✨｜sdxl

[spring fulcrum]

I am trying to get this

but all i get is stuff like this

[late marsh]

Cyborg coffee mug? 😄

[runic lark]

Are there any good comfy tutorials or YouTuber describing how to use inpaint/detailer and such?

[thin nova]

maybe overlay is the wrong effect lol

[late marsh]

I have to stop soon, it's getting too awesome 🥹

[peak dove]

Jazz Carnival using Sytan

[hoary saddle]

ah, maybe i need to upscale the overlay image first to match the output, oops

[nimble heart]

use cpu for seed instead of gpu

also if you're hiresfixing use the same seed for both passes in comfy cause I'm pretty sure that's what auto does

[trim orbit]

love that cyborg lora. i'll have to toy with it

[peak dove]

I d/loaded latest Sytan w/flow, yet it will still not UpScale?!

[trim orbit]

what?! sytan used me in his new layout!?

[runic lark]

How do you hiresfix in comfy? Upscalebymodel?

[trim orbit]

thizizanoUtRAgE

[indigo carbon]

I like it way better with gpu though, makes cleaner results

[peak dove]

Floral Flourish using Sytan

[nimble heart]

then don't complain they're not the same image?

[late marsh]

How different can random numbers be that it produces different kind of results? Like cleaniness of the image

[nimble heart]

hiresfix is just scale then run through sampler again at a lower denoise. So for example if your auto hiresfix is "Latent bicubic" with a denoise of 0.6 you use the latent upscale node set to 'bicubic' followed another ksampler node with denoise set to 0.6

[indigo carbon]

I didn't complain, I just asked if it's possible to reproduce the same results I get from A1111 in ComfyUI.

[nimble heart]

and I just told you

comfy uses cpu seeds so if you want them to match you need to set auto to cpu seeds

[indigo carbon]

Idk, the results from GPU seeds are slightly better than CPU seeds IMO

[nimble heart]

you'll have to open an issue on the comfyui github to support gpu rng then

[vale eagle]

left comfyui right a1111

[indigo carbon]

still not the same as:

[nimble heart]

even those are pretty far apart. Mine are usually similar down to attention noise

[hoary saddle]

you're a genius, "screen" does it at blendfactor = 1

[late marsh]

I think you are imagining things because they are just random numbers, just saying, but whatever suits you. 🤷‍♂️

[vocal rapids]

so real.....

[nimble heart]

I was going to say this but idk if it affects just the seed or also the sampling. some samplers use RNG during their sampling

real 1:1 would be to use Euler on both

with cpu rng

to make sure there's no sampler bias

[indigo carbon]

they listen to the prompts differently. for example: try to import one of my gens into ComfyUI, I can almost guarantee you the A1111 gen would be better.

[nimble heart]

are you using weights?

like (token:0.7)?

[vale eagle]

same prompt

[indigo carbon]

like, please, I can't recreate this type of images I made on A1111 with ComfyUI

[indigo carbon]

nope =\

[nimble heart]

lies

[vale eagle]

I could do the similar result in comfy and a1111

but some are different

[indigo carbon]

do you count (PROMPT) as weights??

[nimble heart]

yes

[indigo carbon]

I didn't use any numbers..

[nimble heart]

it's the same thing

parenthesis are just like 0.95 by default or something

[vale eagle]

Same pormpt and setting in both.

[nimble heart]

and I was saying auto normalizes the weights of all words so they average to 1.0. Comfy uses the literal weights you specify

so if you prompt is slime cube (green:1.2) in auto what it actually conditions is (slime:0.917) (cube:0.917) (green 1.2)

or something like that

[indigo carbon]

ohh, so that's why I couldn't get the same results as I get on A1111 in Comfy?

[nimble heart]

besides the gpu rng thing yes

also im pretty sure auto uses the same seed for both base and hires fix

in your comfy workflow you had a different seed for high res

[indigo carbon]

huh, I'll stay with auto then, IDK

[livid geode]

umm, anyone here got experience running SDXL 1.0 on apple silicon? I read some benchmark numbers and theyre really hard to believe to be true

[nimble heart]

but yea if you do a simple non-weighted prompt like banana taped to wall using cpu rng on auto without any special tweaks they come out basically the same

[vale eagle]

Same prompt happy slime cube bouncing in a field,1111111111111111111111111111111,trending on art station,

[nimble heart]

if you think it's important enough you could maybe look into opening an issue about the gpu seed or attention normalizing or both on comfyui assuming no one else has. Though I get the feeling the answer to the latter is probably 'no'

auto does some other kerjiggery with the prompts too like BREAK and AND keywords + im pretty sure they automatically concat when token limits are reached

[solemn scroll]

Aside from impainting, would you have any advice on how to improve image quality and coherece ? Here are my current generation settings :

[indigo carbon]

damn

[nimble heart]

I'm just speculating. Auto seems to take the more opinionated route, whereas comfy is the make no assumptions route.

[peak dove]

Ravishing Beauty - Sytan w/flow

[indigo carbon]

maybe we could make a custom node that optimizes the prompts like auto does?

[nimble heart]

I considered that

[broken hare]

ddoes anyone have the NMKD superscale file

[nimble heart]

it would be doable by automatically setting up the combine/concat/timescale nodes based on AND BREAK and [] keys in the prompt

the [] specifically would be hella work to parse and turn into a graph though

the noramlizing thing would be easy though

[crisp owl]

I use the 4x_NMKD-Siax_200k upscaler

[vale eagle]

both on 75 characters in prompt

76 characters

80 characters

[nimble heart]

yea if you go above 75 in auto you see the little counter suddenly jump to 140 or 150 cause it concats

[broken hare]

trying to use what is in the Sytan workflow. new one

[nimble heart]

try using the 'concat' conditioning thing in comfyui

see if it gives the same results

[crisp owl]

Yeah I use mine with essentially his same flow

[broken hare]

where can i get it

[crisp owl]

https://huggingface.co/uwg/upscaler/blob/main/ESRGAN/4x_NMKD-Siax_200k.pth

[peak dove]

Mine won't Upscale

[nimble heart]

if you know any python you could probably do it yourself. It wouldn't be hard just a bit time consuming. I personally never used BREAK/AND and [] that much and the normalizing thing doesn't bother me

[broken hare]

thanks...let me give it a shot

[nimble heart]

most of my 1.5 prompts were just like 10 plain words with no negative

[indigo carbon]

I will consider making a Comfy custom node after Comfy supports AIT optimization, until then I'll stay on A1111.

[hoary saddle]

guess "screen" does something with the alpha channel when it sees darker colors, gonna have play a bit more

[vale eagle]

It is not the "token" show in the top right. It is character count. 75 characters still under xx/75

[nimble heart]

that's what I assumed the different results after 75 were about

[nimble heart]

screen is basically more light is more opaque

idk what you're trying to do but maybe overlay would work better. It's closer to the difference in in lightness

Overlay's pretty hardcore though might need a lower factor

[hoary saddle]

is better with a white background

basically just trying to overlay a transparent png on top of final image

[nimble heart]

if you're just watermarking a couple images it might be easier to use GIMP or something since it has a million layer modes

[hoary saddle]

keep thinking it's trasparent where darker colors are even set to 1.0 on blend

[hoary saddle]

got do it all via python for the backend of my web renderer tho, may have to do it all with pillow

was just hoping to get lucky and save a step 🙂

[nimble heart]

you could probably just write a custom image filter node. most overlay filters are like 1 line of math

most of the gimp ones have their formulas right on the GEGL website

[hoary saddle]

will look into it, i have to pull some data from my database to fill the rest of the overlay image anyway

wish gimp was api friendly, tired of using blender for everything haha

oh, hmm, maybe i can do a double overlay with the first blend just a white rectangle

[nimble heart]

3.0 is dropping soon with a new scripting api. maybe new cli too idk

[nimble heart]

https://gegl.org/operations/svg-overlay.html Gegl overlay formula for example

[hoary saddle]

thank you kind sir

[nimble marten]

does emphasis work in ComfyUI?

I can't tell

[nimble heart]

like (token:1.2)?

[vale eagle]

I think the differents came from how comfy/a1111 handle the token limit outside of 75(or 77)? Left is a1111 Right is comfyui. Prompt: happy slime cube bouncing in a field,trending on art station,trending on art station,trending on art station,trending on art station,trending on art station,trending on art station,

[nimble heart]

concat

[nimble marten]

ye

[nimble heart]

yea. you can even use ctrl+arrows to adjust weight like in a1111.
They're weighted differently though. A1111 normalizes all tokens to be near 1.0, comfy inputs the weights you specify literally

[nimble marten]

hmm

[nimble heart]

tl;dr weights needed will be slightly different

[hoary saddle]

bah damn you lol

[nimble heart]

tf you tryna do

[hoary saddle]

lol, got the perfect corner overlay but output is all black hah

[nimble heart]

mask it based on the corner's alpha

since the corner is only a small cube the rest of its alpha shouldn't cover the image

[hoary saddle]

this is what i want, just not having much luck

[nimble heart]

is the rightmost image already composited?

