✨｜sdxl

[marble kindle]

yeah something big is wrong, but idk what most of the time it spits out a random image or noise

[boreal gate]

@ionic gulch thats nice...

[marble kindle]

i think there is an error with my image loader where is doesnt read in the images some reason

[hardy cipher]

come at me, bro

[ionic gulch]

@hardy cipher that node could fit in the space i have for my loras. if i need more that than 4+offset, that could be helpful. have you published it on comfyui manager?

[native knot]

My pimped out new Jordans...

[hardy cipher]

not yet. do I just go to the github and request being added?

[ionic gulch]

i think you have to fork the manager, change the file with the node sources and create a pull request. but in my case the pull request was closed and they added it manually. guess this is easier for them that way.

[ionic gulch]

can you try my workflow and check if all nodes are loaded by the manager?

[hardy cipher]

yeah, just give me a couple minutes. trying to figure out an error I was getting

[eternal fog]

Ok using Revision together with Controlnet, in this case Depth works pretty well

[tepid surge]

Yippe! I think my comfiUI animation graph is working

[tepid surge]

whooa sick armor

[eternal fog]

Interesting mask

Canny and Depth with the same settings

[native knot]

"We are here...
...and we are waiting."

[hardy cipher]

loading now

not gonna lie, didn't have your nodes just yet. but I'll install them now

they showed up in manager under missing nodes

[native knot]

That looks like a chair someone paid too much for on Etsy.

[ionic gulch]

so everything is working fine? you can use all 4 modes (txt2img, img2img, candy, depth)?

[wet nacelle]

Hours later and I'm still seeing this.

[hardy cipher]

well I'd discussed nodes with a couple people previously. should I not show them?

[wet nacelle]

I'm just FUCKING with you. 🤕 🍮

[hardy cipher]

I mean, I also think it's a bit ridiculous, but it works now, lol

[wet nacelle]

Or no one buys it for months and a random woman in her 50s swoops in and casually hands over more than 500 dollars for it.

[wet nacelle]

It's wicked

[eternal fog]

Is there any way to dynamically adjust the amount of inputs in the node. Would make more sense to do it that way if there is.

[hardy cipher]

everything loads fine. but I need to completely reinstall comfy to get the controlnet working though. created a new virtual environment and still no controlnet, so I think I'll take a different approach when I have a minute to do it. but everything looks good

[soft zealot]

Ok So this is going to be a train of 3 messages lets see hwo quick I can get them off 🙂

So still trying to understand how ReVision sort of works and what inputs are doing what

Post 1 (this one) is Normal Textual Prompting & Styling
Post 2 is Textual & Styling Plus ReVision truned on but ConditionZeroing disabled
Post 3 (which is where/how everything up until now has been prodcued) is Textual & Styling , ReVision On,Condition Zeroing enabled (ie only using the image encoding to generate the new image and ignoring any prompt information

[hardy cipher]

that's a real humdinger of a workflow, lol

[wet nacelle]

little to no artifacts

[hard fractal]

That's a cool workflow!

Are you revisioning the refiner also?

[hardy cipher]

yeah, unlike countless others that came before, lol. kind of a hassle to find the sweet spot with 4 clip visions in various places

[hard fractal]

It doesn't work as well, but kinda sorta works.

[soft zealot]

No, Im aonly revisioning the stage 1 sampler (which to most people would be the Base)

[wet nacelle]

I'll share this again because I really like this style.

[hard fractal]

Ah. The refiner only runs for like 8 steps anyways. It's not like feeding anything into it can really help.

Unless...

[hard fractal]

Try a solid black or solid white image.

Maybe it does something

[soft zealot]

and Stage 0 (or PreCOndition as some peole mat call it) only runs for 3 steps

[hardy cipher]

I feel like it sometimes pulls things together at the end

if it looks a bit sloppy

[wet nacelle]

"peole mat"

[soft zealot]

I mean Ok Ive admitted earlier to have sausage fingers but heres a question. What if iDidnt and it was a "condition" ?

[wet nacelle]

It would still be a comedic slip in your typing.

[soft zealot]

its optimised to be used with stringy bits turned off and loads at this screen on my 1440p moonitor 🙂

As you've probably seen its a bit more complex as you zoom out lol

[hardy cipher]

that looks so clean, but I can't bring myself to put flows together like that, at least not for myself. I have to see where everything is going

[hard fractal]

Woah

[hard fractal]

Oooo TIL you can turn noodles off

[soft zealot]

I 've used Node Based tools for work purposes form seevral years now and I find it easy to visualise in my head where things go

Its all about having things in key blocks so you can get a helicopter general view that you can drill into if required

I know for example that if Im having an issue with my text file output I just jump to the circled box and take it from there

[eternal fog]

Is that the PS1 style game LoRA?

[wet nacelle]

no

[tepid surge]

i was thinking ps1 too

[soft zealot]

yup, Comfy only lets you straighten them , these lt you hide them

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

[hard fractal]

Any requests for comfy?

DM me

[hardy cipher]

makes perfect sense. I do tighten things up quite a bit once I have it established. get a bit ocd with it. but I normally have seperate sections

[soft zealot]

my only outsatnding Request for COmfy (which Im sure Ive stuck on his GitHub) is to improve the S&R fucntionality so I can write text files withthe exact same date string as used in an image save file name.

[soft zealot]

damn I saw that bit after I posted lol

[upbeat summit]

yeah I created an issue related to that on the ComfyUI repo

I guess the whole implementation would need to be refactored - S&R, filename handling etc

[soft zealot]

what was your link again?

Mine is

https://github.com/comfyanonymous/ComfyUI/issues/1057

[upbeat summit]

remotely related, but it kinda is connected: https://github.com/comfyanonymous/ComfyUI/issues/1176

I have since customized nodes.py and the Save Image node as well as the S&R Extra Metadata javascript for my needs

[upbeat summit]

If I could do a clean implementation I would submit some code ideas. It's just hacked together. I added a much needed suffix to the Save Image node, suffix can also use S&R.

The counter didn't work how I wanted with multiple images in one queue job. So I removed the counter and name everything by date/time. I moved the whole date/time function to nodes.py since it's not correctly evaluated how I want it when the queue runs.

[boreal gate]

@ionic gulch crashing 😦

was working, but everytime I try to generate something., it crashes the computer...

[eternal fog]

Anyone noticed a bug with memory management on Comfy and the ControlNet Models?

It doesn't seem to unload them once it's finished with them, so you end up getting slowdowns.

Keeps 7.5GB Loaded for me after using them, and if you then try to do anything else after, it overloads the memory

[heady vale]

Yeah, Ive seen that too

[eternal fog]

Hmm it might not be the controlnet, it might be the pre-processor that it's holding there

[ionic gulch]

other workflows work fine? how much vram have you got? are all 4 modes affected? even txt2img?

