✨｜sdxl

[west breach]

less refiner 😄

[upbeat summit]

less refiner is a good thing 🙂 it is very powerful

[west breach]

3%

still getting lots of fail pics haha

e.g.

[upbeat summit]

I ran a couple of prompts earlier:
266 images total
72/266 great
25/266 excellent
I think that's a great ratio to get so many interesting ones. I made sometimes many hundreds more of SD 1.5/2.1 images to get the same. of course it always depends what you are trying to do and your prompt.

[sour obsidian]

oh no

[west breach]

so true, early days of 1.5 was like 1 or 2 in a hundred

[timid sonnet]

wow i like this workflow i have going

[hollow halo]

Combined the Poke Ball and Portal Lora's

[sour obsidian]

this looks great 😄

[sour obsidian]

love me my robos

[upbeat summit]

photo of a battle cyborg fighting against dark cyborg ninjas with chrome skin, on a space station hangar, photorealistic, narrow corridor lights, from the movie "chappie", analog, very grainy, film still, kodak ektar, fujifilm fuji, kodak gold, cinestill 800t, kodak portra, photo taken by thomas hoepker

and a huge negative 😉

[sour obsidian]

I think mine above was just cinematic close up of a cyberpunk war robot in the rain at night or something like that, I have an aversion to using negatives

[upbeat summit]

yeah. totally understand that. but it really helped push fidelity in SD 2.1 so I'm still used to it. but I already generated hundreds of SDXL images without any negative prompt and I'm super pumped 😄

this was an old prompt of mine for SD 2.1

[sour obsidian]

ya haha, def still helps. Others on the team love em, I personally want to remove the need for them entirely. I think we got pretty close haha

[upbeat summit]

Made this with SD 2.1 when it launched:
https://twitter.com/masslevel/status/1597280180069814272

masslevel (@masslevel)

CORE DIFFUSION (2024)
a sci-fi film concept - part 1/4

made with #stablediffusion #StableDiffusion2
raw sd output, no post processing

#AIart #AIArtwork #generativeart

Twitter · Nov 28, 2022, 5:24 PM

[sour obsidian]

oh yea, nice for 2.1 base! Looks great on xl above

[upbeat summit]

yeah base sd 2.1 can do great stuff

[upbeat summit]

very cool world!

but I couldn't do this fidelity using SD 2.1
1920x952 native SDXL

[timid sonnet]

I did these with 2.1

[upbeat summit]

nice stuff! the abstract ones are really great

[dense chasm]

can u help test the celebrity close up photographic images with proper prompts,such as MJ and Kobe Bryant

with the new sdxl 0.9 models

i'm totally confused the new models work bad on celebs,it seems the have a strict limitation of celebs generation

[timid sonnet]

Some of my favorites

[upbeat summit]

from my experience most famous people are like impersonators or stunt doubles in the model. but I've seen some fine-tunings like loras already and it looks like it can be easily tuned to a person.

[upbeat summit]

but sure - lets do it 🙂

[upbeat summit]

I always appreciate a stylish apocalyptic end of the world - nice images 😄

[timid sonnet]

sorry lol

[timid sonnet]

this is what got me into AI art

these are the last 2.1 I'll psot

[dense chasm]

it no doubt the midjourney still works well on celebs,that's a pity if SD can't do the same work, i found D-ID limits the celebs photo upload as well

[wicked frigate]

anything SDXL doesn't do well, after full public release you can just train it in lol

same as happened with SDv1... base 1.5 was very limited compared to what we're used to after having civitai available to download whatever we want

[sour obsidian]

@wicked frigate why are you still awake

[wicked frigate]

i have a destiny raid at 8 am and it's easier to stay awake til then than to sleep and wake up before then

[dense chasm]

what the fuXX, a horse and scenary view renders with the kobe bryant prompt,Imao

[sour obsidian]

this makes me think your prompt not just asking for kobe lol, or something else is going on

[timid sonnet]

ooooo

Unless Kobe was looked at like a black stallion?

[sour obsidian]

finders keepers

[dense chasm]

prompt: a close up photographic style of famous basketball superstar Kobe Bryant

[upbeat summit]

love that

[wicked frigate]

might be files just mixed together in a weird way?

iirc comfyui just shoves images to the first available # if you don't change the prefixes

can drag one of the weird images to comfyui to see what actually generated it

[dense chasm]

idk, loss of control totally

[wicked frigate]

or maybe you did something weird like dropping cfg on base to way too low and setting refiner very high and it got overly creative, or, yknow, something. Can't know without actually seeing the image's workflow

trying that prompt myself on the base gets very consistent results

[dense chasm]

so what's the proper cfg value set for base and refiner models,i always set it up between 5-7

[timid sonnet]

You know one day, we are all going to be old and dying. I don't spend a lot of time on the internet or on my computer anymore, but generating art with everyone on discord in the middle of the night is something I'll never forget doing

[shy kelp]

this is first pass at 1920x952?

[upbeat summit]

it involves the refiner but not much else

and a bit of seed lottery play 😉

[shy kelp]

thats cool it can gen a pic at that resolution that dosent look morphed

[upbeat summit]

getting morphed and duplication as well, but the ratio of good images is pretty good 🙂

[shy kelp]

haha yeah, would have been better with img2img 🙂

[timid sonnet]

that one came out pretty neat

[sour obsidian]

love this type of stuff!

[timid sonnet]

thannk you 🙂

[timid sonnet]

LMAO

[west breach]

sdxl knows what's up

[languid swift]

omg how did you did these!!?? please ^^

[upbeat summit]

great fidelity boost!

[timid sonnet]

Ty 🙂 what is a fidelity boost? 😛

[upbeat summit]

280°

[compact beacon]

💪

[shy kelp]

the new shrek movie is looking real gay

[upbeat summit]

sooo clean!

[stone fossil]

❤️ 🙂

[upbeat summit]

you posted the whole movie 😄 shrek looks good!

[upbeat summit]

I really like the VHS and compressed look. if these are generated with SDXL - nice! I was looking for tokens earlier to apply some VHS tape post-processing effects onto the image, but have not found the right tokens yet

[hazy flower]

Can you share that prompt? Looks so clean 😩

[stone fossil]

A Hand-drawn illustration of a lady with flowing long hair standing atop distant mountains. The lady is depicted from a faraway perspective, but her hair cascades down the mountainside, visible even from afar. The intricate details of the hand-drawn style bring out the beauty of her hair and the grandeur of the mountain landscape.

[compact beacon]

1.5 + vhs lora

[hazy flower]

Thankss

[upbeat summit]

thanks. so I need to make a SDXL VHS lora 😄

[compact beacon]

💪

[sour obsidian]

hmm im curious how far we can push a vhs look with the base

im sure its hiding in there haha

[upbeat summit]

absolutely. I tried earlier. ended up more with a early 90s analog film look but still - post processing tokens is one of my favorite things to do - creating looks etc
prompt: vhs quality effects, one of the ninja turtles working at a pizza place, new york, from a 1980 tv series, analog film, analog distortion, cinematic, tape glitch effect

[upbeat summit]

impossible before!

[upbeat summit]

this is my darkest SDXL space image yet. looking forward to explore this more!

[sour obsidian]

love it!

dig the astrophotography stuff

[smoky patrol]

Oooo! Auroras!

[clever verge]

Beautiful!

[timid sonnet]

@stone fossil its your prompt

[stone fossil]

🙂

You can remove colorful to get rid of that and get more of the back and white stuff.

[timid sonnet]

I only used the positive 🙂

[timid sonnet]

Im going to pm you evil

[stone fossil]

Oh chips.

[sour obsidian]

really love that type of prompt. Not many write like that even though XL loves better descriptions aside from just word spam

[timid sonnet]

its not letting me pm you, can you pm me? I promise it wont be too much of a bother

[stone fossil]

YEah 2.x also did and also Kadinsky.

Maybe even Kandsky even more, 2.2 of that realesed silently btw. 🙂

[stone fossil]

Sure sec.

[clever verge]

I've also found that quality / fidelity is impacted by negative prompts.

[sour obsidian]

yea the openclip models do better at real descriptions

[stone fossil]

Gief me SD XL 1.0 we can rock it all even harder. 🙂 🙂

Cant wait, rubs hands.

