✨｜sdxl

[clever verge]

Yeah, in your custom node.

[nimble heart]

could also be a name thing. maybe naming your field "seed" cues comfy to add a randomize option

[civic sigil]

Yeah this is it

[pure hazel]

Anyone know what this errors means in Comfy UI? Was working but then stopped "RuntimeError: Sizes of tensors must match except in dimension 2. Expected size 154 but got size 77 for tensor number 1 in the list."

[civic sigil]

Iirc Ive looked at the code before

[clever verge]

Interesting behavior. I'll try that.

[grim lodge]

Where do we get the style.csv file from? my stable diffusion folder doesn't have it for some reason

[clever sky]

Insane

[plain flame]

guys, is there any way that i can use the clipdrop 0.9 back again instead of the 1.0 ?

[calm shell]

SDXL 1.0 simple prompt

[clever verge]

@nimble heart @civic sigil Thanks. Worked like a charm. Been trying to figure that out, but way simpler when seed was connected to it's behavior. Which makes sense.

[calm shell]

same background with color and a new subject

[formal kraken]

this is so fun man, just got home from work and the model is out already wild results

[dusty turtle]

my poor SSD with all this prompting today, over 10 drive writes in 1 day....

[formal kraken]

i just got home but what are u guys using as the sampler

for sdxl

[calm shell]

Im using dpm++ 2m

I moved all of comfyUI off of my ssd

[glad fulcrum]

Hmmm. I was loading the refiner and my pc restarted. In A1111

[formal kraken]

i have a whole second power supply i rigged up with a bunch of drives

[formal kraken]

did u update

[glad fulcrum]

Yes

I am updated

[calm shell]

I can't use A1111 with SDXL 1.0 yet

[formal kraken]

y

[calm shell]

Works flawlessly on ComfyUI faster for all other models too

[nimble heart]

full node is here if you wanna peep it
https://github.com/Beinsezii/bsz-cui-extras/blob/master/custom_nodes/bsz-principled-sdxl.py

[clever verge]

Was looking at the code when I saw your screenshot. 😉

[formal kraken]

wow bumping the steps up makes a huge difference on the base model

[nimble heart]

I import the base nodes and use them directly but I've seen other base+refiner nodes that do the conditioning and sampling directly so that's valid too

I just find it easier to just re-use comfy's work

[clever verge]

Yeah and easier to maintain.

[formal kraken]

ive been boomering it and too lazy to set it all up and learn it, i use a text file to save everything lmao

and leaving the webui open forever

[nimble heart]

wdym text file? what are you putting in it?

[eternal fog]

oops lmao, I put Ice Bath and it took it a bit literally and it's frozen her in ice

[formal kraken]

prompts lmao

[nimble heart]

ah yea. I do that for prompts too

[nimble heart]

most people use a Note node that's 500 tiles large to save all theirs

[formal kraken]

how big can u go for resolution

[formal kraken]

in this model

master turd

[static prawn]

same here for me, unusable on auto 1111, 1 image takes 20minutes or more, on comfy 1-2 minutes.

[tribal gale]

You're running into VRAM bottleneck

[nimble heart]

if you do a low res first pass, the 2nd pass can be as high as your GPU can handle. else maybe like +25% max

[static prawn]

but i feel with all the nodes comfy is so much

even for displaying in a structured clean way

[formal kraken]

dam time to get a a100

[static prawn]

im kinda sad i cant use auto 1111

[indigo carbon]

also wait for TensorRT to work for ~twice the speed, even triple

[edgy sierra]

do not use auto1111

[nimble heart]

why have 30 small nodes when you can have one fat really complicated node
#✨｜sdxl message

[nimble heart]

no.

[indigo carbon]

why not?

[nimble heart]

nvidia

[indigo carbon]

ohhh you're an amd boy

[static prawn]

dunno i just dont feel comfortable with using comfy ...

[edgy sierra]

I know... you need to learn how to use it, it's just time. keep it up

[nimble heart]

if you've ever used blender shader nodes or unreal engine blueprints you can learn comfy super fast

comfy you only need to know like 5 or so nodes to get started

[edgy sierra]

it's pretty easy

[static prawn]

i probably just will use artroom for sdxl now, its clean and runs super fast

[azure oxide]

you dont even need to learn how to use comfy, just load up a popular workflow like sytan or serge's and just prompt in those like you would with a1111 lol

[cinder aspen]

What is artroom

[indigo carbon]

I do, I leant Comfy, it's a good UI. but A1111 is more comfortable for me. I would switch entirely to comfy if it will be usable with TensorRT tho

[static prawn]

artroom ai, a standalone program for doing ai

[formal kraken]

fartroom

[cinder aspen]

Lol

[static prawn]

doing really well with sdxl

[nimble heart]

up to about 3.2 it/s on base 1024 so it's working pretty well for me. That's with sub quad cause the card doesnt support torch attention yet.

