voice in my head, is becoming more and more English

[dense tapir]

unet is no problem it is the TE/clip

[oak osprey]

you're probably from the UK, in't ya

[kind quartz]

i know @hearty karma it is result of radiation and all

[bold cave]

is this rihanna

[hearty karma]

can you do multiple of the [ | ] thing?

im assuming yes

[bold cave]

yeh m8

but in a different format

[x:y:z]

[wispy nest]

no, just a random a ebony woman but I'm sure celebs got more weight for each character :D

[dense tapir]

This is partially why I think, as a trainer, SDXL will be even worse. The only way to get better is for SAI to go to a different clip and/or give us the proper training tools made for it. Right now we are using the rubber mallets to hammer in rubber nails that worked on the rubber wood of 1.5 only now we have 2.1 that expects us to use that to build a real house? Ahem, no.

[oak osprey]

ok so 1920x1080 in my SD2.1 base pipeline is definitely much better output than 1280x720. i need to add 768+ to my list so i have something in-between 1280 and 1920 lmao

but this image still resembles an oil painting almost

[bold cave]

yah except when u zoom in

[oak osprey]

the building's all whacked up

[bold cave]

stop doing natives and do some upscaling n00b

[hearty karma]

what cfg are you using

[dense tapir]

I don't get much enjoyment from the genning I derive mine from the training.

[oak osprey]

3.5

[dense tapir]

Can't train I lose interest in SD.

[oak osprey]

@dense tapir lol get ChatGPT to write you a training tool set

[bold cave]

thats the main difference between SD and every other gen out there

if I cant dreambooth it's garbage