[sour obsidian]

soon soon ;D

[stone fossil]

Aye. 🙂

I am happy enough to be able to try 0.9 already local, so no haste.

[west breach]

still on schedule?

[mental pecan]

why can't i load sdxl

any help would be helpful thankyou

[sour obsidian]

as far as I am aware, model is pretty much out of my hands at this point. Working on the next thing now

[clever verge]

You have to use dev-branch with a1111 to make that semi-work. Better of using ComfyUI for now.

[upbeat summit]

whaaatt? haha

austin powers sequencer?

[mental pecan]

thank you will look into that

[west breach]

austin powers with a paragraph from the wikipedia article on vhs

[clever verge]

When you got Comfy running this could be a decent start, load the json: https://github.com/SytanSD/Sytan-SDXL-ComfyUI

[stone fossil]

Ok last one tired of this style by now. 🙂

[west breach]

I'm a bit disappointed it doesn't know who fat bastard is

[sour obsidian]

that is a shame, guess we gotta delay

[west breach]

please do. this is an unacceptable fat bastard

other FAILURES from this thing you call a "model"

[sour obsidian]

omg these are great, I want these on my fridge

[stone fossil]

xD

Perfect.

[timid sonnet]

@sour obsidian

[timid sonnet]

@sour obsidian

[sour obsidian]

"which of our 3 saturns is your favorite"

[upbeat summit]

surfing the scientific education part of the latent space

[timid sonnet]

indeed lol

[sour obsidian]

I desperately need a good cyberpunk style fromsoft game

[upbeat summit]

at least we get some hard sci-fi with armored core

[sour obsidian]

havent been following the new one much, hoping its good on release!

[upbeat summit]

I hope so too

[patent badger]

anyone who's training loras, do you train both the base and refiner?

[upbeat summit]

SD 2.0 / SDXL - same prompt

[patent badger]

what's the prompt?

[upbeat summit]

SDXL made a modern remake of my 1980s sci-fi film concept

[west breach]

huge improvement

[upbeat summit]

SD 2.0 a1111 metadata. it's a nice prompt build. you can make different movies with it. in SD 2.0 it's mostly 1980s styled. in SDXL it looks like it's very modern and high tech

[upbeat summit]

Can I buy upgrades here?

[sour obsidian]

https://tenor.com/view/pfsf1968-gif-18754243

[upbeat summit]

love those 😄

[west breach]

does sdxl not know what a tornado is?

[smoky patrol]

love it

[stone fossil]

He sees one comming but you are viewing the wrong side. 🙂

[smoky patrol]

My....my Warframe senses are tingling

[smoky patrol]

ooo

Death stranding-like

[west breach]

where's the tornado?

[sour obsidian]

I could sure go for some steamed hams

[smoky patrol]

omfg THAT'S PERFECTON

[wicked frigate]

a photo of a tornado

seems to

[smoky patrol]

Dustin wth lmfao I love it ahahahahahahahahaha omg

It's so goopy

[upbeat summit]

looks like you are mixing styles 🙂 real nice

[west breach]

I need to upgrade my kitchen

[stone fossil]

Data center pirate style. 🙂

Rasta styled.

[smoky patrol]

I love all these influences together. Look at that hair

[stone fossil]

UTP hair / fiber hair.

Mix n match ofc. 🙂

[smoky patrol]

Best kind of hair to have

[stone fossil]

I wish lol. 😉

[smoky patrol]

W i g s, my dude. Cosplay yourself to perfection.

Cheese

It's huge

[stone fossil]

Lol cheese mastering, nice.

[smoky patrol]

They're carrying it like a newborn baby, as they should

[upbeat summit]

@west breach inspired me
1980s movie still of a {man|woman} from {Asia|Africa|North America|South America|Antarctica|Europe|Australia} is very {angry|flirty|funny|sad|rude|sad}, from the movie (austin powers:0.3)

[stone fossil]

The more weird the btter imo. 🙂

AI is kinda good in matching things up that do make 0 sense, I love it.

[smoky patrol]

Hahaha, those are all great caricature types!

[stone fossil]

For example try asking for straight curley hair.

[smoky patrol]

Well

[stone fossil]

Or abstract pictures taken by a kodak camera.

[west breach]

love it! 😆

[smoky patrol]

Straight curly hair can be hair that is only curly at the tips. Abstract pictures could be just someone at an art gallery

[stone fossil]

It can be but it can also draw sick stuff depending on the style we use it on say hand-drawn.

[smoky patrol]

Yes

[stone fossil]

Then we get more play room on what it can be. 🙂

[smoky patrol]

Hehe

[stone fossil]

xD

[upbeat summit]

cursed lol

I love it

[smoky patrol]

What in the ever loving hell is that...fashion

[upbeat summit]

those outlines are amazing

you could straight vectorize that

[smoky patrol]

Yes, thin lineart

Very good lineart

Man, lineart is my jam

[west breach]

seems to be struggling with my prompt. still entertaining though

🔥

[upbeat summit]

coherence not 100% but look how clean it is

[smoky patrol]

Nice!

[smoky patrol]

What is it?

[west breach]

Principal Skinner strutting down the street while holding ham (simpsons tv show)

[smoky patrol]

Maybe try a ham

Or a brand of ham

[west breach]

oh yeah, probably should correct the grammar 😆

[wicked frigate]

dude xl always nails styles & aesthetics perfectly, even when it gets confused on structural content, wow

[smoky patrol]

Hormel ham would likely be a good choice

[smoky patrol]

Huzzah! It's a ham! Meat! Thing

Um

Like a balloon

[west breach]

interesting...

[wicked frigate]

ngl it wouldn't be out of line for something like these images to show up in a real simpsons episode. Like as a cutaway gag or something

... less so that last one lol

business homer walked into a different tv show entirely

[west breach]

i'll try glazed ham? that's a thing right?

[upbeat summit]

@sour obsidian explored the latent space a bit. one step closer to a retro films on VHS look 😄
screencap from a retro science fiction film involving elvis presley's stunt double and space cats, broken vhs header look, 1970 british tv show, a {man|woman} from {Asia|Africa|North America|South America|Antarctica|Europe|Australia} a mildly {angry|flirty|funny|sad|rude|sad}

[smoky patrol]

Holding a Hormel ham on a plate with both hands

Holding a Hormel ham on a plate with both hands, view from the side, arms outstretched```

[wicked frigate]

HAHA i got curious and just typed steamed hams into xl to see what it'd generate, and, yeah

nothing else in the prompt just steamed hams and it gets me

[smoky patrol]

AHAHAHAHAHAHAHAHAHA

[west breach]

that bottom left pic 🤣

[smoky patrol]

Glazed and honey hams

[west breach]

oh lisa...

[smoky patrol]

View from slightly above, holding an artisan plate that has a Thanksgiving ham on it

Could try those, too

[wicked frigate]

the correct steamed hams generation strat was obvious in hindsight. Whatever it is, it's perfect

[hazy flower]

Am I the only person that feels sdxl v1.0 getting a bit weird? the result, the anatomy, and the styling, it feels like it's getting harder to create a better result than the previous version

[wicked frigate]

this is the SDXL channel

[compact beacon]

👍

[upbeat summit]

I glitch prompted too far

[wicked frigate]

only version of xl1.0 that's available to the public rn is the discord bot which randomizes a lot of things

[stone fossil]

Perfect.

[upbeat summit]

Herman Raiser!

[west breach]

I think this will have to do

[stone fossil]

Ya man. 🙂

[smoky patrol]

AHAHHAHAHAA

[smoky patrol]

It has a charm tho

[upbeat summit]

definitely. I really like glitch art

[smoky patrol]

I LOVE THE TV AND IT BEING GONE

Also, you can do just view from above

[hazy flower]

I see, but I was trying to create the similar prompt in the v0.9, and idk why but feels like it affects the previous version

[dense chasm]

kobe bryant is succcessfully generated,but Michale Jordan with the same prompt failed to render

[wicked frigate]

er, the spelling might be an issue there

[dense chasm]

how to solve this problem?

[wicked frigate]

michael jordan would be the correct spelling

[stone fossil]

Make it a pokemon card.