[cinder aspen]

Dunno if RTX 2070, 10750h processor and 16gb RAM would handle SDXL well locally

[static prawn]

comfy is just , u wanna add something , then u need a another node , another node and another one, i dont figure out, how all those extra sampler settings work. dunno just overwhelming for me.

[edgy sierra]

yes

[nimble heart]

the official SAI ui uses comfyui as a backend

called swarm or something idk

[cinder aspen]

Its in alpha right now

[azure oxide]

you dont need to add anything lol, lots of people have already done that for you. just load up their workflow and you can just prompt lol

[nimble heart]

all the best software has it's version start with '0'

[cinder aspen]

Yes 😄

[static prawn]

@rootybeer just a lot of additional settings that i dont figure out.

[calm shell]

Artroom is an installable AI renderer, similar to A1111 it's a bit faster but the UI is trash

[cinder aspen]

I dont intend to use this as a main tool and not even as main gen AI tool

[static prawn]

i just dont think auto 1111 will improve a lot bec. most people have better gpus and dont have any issues with it 😅

[nimble heart]

the moment comfyui releases 1.0 someone needs to start over @ 0.1 on a new UI

[clever verge]

There is no need to make it complicated with comfy, just use his sample flow. It will work nicely. Better to put effort into understanding the model. After that you are probably ready to fiddle with the nodes.

[solemn scroll]

SDXL feels like you have LORAs running, except you don't have any.

NSFW stuff aside, it is already good enough as it is, compared to 1.x

[cinder aspen]

Ooh ok thx ^^

[static prawn]

the ui got reworked its not trash.

[nimble heart]

also a1111 has a bunch of weird default settings. Like high vram and whatnot

[static prawn]

they are putting a lot of effort into it

[nimble heart]

so whoever was having 20 minute renders on a1111 you can probably get it up-to-speed

[lean knot]

Is there a comfy Discord server?

[trail obsidian]

Guys do i use the same sampler and steps for high res fix in comfyUI? Whats a good denoise strength, 0.3?

[static prawn]

dunno im running on a GTX 1070, so auto 1111 takes 20 minutes for a gen, comfy only 1-2 , i dont know why the difference is so huge

[cinder aspen]

Id only need this here for specific image generations. Everything else like inpainting and outpainting etc. i do in Photoshop/Firefly

[azure oxide]

he prefers something called matrix instead

[cyan crown]

0.15

[calm shell]

I also use InvokeAI but it has limitations too

[azure oxide]

re:

[trail obsidian]

Ty! and same sampler?

[high skiff]

new changes to my upscale workflow are looking promising

[static prawn]

im really happy with artroom, and its improving day by day, so i cant say anything negative

[high skiff]

it no longer eats skin texture

[static prawn]

had sdxl support directly after it came out

[calm shell]

In Comfy UI I use denoise of 1.0

[nimble heart]

yea it should really eat more than just the skin, else the rest of the body goes to waste

[high skiff]

lmfaoooo

[calm shell]

Artroom crashes on SDXL 1.0

[high skiff]

these are crops, but yeah, its working really good now

[calm shell]

Have to try to updte it. Never got it to work with SDXL 0.9

[high skiff]

eyes are still an issue, but hey

[nimble heart]

left 1024?

[plain flame]

after the update, when i download the image in my mobile browser its in PNG in 3-5 mb with 1024x1024 and when i download it on PC its in JPEG in 200kb with 1024x1024. But with the earlier sdxl clipdrop 0.9 the default images were downloaded in PNG with around 5mb size and 2048x2048 default resolution. The update messedup everything man. Can someone help me what went wrong?

[high skiff]

a crop of 1024x, yes

original image looks like this

[nimble heart]

that's pretty good for just the refiner

[high skiff]

vs a screenshot of the upscaled version

[cyan crown]

use the png i'm uploading to see

[high skiff]

both together so you can jiggle between them to see the differences lol

[trail obsidian]

hm if i go under 0.3 the image gets bad quality

[cyan crown]

open the bunny png and see what i use

[plain flame]

after the update, when i download the image in my clipdrop in mobile browser its in PNG in 3-5 mb with 1024x1024 and when i download it on PC its in JPEG in 200kb with 1024x1024. But with the earlier sdxl clipdrop 0.9 the default images were downloaded in PNG with around 5mb size and 2048x2048 default resolution. The update messedup everything man. Can someone help me what went wrong?