[eternal fog]

Yeah there is definately some weird memory management issue when you load a ControlNet/Depth Model

[hardy cipher]

I uploaded that node if you wanted to try it out. I do believe it works correctly now. But if you do decide to use it please let me know if you encounter any issues.

[hardy cipher]

Got that uploaded. Hoping it all works as it's supposed to 🤞

https://github.com/picturesonpictures/comfy_PoP

[hardy cipher]

are you into the LoRAs?

[wet nacelle]

not at all. just the clip_g

[wet nacelle]

shweet

https://tenor.com/view/jerma-american-psycho-impressive-very-nice-jerma985-gif-25449077

[hardy cipher]

that elon musk image reminds me of my favorite lora.

[wet nacelle]

I have an idea for that actually. I'm thinking of using the example images on that Lora page and using two as a revision base.

I'm using this more of a style merger than a blender.

[polar epoch]

Aye! gonna generate a few tests tomorrow, or if energy enough, a few today and see :P

By the way, got another "for everyone comission" for a better ease of use
And that's for lora's, and models, do you think it'd be hard to make a "load lora/model", or rather "model" preview, that loads like the "load image" you normally use as base image, but it reads from images related to lora's and loads them up in a grid? Similar to "extra networks" on automatic's?

[hardy cipher]

these are the images I was using earlier

[wet nacelle]

I'll use these for later actually. great example.

[hardy cipher]

with that lora actually

[wet nacelle]

Loras are just too odd to work with for me. I'm still geting what I want from everything but Loras. @hardy cipher

[steady grove]

holy mother of mary wtf is that thing xD

[wet nacelle]

@hardy cipherThis is just what I'm doing right now.

"Th is is"

[hardy cipher]

it made so many beautiful images

[wet nacelle]

Trust me. I can make stuff like that with style images like those as a base.

[hardy cipher]

https://abcnews.go.com/blogs/headlines/2012/08/elderly-woman-ruins-19th-century-fresco-in-restoration-attempt

[wet nacelle]

https://tenor.com/view/i-have-no-mouth-and-i-must-scream-i-have-no-mouth-and-i-must-scream-am-gif-26618962

[steady grove]

ah it was a jesus fresco! ruins my whole pun

[hardy cipher]

for sure. same. I just like to get real wacky with it

[wet nacelle]

totes I get what you mean.

[native knot]

That one on the right. 😍

[soft zealot]

pretty damned good innit

[native knot]

Def.

[native knot]

You catch my stormtrooper above?

[soft zealot]

no sorry , I was just heading to bed and saw those 2 had dropped in in a batch I have running (ok they dont all work but thats to be expected)

[soft zealot]

thats crisp 🙂

[native knot]

The one on the right from above is definitely the best of the batch.

[native knot]

And yeah, that trooper came out super nice.

[wet nacelle]

[native knot]

Try this but with something like "layered highlight hair"

[soft zealot]

so you may have misse dthe posts back up there but thes eare being generated purely by the 2 images and are ignoring any prompt details

#✨｜sdxl message

Anyways very late and I should be in bed

[native knot]

Ah...sorry, yeah, I didn't see that.

[native knot]

It's a perfect day to bake a pretty cake.

[wet nacelle]

I was thinking the same thing!

[native knot]

Oh, that hand.

😬

[native knot]

https://giphy.com/gifs/K0AnEB2t2EM

[hardy cipher]

lil nubbin

[wet nacelle]

🤢

[native knot]

The lighting on this one:

[wet nacelle]

sweet

[tepid surge]

Nice, these star wars ones are coming out great

[tepid surge]

My multiple fails trying to get this animation graph to not breed chaos cx (i'm closing in on it)

[native knot]

A wizard is never high. He is exactly as blazed as he means to be.

[glad grove]

shadow wizard money gang 🧙 💲

[native knot]

Straight floatin' that spliff.

[native knot]

I don't think that's how you do that...

[native knot]

I feel like you need to throw Wozniak in there somewhere.

[wet nacelle]

this guy

[steady grove]

smoking. why do i keep seeing on many servers that people can't generate photos of people smoking?

[wet nacelle]

the phat hatter

[steady grove]

iv'e done it. you're doing it

seems to be some meme got a bunch of people convinced of a limitation that isnt there

[native knot]

These last 2 are

[tepid surge]

😆

[heady vale]

mixing art with revision

[glad fulcrum]

Is the revision mixing supposed to be much slower? It's much slower at least for me.

[heady vale]

nah, about the same

[wet nacelle]

Did you make your workflow or use someone else's?

[glad fulcrum]

I downloaded the one which comes with hugging face

[heady vale]

maybe your input images are large?

[glad fulcrum]

oh never mind, It's because it did 4 images at once

[wet nacelle]

ah

rad

[upbeat summit]

beautiful image

[glad grove]

sexy young man who loves to watch the ceiling

[latent otter]

Hi guys, I just got runpod and wanted to train sdxl with my own images using a colab notebook
Anyone know any good ones that are fast and easy?

Also, if I am training a specific scifi object, should I go for LORAs or safetensors? It's safetensors right?

[west breach]

safetensors is the file type. lora is the method of training

[latent otter]

Ah ok, thanks

[autumn forum]

should i release my cinemax model

[upbeat summit]

https://tenor.com/view/yes-no-gif-21780817

[wet nacelle]

A small group of model creators should group together and make the best everything model that can. When it's finished the title of the model should be a literal cryptic code that needs to be solved to get the model name.

[wet nacelle]

I'm gettin there

this just looks like vector art.

[heady vale]

using the refiner model instead

[glad fulcrum]

the depth preview isn't working for me. Not a big deal though.
EDIT: I can see it with a save image node.

[uncut fiber]

maybe found way how get rid of signatures. Just not sure.

[steady grove]

i always put "signature, watermark, logo, signed" in my negs. basically never see them

[uncut fiber]

hm i see them, even printed. Try for example few pictures by Gertrude Abercombie

[wet nacelle]

BURN VICTIM! BURN VICTIM! BURN VICTIM! BURN VICTIM! BURN VICTIM! BURN VICTIM! BURN VICTIM!

[wet nacelle]

It looks like Scott Cawthon renders

[half ivy]

GM fam (includes you @steady grove )
What can I learn today?

[glad grove]

literally me

[tepid surge]

you gotta drop like a "cw: gore" or something if you're gonna spoiler it cx

[latent otter]

I am trying to use Kohya ss to train my lora, but i can't find where to put the class token and such :/

anyone has any idea where they are?