[hoary saddle]

that was photoshop lol

[nimble heart]

the corner thing

or is it separate pieces

[hoary saddle]

composited, single transparent png

[nimble heart]

dont see why you can't just use it directly

[hoary saddle]

i'll use python and coordinates to add text later

[nimble heart]

png alpha is called 'mask' in comfy so use that when compositing

[hoary saddle]

hmm i was playing with mask earlier, will give it another shot

[nimble heart]

if you're using PIL for text anyways you could just

for pixelA, pixelB in image, overlay:
  if pixelB[3] == 1.0:
     pixelA = pixelB 

[hoary saddle]

probably best bet 🙂 think i'll go back to that as the original idea so i can just add the text from the db for the player stats first then pop the corner graphic on already loaded with data

[nimble marten]

[nimble marten]

SDXL even got the text holy crap

[woeful patio]

amazing

[nimble marten]

wait are these SD or real?

[fleet harness]

LOL

right?!

[woeful patio]

no worries tho he's breaking out

[fleet harness]

These are screen grabs from this morning right? lol

XD

[vale eagle]

I would answer you now. A1111 and comfy handle prompt differently after 75 characters that why you would get different result on a1111 and comfy.

[nimble marten]

gotta stop at Minonad first

[late marsh]

https://tenor.com/view/the-universe-tim-and-eric-mind-blown-mind-blown-meme-mind-explosion-mind-explosion-meme-gif-18002878

[fleet harness]

Im dead LOL

[woeful patio]

😂

[hardy cipher]

someting is off. these don't look like beanie babies

[fleet harness]

I laughed so hard I died lol

[late marsh]

Same 😂

[nimble marten]

lol

Lool

[hardy cipher]

I mean, how long will trum even live? I'd be surprised if anything sticks and he actually sees time. he's 77, obese, lives off fast food, wears diapers. probably not going to be a spry 90 year old

[fleet harness]

this is how you lift bro

[urban breach]

perfect technique

[fleet harness]

you know what they say. its all in the wrists

[hardy cipher]

it's all about the form

yes

[nimble marten]

lmao

[urban breach]

it looks like they built him into the table lmfao

[fleet harness]

LOL his jumpsuit would look like that

[charred beacon]

Mustang drivers never change

[fleet harness]

Average Ford Mustang enthusiast...

[woeful patio]

trump after 3 years in jail

[hardy cipher]

oh yeah, that's pretty close to what he looks like nwo

[fleet harness]

in other news... dog guards priceless cermaics from the Han Dynasty; c. 206 B.C. - 220 A.D.

[woeful patio]

aren't those the expensive dogs?

[nimble marten]

🤣🤣

[fleet harness]

Chow Chow? I think they can be

[hardy cipher]

I was trying to make spaghetti but something went wrong

[charred beacon]

you have a created a monster

[fleet harness]

The only way to kill it, is to eat it.

[nimble marten]

that is absolutely terrifying lmao

[fleet harness]

https://tenor.com/view/fooly-cooly-flcl-naota-anime-gif-17776112

good god man

[vocal rapids]

enjoy pineapple pizza , lol

[indigo carbon]

A1111

comfy

[woeful patio]

damn

[fleet harness]

https://tenor.com/view/here-comes-a-new-challenger-street-fighter-video-game-arcade-game-gif-17584657

[nimble marten]

wat

[hardy cipher]

almost

[nimble marten]

the text tho!

[late marsh]

Ok I'm pissing honey now ❤️ 🥹

[hardy cipher]

weirdly enough I just saw an article earlier today about how the john hughes movies promoted bigotry. I don't know if that means they're getting cancelled or what

[woeful patio]

place your bets

[hardy cipher]

jeez, they're yoked

[hardy cipher]

that kerning, lol

[nimble marten]

I mean it's not terrible kerning, lol

"A 1960's movie poster for The Matrix as if it was filmed in the 1960's"

[left coral]

is lycoris working for SDXL?

[hardy cipher]

that's pretty quality

[nimble marten]

ikr

this is insanity

[hardy cipher]

I can't decide if I prefer to make mind blowingly good or terrible things with this technlogy

[fleet harness]

https://tenor.com/view/bet-casino-gambling-victor-polizzi-john-magaro-gif-15213368

[fleet harness]

[fleet harness]

ancient artifacts are fun to make IG

Opium Pipe with Dragon Monkey, China, circa 1900

[nimble marten]

"A movie poster about the corrupt chicken farming industry, text: "Mother Cluckers""

[brazen patrol]

West Side Story already beat you to creative ways to swear in the last line of the song, "Gee Officer Crupke... Crup you!"

[copper rose]

gdi the refiner destroys my loras

[hardy cipher]

I like the vintage

vintage images

these scenes don't bring me joy when I look at them

[nimble marten]

"A movie poster about the illegal cheese mines on the moon"
This looks pretty good compositionally

lmao

[hardy cipher]

nice

cheesse

[nimble marten]

CHEEESSEEE

[hardy cipher]

I got one like that

[nimble marten]

uhm what

lmao

this looks like a terrible photoshop

[hardy cipher]

ugh,I'm really fed up with whatever this aiohttp program is

File "C:\Python311\Lib\site-packages\aiohttp\web_urldispatcher.py", line 348, in add_route
raise RuntimeError(
RuntimeError: Added route will never be executed, method HEAD is already registered
NativeCommandExitException: Program "main.py" ended with non-zero exit code: 1.

or library

whatever

[hardy cipher]

that cheesy moon van is pretty legit

[trim orbit]

they're whalers on the moon!

[hoary saddle]

is something missing with the comfy API example using python?
it always dies on the second attempt, first comes out fine, then i have to kill it and restart comfy
seems like maybe it's leaving the connection open and just can't start the second render

[hardy cipher]

moon mining cheese might be rough, but it'd beat living in an existential bean nightmare.

[nimble marten]

I like how the gloves and helmet visor are yellow like cheese

[trim orbit]

visor looks cheese lined so the astronaut has continuous nutrition

[nimble marten]

yes

[hoary saddle]

nm, found an answer

https://github.com/comfyanonymous/ComfyUI/issues/1019

[nimble marten]

and the space suit is made out of cheese cloth 😄

[hardy cipher]

top notch attention to detail

[vale eagle]

Sad Comfy

[nimble marten]

lmao "A documentary movie poster about the lost city of cheese"

[indigo carbon]

this is with comfy, I'm working on a workflow

[vale eagle]

You might create an issue for that. If you certainly confirmed comfy produced image is not as good as A1111

[nimble marten]

i find Comfy currently produces better quality images than Auto1111

and is a tad faster

[indigo carbon]

I am working on a comfy workflow, so far it seems to be about as good as A1111

[hardy cipher]

I find that comfy is now broken because of this aiohttp thing. might just nokethe entier install and start over

[nimble marten]

not sure what that is tho

i dunno much about comfy

[vale eagle]

Great. I am interested in that.

[hardy cipher]

well, just this error over and over and over "RuntimeError: Added route will never be executed, method HEAD is already registered
NativeCommandExitException: Program "main.py" ended with non-zero exit code: 1."

and not sure how to rectify. created a new virtual environment, same thing

so might just delete everything but images and models and start over

[vale eagle]

It is hard to tell. I am not saying comfy is better than A1111 or another side. But they are producing different images after 75 characters in the prompts

[unreal plover]

so good theme

[nimble marten]

lmao wat

[mellow tendon]

I have had to recopy my python_embeded folder back over from the download 7zip a few times in the last few days (so almost everything) as adding modules kept breaking my install and I couldn't fix it.