[west breach]

see it can do michael jordan

[wicked frigate]

or niciheal jordoan as is apparently how xl thinks his signature should be lol

[dense chasm]

are u kidding me, take a closer look,this guy is someone else,not looks like MJ at all

[wicked frigate]

i'm not really a big basketball fan, but, idk looks pretty close to me , other than the intensity of the lighting/coloration

[smoky patrol]

yeah, that's pretty close

[dense chasm]

i spell the name correct,but the render is not exactly the same person with MJ

[wicked frigate]

yes

[hazy flower]

Do you know something about this? @wicked frigate

[wicked frigate]

oh god the queue without a login is long lol

[stone fossil]

Noice. 🙂

[wicked frigate]

I... don't know what you're expecting?

the only version of xl 1.0 you have access to randomizes params and limits your access

[smoky patrol]

Works for me, too

[wicked frigate]

it'll be available for full testing and comparison soon™️

[smoky patrol]

and yeah, super long rn lol

[wicked frigate]

why can't you make an account?

[dense chasm]

i still can't generate MJ precisely,it's so weird

[hazy flower]

Idk, just feels like the art model was different from bot v0.9

[west breach]

probably a good thing it can't do celebrities exactly how they look

[hazy flower]

But np, maybe the prompt I used is not good enough to get better result

[wicked frigate]

it is different yes

a lot more gets changed than just the base model

eg bot previously had the refiner on mostly, but current version does not

yeah okay rip that one sounds like some form of server iss-

[hazy flower]

Hoouuu, I see

[wicked frigate]

please stop spam pinging me

[wicked frigate]

-server issue that would need the clipdrop dev team to look into it

[upbeat summit]

will be back later. have fun all!

[wicked frigate]

can you access the "Contact us" button on the bottom of the clipdrop site?

[dense chasm]

is this 0.9 model?u can give it a shot on sdxl

[smoky patrol]

Have a good day, mass!

[west breach]

it's sdxl 0.9

[smoky patrol]

Haha! I love the Pokemon cards.

[upbeat summit]

you, too! actually I need to get some sleep. I've been playing the seed lottery for over 20h :D. having so much fun

[smoky patrol]

totally a great time ,but get some sleep!!!!!

[shy kelp]

is 1.0 on clipdrop now?

[paper phoenix]

you say incorrect amount of fingers and hands, i say he has the pick of destiny in his back pocket.

[west breach]

maybe someone is giving a helping hand? I guess we'll never know...

noticed I occasionally get these weird blotchy areas in an image

[stone fossil]

rough sea.

[languid scarab]

Hi! Yes! ,-

[ionic dragon]

can i dm?

[delicate grotto]

A1111 cant use 0.9?

[trim orbit]

It can. there is an update for it. I think for now, comfyui is the best way to use 0.9. The way you can get it set up on that side, gives more technical control over the model. It has two text encoders which require different style inputs. A linguistic descriptor, and then additional support tokens for the second prompt.

I don't think A1's ui have the ability to input two prompts or set up the refiner model to chain with the base generations

[delicate grotto]

i didn't see any update, but there is a branch with sdxl support

question is can this branch also use 1.x?

[smoky patrol]

https://cdn.discordapp.com/attachments/1101178483608137918/1129774923217387661/a_woman_running_on_a_road_view_from_the_side_tornado_in_the_background_steps-30_style-Photographic_seed-0ts-1689429720_idx-0.png https://cdn.discordapp.com/attachments/1101178483608137918/1129774680065192048/a_woman_running_on_a_road_view_from_the_side_tornado_in_the_background_steps-45_style-Photographic_seed-0ts-1689429670_idx-0.png

[amber fulcrum]

Honey, I Shrunk the Kids 2024 by Pixar

[smoky patrol]

https://cdn.discordapp.com/attachments/1101178483608137918/1129776735924260924/a_woman_running_on_a_road_away_from_a_tornado_with_the_tornado_in_the_foreground_and_woman_in_the_background_steps-51_seed-0ts-1689430126_idx-0.png

[amber fulcrum]

Nike. "Just do It!" 🤣

[smoky patrol]

https://cdn.discordapp.com/attachments/1101178483608137918/1129776628755603456/a_woman_dancing_in_a_ball_room_with_roses_in_the_foreground_and_mirrors_in_the_background_steps-37_style-Watercolor_seed-0ts-1689430106_idx-0.png https://cdn.discordapp.com/attachments/1101178483608137918/1129776931634692196/a_woman_dancing_in_a_ball_room_with_roses_in_the_foreground_and_mirrors_in_the_background_steps-33_style-Watercolor_seed-0ts-1689430186_idx-0.png

[novel canyon]

Yeah you have to run git switch sdxl or something similar

[delicate grotto]

yeah, i saw that
can the switched branch run the regular models?

[smoky patrol]

OMG SO PRECIOUS NEED

[trim orbit]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/tree/sdxl yeah its a branch, mb.

https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/11757 heres the notes on it. looks like it does old models still

[uneven dove]

smooth like soft serve, no texture

[wicked frigate]

that's from 0.9ish base cause it's the file i got sitting around on local

[novel canyon]

I haven't tried it out. I can't run any of the regular embeddings on the new branch though

[trim orbit]

real world tornado. for comparison. While generations could be a little better, i don't think "smooth" is an apt criticism. Whhirling high speed wind tends to smooth over a camera exposure no matter what you do

[delicate grotto]

yeah he did mention that embeddings dont work on 0.9

[trim orbit]

regular embeddings? you mean, 1.5 embeddings?

[novel canyon]

how do you tell the difference? I'm using easynegative

[uneven dove]

i live in canada haha i have seen them in person and on video

its okay that the image has an issue you dont have to leap to google images to defend it

[trim orbit]

rename the files so they say "1.5" or "xl" in the filename. Alternatively, the UI will cache all compatible embeddings on model load, so if you use the extra networks button, they'll all show in there

you need to train new XL embeddings with kohya's newest version, to use them on XL

[delicate grotto]

closest i've been to a tornado was ~40 miles away, 5 10m waves in the water at ~30miles away from it

man it was scary AS ALL HELL

bit closer and we would drown

[trim orbit]

I actually said the image could be better. I live in Canada too. We get 1 major tornado every decade or two? National news when thye happen here

[delicate grotto]

how do tornados look up close?

[uneven dove]

yeah we were at an intersection in the prairies and saw it off in the distance and my mom yelled at my dad to drive and the light was red so he didnt want to

[delicate grotto]

ahahahahaha the man the legend

[uneven dove]

in saskie we get a lot of tornadoes and ontario gets derecho storms...

[delicate grotto]

got any pics?

[uneven dove]

i was like, 6

[delicate grotto]

would be awesome to make an embd out of them

ah

[trim orbit]

Sure but they'd be smoothed over if he did

[uneven dove]

ok boomer you just got here, dont make enemies already

[trim orbit]

even with highest iso, it's hard not to motion blur whirling wind

[trim orbit]

have actually been here since october. nice to be back. we've talked lots.

[uneven dove]

i will do it on a damn polaroid

[delicate grotto]

werent you in 2.1 for like half a year? why do you have the new icon?

[trim orbit]

again, i'm not saying the sdxl generation is good. just that "smooth" may be the wrong qualifier for why it's not good

[uneven dove]

smooth is just what sdxl does tho

[trim orbit]

left because reasons

[uneven dove]

its the ideal denoising framework

[delicate grotto]

ah ok, well happy to see ya back

[uneven dove]

ideal denoising means soft serve

the tornado itself isnt even the biggest problem in that image

its the grass, the fence

[trim orbit]

I get it. You're hung up on smoothness.

BUT, this criticism for a motion blurred tornado photo may not be apt.

capice?

[uneven dove]

nope, because its a continuous time model with time dimension to it...

[trim orbit]

"Details matter in an investigation" - Jack Reacher. Just binged that show last week so the quotes are perculating

[uneven dove]

'Be careful what you wish for' - Mike Hunt

[wicked frigate]

[uneven dove]

i would eat that tornado

like a miniature effigy version

[trim orbit]

Maybe that's the reason all images look "smoothed" sure. I'm not denying that XL has a smoothness to it. You're right about that.