[wet nacelle]

I am stupid

FUCK

[glad fulcrum]

trying basic revision on a gorilla image and the outputs are all men... Can't get a gorilla

[wet nacelle]

chnage the weights

[wet nacelle]

what I rec is that you chnage the weights so that they are in favor of the image you want the most to blend with.

[glad fulcrum]

I'm only using 1 image.

[wet nacelle]

ah shiiiiii

I haven't been able to fuck with that yet

[wet nacelle]

annnndd it's back to this

[native knot]

It's like someone is wearing a Tom Hanks skin suit.

[polar epoch]

Has there been made a addon for comfy to see updated previews of the generations so i can cancel sooner/see live progress?

[wet nacelle]

Is that leonardo dicaprio?

[native knot]

lol, no

Still T.Hanks.

[wet nacelle]

oh shit wait

that's blackface

wait?

I don't care

[tepid surge]

YEES, I think i finally figured this out

[tepid surge]

I'll need to figure out and add interpolated frames, but here's the json if anyone wants it, it work in progress (will make another, better Commented one for #🎥｜animation )

[minor spruce]

Trying to use the refiner, and it's just outputting the same image... not sure where I'm going wrong? Any ideas? Thank you 🙏

[heady vale]

fixed seed

[minor spruce]

Thank you, I've kicked out the fixed seed on refiner, but still no dice...

[heady vale]

on the Ksampler its still fixed

[minor spruce]

how else would I generate fixed images though on the regular model?

[heady vale]

if you want new images change to randomize

[tepid surge]

I think your 'start_at_step' and 'steps' are both 20

[minor spruce]

what should It be?

[tepid surge]

i think i remeber seeing someone do 12 steps on the base model then start at 12, to 20

[glad fulcrum]

This is weird

It just does not want to do a gorilla

I tried many prompts and settings by now

[visual glade]

the conditioning zero out removes the text prompt

[glad fulcrum]

oh

how can I bypass the node?

[eternal fog]

RIght click > Bypass?

[visual glade]

ctrl-b or right click -> bypass

[tepid surge]

okay! i did it: 'start at' is where ever you finish generating on the base xl model and then 'steps' is where it'll go too up to a maximum of what 'max steps' is set to

[eternal fog]

They actually have their steps at 20 and their start at steps at 40, so it's not going to do anything.

[native knot]

Be sure to pick up your exclusive Ryan Seacrest action figure. In stores today!

[minor spruce]

ohhh let me check that! thank you

[tepid surge]

Yeah i think it was "Scott Detweiler" who i saw do the 12 then refine to 20, I actually like the base model more then post-refining but I still fresh with xl cx

[eternal fog]

Yeah you can do that, can do more or less depending. Although to be honest, with the particular image noel has on there, it won't make much of a difference.

It's more useful for sorting things like eyes out. But it's not always needed.

I've stopped using it really.

Especially with custom models

[hardy cipher]

aikido bear

[heady vale]

what stable diffusion looks like to SDXL

[minor spruce]

Interesting, better to just skip it and run more steps + upscale?

[eternal fog]

Depends on the image, for that one you provided an example of I don't see it making much of a difference.

[native knot]

Clinton?

[minor spruce]

Thanks, challenge for me is I'm not even getting any difference whatsoever which is weird

[eternal fog]

With the screenshot you posted, you had your start at step on 40 and your total steps on 20. So it isn't going to do anything.

[minor spruce]

You're right, I was fiddling with numbers. Here's 10 and 10, same exact image/output

[hardy cipher]

so close to greatness

[eternal fog]

That still won't do anything. You are telling it you want 10 total steps, but to start on step 10, so you are doing 0 steps

Set it to something like Steps 30, Start at 15.

[native knot]

I'm dyin' over here... 🤣

[eternal fog]

@minor spruce Actually use it like this. Set the Base KSampler to 30 steps, start at 0 and end at 20.

Then set the Refiner sampler to 30 steps, start at 20 and end at 30. Also disable add noise.

Just changed a number as I fat fingered the wrong one

[eternal fog]

If you really want you could do 0 - 10 on base and 10 - 20 on refiner, but then you don't really give the base enough time to make the image and the refiner will just screw it up

[hardy cipher]

yeah, I've never once had positive results taking that approach

[hardy cipher]

maybe if I took the time to run the refiner through some loras and have it skip some end steps

[eternal fog]

I mean that's how you are supposed to do it, you generate an image on the base and leave the left over noise, then you move that still noisy image into the refiner and let it finish the image off.

But I, when using it, tend to do something like 0 - 30 on the base and then maybe 30-35 on the refiner

[hardy cipher]

wat?

refiner is supposed to do like 20 percent of the steps or denoising

not saying that's set in stone

[eternal fog]

There's no set value you have to run the refiner for. I've never done it as a percentage.

I've found most of the time a small amount of steps works better. You go too far and it starts smoothing things out.

[hardy cipher]

I don't think there's have to anything. but I now I don't like the results of heavy refiner usage

[eternal fog]

Yeah, if you do too many steps, it wil make the eyes look brilliant, but it will start melting everything else.

[hardy cipher]

might be able to get away with it if you set it to low denoising or something

and then send it somewhere else from there

I have no doubt there are exceptions to this

I just haven't found them myself

not bad. I haven't had great luck with that approach. usually end up having base and refiner overlap on a few steps in then middle

[minor spruce]

Sorry, previous post was wrong (with noise on)... nice outcome now!

[glad grove]

crar i powties fr 👍🏿

[native knot]

ISRASTED

Next one is impressive... got the whole name.

[glad grove]

raied 🫡

[native knot]

The fact that I can get the "WANTED" to be correct most of the time is pretty impressive.

I didn't realize how much I needed AI generated Wanted posters.

[hardy cipher]

do you think we'll get proper hands or proper words first?

[native knot]

Who knows... I mean, it's getting better all the time on both fronts.

[hardy cipher]

not sure how they'll really fix the hand thing completely until the model knows how human anatomy works

[native knot]

FRAGSTAR:

[hardy cipher]

that's just a picture

[clear flicker]

Hey could anyone help me out I just am trying Comfy for first time and successfully used the controlnet LORAs recently released... now all I can't figure out is how to add my own custom LORA to those node workflows? I am using the workflow given for controlnet LORAs canny in this example. Like for example say I made a custom lora for myself. I see how to use control net now but how do I also use the LORA I made of myself...if this makes sense

[native knot]

Don't worry, the gorilla is wanted, too:

[hardy cipher]

double click in the empty space and search for "lora"

should have a lora model loader

you don't need to add it to the prompt. at least not the word "lora"

[clear flicker]

cool that was probably the easiest answer i ever got and maybe my dumbest question yet

[hardy cipher]

you'll see the checkpoint loader has model and clip as output

lora has model and clip as input and then as output as well

[clear flicker]

awesome this is gonna be badass with sdxl

[native knot]

Keep in mind you can only use LORAs made for SDXL with SDXL.