[late marsh]

Oh yeah! I had cyborg style lora enabled at 0.7, magnificent

[brazen patrol]

so I think there will be those who wonder about the big commercial rival, and I will say this. A couple of months ago I spent a good deal of time trying to make an image of a Storm in a Teacup (with a floundering ship to illustrate it) and and this was the typical set I got:

Now, after a dozen tries in SD I got the equally impressive :

[hardy cipher]

not a bad idea. I'll try that out when I have a minute. I got used to wiping installs with a1111. probably installed that six or seven times, lol. easier for me than trying to resolve all the issues

[nimble marten]

the text is getting insane!
"A movie poster about Batman and the Cheese Goblin"

[mellow tendon]

You should add some of these to the Civitai Galley for this Lora, it is looking pretty empty right now,

[late marsh]

Oh yeah, I always forget 😅

[hardy cipher]

some of my last pre-broken comfy images, lol

I don't really know why it made that

[spring fulcrum]

Hey everyone look... It got the eyes right for once... lol

[woeful patio]

I asked for a labradoodle dog and it came up with this, not quite sure what happened lol

[hardy cipher]

I think it did drugs

[unreal plover]

is it available on civitai or huggingface ?

this one ? https://civitai.com/models/119405/cyborg-style-sdxl-or-goofy-ai

[late marsh]

Yes!

[unreal plover]

ok i think it is the good one, thanks

[late marsh]

Just uploaded images

[nimble marten]

hehe "Call of Doody"

[true wing]

Any news on the control net front?

[late marsh]

Done!

[hard monolith]

is A1111 good enough for SDXL or do I need to go with comfyUI? any suggestions?

[indigo carbon]

idk, the best results I managed to get out of SDXL are with A1111+refiner extension. but I might be biased, idk

[true wing]

I couldn't get it to run SDXL at all when it first dropped. I've heard there have been some fixes but Comfy runs SDXL at 1024 faster than A1111 ran 1.5 at 512 for me

I usually just use a premade workflow and tweak it. It's not too hard to figure out what the nodes do

[nimble marten]

not sure what I expected when I asked for someone stuck in a giant bag of doritoes with only their legs visible...

[indigo carbon]

your A1111 isn't optimized then. but I must agree that Comfyui has the highest potential.

[true wing]

mmm Dorito's breakfast cereal, my favorite. Just pour mountain dew on it I guess

[nimble marten]

A complete part of this balanced breakfast!

[true wing]

Maybe. I ran A1111 pretty much exclusives while 1.5 was out and tried a bunch of different settings and tweaks and it never ran faster than Comfy's prebuilt workflow for SDXL. There certainly could have been something I was missing but that's just my experience. Also Comfy has never randomly crashed or bugged out on me, so that's a plus

[soft bone]

anyone managed to fix artifacts like this in lora training?

wonky af

[true wing]

It's like each half of his head has a slightly different pose...

[urban fjord]

What artifacts are you thinking about? You can try to reduce the max time-steps to 800 or 900.

[true wing]

Yeah if you cover up either half of the picture it looks fine, it's split almost literally down the middle. Weird.

[soft bone]

This was the only epoch with good enough likeness. It's like if you want correct likeness you're guaranteed artifacts just like 1.5 and 2.1. If I don't care who the person ends up looking like then its easy

[urban fjord]

I'm talking about this setting here, not number of epochs.

[soft bone]

yeah most of them are like that

[urban fjord]

Oh, the good old misalligned eyes issue.

[soft bone]

what sort of visual issue does this usually address?

[spring fulcrum]

What do yall think

[true wing]

That's so strange, I wonder what causes that? I heard someone talking about the rendering doing some sort of cropped image training in the process so I wonder if that's what's messing it up? I haven't trained anything yet so it's just a guess

[nimble marten]

Doritos Cereal with Mt. Dew Berry Blasts!

[soft bone]

image bucketing maybe?

[urban fjord]

I'm testing stuff here, but the idea is to exert less influence at the end of the inference.

[soft bone]

so you could probably also expect less details like in skin

[true wing]

If she's got pink hair you should try to get the flowers to not also be pink. White or pale blue could be good options.

[true wing]

I think that was it. There's something about it in the information stability released about how they trained the model, I know that much

[spring fulcrum]

good idea how about this?

[true wing]

Still a bit too much pink because the kemono swapped but it's better separated. I'd say it's a better composition.

the other one kind of blurred together because of all the pink overlap

[nimble marten]

Aww ye. Gimme that chesesubeer cereal

[true wing]

It sounds and looks like it would taste like dog food, or actually just be dog food. Like when they put "real steak and gravy flavor" on the bag or something.

[nimble marten]

lmao ikr

cheesbreger

creeal

it's so close!

[true wing]

Can't go without your daily dose of Ducl Afefflt

[nimble marten]

obviously, lmao

[brazen patrol]

so what was the prompt out of curiosity?

[high skiff]

Holy crap, I was not expecting my workflow to go under the amount of support that it did in that little time

[nimble marten]

"A box of cheeseburger flavored cereal"

[true wing]

Congrats I heard you released 1.0. I haven't gotten around to trying it yet but I've heard it's good and I know 0.5 was good.

[brazen patrol]

that is it?

[nimble marten]

ya, lol

[marble dew]

https://tenor.com/view/harry-potter-dumbledore-pensieve-magic-gif-17242310

[high skiff]

I released 1.0, and it has a very high quality 2x upscale workflow built into it with high res fixing

[brazen patrol]

Well, tit for tat. here is my image with w/flow:

[nimble marten]

that's dope as heck!

[brazen patrol]

🙂

[true wing]

"Can I get some service over here?"

[spring fulcrum]

Sytan 1.0 Awesomeness Proof right here!!!

[soft bone]

epoch 1-7. the problem is present in every one. has to be a different param thats fucked

[true wing]

That style is awesome. It's got an interesting mix of almost plant looking elements, humanity, splatter painting, and fire. Very dramatic but also stoic, I like it.

[nimble marten]

i'm not even trying with my prompts, lmao

[woeful patio]

yeah it works really well

[wet glacier]

@high skiff your 1.0 workflow is the bomb, bravo👍

[nimble marten]

this is some SCP type shit what the heck

"A surreal Polaroid photo of an 80's office flooded with spaghetti and meatballs"

[true wing]

I mean you did ask for a surreal polaroid, what did you expect?

[still dove]

which workflow is that btw?

[high skiff]

https://www.reddit.com/r/StableDiffusion/comments/15ed0o8/sytans_sdxl_offical_comyfui_10_workflow_with/

:D

[true wing]

It's Sytan's SDXL workflow lol

[high skiff]

I was not expecting it to blow up as intense as it has 😅

[eternal fog]

@high skiff I managed to have a look at your workflow. You are doing pretty much the same I was doing. But I think with better prompts.

[woeful patio]

I mean you've made some really clean looking images, some of the best i've seen

[high skiff]

thank youuu

[true wing]

Tons of people are looking to switch to comfy and want premade workflows and yours is definitely one of the better ones. I think it's well deserved.

[rustic garnet]

did you tried several prompts and seeds? Sometimes its just bad luck with a specific seed

[high skiff]

god, I feel so much better now haha

[eternal fog]

Oh and a better sampler I think

I'd got fixed on using DPM SDE

[high skiff]

I was so worried I wouldn't find a good upscale workflow, but I was able to do it using a second pass of mixed diffusion

[eternal fog]

And it doesn't seem as good

[soft bone]

yeah tried plenty:/

[woeful patio]

this one in particular is awesome

[true wing]

everytime your Italian grandma asks if you're hungry and you're like "I could eat a little, but just a little ok?"

[nimble marten]

lmao

[soft bone]

will everything break if i turn off the refiner node

[high skiff]

oh, I have a 4x upscale example of that one!

4096x4096

[ornate elk]

omg berks

[nimble marten]

When you're trying to go home but the boss gives you a pile of spaghetti to sort first

[true wing]

The refiner isn't necessary, you might need to reroute a couple of nodes depending on your workflow but whatever you're using should work fine without it.

[soft bone]

yeah i meant on sytans specifically

[eternal fog]

@high skiff I think those negatives you use are doing a lot of work as well.

I don't use them as much and it definately makes worse output.

[hoary saddle]

is there a node for "save as" so i can use a variable to save in a directory other than the comfy output folder?

[eternal fog]

Like with the negative prompts, Left is with your negatives and right is without.

[clever verge]

The owl beak seems to be made out of fur. 😍

[crisp owl]

https://blenderneko.github.io/ComfyUI-docs/Interface/SaveFileFormatting/

[high skiff]

oh weird, I use faily light negatives

[tribal jackal]

Mindlessly scrolling /r/stablediffusion... see wall of text... see tiger picture... huh okay scroll back up. Oh it the guide everyone's been bugging @high skiff to release. Better hunker down lol

[fleet harness]

The statue of Laocoön and His Sons, has been one of the most famous ancient sculptures since it was excavated in Rome in 1506 and put on public display in the Vatican Museums,where it remains today

[ionic dragon]

why did the image mess up?