BUTTTT, a tornado, a photo of one, may look smooth no matter what. Coming from diffusion models or cmos sensors. That's all i'm saying. Softserve looking tornado would be a better way to describe why it looks fake. The smoothness actually benefits the photo in this case

the gigantic dairy queen swirls don't

[uneven dove]

again look at the rest of the image

[trim orbit]

If it was about the rest of the image, why did you flex about me never seeing a tornado as your first knee jerk response?

[uneven dove]

2.x goes crispy when it breaks and sdxl goes uh, soft serve incoherence, 'artsy'

[delicate grotto]

thing is that tornadoes aren't something that is usually captured in pictures
and those pictures that do have it most likely lack focus - hence the smoothing problem imo

[trim orbit]

I don't think i'm trying to make enemies but you seem lioke you're taking this very personal

i appologize and i'll just not interact less. nice to be back guys

[uneven dove]

it was about multiple things and go look at the video from the Alberta tornado last week

[trim orbit]

Those were national headline news. I saw them chief. thanks.

[delicate grotto]

have you tried making embedding yet?

[trim orbit]

Made a few loras but haven't had free days off to try embeddings yet since the update came out

[uneven dove]

anyway not sure what that was but moving on, blocked, etc lol

[delicate grotto]

sweet, how did the loras go?

[trim orbit]

surprisingly well! i did some friends and it really captured them well

[delicate grotto]

im guessing it still takes ~4 pics?

[trim orbit]

faces weren't so great. I was using full body photos and wasn't cropping to the faces. i can improve the data set here a lot i think.

i use 10-20 for all my loras. Some of my datasets have 50 images but i think that's not needed in my experience.

[delicate grotto]

have you tried to take pictures around them? one of the lower body, one of the upper and one of the face and repeat a few times?
it usually screws up, but sometimes gets the person well if it is prompted well

[trim orbit]

anyways. i don't mean to piss anyone off. i've already caught a block from a prominent regular because i talked about motion blur. going to get going. I'll pop in as conservatively as I can, but i lack self control so we might see you more still. This is exciting times to be involved in AI communities

[trim orbit]

i've only ever tried 1.5 loras with full body shots and cropped closeups. I think showing a photo of legs would be weird and i just got a bad intuition about that

[delicate grotto]

i didn't see you doing anything wrong tbh

[trim orbit]

yeh me neither, but i got blocked and probably reported. mod showed up if you noticed

[delicate grotto]

meh, you're fine

don't get caught in every small detail

[trim orbit]

pissing regulars off is never something that leaves me feeling fine. Regulars get all ornery when threatened

[quartz sequoia]

anyone tested how fast a 3090 generates on SDXL

[trim orbit]

shiut. self control. gotta go

[delicate grotto]

you are a regular yourself

[boreal bough]

hello~

[uneven dove]

mornin

[autumn forum]

Mornin to ya

[autumn forum]

It’s good. Pretty fast I get 3-5 it/sec

[uneven dove]

its fine but varies, deffo better than a 3070 and way better than a 3060

for inference you likely wont see any benefit from 3090 over a 3080

[quartz sequoia]

around how many seconds for 25 samples?

[uneven dove]

depends your settings lol

no one can say for sure

probably 2-3 min tho

[quartz sequoia]

cuz i have a 1070.... that takes 2min and 30 seconds for SDXL

[autumn forum]

For me 25 steps would probably take 20 sec

[boreal bough]

batch 8, should give it a good heads up though, based on my batch tests

[uneven dove]

on my 4090 i get 4 images in about 25s at 1152x768

[quartz sequoia]

for a 1024 img?

[autumn forum]

Yes

[uneven dove]

sdxl is a minimum of 1024

[quartz sequoia]

y'all need to get on my m1 macbook pro speed 45min for 1 img, 1024res.

the higher the number the better obviously

[uneven dove]

thats why i didnt get a mac

[boreal bough]

my condolences

[quartz sequoia]

but im considering upgrading my 1070 to a 3090 for my pc

[uneven dove]

using mps?

[quartz sequoia]

but im seeing people say 3 seconds

10 seconds

now 20

[autumn forum]

I moved away from MacBook when I started gaming lol. But mac is very good for photo editing!

[delicate grotto]

the dude who made macbooks needs to be excecuted

[uneven dove]

yes, it varies

[boreal bough]

has its uses - just not ML or AI gen

[visual glade]

if I do --gpu-only --bf16-vae 20 steps on the base only on my 3090TI is: Prompt executed in 4.86 seconds

[quartz sequoia]

i know that, but 25 samples, 1024, euler a

[uneven dove]

i do 4 images at once, and it takes 25 seconds on 4090

[delicate grotto]

i tried it with photoshop, clip studio, blender, gaming etc...
sadly it was inferior in everything to my 7 year old laptop

[autumn forum]

Frrl?? I should try those

[boreal bough]

should we do high vram cli setting, or does the ui automate best settings on 3090/4090?

[quartz sequoia]

could you provide ur results?

[uneven dove]

euler a is your problem, try ddim

[quartz sequoia]

would that perhaps maybe... maybe speed up my 1070 by the slightest inch?

[visual glade]

if you have the vram do: --gpu-only, highvram is only enabled by default if you have more vram than regular ram

no

[delicate grotto]

unipc might work too

[uneven dove]

not in a1111 ahahahaha

[autumn forum]

I just woke up man lol just gotta trust me

[quartz sequoia]

i used it as just my normal sampler.. i switch between them.. im just comparing atm with what iv tried.. i dont want to wait another 2min and 30 seconds again

[boreal bough]

did literally all of us just get up?

[uneven dove]

ddim 20 steps only

[quartz sequoia]

i was talking about comfy settings he just wrote 😂 , like once you test that out

[autumn forum]

Oh yeah for sure I can give you my results when I try it. It may make my vram spike and fail who knows

[uneven dove]

vram smoke escapes

[visual glade]

--gpu-only only works well if you have lots of vram and --bf16-vae works only on 3000 series and up with nightly pytorch

[uneven dove]

comfy bf16 is only 3080 and 3090

[quartz sequoia]

idk what happens on auto111.. tile controlnet used to work... but now i keep running out of vram on my 1070...

not sure how to use controlnet on comfy yet

[uneven dove]

not 3070 or under

[visual glade]

oh then only 3080 and up

[trim orbit]

I have got a roop custom node for comfy UI, and wired that onto my SDXL generations. Still learning a lot here and doing this was a cool way to figure out how to use Comfy's node editor. Only thing is, i don't think this custom node has a gfpgan step. might have to do that myself

[autumn forum]

I have a 3090 would I see a benefit?

[uneven dove]

idk what benefit it would be, bf16 still uses more vram

it could be a hair faster

if it is hardware accelerated, bf16 is about 3pct slower than hardware accelerated fp16 on rtx cards

but i havent done a lot of comparison for fp32 vs bf16

[visual glade]

this is without it, vae in fp32: Prompt executed in 4.97 seconds

[uneven dove]

it could be way better than that on certain cards

[visual glade]

so it's a tiny bit faster overall

[uneven dove]

true, its hardware specific tho so next gen gpu could support bfloat even better

i doubt it because its machine learning specific and nvidia hates their users

[quartz sequoia]

honestly anything better than 2min and 30 seconds

or 45

[uneven dove]

youre already beating a 1660, which takes 4 minutes

[quartz sequoia]

yeah but found a used 3090 for 650

fuck it why not

[uneven dove]

go for it

you can even train sdxl with that

have fun, ma boy

[arctic bloom]

i had error with --dont-upcast-attention and --bf16-vae on 3060

  File "/tmp/ComfyUI/comfy/ldm/modules/diffusionmodules/model.py", line 58, in forward                                  
    x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")                                            
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/lib/python3.11/site-packages/torch/nn/functional.py", line 3931, in interpolate
    return torch._C._nn.upsample_nearest2d(input, output_size, scale_factors)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: "upsample_nearest2d_out_frame" not implemented for 'BFloat16'

[visual glade]

that's because you don't have nightly pytorch

[arctic bloom]

i see, thanks. atm on 2.0.1

[uneven dove]

comfy relying on a library =p

[boreal bough]

any quick command to update it? or do I need to manually run the python embed

[visual glade]

on windows I would suggest downloading the standalone with it: https://github.com/comfyanonymous/ComfyUI/releases

[boreal bough]

ah noice ❤️

[vagrant plover]

@valid valley hey how aer you

@dull nimbus hehe i see you

[void skiff]

What shall we do when there is prompts like "a girl holding her vag***" and folk trying to make images of small girls in adult situations? I seen several of them rolling past in the bot-1 cannel, the images been clean, but it still try to make bad images.