[clear flicker]

yeah for sure I got those made already

I am not sure what the strength for clip and model means I feel like in the past I just had one strength to worry abou t

[hardy cipher]

this is the lora setup you want

[hardy cipher]

well just leave them similar if in doubt

[native knot]

Nothing out of the ordinary going on there.

[hardy cipher]

gotta say, at least you're asking reasonable questions. I think you'll pick up on it pretty quickly

lots of things seem obvious in hindsight

[native knot]

lots of things seem obvious in hindsight
Quote of the decade right there.

[hardy cipher]

https://tenor.com/view/sensei-disappointed-sensei-mad-sensei-upset-sensei-gif-12003855

soul, what are your 30 favorite sdxl loras?

[native knot]

lol.... 30?!

[hardy cipher]

yes, I need to fill these

[clear flicker]

yeah i gotta do a little research

[native knot]

https://giphy.com/gifs/aint-nobody-got-time-for-that-gif-10PcMWwtZSYk2k

[clear flicker]

actually comfy seems kinda cool because at least by using nodes u understand what is going on under hood btter

[hardy cipher]

correct. and it's like playing with ai legos or something

[clear flicker]

i was so happy i had just mastered dependencies in python and shit lol I feel like 90% of what I was learning was not AI related as much as bug and driver releated or performance searching etc. with GUIs

[hardy cipher]

I guess it just doesn't click with some people, but personally I like it way more

because you can connect things in basically endless combinations

things that barely make sense, that's my favorite approach

you can do the same things with a1111 technically, but it'd be ridiculously tedious

[native knot]

Zardu Hasselfrau

[dusk mica]

Is ClipTextEncoder better for SdXl? Just realized i have been using the normal one 💀

[hardy cipher]

what do you mean exactly?

I'm going to load 30 loras because I can

https://tenor.com/view/blessed-gif-21519023

feeling blessed

[crisp owl]

It's the designed way to utilize the full potential of SDXL, so technically yes.
Can you still get nice images not utilizing it? Absolutely
(speaking to splitting to G and L)

[clear flicker]

k one last question is that lorastackloader a custom node if what is the name so i can snag it

[vocal stream]

are the default controlnet workflows stability posted for sdxl just broken on Windows?
'control-lora/control-lora-canny-rank256.safetensors' not in ['control-lora\\control-lora-canny-rank256.safetensors',

[native knot]

Get ComfyUI Manager. That way when you load someone's workflows based on images or .json files, you can easily download the missing nodes.

[hardy cipher]

I guess loading 30 loras might take a little while aye?

only added 10 so far

[native knot]

https://giphy.com/gifs/HallmarkChannel-mtp105-1rQSu7v6dAwLZg3zDM

[hardy cipher]

very nice

that's not the result of 30 loras, but it is quite easy on the eyes

[frosty pine]

I think SDXL is 80% into perfection

[proper depot]

Still a pain to get eye colour and skin colour right 😒

[frosty pine]

i start SD before it was 20% quality, then 1.5 around 40-60%

and now its 80%,

[hardy cipher]

how are you quantifying these percentages?

[frosty pine]

for me 100% if it really can give result as what we tell/ask them to do

based on the proportion and perfections of the images

[hardy cipher]

seems a bit suspect, but I'll allow it

[frosty pine]

this is near perfect images

but i told her to be backward flip baseball hat and with mecha in the background

but overall its convey the idea

[vocal stream]

SDXL is definitely worse at following complex prompt than I hoped but it's still maybe 5-10% better than 1.5 and higher res

[frosty pine]

1.5 cant achieve this

[glad grove]

it can,thats just a normal average img

[hardy cipher]

I'm 12 percent better than stable diffusions 1.5

[frosty pine]

no, it look suck I tried it

[hardy cipher]

and 16 percent better than a honda civic

[glad grove]

https://tenor.com/view/ez-noob-skill-issue-simply-a-difference-in-skill-gif-20780203

[frosty pine]

and will disfigured

[glad grove]

skill issue

[frosty pine]

I just use common clip drop, not locals

just basic not using own models or else

not training to, I comment here based on basic skill usage not lab pro usage lol

I bet next year its perfection

[hardy cipher]

if you don't know how to drive a car it won't drive well

[frosty pine]

or end of this year if the dev really working on it

[native knot]

Sydney Sweeney as Jessica Rabbit...or at least one could dream.

[frosty pine]

I know how to drive the car I think, just basic sedan. but of course not some Initial D sliding lol

[hardy cipher]

I feel like a lot of new people, not saying you exactly, misunderstand and think some people get bangers ever time they run it

but depending on the seed and all that, it could just be a dud through no fault of your own

[frosty pine]

it can be a dud, but not as common as before

[hardy cipher]

yes, higher win rate than 1.5

[frosty pine]

it is

thats why i said 80% chances lol

near perfect, bet nex year wll be 100%, or we already achieve stagnation

[glad grove]

still waitin for the perfect img u mentioned

meanwhile ill post my perfect img

[frosty pine]

not bad

1.5 will be hard to achieve this

[glad grove]

i can do that on 1.5

[frosty pine]

try it compare it here

[glad grove]

but i would have to use inpaiting

[frosty pine]

see

intricateds

[vocal stream]

given that 1.5-2 was a downgrade and 1.5-sdxl was a small update (outside of higher res) I really doubt we'll get anything like 100% for complex prompts by next year

[frosty pine]

2 years?

or never?

[vocal stream]

I suspect we can get things way better in like 6 months but only on consumer hadware

a bit hard to get the text understanding part to be cutting edge when you have to use a tiny language model (compared to gpt) to run on consumer devices

[frosty pine]

if consumer hardware better, just minus 10% on regular website

free website usually just lost around 10-20% from local

[vocal stream]

yeh in 2 years we might get a 5090 that has like 30% more vram so we'll be constrained for ages on the consumer side

[frosty pine]

10% maybe

[glad grove]

speakin of perfect imgs

[frosty pine]

nicee, care to share the prompt?