[nimble marten]

"My dog loves to sit at the table while eating spaghetti"

Not sure if doggo is eating it or vomiting it 🤣🤣

[crisp owl]

He'd probably still eat it anyways

[ionic dragon]

i was running the upscaled image through refiner?

[high skiff]

hahaha, yeah

[woeful patio]

some trippy shit

[clever verge]

By this point I feel that there are no single workflow to rule them all. You basically have to adapt to the image and style you want to diffuse.

[nimble marten]

good doggo gonna get the pizza

[spark bear]

is NMKD Superscale better than the Siax_200k version?

[high skiff]

I did not test against that one, but I did test about 20 different pixel upscalers for my workflow

[shell ice]

Hi. I'm a bit new to using ComfyUI after switching over from A1111. Is there a particular node that does the job of "Inpainting conditioning mask strength"?

[ionic dragon]

@high skiff can you explain the 3 samps used in your workflow?

[eternal fog]

lmao that's annoying. Yeah most of my issues were partially from choice of sampler, but mostly shit prompting

[ionic dragon]

i mean Ksamps

[mellow ether]

You are making me hungry. 😅

[high skiff]

sampler one, base, it samples most of the steps, stopping before the final (high detail) steps

Sampler two is a refiner sampler, which takes the incomplete Latent and its residual noise and refines from there, only changing the high detail/resolution parts

that gets upscaled, down sampled, and encoded

third sampler is a base sampler that only does img2img on the high resolution parts of the image. In the case of how I shipped it, the final 33% of the steps

[late marsh]

Flying spaghetti monster!

[high skiff]

LOVE

[hoary saddle]

not seeing a way to save outside the comfy output folder

[high skiff]

Pastafarian moment

[late marsh]

Yes 😄

[trim orbit]

may his noodly appendage bless us all

[high skiff]

https://tenor.com/view/spaghetti-monster-flying-gif-20812422

[late marsh]

https://tenor.com/view/fsm-spaghetti-monster-gif-5611270

[high skiff]

HE BOILED FOR OUR SINS

[mellow ether]

prompt: ((colossal):1.3), ((bio mechanical monstrosity)1.5), ((spaghetti and meatballs and lasagna)0.01) on an apocalyptic landscape

[late marsh]

If that would've been made with 1.5 model I'd add Biomechanicals lora to that and test it out 😄

[sharp robin]

❣️

[stone fossil]

❤️ XL.

[clever verge]

Beautiful coloring

[mellow ether]

"bio mechanical monstrosity" seems to work out well in many models. 👍

[eternal fog]

@high skiff Did you notice in any of your testing at all that if you use dpmpp_2m_sde/normal as the refiner scheduler it wrecks the image?

It's fine on karras, but for some reason if I use it with normal as a refiner step, it doesn't clear the noise properly.

[nimble heart]

it's both 2m samplers. I opened an issue on the comfy repo for this

[late marsh]

Yeah probably.. try adding H.R. Giger style in it 🙂

[high skiff]

yes, there are lots of combinations that do that

Thats why I tested and found that DDIM/Normal is the most reliable and least volatile experience

[nimble heart]

it affects both SDXL base and refiner

[eternal fog]

It's just on the refiner though right?

And Normal.

It works fine on base and if you use Karras.

[ionic dragon]

oh, so the 3rd samp does hires fix?

[eternal fog]

Oh it seems perfectly fine on base for me. It only effects the refiner.

[high skiff]

my workflow was made specifically for DDIM

[nimble heart]

normal?

or karras

[eternal fog]

normal, karras is fine.

[nimble heart]

so normal does or doesn't work on bae for you'

base

[eternal fog]

Normal works fine on base, not on refiner.

Well, actually hold that thought

[urban breach]

lol

[nimble heart]

cause on sdxl base dpmpp_2m and its sde version are broken using non-karras schedulers. Examples
https://github.com/comfyanonymous/ComfyUI/issues/1034

I also had masslevel and caith confirm it

[vale eagle]

@high skiff Very interest on your upscale method. I think I could update my mixsampler using your upscale idea of contrast fix. It turns out the image going to be darker after multiple loops.

[eternal fog]

Ok it was "working" because I wasn't finishing the steps before passing to refiner.

If you do allow it to finish, yes, it wrecks the image on Normal.

[high skiff]

you can adjust the values after every loop to get some cool lightening or darkening effects :>

[soft bone]

@high skiff if I load a face lora in your workflow and disable refiner.. am i still losing likeness to the mixed diffusion step or does it still use the same models and prompts for that?

[nimble heart]

yea the last step fucks it up on the sdxl architecutres for some reason. doesn't happen in 1.5 or 2.1

[high skiff]

if you load in a LoRA, only attach it to the base model, and do not touch the refiner at all

it should still work, assuming you use the upscaler

[nimble heart]

a1111 had something like that happen too

[mellow ether]

I have found the Giger style in sdxl to be rather strong, I have had to dilute it with other styles. It's a good mixer.

[soft bone]

well i have to disable refiner cuz it ruins likeness of the lora subject

[high skiff]

and the upscale will bring it back

Just trust me, give it a try? I made it that way on purpose

if its an extreme LoRA, then it might mess it up too much

but for any sensible LoRA, it will work just fine

[soft bone]

oh i get it. mixed diffusion is img2img. excellent. I'll try again

[eternal fog]

The refiner can properly bork LoRas

[vale eagle]

yeah, I also noticed that.

[ionic dragon]

Emad testing controlnet, hope it will be out
https://twitter.com/EMostaque/status/1685104715514716160

Emad (@EMostaque)

Spot the interesting things 👀

#SDXL

Likes

384

Twitter · Jul 29, 2023, 1:47 AM

[eternal fog]

Just skip the refiner step and let the base finish all the steps if it's causing an issue.

[vale eagle]

after the upscale, the lora give the likeness back to the image

[nimble heart]

what if you only run the refiner for like 5 or 10%

[hardy cipher]

ugh, stupid nsted nodes breaking things

[nimble heart]

don't use that extension then

[hardy cipher]

I figured it out though. ONE line of code to break it all

[nimble heart]

just write your own node 👍

[mellow ether]

0.5 Giger, 1.3 dali

[candid spindle]

has anyone else tried, letting the refiner do most of the work? Left is the refinded output and right is the base output.

[hardy cipher]

how very passive aggressive of you

[nimble heart]

genuine suggestion

[hardy cipher]

anyway, I just #ed out one line of code

and it was fine

[eternal fog]

Do that on a close up face and watch it melt lol

[nimble heart]

that's what I did

[hardy cipher]

ahh, well you know, I have thought about it

setattr(server.PromptServer, "add_routes", add_routes_wrapper)

[nimble heart]

you don't need to know anything about torch or whatnot you can just re-use the existing comfyui nodes

[nimble heart]

escape your pound signs

[mellow ether]

the previous one was just plain bio mechanical monstrosity, here it's spaghettified

[candid spindle]

well

[hardy cipher]

I thought about combining some nodes that, at least for my purposes, don't need to be seperate

[eternal fog]

Yep lmao

[vale eagle]

Kind of learned a few tensor to output multiple images

[candid spindle]

beauty is subjective

[nimble heart]

I've made it this far not knowing what a tensor is so I'd say great success

[hardy cipher]

tensors are those things that cause errors

becaues they're a million different sizes

[nimble heart]

I wanted to port my image calculator library to comfy but when saw that the image type was a tensor of pixels I gave up

[hardy cipher]

I want to use the boolean and math nodes to create some logic stuff. not sure exactly what. but I really haven't seen any workflows that do it

[civic sigil]

I just based all my custom nodes off that Psuedo_hdr one that converts it to PIL and back

[vale eagle]

Infinite dead-loop-workflow

[nimble heart]

if I can just get a list of floats instead of a tensor of floats that'd be nice

[hardy cipher]

so a tensor is comparable to a vector or a matrix correct?

but then way higher dimensions

[nimble heart]

as long as I can get a tensor of floats my library can do the colorspace conversions

[hardy cipher]

not sure how to actually apply that

hoping my new chldren of the bean prompt works out. I wrote it a small essay

[civic sigil]

You can treat it as an array, like 1920 * 1080 * 3

dimensions

[nimble heart]

wat

like * 3 being linear RGB?

[civic sigil]

Yeah

[hardy cipher]

people that say long prompts don't work well are just lying to themselves. writhe a small noveil. it's the way

[nimble heart]

after the decode ofc

[civic sigil]

Like img[100,100,2] will give the blue channel value of the 100,100 pixel

[nimble heart]

brb porting my image calculator to comfyui

[hardy cipher]

I saw that there are nodes to play with that data, change dimensions, swap values. but I can't think of anyhting I could do in that regard that wouldn't just break it

[civic sigil]

But comfyui passes images in batches so it would be more like img[0,100,100,2]

[sweet abyss]

can anyone tell me please how do i install this https://github.com/wcde/sd-webui-refiner extension ?