[trail obsidian]

ok but how many steps are you using?

[nimble heart]

@high skiff wonder how doing the misalign step would look using only the refiner. Have a refiner node that just adds 30 steps of noise but doesn't sample, then the 2nd node samples @ 29 steps

[cyan crown]

if you open the png in comfyui you'll see all the parameters

[high skiff]

no idea, give it a shot :P

[nimble heart]

noooo another thing to code into my omega node

[cyan crown]

50. But open png in comfy and see all the config

[high skiff]

hehehe

[trail obsidian]

hm when i drag it in nothing happens do i need to do something special?

[cyan crown]

just uploaded json

[west breach]

Morning peeps, here are some male celebs I generated overnight with my 'grumpy style' lora. Turned out pretty good I think

[cyan crown]

also if someone needs help I made a quick tutorial

https://www.youtube.com/watch?v=29hKqm92Rvo&t=53s

[trail obsidian]

TY

[upbeat summit]

good morning 🙂

[upbeat summit]

ripped eddie 😄

[west breach]

he means business! 😄

[nimble heart]

guys dont accidentally swap your base and refiner models you get Kaminoans

[trail obsidian]

oh but you are doing the normal upscale not latent which give much better details?

[orchid oar]

I love him

[upbeat summit]

you should really switch your positive and negative prompt to enhance the quality even more

[solemn scroll]

Hi everyone, do you thing the steps are most important with the refiner or the base? rn I'm getting mid results when looking for some cyberpunky stuff (20 steps each)

[static prawn]

u need the beta

the beta has a completely changed ui

[trail obsidian]

why does the latent high res fix take sooo much loger than the first normal pass?

[nimble heart]

vram just hit 23.8 / 24.0

[native moon]

hey when i switch models in automatic 1111 it uses a ton of system ram how can i stop that?

[west breach]

Because it's 4x or 2x the size? Maybe it's going into shared memory?

[solemn scroll]

A1111 or 4K generation?

[nimble heart]

2k on an AMD card lol

[trail obsidian]

Yeah probably it the 2x lol its late

takes like 2min to generate a 3k image

[native moon]

like when i switch from the bais model to the refiner it eats up all my system ran and crashes my pc

[trail obsidian]

ComfyUI is doing work starting to love it

[nimble heart]

I should try 3k

[trail obsidian]

i hate the gradio bs

[west breach]

would be like generating 9 normal sized images

[native moon]

can someone please help me

[trail obsidian]

its pretty dope doing widescreen 1024x1024

[trail obsidian]

In what way?

[nimble heart]

dope yes but it's gonna have to unload models while working which means slow AF

[native moon]

in automatic 11 11 when i swich to the refiner model it eats up all my system ram and crashes my pc

[trail obsidian]

oh dont use that

for SDXL its utter trash

use ComfyUI

its like 10x faster

[honest flint]

lmao, not exactly what i was going for but interesting none the less

[trail obsidian]

A1111 didnt even load for me others need like 5 min per pic

[native moon]

why is it faster? doesnt it use the same underling code

[nimble heart]

--medvram should really be the default in a1111

[trail obsidian]

A1111 isnt ready and badly maintained atm

[honest flint]

its not really faster at least by anything noticable, its just alot of comfy fan boys in this channel

[trail obsidian]

at least thats what everyone says switched to Comfy and its blasting pictures

[royal tangle]

Not sure why such hate for A1111 it's a different tool it's working fine for me.

[trail obsidian]

can even do fullHD into 4k

[trail obsidian]

Because it doesnt work for many and is very very very slow

[indigo carbon]

hey guys, I'm not updated- is 11.8 still faster than 12.1?(CUDAtoolkit)

[trail obsidian]

on comfy i can do 12 sec gens on euler for example in 1024 and i have only a 2080ti

[honest flint]

A1111 works fine for me

[nimble heart]

depends what you're doing. If you're not using Xformers comfy is a noticable chunk faster

[trail obsidian]

comfy needs less vram

[honest flint]

wth would you not use xformers?