[glad grove]

this isnt perfect,no steven seagal

[frosty pine]

make steven vs chuck

[frosty pine]

wanna compare it using clipDrops

[glad grove]

(masterpiece,best quality,official art),1girl,solo,puffer jacket,open jacket, navel piercing, tight pants, denim shorts, torn legwear, eyewear on head, tanlines, and the negative : worst quality, low quality:1.4),

[frosty pine]

it will be perfct 100% if we prompt "cyberpunk girl holding weapon aiming with mecha in the background 3/4 view" with exact representation of it. righ now it just will give us a great cyberpunk girl sideview looking at us thats it

lol

[half ivy]

how does one go about learning about this maths thing? o0

[hardy cipher]

secret

[frosty pine]

this what I got, not bad but of course not as good as local trained

[frosty pine]

anime style

[glad grove]

compared to base 1.5 its rly good

[half ivy]

broooo

[glad grove]

perfect tbh

[frosty pine]

it is

[half ivy]

is it like, 1 latent image + 2 prompts + other settings = 1 diffused image

[frosty pine]

60% better

[half ivy]

SOS

[tepid surge]

Who is this character cx ive been seeing them pop up all day and now they're fighting a gorilla xD

[glad grove]

steven seagal

[frosty pine]

lol

[glad grove]

https://tenor.com/view/steven-seagal-eating-carrot-healthy-living-gif-9753152

[frosty pine]

dude you dont know SEAGAL?????

best man ever lived on this planet

[tepid surge]

I am now going to educate myself cx

[dusk mica]

Season 1 of 10

💀

[frosty pine]

this a perfect images. thats why sdxl is 80% to perfection

[frosty pine]

The truck is badass

[hardy cipher]

truck lora with goat lora equals divinity

[frosty pine]

the tires and proportion is great

i dont know how to use lora

not local

[half ivy]

thats baaaa'd ass

[frosty pine]

sdxl proportion is great

[tepid surge]

Yaay i got film interpolation working

[hardy cipher]

https://tenor.com/view/proud-ron-swanson-nick-offerman-parks-and-rec-so-proud-of-you-gif-4531083

[tepid surge]

Hey! i'm a tiny apple that'll...! rot into the soil cx just you wait!

[uncut fiber]

brum brum brum

[hardy cipher]

and then you'll make crappy crap apples because all the good apple trees are clones

[tepid surge]

https://tenor.com/view/rikka-takanashi-rikka-takanashi-math-numericals-gif-22074535

[hardy cipher]

you made that?

[glad grove]

kyo ani made it 🤓

[hardy cipher]

anyone know what ngms means?

I'm sure I could figure it out, but figured I'd ask

it's one of the parameters of generated images on civitai

[tepid surge]

Nah, just pertinate gif
I messing with film interpolation atm

[glad grove]

sexy

[hardy cipher]

I figured out the formula

[uncut fiber]

me as well playing with videos... But it is something to spend very much time

[glad grove]

😬

[uncut fiber]

i make a tyson, tried him eating ear, but it didnt worked

[crisp owl]

"whispering sweet nothings while simultaneously munching on an ear"

[hardy cipher]

I tried michael jordan and it looked nothing like him at all. wonder if they were scared he'd sue

or coincidence

[hardy cipher]

this isn't rocky dennis 😭

[uncut fiber]

This kissing, it used to be absolutely normal, when Breznev or some communist leader arrived 😄

They swapped their pants 😄

[glad grove]

🤣

[dense chasm]

some celebs have to be trained with LoRA,Christoph Nolan is another example

web crawler data from e-commerce site is interesting ,RNN and LSMT predict for custmers emotion when anylize with different comments

[uncut fiber]

Poe by Jackson Pollock 🙂

[void zenith]

No way bro

I put a pen up my ass right, it got stuck and I had to be put to sleep in hospital to remove it

[void zenith]

Gay

[noble shoal]

Relax, its just a nice garden plant 🤷

[void zenith]

Its cool bro, btw my asshole hurts and im done in hospital

[void zenith]

Wow

So cool

[tepid surge]

that's the final dark souls boss up there x.x

cx

[void zenith]

Idk What it is but cool picture

I put it as phone background?

[tepid surge]

go for it idc

[void zenith]

Tnx bro

I use it now

[void zenith]

Bro im not gay btw

[tepid surge]

We're just flipping your social autonomic systems cx

[void zenith]

What

[quartz sequoia]

is it possible to create a custom Ui based on a Comfy node workflow backend setup?

like the ui customizes in terms of input buttons or whatever based on how i built my comfy workflow?

[noble shoal]

https://github.com/space-nuko/ComfyBox

[quartz sequoia]

thaank you

[noble shoal]

You are welcome

[quartz sequoia]

it hasnt been updated.. in awhile 👀

[noble shoal]

I haven't used it myself. Maybe it was flawless, so it doesn't needed updates? 😄 . I think comfyui has also an api.

[tepid surge]

they might have been working with stability to get it integrated or something? dont quote me though

[uncut fiber]

official app Stableswarm-UI is front end for Comfui, isnt? I would wait and see what they will do.

[quartz sequoia]

well it is what i was assuming stableswarmUi is

[prime barn]

WiFi

[quartz sequoia]

but i dont think thats techincally customizable

[noble shoal]

Well, yeah. Talking about that. You could just use stable swarm. The repro says: "In-UI Comfy workflow Save&Load, and metadata edit (ie to reconfigure which parameters to send over or not)"

[tepid surge]

So Vogue! 🎉

[quartz sequoia]

well stable swarm doesnt want to tinstall

[tepid surge]

i still need to hop on the stable swarm train, Just got ComfyUI and back into generating with Stable XL (so much progress kinda crazyyy)

[wicked frigate]

if you have trouble with it, feel free to ask in #🐝｜swarm-ui

it should work for you needs pretty well

[tepid surge]

it is time, to make contact

...greetings

[timid garnet]

"Model's level, in one generation, with one word as positive prompt."
https://civitai.com/images/2120399?period=AllTime&periodMode=published&sort=Newest&view=categories&excludeCrossPosts=false&modelVersionId=140608&modelId=119229&postId=519267

[halcyon tusk]

hey let's keep it civil here, please be respectful to everyone.

[timid garnet]

https://tenor.com/view/spider-man-we-one-gif-18212100

Are you my spanish twin ?

[halcyon tusk]

hahaha seems so!

on my timeline we don't use "i" tho

[timid garnet]

Yeah, y is level one, i'm level two

Asking for 'Creativity' as positive prompt.
Duude, i love this model

[tepid surge]

||I'll need to learn how to work with* human faces soon||

[timid garnet]

Nice ! What do you mean by 'to work human face' here ?

[tepid surge]

ah, with* woops

and like to do touch ups

[timid garnet]

Also, are you using a1111 webui ?!

Cause your VAE is obviously broken

Here, download "sdxl_vae.safetensors", and save it in you models/VAE folder, replace your existing one (rename/archive the old one)
https://huggingface.co/stabilityai/sdxl-vae/tree/main

This is what i mean

[tepid surge]

ooo good to know! I'm using Comfy and i'm just pulling the pin from the xl base, and I see the vea one must be the one with the embedded one or something?