[late marsh]

Randoms 😛

[urban breach]

so angry 😂

[woeful patio]

in a1111 click on extentions and klick on load, then search for the refirner extension and click install, then klick restart gui and it should be there at the bottom of txt2img

[thick harness]

SDXL v0.9 is wayyyyyy better than v1.0

i dont think the experiment of fine-tunning according to votations on discord worked properly

[trim orbit]

what tests have you done?

[twin bluff]

my favorite blonde... see resolution 🙂 sdxl is omg wow

[eternal fog]

How did you determine this?

[trim orbit]

accurate data is important during investigations

[thick harness]

using the same prompts

[vale eagle]

@high skiff I have a suggestion to your upscale workflow. If lora attached to base model, you might use another lora loader to adjust the lora strength and reduce the denoising strength of 3rd sampler which might be add too much lora effect to the image.

[eternal fog]

There's your problem.

The prompting style has changed a little between versions. At least in my experience.

[thick harness]

using the same library of prompts i had made and coparing results

[trim orbit]

i'm just joshing you. 0.9 is actually MORE refined in a lot of ways. in a way that makes it HARDER to refine better .... so they gave us one with more knowledge that's looser. edit: mb on 1.0 i meant 0.9

throw a lora at it. like even just the offset noise one. It ties thigns up a lot tighter every time

[cyan crown]

in Comfyui what sampler are u using?

[high skiff]

that would be something a person could decide to do, but its intrusive for people not using LoRA's

[thick harness]

im using the Clipdrop version

[vale eagle]

yeah, it could be adjust

[trim orbit]

ok

[scarlet pine]

So using --medvram command for A1111, my generation takes 31 seconds. Removing only that command and doing the same image, the generation takes 8 min. 57 sec.

[thick harness]

any notebook with weights i can use? lol

[scarlet pine]

Any idea why I'm getting such drastic time difference without --medvram? 10gb 3080

[eternal fog]

You on the latest Nvidia Drivers?

[scarlet pine]

Yep

[trim orbit]

because it tries to fill your memory up too much and you get clocked down to the mainbus speed

[eternal fog]

It will be running out of VRAM, so it offloads it into RAM and it's super slow when it does that.

[molten gull]

postive naked and negative nude? interesting combo

[soft zealot]

if it works lol

[molten gull]

yeah, in the end it's just words, i got that a lot

[solemn scroll]

Hi, what's the current best way of running a fix faces node with comfy for SDXL, please?

[mellow ether]

Would you like your toast, burnt, carbon slab, or molten brimstone?

[high skiff]

yes

[scarlet pine]

I mean, is there some downside to using --medvram? I'm not seeing much different in the images that came out

[runic lark]

I tryed the upscale to a new ksampler in comfyui, but starting with 512x512 is never a good option i guess? You have to start with 1024 and upscale it if you want to use hires fix?

[solemn scroll]

It's impressive how the reflection is handled so well.

[mellow ether]

added an 'unreal engine' to the prompt.

[eternal fog]

512x512 will just generate a mess, it's not designed for it.

[trim orbit]

if you've got enough ram, for full vram support, it's a little slower

in theory

[soft bone]

are you saying you want the lora to be less present on third pass? cuz I need it stronger tbh

[scarlet pine]

Welp. Hopefully one day 😛

Guess I'll stick to --medvram for now

[mellow ether]

yeah, 512 isn't great, better than 1.5 is at 256 though. 1024 is what it wants. 768 is ok.

[soft zealot]

Disclaimer :: I'm as guilty of this as the next person.

Why do we insist on calling it a Hi Res Fix ?

Why dont we just say upscaled ?

[trim orbit]

honestly i have 16 gb and it works better for me in most cases with --medvram. i think it just forces the engine to use optimizations which are essential for sdxl

[scarlet pine]

I'm not seeing many people mention the command online, but I see many people complain about slow speeds. Maybe why

[thick harness]

did this with SDXL v0.9 ...but the prompts dont work anymore with v1.0 ...has anyone found a logical connector, word or use of 'comma/parenthesis punctuation to solve this?

[trim orbit]

i use comfyui more often but i come back to a1 just to see how its coming along

[vale eagle]

Because it is strong enough before the upscale process in my case

I have the image bypassed multiple base-refiner loop which added enough lora likeness to the image

[soft zealot]

is that in ComfyUI or A1111 ?

I'm sure I read somewhere that --normalvram in ComfyUI is equivalent to --medvram in A1111

But I could be wrong, I frequently am 🙂

[soft bone]

@high skiff likeness came back and mutations fixed. fuckin dope. low hit rate though, 3rd pass often adds extra lips or tumors

[mellow ether]

Maybe it's just me, but I kinda liked 1.5's failure sat text. But sdxl 1.0 is much better at text.

[thick harness]

beautiful!

[soft zealot]

it's almost like Olivios younger slimmer brother 🙂

[soft bone]

its hilarious you say that because this morning i was asking this guy what celebrity he thinks he looks like so i can use it in training captions. thanks for the help🤣

time to see if olivio is in sdxl

[trim orbit]

yeah, i only use the --medvram for auto1111. comfy comes with good optimizations and engineering out of the box seemingly

[thick harness]

if v0.9 was a scooter skate... v1.0 is not exactly a motorcycle, its more like a screwdriver

or a soccer ball xD

[spring fulcrum]

Are there any sites where we can actually post AI Art to make money? Does anyone have any reccommendations on sites if they do exist?

[soft zealot]

Ok SDXL smoked crack on this particular one

[mellow ether]

"I do not like green eggs and ham. I do not like them, SDXL-I-Am." .... can't seem to make it generate actual green eggs and green ham!

[soft zealot]

compare apples and apples not apples & oranges

ie

Compare SDXL1 wirh base vanilla SD1.5 not with finetuned SD1.5 output

[trim orbit]

he's comparing 0.9 to 1.0 clipdrop. it's caused a lot of sour people over there

subjectively there seems to be a lot of loud preference for the 0.9 implementation they had

[thick harness]

i was taught this at business school and its the standard for making digital products, software, apps and products in silicon valley, hope you enjoy meditating on this is a lot of fun

[mellow ether]

My green eggs & ham project has gone awry.

[urban breach]

apparently it does not like to make Dr. Seuss lol

[urban breach]

who is this guy lmao

[soft bone]

@high skiff wots going on m8? cant fix this. It does this with UMD start_at_step 18-25 and i cant go higher. The part of the image that worked looks fucking incredible tho

[high skiff]

I am not sure, I have not had a single failed gen in like... over 1000 with the upscale

[soft bone]

I'm sure it's my lora

[high skiff]

oh wait

let me see your nodes

[cyan crown]

so finnaly I trained a Lora...but I have to use high weights.....why?

[high skiff]

undertrained

[soft bone]

send nodes?

one sec

[urban breach]

oh i'm an idiot, Dr Suess wasn't the actual character lol

[high skiff]

lmao

yeah haha

[soft bone]

@high skiff

[high skiff]

ok, yeah, likely that is a LoRA issue

[soft bone]

f

[urban breach]

the good news is, we're gonna feature that guy's face in the next L4D game

[soft bone]

he'd be way too excited about that

[tribal jackal]

It would suck to get punched with that lol

[fleet harness]

🤣

definitely

[cinder aspen]

They really dont make it easy to install SD etc

^^

[tribal jackal]

Mostly we're just spoiled by install wizards these days

[fleet harness]

I like how you istall SD

[tribal jackal]

A little more fluency would be nice, but that's the nature of decentralized development

[fleet harness]

makes me feel like a hacker lol

.>

[urban breach]

this sam-i-am guy sucks, he refuses to generate

[cinder aspen]

I installed ComfyUI, now trying to get to SDXL

[high skiff]

gereen eggs and damn

[fleet harness]

https://tenor.com/view/rock-climbing-rise-song-going-up-hiking-climbing-the-mountain-gif-25792523

youre almost there

[tribal jackal]

just click the download button next to the sdxl base and refiner safetensors on the github page

then move them to the checkpoints folder in comfyui

[cinder aspen]

Yeah downloadikg Base now and then Refiner

[cinder aspen]

I dont even have Pytorch on my PC yet tho and i dont have any idea about VAE and getting Controlnet yet

[tribal jackal]

all in due time

[twin bluff]

i really need controlnet pose with sdxl... any news on it?