[sharp robin]

Prices are plumeting. I can buy one for 600 rn. Let’s go.

[cold mica]

We have a weekly Stable Diffusion art contest over at #1087493421209485393 if you guys are interested in a challenge! Runs throughout the weekend

[quartz sequoia]

they barely exist where i live.. i found 1 by luck

[boreal bough]

finally figured out why I was on 3gb vram overhead
Nvidia Broadcast -> 1.5gb vram
2k monitor -> 200mb vram each
4k monitor -> 400~600mb each

but biggest issue was obviously nvidia broadcast

[cold mica]

You can report them by reacting with ⚠️ or right click > apps > report to staff!

[sharp robin]

What do we win?

[sharp robin]

This doesn’t work.

[boreal bough]

on the bot only. made this same mistake yesterday

[cold mica]

Nitro, DS credits, the elusive Stable Society role, showcased on the server and shoutout on our socials ^^

Also it's just fun to participate and discuss with folks

[boreal bough]

didnt realize bot has its own reportability x_x

[uneven dove]

sometimes the App -> Report to Staff is broken and i think it's because the bot has a worker thread that crashes/does not reconnect

[cold mica]

Yeah that's a discord thing unfortunately! Reacting with ⚠️ is the best option

[void skiff]

I thought a bot should spot that, I once got a warning for use "bass".

[cold mica]

It sends the message directly to our review list and we can take quicker action

[sharp robin]

That’s good to know

[void skiff]

I am old old old and not so custom to use discord, but I should learn to use reactions, I mostly stick to a thumb up and a heart.

[cold mica]

That's all you need Kenny

[boreal bough]

#1087493722645725184 message
no regrets XD while there were better from last night, I'm not sure where tomato horror stops being acceptable

[cold mica]

Aww haha the aesthetic is definitely horror but it looks endearing with the lil tomato

[trim orbit]

I've got a theory. The leak 0.9 was strategic so that people who were concerned about how it trains but for various reasons didn't want to sign the research license, could get up to speed before the official 1.0 drop. just a theory though. going to be very very cool when 1.0 drops and hits the ground running

[boreal bough]

it was a whole movie plot I went through XD

[sharp robin]

Movie stuff is all 9:16 right?

[trim orbit]

16:9 but yeah

[sharp robin]

[boreal bough]

1920x1024 gets better results than 1080. just a heads up

[trim orbit]

it varies a little but that's the general format. 16:10 is good too

[cold mica]

2.39:1 as well if you can manage to get that AR 👀

[autumn forum]

I will definitely check this out!

[uneven dove]

frisky business

[quartz sequoia]

are weights normalized in comfy?

[smoky patrol]

Whatever was this plot?

[cold mica]

I'll spitball

attack of the sentient tomatoes and it's all up to one suburban house cat to stop this madness

[smoky patrol]

Sign me up! I'm ready to watch it in theaters! LFG

If it's anything like Attack of the Killer Clowns, we're in for a good B movie flick, esp if Rifftrax comes along

[boreal bough]

from cat horror, to obtaining infinity stones, to can infinity stones be turned into tomatoes (and would they lose their power), to nop, infinity tomatoes, to trying to run infinity tomatoes in fp16, to int8 infinity tomatoes causing a NAN, end of world

[cold mica]

It can be "gory" without being gory

[smoky patrol]

Somehow, I know a certain cat that's perfect for this movie

who is currently, I think, outside the door, batting something around

[lusty raptor]

lol i have the indie movie documentary cover for this #🏅｜pantheon message

[boreal bough]

love it 🤣

[lusty raptor]

grainy "soup" though 😦

[boreal bough]

too many film/still prompts? one is max to avoid grain

oh, saw the prompt. nah, that's just rng settings doing its magic XD

[lusty raptor]

it's an older version, sdxl sometimes has problems with spilled liquids looking more like dirt. I had a series of "epic shot of a woman crying over spilled milk" and it had a similar thing going on

[smoky patrol]

Someone like...

https://cdn.discordapp.com/attachments/1101178530865352815/1129802222348275794/a_black_cat_photobombing_a_picture_steps-55_seed-0ts-1689436221_idx-0.png https://cdn.discordapp.com/attachments/1101178553900478464/1129802746808242186/a_black_cat_photobombing_a_picture_looking_out_the_window_wanting_attention_in_strong_sunlight_blue_pillow_with_cat_hair_on_it_in_the_background_steps-38_seed-0ts-1689436341_idx-0.png

this kind of cat

[trim orbit]

just wanted to clarify. earlier i mis spoke and said i'd been here since october. just now remembered that i started the hobby in october and found this server in december when i geared up. @delicate grotto

[smoky patrol]

I love the little basket, ahahahaha! ❤️

[urban fjord]

Running SDXL on an A100 is so much nicer than on my 3060, if only they were several magnitudes cheaper 😄

[boreal bough]

vae in bf16 is amazing x_x

vae decoding reduced to 0.2 seconds

regardless of batch size

[urban fjord]

So it is better than fp16?

[boreal bough]

on 4090

[boreal bough]

prob only worth it on 3080/3090 and 40xx equivalent

[sudden cliff]

I don't know much about vae, how's the loss in precision?

[boreal bough]

currently doing lithograf style - where I notice literally no difference

[lusty raptor]

reminds me that i need to bump my install to a torch nightly too

[urban fjord]

Is there anything that needs to be done to run ComfyUI effectively on an A100 GPU?

I seem to have huge performance issues when trying test my LoRA.

[boreal bough]

A100 gpu floating in the air, a gift from the gods, lithograph, risograph

[visual glade]

are you using --gpu-only?

[urban fjord]

I used highvram, I can try --gpu-only.

[delicate grotto]

point stays the same, you've been here for a long while

[urban fjord]

CPU is running constantly at 100% when in a LoRA workflow. I might just have to run on a CPU heavier machine next time.

[visual glade]

I can fix that

[sudden cliff]

comfy, is prompt weighting for sdxl in at all, like for clip_l?

[visual glade]

it's in but the algorithm is the same as SD2.x and 1.x

[sudden cliff]

Ah so may be borked

[visual glade]

and it seems to work less on SDXL

the algorithm does what it's supposed to but it looks like the effect is nowhere near the other model versions

[sudden cliff]

I heard that being said but thought it was experimental not already in, hmmm

[urban fjord]

But yeah, even on gpu-only it seems to be slowing down before stepping in the k-sampler node despite having more than enough VRAM and RAM, only CPU is struggling. And it slows down in both text-conditioning nodes too. Terminal window isn't giving much of any info. Other than missing {'cond_stage_model.clip_g.transformer.text_model.embeddings.position_ids'} at base model load which might be my fault for missing an instruction.

[visual glade]

yeah I figured out how to fix it, just wait a bit for me to make sure it works and I'll push it

[urban fjord]

Ah sorry, I just wanted to give the additional info I could give in case it was necessary.

[uneven dove]

in general, SDXL responds poorly to prompts

the prompt embeds not working as well, unfortunately tracks with that

a long prompt, under the 77 token limit, seems to ignore everything but the most heavily-weighted terms

[visual glade]

I have seen evidence that making a model follow prompts less leads to people liking the images better

[urban fjord]

It should be optional.

[uneven dove]

well yeah. the aesthetic score on the base model results in that.

[urban fjord]

Sure it is nice to have a model that filters out noise from the prompt but you shouldn't make a worse model just because some people never learn to prompt properly.

[uneven dove]

^

SDXL just can't do complex subjects, which makes me like its images less

i need to get the prompt weighting implementation going on my bot though, so i can under-emphasize the terms that "take over the prompt"

[sudden cliff]

compared to what? Seems kind of middle of the road between SD and DF IF in that way

[uneven dove]

how about just compared to Bing Image Gen?