[timid garnet]

colors artefact, it's a wrong VAE. StabilityAI uptated it a few weeks ago

Also, wait

Don't use the link above ... The orignal VAE is way too long

Causing the freeze at the end on the generation

Let me find you a good one (fp16)

This one !
https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl_vae.safetensors

The end of the generation will be barely instant

like 10 times faster

*reroll the exact same pic with same prompt/seed, to see the diff

[tepid surge]

also does this one mean it has it embeeded or something i remeber hearing somthing about that, but just dont know what the diff is between the two, also thank you btw xwx!

[timid garnet]

Delete these failures, both lol.
The new quality when zooming will be highly improved. And way faster with a decent resolution (1024*1024 or more)

And if you wanna a good model. Try ZavyChromaXL ((:

[timid garnet]

Is this a MicroShip ?

[stone fossil]

It need to be there else the circuit is not complete.

[timid garnet]

Aww, i forgot. Delete --no-half-vae from the launching batch !
(Need to fully restart the webui)
We aren't Amazon, half precision VAE is fine for us !

Only 'Lofi' as prompt

[timid garnet]

Wan't wait for CN on a1111 webui

[tepid surge]

Nice 🤘
as a thank you: https://www.youtube.com/watch?v=ty0EezB5qkI

[timid garnet]

This stupid thing keep spaming "CN not compatible, disabling" ... even with good models

[halcyon tusk]

chill out other me!

[timid garnet]

Awwwww, and good music ! Much love buddy

[timid garnet]

[halcyon tusk]

https://tenor.com/view/el-risitas-juan-joya-borja-ratones-coloraos-laugh-meme-laughing-man-gif-24899295

risitas to the rescue

[timid garnet]

ahahah, always !

[dense chasm]

ComfyUI can be used to make inferences on controlnet

[timid garnet]

Yeah, but i don't need all this workflow ui architecture... I didn't touch it yet. Also i'm a pure txt2img guy, seing CN as a cheat, kind of

But cool tool, for sure

Especially with Deforum, tho. Not really txt2img

[tepid surge]

before I started fixing my graph cx

[timid garnet]

[timid garnet]

drum rolls 🥁

[prime barn]

Hi

[wicked frigate]

what if i told you controlnet in swarmui is as easy as

upload image, select the model, and go

even has a nice UI listing the models you got with metadata

[tepid surge]

Gotta read the text of the msg too

[timid garnet]

Hum, didn't know this one

[prime barn]

Hshsgsghs

[timid garnet]

lol, is it supposed to be the same pic ? xD
No colors artefact, seems good to me !

And probably way faster

not stucking the webui into 404 for the api, LOL

[tepid surge]

no that other pic was before i made the changes
thx again btw xwx

[timid garnet]

Lofi

[timid garnet]

https://civitai.com/models/119229/zavychromaxl?modelVersionId=140608

[tepid surge]

OKAY OKAY i'll go get rn

[timid garnet]

My last posted pictures with this model !
https://civitai.com/user/wizz13150/images

[timid garnet]

Sweet, would be crazy with the linked model above too

It knows how to write his name

One of my favorite pic !

[tepid surge]

Same prompt as this one

[timid garnet]

Give me your prompt lol, to test

[stone fossil]

Eh well Im kinda cheating.

I use a fine tuned model and a self made lora for this one.

[timid garnet]

ahh, i was bugged. With the new model

Not the same sh*t, right

[stone fossil]

So the prompt alone will be nothing heh.

[timid garnet]

Don't underestimate this model, LOL

no need lora here, for most things

[stone fossil]

Most yes.

[stone fossil]

This wont, nor real good ice styles for example.

So I lora em.

[tepid surge]

The lighting's def better so far x.x

[timid garnet]

Yeah, as you said, depends the finetuned model. (trained, not merged)

Train the model with the same pics, don't need the Lora anymore

[timid garnet]

Summer Collection

[timid garnet]

Sexy ! xD
More a realism model, a good prompt will fix this lmao

[tepid surge]

that could like so be made (love the brat style-like seats >u<)

[stone fossil]

Yup.

GL on that one tho. 😛

[flat nymph]

Have anyone already trained SDXL on runpod? Are you satisfied with the result?
(Has a tough choice: try SDXL or v1.5 model training)

[stone fossil]

Yep me lora.

[tepid surge]

Yeah the lighting and portate style stuff's kicking butt cx (i def need to learn how to work with faces though cx)

[timid garnet]

Yeah, no real secret with faces. It's how the model is trained, with 1024*1024 pics.
OR the face is highly detailed, and close
OR the face is not, and more far
So if you use "highly detailed face", the model will generate a pic with a closer face

You see the logic !?

So, "highly detailed face" is the way, but will more be like a portrait

OR, ofc, higher res

More pixels = more detailed

[stone fossil]

These are good questions hehe.

And yes.

On last one.

[tepid surge]

Ah nice will add, yee i add dynamic lighting and vibrant to a lot of mine

[timid garnet]

Also use a related ratio, to help the model.
More 768x1216 if full body, more 1216x768 if landscape ..

[timid garnet]

Ho, great indeed

[timid garnet]

Very realistic, nice one too

[timid garnet]

Dude, this is cool as f

[tepid surge]

prompt was like "robot holding a sun" or something like that

but i also get ones like this where it's like okay what's going on cx

Oh yes posing for fucking vogue with a banana xD

[timid garnet]

Magic one

[untold crescent]

is there already easy way to get auto1111 gen SDXL base model at same speed as comfyui on 8GB VRAM?

when SDXL launched, comfy was notably faster, which I assume is because it runs with different parameters by default on such setup

[timid garnet]

Yeah, as you said, probably other params. You can use the token merging ratio (0.3) and the negative guidance (1-2), in settings/optimizations

No secret about the speed.