[vale eagle]

soon

[polar epoch]

One thing i honestly miss with auto over comfyUI, is "extra networks" so i can see a preview of the damn models, as well as PNG info to easily get prompts/seed from an image which won't rearrange my entire workflow lol

[cinder aspen]

Firefly is my main tool for this kind of work so dont judge me xD

[tribal jackal]

The technological barrier to entry is worth it

[cyan crown]

Me too

[fleet harness]

its like reaching enlightenment

[tribal jackal]

Firefly is nice, but it's still far behind and seems to have plateaued

[cinder aspen]

Im trying to learn to program anyway

[soft zealot]

@hard fractal even more like a "Proper" UI if I turn the links off lol

https://github.com/failfa-st/failfast-comfyui-extensions

[cinder aspen]

Its different with different audience and purpose plus still in beta with all the other features to be released. Irrelevant for average prompter, not so for target audience like me

[tribal jackal]

Also if you try to cancel your photoshop subscription, they threaten you with the early termination fee (varies by plan). I went to cancel anyway and they gave me two free months "to rethink it" and definitely not "to reduce the early termination fee"

[cinder aspen]

Oh i know about the fee but i wont cancel anyway

[vale eagle]

The link off idea is brilliant

[tribal jackal]

Microsoft Edge, of all things, does what Adobe Express can do and they've natively integrated Dall-E with it.

Combine that with Photopea and Gimp...

I'm set.

[cinder aspen]

Yeah everyone to their needs but i bound myself with Adobe ecosystem tbh

And there are more

[tribal jackal]

got lucky with a student discount on a CS6 suite right before they switched to cloud services way back

[nimble heart]

is there an easy way to get a ctypes pointer from a torch tensor cause I'm not finding it

[tribal jackal]

Managed to avoid them until Firefly

[cinder aspen]

for generative AI as of now only Adobe ecosystem is important for me because Autodesk and Maxon dont have those (quite) yet

I cant ask for more now either

@tribal jackal after i got Base and Refiner i guess i will need to look for controlnet and VAE either?

[soft zealot]

Just cleans it up, sure its not very helpful when designing......................

[cinder aspen]

I really dont need much from SD at all, all i need is generating some specific images and for a temporary time controlnet with the doodle thing

[vale eagle]

for the people who only want to generate image by designed workflow, it is too good to use

[soft zealot]

or if like me you enjoy fiddling but also want a cleaner interface for when you've done fiddling

[nimble marten]

Batman has a bit of a moose knuckle going on 🤣

[fleet harness]

That's not batman, that's the cheese goblin

[nimble marten]

lmaoo

smuggling it in his pants

[tribal jackal]

I installed mine using "git clone" and a link to the github repo in my terminal. Both with ComfyUI and Auto1111 web ui's. With your method, I would check the github page for a VAE and do a similar manual relocation as before. For ControlNet... I'm not sure which interface you're using, but...

[nimble marten]

I mean, have you ever seen the Cheese Goblin and Batman in the same room together?

mhmmm

[tribal jackal]

@cinder aspen I have some controlnet nodes from some package I installed but I don't remember how lol

in comfyui.

For Auto1111 it's just a matter of navigating to the Extensions tab and finding it

thenr estarting UI

[cinder aspen]

@tribal jackal i have ComfyUI

[mellow tendon]

@high skiff I managed to break your workflow to only produced mangled images, but I hadn't really played around with any of the settings so I have no idea why it broke (I added in a Lora and took it out again to see if it was that).
Anyway if I reload from the .json file it is fine again. I just thought it was interesting and it's bugging me not knowing how I broke it!

[eternal fog]

I think sometimes LoRAs get stuck and don't unload properly.

[spring fulcrum]

That looks like some SD 1.5 shit there.

[autumn forum]

Looks like you changed the clip width and heights lol

[mellow tendon]

Ahh I did wonder if it was that.

[polar epoch]

Too high native res by the looks of it. Make it native, then highres upscale it.

[mellow tendon]

Yeah it looks like that, but it is 1024X1024 that's why it was interesting.

[polar epoch]

Here's 1440p ultrawide, take res divided by 2, then ultrasharp upscale by 4, then boom, easy 2880p :P

[high skiff]

Oh wow, that's pretty impressive

You know, the entire community is impressive and how severely they can mess up my workflow in ways that I never considered lol

I should start hiring random community members to help test my workflows and see if they can find ways to break them, cuz I ran well over 2,000 images through this workflow and never even had a single hiccup lol

[digital shoal]

when upscaling (latent) with sdxl + refiner, do you plug the refiner into the 3rd pass or the base model? I keep getting weird results

[high skiff]

This is a joke, I have no money to hire anybody with lol

[glad flint]

me VS the desire to create a conlang based around random elegant "characters" in the imitation cursive script sdxl can put out

[polar epoch]

With SDXL base too?

Oh boi, using 1.5 model with SDXL refiner. I'm gonna have to do a few tweaks

[lilac raven]

With the upgrade from 16 to 64 GB, I can use the refiner, finally

[glad flint]

you got more RAM? 🤔 I'm at 16GB and it's tight

[lilac raven]

Yes, It was dying when loading the refiner

[void skiff]

This is odd, I made this prompt "computer generated art, a strong space boy commander dressed in slim silver spacesuits is sitting in a small futuristic tech chair , seen from 45 degree side angle, clear visor . graphical render, digital smooth clean colors, mildly detailed, clean shading".

Change "girl" to "boy" and you get a solid visor, can some help me get it clear so I can see a face?

[glad flint]

You using Auto?

[twin bluff]

i can't get sdxl to make full body from feet (shoes) to head (hair), whatever resolution set. works 1 in 20 times. standard negative "out of frame" doesn't help

[lilac raven]

Comfy, Auto, both died

[polar epoch]

You'd preferably want all within video memory, otherwise speeds are crippled to a 10'th

[lilac raven]

64 GB of RAM is way cheaper than 64 GB of VRAM

[runic hatch]

just a quick question. seems on googel colab is no problem tu run sdxl 1.0 with comfy ui, but is there no way to run SDXL with A1111?

thank you

[glad flint]

How high does the RAM usage get if you give it 64GB? I was thinking about doubling my ram to 32GB

[mellow tendon]

Yes SDXL base, the workflow metadata is in the image I uploaded, if you wanted a workflow that makes twins with multiple limbs for some reason...

[lilac raven]

Right now, ComfyUI is using 17 GB of memory

[soft zealot]

I have 64Gb system RAM, whilst running ComfyUI with SDXL i ytypically hover between 28 & 32Gb used (thats witha load of other stuff open as well inc chrome & excel

[eternal fog]

Has anyone manage to get wings from a person? Every time I try, it always puts them on the wrong side lol

It consistently puts the wings in front

[polar epoch]

So damn annoying. VAE encode won't cancel until done

[eternal fog]

Every time lmao

[upbeat summit]

I don't remember the correct tokens, but search for angels with wings in the bot channels. it's solvable at least with some prompts

[eternal fog]

if they are looking forwards it works

But if they are looking away, it doesn't want to do it

[lilac raven]

It seems it has leaned that wings are behind people, not that wings are on the back of people 🤔

[brazen patrol]

Plateaued on what? It's inpainting is exceptional

[mellow tendon]

I accidentally got an image of someone with wings just now (nothing mentioning wings in the prompt) but they were also looking towards the camera, and I see it does that just fine.

[south igloo]

Did the high res fix workflow come out?

[high skiff]

yes

https://www.reddit.com/r/StableDiffusion/comments/15ed0o8/sytans_sdxl_offical_comyfui_10_workflow_with/

[boreal bough]

good morning yall!
(Nothing like sleeping till midnight) 🤣

[ember cargo]

Working on a much much better Blame! manga Lora, with a much better and curated dataset (took some hours do prepare, and more to come). This is just 1h training

[south igloo]

Thanks a lot I'll try it later today!

[tribal jackal]

Don't get me wrong. I know they've got a lot more planned down the road. I'm just not going to pay monthly until they've finished developing beyond text prompts editing video.

[wet nacelle]

Does anyone know how I can fix the embeding issue?

[tribal jackal]

and by that time I figure runway will be on like gen5

hyperbole, gen3 minimum

[mellow tendon]

I figure out what it did wrong (I was an idiot), I put the refiner model in the base slot and the base model in the refiner slot when I was playing about with changing the models.
Although I would argue that having the Base node above the refiner would make more sense (in my head at least!).