[sudden cliff]

The prompt weighting right now I can't down-power those powerful terms at all

[sudden cliff]

That's one of the better ones in that way for sure

[uneven dove]

yep, and SDXL has two text encoders. we don't know why. it didn't help with prompt comprehension lmao

it was claimed that T5 didn't work but i don't see how they can know that when they didn't fully train the model on T5, it just must have been small toy model experiments

currently it's not looking like the dual text encoder solution is doing the job, so, i don't see how T5 would have been any worse other than in efficiency

[sudden cliff]

Curious if there's a specific area of sore spot you wanted to create and can't?

[uneven dove]

its a general issue

say you try and create an 'anthro version of <animal>' but that animal is too powerfully weighted, and 'anthro' just never gets picked up

[urban fjord]

I see some benefits of the dual text-encoder, or rather giving different prompts to both but nothing is conclusive yet and it is impossible for me to test against a model that only has one encoder as that doesn't exist.

[sudden cliff]

I agree with the inflexibility but I'd rate it at like 0.9 of Dall-E2.x and 0.5 as good as DF IF

[sudden cliff]

Yeah animals are overpowered for sure

[urban fjord]

Anthro as in furry?

[uneven dove]

it happens with cars, too

[sudden cliff]

With cars what I do is lower my CFG

[uneven dove]

there's WAY too many images of cars in the dataset

i constantly run at like 1.7-3.4 CFG, how much lower do i need it?

[sudden cliff]

Oh that's really low

[uneven dove]

trust me i split it between 7.5 and low like that

i see it all

[sudden cliff]

I was able to do the car things I wanted to do (for now) at CFG 4

but at 7 it was overbaked city

[uneven dove]

tried to make a porsche 911 spyder version of the M1 Abrahms tank

two overfitted subjects battling each other for supremacy

[sudden cliff]

The Abrams is overbaked for sure 😄

I have to do a CFG 5 just to not get deep fried

[uneven dove]

but you can't get them to combine

this is 3.4 CFG

[sudden cliff]

BTW have you ever gotten that to work in Dall-E? I just tried and it's just a Porsche

[urban fjord]

Does anyone use Dall-E?

[boreal bough]

underweighting works really well and is important. words like disney need to be turned into (disney:0.7), or else they'll overkill the rest of the prompt

[sudden cliff]

I haven't gotten the underweighting to work so well... but I only just tried now

[sudden cliff]

pseudoterminalx, why I mentioned it

[boreal bough]

maybe in the short run, but in the long run, thats how you end up with gpt-4, which also corrects you, and gives you what you actually want. -which is important for the majority of users users

keep in mind the community here represents the absolute minority

[sudden cliff]

I'd be really disappointed if it was more inflexible than SD, so I think they're still headed the right direction

[uneven dove]

so far so good

[sudden cliff]

Can't share it, national secret

[uneven dove]

ok, that's amazing

Bing wins again, really?

welding babies, porsche tanks..

[boreal bough]

welding babies

I... have questions

[uneven dove]

SDXL wouldn't even do that

i might need prompt segmentation to make it work

[shy kelp]

when people rate images they rarely account what that image will be used for. not following a prompt can be good for making good looking images that aren't what the user need

[urban fjord]

Yeah I don't vote for broken outputs even if it follows the prompt better than a good looking output

And when people write an essay as a prompt in showdown I just vote for what looks nice.

[boreal bough]

do I need to be feel bad now? xD

[uneven dove]

god it has E.D.

someone get that car some Cialis, quick

[urban fjord]

When using styles you do not even know how well it follows the prompt as you do not know the prompt.

[uneven dove]

i love how it's driving on donut tyres and has misaligned headlamps

this is like a video from youtube where some guy in India made a DIY ferrari out of garbage

[sudden cliff]

teased the prompt a bit
gonna keep trying, I want to see it too

[uneven dove]

oh i never knew i want to see someone tokyo drift an Abrams M1

[amber fulcrum]

Addicted to these ones - would love to see some retro 80/90s vibe animation from pixar

[lusty raptor]

i had some similar retro cars+military stuff in 1.5 doing it this way

[sudden cliff]

Putting it on a CRT makes it much nicer

[uneven dove]

sorry pal we moved on from amazingly hand drawn and curated artwork, we're into the era of sterile 3D models with no life to them 😁

[amber fulcrum]

I actually did liked some of 3d animations - what left is mostly anime

[uneven dove]

i miss the early Simpsons art style where they would, for example, do a "pan" shot down from the attic of the house, down to the basement, showing you cutaways of each floor as they go past, incl stuff hidden under the floorboards. they still do that in newer episodes, but the perspective has 100% straight lines and no "warped" perspective that was common to hand drawn art

[amber fulcrum]

Yeah - I am dreaming of making by own CRT monitor - like repairing some old one but putting into nostalgia/retro vibe DIY case

[amber fulcrum]

my favourite animation was peter pan. Magical.

[sudden cliff]

I like this one

[boreal bough]

tank made by (mercedes:1.4) racing down akihabara with police in background, neon punk, tokyo drift :: City center dripping with black ink and black slime in the background, lights reflecting gasoline colors:: Bojan Jevtic + Ashley Wood :: maximalist intricate detailed :: ray tracing :: hyperdetailed, maximalist, psychedelic, post-apocalyptic, photorealistic, 64k resolution concept art, dynamic lighting, trending on Artstation :: hybrid car made by mercedes :: bmw style tank

went all out on prompt there

[uneven dove]

nice, i liked Pinnochio and the art within that one

it's so immersive, and terrifying

oh, or An American Tail

[delicate grotto]

try adding planetside 2 tank - somewhere in the start

[uneven dove]

brilliant drawings in there

[delicate grotto]

looks like a stargate

[boreal bough]

and the finished image

[delicate grotto]

mind doing 10 steps on this pic again, and on step 10 add stargate?

[hollow halo]

How do I add something mid gen?

[delicate grotto]

what ui do you use?

[hollow halo]

ComfyUI

[sudden cliff]

OK I'd call SDXL the winner

[delicate grotto]

mmmm 2 sec, ill see if comfy supports that

[hollow halo]

I think maybe if I just feed it an extra Ksampler to it that starts at step 10

[eternal fog]

I was going to suggest that ^

do 0-10 steps with 1 prompt, pass the noise across to another and do 10-20 with another prompt

[delicate grotto]

how is the 0.9 with anime?

[boreal bough]

oh well XD :: prompts might have hit the limit here

[eternal fog]

That's from an early version of WDXL. It's extremely finicky at the moment, most of what it generates is not good, unless you get very very specific prompts.

[sudden cliff]

so street racing?

drifto?

[uneven dove]

how? the bing output has way more details

[delicate grotto]

what ui do you use?
im assuming comfy too right?

[boreal bough]

car brand -> tank -> tokyo drift was the theme

[eternal fog]

Yeah

[uneven dove]

i got four images on one shot in Bing and they were each detailed and amazing and you had to work super hard to get a mediocre result in SDXL, but, SDXL wins?

[delicate grotto]

does it download extra stuff when you set it up?

[sudden cliff]

Lol do you work for OpenAI? that's crazy

[eternal fog]

I just linked it to the venv I use with Auto1111 so no

[uneven dove]

another Bing result

@sudden cliff uh, no, i just have functioning eyes

[delicate grotto]

so that is an option, great

how do you link it? im sorry for asking lots of stuff today

[sudden cliff]

It's a green porsche...

[eternal fog]

Activate the Auto1111 venv before you launch comfyui

I've just made a batch file

D:\Code\Stable-Diffusion\AUTOMATIC1111\stable-diffusion-webui\venv\Scripts\activate && python main.py

[uneven dove]

SDXL, same prompt

[delicate grotto]

umm what do you mean activate the venv?

[uneven dove]

zero-detail tank

oh, it has a porsche logo on the front

my bad, it worked

[uneven dove]

#✨｜sdxl message this

[delicate grotto]

im not that fluent with what this means, i get into cmd and write my path?