Also a fp16 VAE

And depends the sampler, don't use DPM++ SDE karras, for example, way too long

prefer 2M karras

[untold crescent]

it could be other things too, because without medvram setting on auto1111 it would be super slow.. I dunno if comfy uses same solution for vram saving

[timid garnet]

medvram is slower, indeed. Not sure how comfyui deal with vram. Probably not magic on this point

[untold crescent]

thing is, you can't run SDXL with auto1111 and 8GB VRAM without medvram

it's going to take minutes to gen single image then since it goes over 8GB in usage

[timid garnet]

Yeah, even my rt3060 isn't enough at this point, huh. Some optimizations should probably come for the 12gb cards, not sure for 8gb

[untold crescent]

comfy has been good in this regard, but I can't help to think the same improvements on base settings could be done on any webui really

mainly asking here is there's already some solution, not really looking to reverse engineer the differences myself

[timid garnet]

Yep, would need some further analysis of your workflow

or a comfyui thug

[untold crescent]

I don't really like iterating with comfy.. it does the job but that spaghetti isn't my thing 😄

it's great if you need something custom

[timid garnet]

https://tenor.com/view/seth-meyers-right-am-i-right-totally-same-gif-9191853

it's like

https://tenor.com/view/papercut-hand-fingers-paper-wound-gif-4860085

"creativity" as prompt

[timid garnet]

ah, try this half precision vae, if you don't yet.
https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl_vae.safetensors
Put the safetensors into the models/VAE folder
then remove --no-half-vae from the start batch

The end of the generation (the -way too long- 100% duration with full precision VAE) will be way faster (like 10 times, regarding the resolution)

[untold crescent]

hmmm, VAE stage isn't super long on comfy itself

I mean, it doesn't make these solution same speed still

[timid garnet]

If you generate a pic in the webui, at 1024*1024, you don't see a delay between the last step and the pic coming ?

[untold crescent]

I do, but the difference between auto and comfy is still bigger

[timid garnet]

For me, using the full precision vae, for example at 1024x1024, i get 20 sec to do all steps, and the pic is ready in almost 30 sec.

This delay is the vae working on the pic

Using a fp16 vae will greatly reduce this duration. Barely to 0 for a 1024*1024 pic

Give it a try ((:

You won't be disappointed, in the worse case lol

[tepid surge]

wow xox

[untold crescent]

I don't see any improvement on auto111 for using that fp16 vae

it's loaded (saved it next to the model so it autoloads)

also removed that --no-half-vae param

still taking 37 seconds per image

[uncut fiber]

@untold crescent what gpu?

[untold crescent]

2070 super.. I get the VRAM is issue here, just trying to figure how to get auto1111 as fast as comfyui with SDXL

[uncut fiber]

o.k. i believe with next update. Hope it will works.

[timid garnet]

repost of one of my most more cleaner pic

not even upscaled, tho

[untold crescent]

same model, same fp16 vae, same gen settings, takes 18 seconds to gen on comfyui

on auto1111 37s

it shouldn't be twice slower

[tepid surge]

yeah i had the same prob on auto 1111 i'm sure it'll get sorted out soon enough

[timid garnet]

upscaled :

cool, even better

[tepid surge]

i have a 3060 with 12gb

[untold crescent]

could try on vlads repo I suppose

[timid garnet]

That's a lot ..

[untold crescent]

it is, but not 37s 😄

SD1.5 gens at it's lower resolution in few seconds

but SDXL isn't super happy about those lower resolutions

[tepid surge]

used "final fantasy" as a part of the prompt

[timid garnet]

At this point, i have the feeling that i'm able to see when a pic is using ZavyChromaXL or not

[untold crescent]

auto1111 also gets stuck on first gen forever

once it gets past it, it's that 37-40s per image 😄

ok, I'll try vlad's

[timid garnet]

did you remove the arg from the batch ?

[untold crescent]

yes

[untold crescent]

^

[timid garnet]

bah, this is just a fork from a1111 if i'm not confusing

[untold crescent]

it is but it has bunch of different approaches to things

[timid garnet]

hum, no reason

[untold crescent]

it had SDXL support before auto1111

[soft zealot]

theres no benefit in generating below 1024 with SDXL and most of the time it wil simply mess up the image

Same as going too big will cause image quality issues.

Theres a table somewhere of recpommended Generation sizes to use with SDXL

[untold crescent]

yeah I figured

[timid garnet]

When i slown down, i always finally notice something like a bad sampler or something

in fine

[timid garnet]

Join Zavy's discord lol, there is a bot on it

link on civitai

[tepid surge]

i dont like to be in too many discord cx

[timid garnet]

Wise words !

[tepid surge]

fr fr it def hits faces more studio like 💯

[timid garnet]

Wait, i have an awesome prompt

Pikachu, Breathtaking, 8k resolution, extremely detailed, beautiful, establishing shot, artistic, hyperrealistic, octane render, cinematic lighting, dramatic lighting, masterpiece, light brazen, extremely detailed and beautiful face

Just replace Pikachu with anything, and show me

This octane render is pretty crazy, for me

(I have 2 different negative prompt for this)

[tepid surge]

uh oh cx will do a few runs cx

[copper spire]

Hello world ! I was just wondering if it is possible to use Sdxl 1.0 with automatic1111 on colab or it s just locally or the previous version of stable diffusion ?

[soft zealot]

I replaced Pikachu with your mom

[tepid surge]

there's prob a colab somewhere by now if you search it

[soft zealot]

It should work as well (or badly) on Colab as it does locally

[timid garnet]

(deformed, distorted, disfigured:1.3), 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. tattoo, watermark, text,
or
(worst quality:2), (low quality:2), (normal quality:2), lowres, easynegative, signature, watermark, username, cropped

here they are

[timid garnet]

No joke, dig further with this prompt, and the second negative prompt

Crazy ones

And pretty universal

[copper spire]

Ok ok thanks i m gonna give it a try 🙂

[tepid surge]

This isnt your prompt Wizz, just had to finish this one casue it was so intense cx

[boreal gate]

2

[soft zealot]

why, will it enable an ulocked level?

No offence but the way this is all phrased this sounds like the sort of spam clickbait stuff that you get on other social media sites

[timid garnet]

My second negative prompt fit pretty well this model

And the first one for full body things

👍.

[timid garnet]

Try it

with this neg :
(worst quality:2), (low quality:2), (normal quality:2), lowres, easynegative, signature, watermark, username, cropped

[soft zealot]

I also replaced the word pikachu with cockwomble

[timid garnet]

With ZavyChromaXL model, of course !

Not sure for others

I gave up, only using this one

[soft zealot]

are you shilling for them or something?

Sounds like it from my seat

[boreal gate]

@ionic gulch img2img, candy causes the computer to crash., 12vram

[tepid surge]

"Street vender" how melancholic cx

"sad fruit for sale"

[soft zealot]

took out all the additional crap and started with

"your mom, Breathtaking, extremely detailed and beautiful face"

Much better (although I accept the styles will always add stuff back in)

[tepid surge]

Luxury hurrican fruit

[tepid surge]

the materials so nice!

[timid garnet]

lol, dude , i'm obviously shilling, but it's totally a personnal pov. Deal with it.
Here is what i mean with this prompt !