[boreal bough]

shape mismatch means you either threw it into the wrong model, or that you're using an embedding from 1.5 on sdxl

[brazen patrol]

Pay? It is free. firefly.adobe.com

[wet nacelle]

How can I check through all of that to fix it?

[ember cargo]

quite cool how the style now works on whatever you want more or less? prompt is just "mountain landscape with a cyberpunk skycraper in the middle by nihei tsutomu"

[boreal bough]

link me the civitai or huggingface page

[wet nacelle]

https://civitai.com/models/72437?modelVersionId=77173

[south igloo]

Just wondering how big can I go with 24gb vram and your workflow? Examples say 2.5x for 10gb so I assume 5-6x?

[boreal bough]

needs to say "sdxl" there, else it wont work

[lilac wren]

@high skiff your workflow works very well !

[wet nacelle]

@wet nacellehttps://www.youtube.com/watch?v=zyvPtZdS4tI He is using it and has it working in XL.@boreal bough

[high skiff]

Its looking all over the place

I have seen some 10GB people say that 2048x uses like 17 GB

I myself was able to do 2.5x on 10GB VRAM no problem

my 3090 gets here in a couple days, I can do more testing from there

24GB VRAm should be able to do 4x, from my tests on my 3080... so I will have to see

[high skiff]

glad you like it, I worked hard on it haha

[wet nacelle]

@boreal boughIf you want to go through it and see if you can fix the issue.

[south igloo]

Ok I'll give it a test in a few hours I have 64gb ram and a 4090 and a clean comfy install to avoid any issues

Thanks a lot for making this

[lilac wren]

I manage to have very good results otherwise, but not with so few steps and cfg scale, well played

[south igloo]

Was the one huge thing keeping me away from comfy

[high skiff]

no problem, just glad to share

[south igloo]

Pretty much high res fixed every image in 1111 as the results are night and day compared to not doing it

[boreal bough]

one of the top comments explained it

[lilac raven]

How do I get the information about a ComfyUI image without messing the layout?

[wet nacelle]

I clearly am the common YouTube viewer then.

thank you for pointing out the obvious.

I can't believe I pride myself in pointing out stuff like that.

[brazen patrol]

remember you can simply reload the last saved image in your Comfy output to revert

[queen ether]

.....don't forget to change your .txt files to .captions files lest you waste 3 days of troubleshooting like me........

[hoary saddle]

Comfy API script working with my webui. Finally! Celebrating with some dinner.

[boreal bough]

for kohya training?

[queen ether]

@boreal bough yes 😦

[hoary saddle]

Just gotta figure out my data and overlay now.

[boreal bough]

--caption_extension=".txt"

depending on if you use cli or or the gui

[runic lark]

Still trying out upscaling. How would you create a solution to make a loop for the second Ksampler into multiple images?

[wise edge]

Could do with some skin texture, but dang this thing can do some pretty convincing photography

[autumn forum]

caith should they be .caption or .txt? i put ine as .txt because thats the type of file. thats correct right?

[urban fjord]

".txt" Yes

caption extension is what extension you're using for your captions.

[autumn forum]

okay cool thats what makes sense didnt know if it was something weird.

[wet nacelle]

I followed his directions and still am getting erros.

Do you think you can just install both the model and my setup to see if you can solve the issue?

[wet nacelle]

https://civitai.com/models/72437?modelVersionId=77173

[autumn forum]

but thats a 1.5 embedding.

[wet nacelle]

https://www.youtube.com/watch?v=zyvPtZdS4tI

[hushed forge]

1.5 embeddings work with SDXL???

[autumn forum]

no

[wet nacelle]

yes....

[autumn forum]

then why isnt it working for you? lol

[wet nacelle]

https://www.youtube.com/watch?v=zyvPtZdS4tI https://www.youtube.com/watch?v=zyvPtZdS4tI https://www.youtube.com/watch?v=zyvPtZdS4tI https://www.youtube.com/watch?v=zyvPtZdS4tI https://www.youtube.com/watch?v=zyvPtZdS4tI https://www.youtube.com/watch?v=zyvPtZdS4tI https://www.youtube.com/watch?v=zyvPtZdS4tI

[boreal bough]

let's rephrase that. 1.5 embeddings shouldn't work

[void skiff]

Just a fun note. I downloaded a extension for Automatic1111 Web-ui that is called "SDXL Styles", one style is called "Digital Art", but I thougt AI was like digital art as it is 😄

[autumn forum]

are you trying to run it through a sdxl clip node or the basic prompt clip node?

[wet nacelle]

Exactly. They have been proven to by this individual to.

[wet nacelle]

1.5 as shown in the text instructions. The issue is that.

He does not explain further.

[boreal bough]

cant see it cause I'm missing the custom nodes. but you did put the embedding in the negative prompt, right?

[hot rover]

Dude, your work is much appreciated. I was trying to put something together like that and you made it so much better than I ever could have. Do you have any plans to add a segmented and refined face fix before doing the hi-res fix stage? Faces look pretty cooked from the start with SDXL (specifically eyes) and I’m wondering if giving the faces another pass exclusively before the hi-res stage would help that or make things worse.

[boreal bough]

cause the base model & refiner model are not the same in terms of clip/prompts

[wet nacelle]

In the video he color codes it so that it is a stark difference. I should be yes.

[high skiff]

the high res fix alone helps a ton with that

[autumn forum]

only time ive seen that type of file for sdxl was with 0.9 or something

[boreal bough]

well could be many things. he was still using the pruned 0.9 model. so 🤷‍♂️

[polar epoch]

Ain't easy to find a damn solvent to make SDXL refiner talk with SD 1.5

[high skiff]

oh also, side note

there is a 1.0 base model with the 0.9 VAE baked in that we can use to get rid of the 1.0 VAE issues

[boreal bough]

I can tell you that officially, nothing 1.5 trained is supported with sdxl

[copper rose]

Hey I tried your workflow and my results were kinda bad sadly. I think it needs more training. I can send you the X/Y plot if you're interested in taking a look compared to my previous SDXL lora

[wet nacelle]

Ah I see now. That likely is the issue.

I'll disregard it then.

He has tutorials showing hand made graphs that. I thought since the video was uploaded not long after the 1.0 it was made as a 1.0 workflow video.

[copper rose]

Most other SDXL LoRas being released are 800-900mb and the ones I produced with your technique are only 40mb. I assume its in relation to the network rank and alpha

[glad flint]

all the download links here that I've tried, are monthly bandwidth exceeded pages
https://nmkd.de/?esrgan

[urban fjord]

Yes size is related to dimensions, though not sure where you've got the 800-800 mb number from.

[glad flint]

found what I needed on huggingface

[high skiff]

You know, I have gotten a lot of good from SDXL, but I cannot get it to do good digital art

[charred beacon]

How 'bout some meat sonny boy

[copper rose]

800-900mbs are the usual ones I see on civitai rn almost all are that size I'm seeing, potentially because the standard network alpha rank used is 128

[cinder aspen]

wdym?

i mean how is it not good

[high skiff]

It all just looks like low detail photographs

[autumn forum]

yeah they were using the wrong network rank. probably from a youtube video or something. 40mb is these size im getting with rank 8 and work just as well as the 800mb ones with rank 128

[urban fjord]

I thought people stopped using 128 dim due to VRAM constraints.

[thick harness]

PROMPT: anatomy, hyper-detailed art by (artist A), (artist B), 4D cinema

[high skiff]

this is what digital art looks life for me

its just kinda ugly in general, and I can't seem to get it to look any better

[cinder aspen]

thats some bad news for me lol

even comics for example?

[crisp owl]

likewise, couldn't figure out why my generations were coming out with that general styling, then realized it was that tagging

[boreal bough]

yes please!

[cinder aspen]

this actually looks good imo

[boreal bough]

also 1 image + its caption please

[high skiff]

to each their own I guess, i think it looks terrible haha

[tough vigil]

Does not work for me in A1111 with LyCORIS extension loaded.

[boreal bough]

exactly. network dimensions determines how many parameters it can train, and therefore also size. but at the same, time, more does not automatically equal better

[cinder aspen]

one of the main things im interested in is how it is with comics style

[boreal bough]

keep in mind my settings aren't intended to massively overfit data - so if you're trying to make an elon musk lora clone, that's not gonna work, due to a lot of reasons

[boreal bough]

nah, that was just poor optimization settings being shown in the youtube videos. if you do the settings right, then bigger dim is relatively harmless on vram, from my tests

especially if you go all bf16 like many now do

[urban fjord]

I really need some updated settings then.

What does the precision you set while configuring accelerate affect?

[grim lodge]

Do I have to download a new SD or does updating the Automatic1111 and just placing the sdxl safetensor into the model folder works fine?