[eternal fog]

the activate at the end of the path is a command. It tells it to use that python environment

[sudden cliff]

ok well show me a single image that's not just a painted porsche with weird tubes attached to it and I'll say this isn't a Bing fail

[boreal bough]

@uneven dove ufff porsche has some genuinely weird weighting x_x messed up my prompts that worked on every other brand

[uneven dove]

my point isn't that Bing is the best image generator, it's that it has better prompt comprehension for heavily weighted subjects. and if you don't like that, i don't know what to tell you

SDXL really sucks at combining overly weighted subjects

[sudden cliff]

I think it's marginally better at that, so I can agree

[uneven dove]

and there's workarounds for it in SDXL

but you don't have to do those on Bing

[sudden cliff]

Yep

[boreal bough]

porsche has... background weights? wth XD

[hollow halo]

Figured out the workflow i think, going to try that one again @delicate grotto

[uneven dove]

i like the holes in the hood on the Bing version, reminds me of the simpsons

"What... keeps... DOING THAT?"
"Those are speed holes."

[uneven dove]

that's a very nice tank, Irish, but it's no barn

[delicate grotto]

you can use this to delay or to remove a prompt, or to lower a prompts strengh across time i guess

[glad fulcrum]

Suddenly my pc shuts down when trying to open automatic1111 🥺

[hollow halo]

Here's that one with stargate added 10 steps in, I just added it to the beginning of the prompt

Oh wait it wasn't the same seed lol

[delicate grotto]

a bat file like this could work?

@echo off

D:\Code\Stable-Diffusion\AUTOMATIC1111\stable-diffusion-webui\venv\Scripts\activate && python main.py

call start.bat

[uneven dove]

multiple people get blue screen

[eternal fog]

You don't need @echo off and you don't need to call the bat your running it from

[eternal fog]

Put what I posted in a bat file in the comfyui folder and put in the path to your Auto1111 Python venv activate @delicate grotto

[glad fulcrum]

Trying comfyUI now. Works well. Must be an Auto1111 problem

[eternal fog]

WDXL really likes to do watercoloury type stuff

[boreal bough]

cabrio tank by bmw

(porsche has so much bad bias, its hard to even list all of it)

[hollow halo]

Batmobile just came out of a portal

[amber fulcrum]

what do you trying to achieve? anime?

[uneven dove]

damn, i really like wheeled tanks lmao

[hollow halo]

Not just realistic looking portals today

[boreal bough]

same prompt with only 'bmw' -> 'porsche'
kills all style, background is more baked in than the porsche itself

[eternal fog]

Just testing out what they have for the WDXL model at the moment, so yeah anime stuff.

It's still quite low on the training

[compact flax]

The Ultimate Hoedown of Ultimate Destiny

[delicate grotto]

@eternal fog mah man

[amber fulcrum]

Like the perspective of this shot

[delicate grotto]

any way to save prompts in the comfy?

[uneven dove]

@sour obsidian it's almost like indiscriminately training on 12 billion images that SAI doesn't have the rights to, is a bad thing

[eternal fog]

The entire node setup gets saved into the image metadata

Just drag and drop the image onto the page and it will configure it all for you

[boreal bough]

no, but you could do it like me
each green note is a prompt I saved in a note,
each red is a negative prompt I saved in a note

[delicate grotto]

ah, that is useful

[eternal fog]

This is true scatter brain territory

[boreal bough]

gonna get a mindmap custom node XD then its gonna get even worse!

[delicate grotto]

great thanks

[delicate grotto]

yeah, seems like notes are the way to go

[eternal fog]

hmm weird, I've suddenly started getting black images, but only on Euler A - Normal

[boreal bough]

I used to use the saved images as well, but it was painful, cause they didn't reflect the changes I added to my ui x_x and loading in to just extract the prompt is a bit painful

[glad fulcrum]

Is it normal to do this?

[eternal fog]

Yeah this is my only gripe with it

[glad fulcrum]

I feel like its taking too long between diffusion

[delicate grotto]

well, 6gb vram can do 1.25it/s for 1024

[boreal bough]

depends on your vram, but normally no

[glad fulcrum]

I have 12

[boreal bough]

then no

[glad fulcrum]

How can I fix it then?

[boreal bough]

oh wait, you're going over 1024x1024 -> you sure you're not running out of vram?

[eternal fog]

Isn't 12GB VRAM that horrible spot where the memory management has a fit and can't deal with it properly.

And they aren't running out of VRAM, I can do a lot higher than that with 10GB VRAM

[boreal bough]

do a 1024x768 and see if it speeds things up by around x4

[glad fulcrum]

I am not using all my VRAM

still it does that.

[eternal fog]

Are you just talking about the speed?

[glad fulcrum]

it loads something before diffusion

[high skiff]

WOOOOO

new PC gets here today

[boreal bough]

no 'fix'. just method of elimination of finding out how not to fix it

[eternal fog]

If so are you changing your prompt between these runs?

Changing the prompt makes it have to reload the text encoder

So it will take longer

[high skiff]

hello everybody

[compact flax]

The Sacking of Goldman-Sachs

[delicate grotto]

anyway to show cheap image generation mid gen?

[autumn forum]

Hello sytan 👋

[high skiff]

[glad fulcrum]

well inerestingly enough, it loaded faster now with 1024x1024

I will do more testing

[eternal fog]

There's a command for it I believe, run the main.py with --help and it will give you a list of commands

[glad fulcrum]

like 5 seconds before diffusion started, still not inmediate

[high skiff]

nice!

currently trying to repair my neutered A1111 install lmao

[uneven dove]

sytan's daily self-doxing is done

[compact flax]

The Embargo of Wells-Fargo

[delicate grotto]

there is a preview node, just cant understand how to connect it to the latent

[uneven dove]

"the rare and elusive auto-doxxing"

[boreal bough]

keep in mind background stuff also uses vram. since 12gb is just on the edge of 'fast', be careful to not tip it over

[high skiff]

moving my python install broke so much stuff

[eternal fog]

Preview just shows you the end result, it's not a live preview

[uneven dove]

did bill gates appear and grant you 3 wishes and you used them for a CPU, PSU, and GPU, and not a working User account location? nooooob

[high skiff]

hey, I will not be making that mistake with my new PC lmao

[lusty raptor]

protip: two letter username, makes life so easy

[eternal fog]

Your new username on Windows should match your discord surely ⛧ Sytan ⛧

[compact flax]

Every knows you make two wishes then make a wish for three more wishes

[high skiff]

I just need to figure out how the hell I am gonna keep my current PC turned on in order to transfer info between them lol

[high skiff]

if the unicode characters were legal, I totally would lol

[delicate grotto]

it does preview, but only on the last step...

[uneven dove]

just make sure it's not 'cd'

[delicate grotto]

mmm

[boreal bough]

try to make your name a code break space block using Alt+255

[eternal fog]

Look at the commands, pretty sure there is one for live previews

[delicate grotto]

dunno if i can call it preview

[lusty raptor]

is that a thing? i mean, of course it is... ugh

[boreal bough]

surely that will not cause issues XD

[high skiff]

I will just make it Sytan lol

[uneven dove]

@high skiff you should make your new username

' `_` '

[high skiff]

screw you lol

[boreal bough]

ヽ(´ー｀)ノ

[high skiff]

new username:

¯_(ツ)_/¯

[uneven dove]

can't even access it in Linux live env

[boreal bough]

(｡•︿•｡)

[uneven dove]

Welcome to Windows. Enter a username:

.

[compact flax]

Make your username :(){ :|:& };: and then type that in as a Linux command

[high skiff]

what if I make my user user

[boreal bough]

dont. genuinely dont

[uneven dove]

make sure to wrap it in `s so that it gets executed if someone tries to encapsulate it in "s instead of 's

[high skiff]

I won't lol

its gonna be Sytan

[boreal bough]

XD

[lusty raptor]

be sure to make a backup user with different case

[compact flax]

[high skiff]

its a whole new PC

[boreal bough]

user/guest/admin (and localized translations of user) - are off limit names... uff the pain that has caused me in the past while in IT

[compact flax]

Why? What the fuck Microsoft.

[delicate grotto]

--help crashes it

[high skiff]

cool, my auto is fixed

[visual glade]

can you try the latest I might have fixed it

[eternal fog]

You're doing something wrong then

[delicate grotto]

yep

[compact flax]

I know Guest is a built-in account, but user and admin, go fuck yourself Microsoft

[visual glade]

what error does it give you when it crashes?