You can't tell about a model you don't know. Chill out too

I only replaced the first word, in every pic

[soft zealot]

I see nothing special 🙂

[timid garnet]

Sure

https://tenor.com/view/jennifer-lawrence-the-mocking-jay-the-hunger-games-hunger-games-mocking-jay-gif-5701115

[tepid surge]

explain this

cx

[timid garnet]

almost centered head

[soft zealot]

the first three lines are what I entered in the input prompts, everything else added by the styles selected

Here's "your mom, Breathtaking, extremely detailed and beautiful face" again with the same seed and other parameter s just with a different model

[ionic gulch]

i'm working on adding revision to my workflow without giving up the other options like img2img. revsion + img2img looks very promising already - i think with the right settings this can give you better results than revision with two input images.
https://i.imgur.com/ve3OvNh.jpg

[timid garnet]

wow, impressive model

NB: I didn't have to generate xx pics to get those above, only a batch of 4, in a discord bot

But anyway, i'm not sure on what we disagree

[ionic gulch]

if the whole system crashes that sounds like a more serious problem. a workflow alone should not be able to crash your computer, only run out of vram/ram.

[timid garnet]

As a cool experiment, try some batch using only the model's name

Results are pretty crazy for me

single generation, batch 4

(using a discord bot, neg and lora is default)

[tepid surge]

1 v 1

[timid garnet]

hum, i tried, not able to get a decent snake (body) easily. Was trying to produce an Orobouros

dinosaurs are quite complicate too

Ectoplasm Lich video, using deforum and parseq (increasing seed)
(SOUND! SOUND!)

[tepid surge]

I could do with that on a stage screen while at concert fs fs cx

ai so cool

[soft zealot]

Im not overly taken with it ATM TBH , had it installed for a couiple of days so maybe havent fully playe dwith it yet

[ionic gulch]

those Boolean switch nodes are stupid. 0=b 1=a for them. luckily i noticed while playing with the conditioning input switch. this means my candy/depth options used the wrong images...

[soft zealot]

yup, thats a mindfuck lol

[tepid surge]

Writes this down

thank you

cx

[timid garnet]

Yeah, joking buddy 😘
All relative

[soft zealot]

I'm presuming you looked at the image filename as that contains the model(s) used as well as the styles selected?

[timid garnet]

also mixing creature is fun, using the syntax (rat:blobfish:0.75)

[timid garnet]

lol nop. Not this kind of guy

[soft zealot]

maybe you should ;o)

[timid garnet]

😭

[ionic gulch]

revision + candy also gives nice resulits:
https://i.imgur.com/RAvTvwx.jpg

[timid garnet]

What revision means here ? 🤔

[ionic gulch]

@timid garnet using the new clip vision option instead of a text positive. in the example the voodoo lady is used instead of the text prompt and mixed with the alien thorough control net candy

[soft zealot]

Think of it more as ReVision.

Its a process of using clip ewncoded images instead of textual prompts

[timid garnet]

A mixing process

[timid garnet]

hum, i'll dig this

Yeah, i don't see the links, but yeah okay got it now

cool

[soft zealot]

Its not a "mixing" Process.

You can use one or more images to build a prompt without using any textual input as the Clip Vision Process reads the image for its content and outputs it as a conditioned input to a sampler

[timid garnet]

I'm using CN only in deforum, so waiting so hard for updates

[timid garnet]

It is exactly what i mean by "mixing process" !

https://tenor.com/view/hillbilly-gif-19811957

jk jk

[ionic gulch]

the mixing happens trough candy control net in my example above. or through img2img in the one i posted a little bit earlier. clip vision replaces the text prompt completely in my workflow, as mixing txt + revision didn't give me consistent results.

[timid garnet]

Thanks for the elaborated explanations. You just saved me an hour of my life searching for it

[soft zealot]

or you can just use it in a multistep txt2img process as I do 🙂

[timid garnet]

*sry if i'm a 'lil rude lol, internet contaminated me. I'll shut it up

[timid garnet]

[tepid surge]

cw: minor blood

[timid garnet]

lol

I also have a funny (not really, the phrasing is) prompt, let me find it

[untold crescent]

SD.Next with diffusers backends seems way faster than Auto1111

[ionic gulch]

i'm still using the txt prompt for the refiner part, so i already have something like that included. my goal is to fit all inputs on a single screen (at least on my monitor) so i can't go too crazy with the options. fitting txt2img, img2img, candy, depth and revision + 3x lora + offset lora + 4 refiner strenghts is already a lot :)

[timid garnet]

landscape of holographic futuristic city made by fractal zentangle colorful moire pattern art, like 3d sound waves
(worst quality:2), (low quality:2), (normal quality:2), lowres, easynegative, signature, watermark, username, cropped

[untold crescent]

taking 22-23s on this gpu

comfy is this faster by 5ish seconds

but nothing like almost 40s on auto1111

[soft zealot]

screen grab is what loads in the bowser window on my 1440p monitor

[tepid surge]

i getting 27~43s an image dependeing on the sampler comfiii

[timid garnet]

Don't tell me you're using the refiner with a finetuned model

[soft zealot]

sigh

why oh why is everyone obseesed about speed to generate or iterations /second etc.

Does it really really matter if it takes 20 seconds or 40 seconds or even 60 seconds?

(thats a rhetorical answer to which the answer is "no it doesnt bloody matter")

Ok if its taking minutes to generate a bog standard 1024x 1024 inage thats slightly different but still not the end of the world

[ionic gulch]

i'm using refiner with all models, but always create 70%, 80%, 90%, 100% base - so i can see if it helped or if the 100% base is better

in a lot of cases refiner helps - in some cases different elements are better or worse with refiner - if you have 4 options you could easily pick the good parts of 2 or 3 variants in photoshop or some other image editing app.

[tepid surge]

that'd be 40-60 if I used the refiner, I on a 4060

I'm having too much fun generating

seriously this xl model's fantastic cx

[timid garnet]

I mean, redo the same pic with the same model, would probably be the same. That's what i have in mind. I see it as a model for the imperfect base model. This last model being only for training purpose, in fine, not to use as it is.

So, totally useless from my pov. Not mentionning that finetuned/trained models basically don't need the refiner at all.
Never used the refiner

[timid garnet]

https://tenor.com/view/seth-meyers-right-am-i-right-totally-same-gif-9191853

at least i'm not shilling a very bad one

[ionic gulch]

if you create all 4 images at once, you can save steps, as you can use the first 70% for all 4 variants, and after that just need double the steps instead of 3x the steps. yet you still need 4x decode vae - so overall it still takes some time to create all 4 variants (not too bad on my 4090)

[timid garnet]

I 'memba my first model. omg

[timid garnet]

🤔
Hum, never thought about this concept

Clever in some situations

Even more if you can choose when you 'cut' the canva

hum

ah, the wowifier fit pretty well this model too ! and details lora ofc

[rugged crow]

which one is this

[timid garnet]

Huh, i will really looks like shilling if i repost the link, LOL !