[heady vale]

if you have 1.5 and higher version of A1111 then just drop the safetensor in the folder

[mellow tendon]

Yeah, I am glad I got 500Mb Fibre Broadband and a 4TB NVME recently 🙂

[sweet bane]

trying to get the feel of the deathly hallows , the tale of the three brothers

[thin nova]

I got a 4TB HDD, shoulda gotten a solid state 😔

[digital shoal]

@high skiff your 1.0 workflow works amazingly, well done!!

[mellow tendon]

As a RTX 3090 owner, You will love it for SD!
I am just hoping Nvidia make a consumer card with even more RAM for their 5000 series.

[glad fulcrum]

Which one?

[high skiff]

I had a 3090 for a few days before it died

[crisp owl]

https://github.com/SytanSD/Sytan-SDXL-ComfyUI

[spring fulcrum]

Alternate poses... also seem to work well.

[glad fulcrum]

Ty

[cinder aspen]

I guess the 50 series will have more VRAM

[crisp owl]

of course

[midnight shuttle]

I hope so but Nvidia wants to sell datacenter cards so less VRAM in consumer cards means more datacenter sales. Plus the extra value of forcing all AI into data centers.

[spring fulcrum]

Honestly I don't think NVIDIA is going to move much on VRAM... They dont seem to care about the consumer anymore

[nimble heart]

i had to learn some numpy to turn it into a compatible bytes object but I'll be damned it worked

[cinder aspen]

Data centers are a whole different league

For the consumers what we have and might get is enough

[midnight shuttle]

Nvidia's dream is all AI in data centers and all gaming streamed from data centers.

[spring fulcrum]

I heard they might even be cancelling their Titan card for the 4000 series which would have a lot of vram

[cinder aspen]

Nvidia for sure is not going to make exceptional consumers to be the norm

That is reserved for professionals

For example you d find barely a non professional buying a A100

[midnight shuttle]

If they put 48 GB on a consumer card many people will bu that for AI with a price of maybe $1500. But if they only put 48 GB on datacenter cards then they can sell those for $5000.

[trim orbit]

going to make?

nvidia's have been the prosumer choice for a while now

[midnight shuttle]

It is also a lot easier to manage AI alignment when it is in a datacenter versus on individual PCs.

[cinder aspen]

48 GB 1.500$ for a consumer card. No way

Not even AMD would do that

48 GB is not even made for consumers

Barely any consumer would consider this one

[trim orbit]

naw. $5000 is not a road block for any of that at all

[midnight shuttle]

Not today. But 10 years ago everyone would say 8 GB is not made for consumers.

[trim orbit]

not even a speed hump

[cinder aspen]

Then they will make you such a consumer card...in 10 years

But not at this point

[midnight shuttle]

What I am trying to say is that Nvidia has business reasons for not putting a lot of VRAM on consumers cards. That is all.

[trim orbit]

theres no management of what training scripts people are running on cloud h100s

its the wild west out here

[cinder aspen]

Which is their right. It has to be rentable

[midnight shuttle]

Everything is wild west when it is new.

[heady vale]

cards like the A100 get HBM2/3 vram so vram overall will naturally get larger with higher density chips

[trim orbit]

just your claim about managing alignment is some qanon nonsense. nobody is making any effort to manage that. everything is in an effort to shore up their brands

[boreal bough]

A6000 used to be prosumer. but with updat4ed prices that's no longer a thing eitehr 🥲

[midnight shuttle]

Easiest alignment is when only corporations have AI. Next easiest is when AI is in datacenters because it is at least monitorable by the owner. Hardest for alignment is AI on individual PCs. And alignment is a very important topic in AI.

[trim orbit]

nobodys putting ai in the datacenters to monitor it. that's an overhead cost that they won't bother with

[cinder aspen]

How much is it at this point? lol

[midnight shuttle]

Do not associate me with that group. I am merely talking about the industry considerations.

[boreal bough]

@trim orbit would love your input ❤️
#🔧｜finetune message

[west breach]

https://github.com/bash-j/mikey_nodes

[trim orbit]

i'll come back to this! expect it soon

[winged tree]

they don't put 48GB on consumer GPUs yet beccause there isn't a market for it

[boreal bough]

weeeee

[wet nacelle]

Hey I'm trying to set up an upscaler set up with a comfy ui workflow I'm using and encountering an error that I don't how to fix.

[spring fulcrum]

There is now

[winged tree]

sure, AI enthusiasts will buy them, but 99% of their sales will be coin mining and gaming

that and yeah, use your production on the $5,000 server cards because that demand is skyrocketing

[trim orbit]

nobody is using gpus for coin mining since last year

[glad grove]

nobody? damn

[trim orbit]

if they are they're dumb about it

[boreal bough]

etherium change kinda spelled the end for profitable gpu mining

bit of a ripple effect

[trim orbit]

the minepocolypse

[winged tree]

not talking btc

there's still coins for it

[trim orbit]

ethereum bro i know. btc has been YEARS. like asics took over there

[winged tree]

but anyway, it was just using that to avoid being corrected.. holy

[glad grove]

what about that coin they use to mine on HDD's are they still minin that?

[trim orbit]

the entire market crashed and it's held up with proof of stake now. since all other POW chains were tied to ethereum's economy they all fell. it's simply not profitable to gpu mine anymore

[wet nacelle]

Does no one know how to fix this?

[winged tree]

your models aren't compatable

[trim orbit]

ai sales took nvidia to a trillion dollar company. i think it's safe to say its their bread and butter now

coin probably drives 0.01% of their sales now

[spring fulcrum]

gotta go on the hunt for some brand new crap coin to pump and dump to make some quick money and get out.

[winged tree]

they're dominating the server cards

[wet nacelle]

How does that make sense? Does a model need to be a specific way to be allowed to upscale?

[boreal bough]

'profitable' coin mining. as in you would recoup the cost of buying the gpu + electricity, usually in under 6 months.
now it's just hackers that use infected pcs to mine, people who misapproriate universe pc labs, or people who get access to high end equipment during downtime. Nothing that can be considered 'average joe' level of interaction with it

[digital shoal]

@high skiff can you explain the "stop at step 1000" parts in your workflow? I am not understanding exactly why its set up that way

[winged tree]

not sure on upscaling but this usually happens when i use a differnt model for the upscaled one vs the one that made the latents

[winged tree]

wasn't even the point of the message but hey

[spring fulcrum]

It stops at whatever the last step you set is... it doesnt go to 1000 steps unless you set 1000 steps

[high skiff]

what Scorp said

[wet nacelle]

This is also here in the graph. It could be a reason.

It's that res then downscaled to 1024

[boreal bough]

ah sry, think I misread it

[spring fulcrum]

that number is not very relevant to your outcome

[winged tree]

not really, but we were just talking about nvidia sales and the qanon conspiracies

[midnight shuttle]

Is there a limit to the fan-out of Comfy nodes? The workflow I'm building needed the width and height to be connected to 8 places but I could only get 6 of them to connect.

[west breach]

Maybe try a reroute node to split one pipe into two?

[midnight shuttle]

That's a good idea. I'll try that. I wonder if there is a limit or if I just messed up somehow.

[copper rose]

Sent the files 😄

[brazen patrol]

actually, this seems logical but is wrong. Nvidia is using a loophole in the law to prevent datacenters from using consumer cards. They cannot by law prevent them from buy the cards, but you may not know the drivers have a cluase declaring they cannot be used for datacenters

it is why only private one-on-one services like Vast.Ai even exist

[vernal cloak]

can anyone tell me what Clip_G & Clip_L does in the SDXL text encoder prompt boxes?

[brazen patrol]

and why Amazon, Google, and you name it cannot use a 4090

[winged tree]

don't quote me, but pretty sure Clip_G is for the natural language prompt, and Clip_L is the supporting tags

[midnight shuttle]

Oh, that's interesting and I didn't know that. It does show how they want to keep that market division in place. Someone else said that was a Qanon conspiracy but it's just good for profits.

[brazen patrol]

they can for their own uses, but not to sell out

[winged tree]

also massive datacenter expansions wont wreck their consumer market letting AMD get into it

[vernal cloak]

Ahh... thx m8. That does make sense actually.

[brazen patrol]

No, it is how the enforce sales of their server cards at triple the price for only a 20% speed diff at times

[midnight shuttle]

That's what I've been trying to say.

[dusty turtle]

clip G understands natural language, L is like old SD prompting

[brazen patrol]

By all means buy a 4090, but if you use our drivers we will sue. and don't think they won't. They have done it many times