[delicate grotto]

im running a bat to use the venv of a1111, i use this:
E:\stable_diffusion\stable-diffusion-webui\venv\Scripts\activate && python main.py --help

[uneven dove]

the CEO of a company i had as a client was named David Guest, and he wanted his last name as his username everywhere, i hated that guy

[delicate grotto]

its unrelated to comfy

[eternal fog]

It's not crashing @delicate grotto It's closing the window because it's finished

[delicate grotto]

guessed i screwed up somewhere

[eternal fog]

You need to run the start.bat within CMD if you don't want it to dothat

If you double click it will just exit once it's finished running

[visual glade]

yeah that's why my comfyui .bat files have a "pause" in them

[vital wolf]

Stupid question but how do you interrupt generation midway in comfyui? @visual glade

[eternal fog]

See queue under the queue prompt button and then click cancel on the job

[visual glade]

yeah in "view queue" you can remove stuff from the queue and cancel the current one

I mean "see queue"

[vital wolf]

Man, I checked everywhere except "view queue", so stupid and thanks!

[urban fjord]

I can't right now but will test later and report back.

[high skiff]

what issues are being checked in the new comfy UI update?

[high skiff]

PEEPER

[eternal fog]

Thats one big eye

[high skiff]

PEEP PER

[hollow halo]

Here's another lol

[eternal fog]

that I dislike

[high skiff]

upsetting lol

[hollow halo]

The area around the eye and the pupil bother me lol

[uneven dove]

the iris?

[violet shale]

what even is sdxl

[hollow halo]

nah the extra eyelashes

[uneven dove]

that's just facial hair / stubble

[hollow halo]

I guess I've just never been so close to someone's eye lol

[uneven dove]

doesn't everyone's beard grow up to their eyes?

and then around the forehead

and merge with the top

[eternal fog]

howdy

[uneven dove]

wonder if SDXL can reproduce the image of that guy with the beard loop around his head like a lions' mane

[hollow halo]

Last eye gen I'm doing right now, feels like my main monitor is just staring at me

[delicate grotto]

can comfy train embeddings and such already?

[uneven dove]

ah that looks like they have Lupus

[hollow halo]

Is that from the pattern in their eye that you say that?

[uneven dove]

lion-man vs man-lion

[uneven dove]

it's 5 seasons of House, MD. that makes me say that

[hollow halo]

Fair lol

[winter raptor]

How would you guys recommend creating an image of 1170x2532 resolution in Comfy? I would like to do it by upscaling by a certain amount using the 4x-ultrasharp upscale model, but I can only do it by 4x. Is there a way to specify the new resolution I want when upscaling?

[eternal fog]

Just use an upscale image node after to make it smaller again

[glad fulcrum]

How can I avoid saving the base image?

[eternal fog]

Dont put it into a save image node

[icy brook]

have you used the regular finetuner the same person made for colab? I don't get it to save ckpt's/safetensors anymore. and they won't respond when I post errors or contact anywhere.

[urban fjord]

@visual glade It is much better now but there are still issues. Even with gpu-only it lags in the ksampler and when changing LoRA weights there are some lag in the text-conditioning node but a lot less than before and I guess that's just how long it takes to change the LoRA settings.

[upbeat summit]

Hi all

[delicate grotto]

oi mass, how ya doing?

[winter raptor]

Hi mass

[eternal fog]

I swear SDXL does not want to work for me today lol. I keep getting gens with loads of artifacting

[delicate grotto]

maybe do the 2.1 tech?

[hollow halo]

Are you using the same sampler on refiner and base?

[eternal fog]

Yeah it's not all images, it's just happening sometimes

[hollow halo]

I noticed it happens if denoising isn't set to 1 for base

[eternal fog]

Like this is the exact same but with a different seed

[hollow halo]

unless you're using advanced then idk

[winter raptor]

I'll give it a try. Been trying to figure out a good way to do it without reducing the original resolution too much, because I can't upscale past 8k and then downscale after that

[eternal fog]

It's just randomly being ass lol

[hollow halo]

lol

[winter raptor]

2.1 tech?

[lament rune]

yeah what is that!

[glad fulcrum]

wait, my base images are better than my refined ones with my current prompt

[delicate grotto]

oh boi, prepare to see your eyes fucked
here i go

bad drawing, bad painting, horribly drawn, bad hands, (etc...), bad drawing, bad painting, horribly drawn, bad hands, (etc...), bad drawing, bad painting, horribly drawn, bad hands, (etc...),

[delicate grotto]

3 times, shit ton of negatives

the more the merrier

[eternal fog]

The stupid prompts don't work and it also doesn't make that much difference in 2.1

[upbeat summit]

just got up and can't wait to see what you've all made

[eternal fog]

You can get the same effects without going mental on your keyboard

[glad fulcrum]

the refined one looks like with dots

[winter raptor]

ohhhhh, negatives. I thought you meant somehow incorporating the 2.1 model lol

[delicate grotto]

just got 0.9, can't wait to see what i can make

[uneven dove]

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

this is the most powerful non-trained negative embed for 2.x

fwiw

[upbeat summit]

have fun!

[delicate grotto]

nah, the tech for 2.0 was to use negatives to build your pic instead of positives
2.1 for a goddamn reason likes when you repeat to it 3 times what you don't want to have

we had like ~500 different prompts in the negatives

maybe 2 5 in positive

[clever verge]

Is this supported {red|purple|orange:1.3}?

[uneven dove]

that's just because no one did prompt weighting properly.

trust me, as someone who spent a lot of effort fine-tuning 2.1, it does better with more precise and shorter prompts, just like SDXL.

[winter raptor]

What negative prompt have you been using in SDXL?

[delicate grotto]

we tried that for months

i have no clue why short worked so badle, it should have worked well

[uneven dove]

none, or the one i posted earlier

i use aesthetic scores and positive prompts to direct SDXL

[cloud silo]

a gigantic mouse, robotic, cybernetic, outdoor installation art,by thomasz sientoswki

[uneven dove]

idk, a lot of fine-tuning techniques that came about for 2.x are just recent discoveries that improved the results. StabilityAI is to blame for that, too

[delicate grotto]

mmm more like laion

their set was trash

[uneven dove]

i've used LAION data to improve 2.x

[urban fjord]

The Laion dataset was never intended to be used unfiltered.

[uneven dove]

SDXL is trained on LAION data

you're blaming the people that put something together that SAI used improperly

please do more research before saying stuff like that

[delicate grotto]

yep, this time though they used a great way to improve it with a mass scale rating that improved it

[uneven dove]

no they didn't lmao

[delicate grotto]

yuval was doing a great job at it

[uneven dove]

they used LAION's own aesthetics scores

[delicate grotto]

heard about pickapic?

[uneven dove]

i don't care what they used to generate the aesthetic scorer, it's the same exact scores

you can visualise their meaning, here: http://captions.christoph-schuhmann.de/aesthetic_viz_laion_sac+logos+ava1-l14-linearMSE-en-2.37B.html

[delicate grotto]

well, basically they used the dataset that he made to improve it

[uneven dove]

they also used the same exact NSFW filtering as before, and it's really bad

look at laion's dataset, filtered so that the NSFW column is "Definitely", tis a bunch of beautiful photos of faces, hardware store inventory (screwdriver shafts lol)

i've been going over this stuff with a fine-tooth comb for about 3 months now

Stability is afraid to innovate, so they tend to stick to what others have already researched, and combine others techniques

it's like an unofficial R&D branch of NVIDIA

[lament rune]

Full agreement with this. NSFW filter seems exact same as 2.1

[uneven dove]

yep, i think it's the same ancient NSFW scorer that everyone put together on that Github repo with like, 65 million NSFW images

you can tell it's the same one because the Clipdrop results get blurred waaaay more often than they should be

[true wing]

Honestly I don't give a shit how they did it, SDXL is a much better base model than 1.5. Sure 1.5 had more NSFW out of the gate but vanilla 1.5 was more likely to give you some body horror monstrosity than anything 'useful'. People can add NSFW with fine tuning and LORAs, there just needs to be enough incentive to actually do that which I'm sure SDXL's quality will cover.

[lament rune]

That's what Stability is banking on

[uneven dove]

dude i'm glad SDXL doesn't do NSFW, i'm just pointing out that SDXL is pretty much "same old, same old"

[lament rune]

I agree that the speed improvements and quality will make everyone make the jump