[nimble heart]

amd

[trail obsidian]

and is much faster thats just a fact i havent used comfy before today

i tried it and its night and day

[lean knot]

I'm not sure that the time saved due to performance is used at all

[trail obsidian]

do what you will

[lean knot]

because it feels like you have to tweak a lot more

[native moon]

this does not solve the proplem. my proplem is the system ram and not the vram

[honest flint]

i use comfy and a1111 depending on what im doing and comfy isn't noticably faster on my system

[lean knot]

Trying to get something similar to ADetailer with ControlNet and it's a complete pain in the ass

[nimble heart]

16 gb?

[orchid oar]

I prefere a1111 but comfy is way faster

[solemn scroll]

I strongly approve. Currently running Base and refiner

[native moon]

no 32

[nimble heart]

dafuq

I never go above like 22 gigs

in comfy at least

[native moon]

it is like at 18 normaly and when i switch odel it shots up until my pc crahes

or i stop it bevore that

[nimble heart]

....does it not unload the old model before loading the new one?

[trail obsidian]

Yeah its crazy i tried A1111 and it legit needs 3x the time for generating AND more vram. i dont know why people hat on comfy it looks so clean the gradio stuff was always laggy

[nimble heart]

try setting your model caches all to 0 in the a1111 settings page

[native moon]

i am nut sure but i think so. (in system ram)

[trail obsidian]

I can generate 12sec pics with comfy of course the full 4k workflow with high res takes longer with the best samplers

[honest flint]

cause different people like different things? and i dont know whats up with your system and others but these 1 billion times faster claims are crazy

[trail obsidian]

Comfy is like 1 click download took me 30sec to install at least try it tbh

[trail obsidian]

I mean check reddit threads. its not only my opinion or tests

Not here to argue works awesome for me im happy just trying to help

[honest flint]

i dont need to check reddit, i have both installed and on mine they performan the same at about 3 it/s

[rough rain]

hmm

[native moon]

i will also try it but i want to get 1111 to work its clearly not normal behavior

[trail obsidian]

Ok but other people and the one i was replying to having issues. Not sure why you come and say hey it works for me

[native moon]

is tcmaloc important?

[old mulch]

UncomfyUI sux

[solemn scroll]

Some people on /sdg/ where explaining about how there were some contrast issues at some point. It's also some kind of weird FOSS chauvinism, just like with linux distros I guess lol

[trail obsidian]

Nah wayyy cleaner gradio is laggy af

[honest flint]

cause if they like A1111 maybe they can figure out whats wrong instead of moving to something like comfyui which can be way more complex

[trail obsidian]

whatever im back making 4k pics 🙂 have a good one guys!

Many said A1111 isnt yet officially ready else try github issue thread

thats why i said use another software vladmatic works too

[nimble heart]

@high skiff kinda see what you mean by eating skin texture when using only refiner on upscale. This is just bicubic pixel upscaling. Sounds silly but I wonder if you could just run a regular image sharpening algorithm from PIL or something instead of a whole ass AI upscale since the refiner just needs the extra small detail contrast.

[honest flint]

it can improve performance and memory usage

[native moon]

i cant install it with pip

[high skiff]

oh wow, that is a lot of image change

[nimble heart]

only 0.1 denoise, but I also change the clip target resolution to match the output

so first is 1024 target, second is 2048 target

[high skiff]

damn

[honest flint]

its a c library, if you are using wsl or linux you will probably want to install it through your os package manager

[high skiff]

let me try that I guess lol

[autumn forum]

i believe the target should be 1024 but the others can be 4096.

[nimble heart]

I always target what I want the final image to be

seems to work well enough

[autumn forum]

yes, i think that is the proper way to do it.

[native moon]

ther the package cant be located

[honest flint]

what os are you running?

[native moon]

linux

[honest flint]

do you use apt, rpm, yum, apk ?

[native moon]

apt

[nimble heart]

tcmalloc is part of the google perf tools I think

doesn't really do anything for memory usage though

[honest flint]

try libtcmalloc-minimal4

but yeh its also inside the google perf tools package

[native moon]

what is it called

[nimble heart]

just search "perf" on your package manager

[zealous horizon]

SDXL can finally generate skateboards that don't look like alien creatures eating your feet 🙂

[honest flint]

google-perftools

[lilac raven]

When using --lowvram on automatic, it's a must. Otherwise, the leak will kill the system after some generations

[nimble heart]

I think tcmalloc is a bandaid fix, I don't think it actually solves the memory issues a1111 has

just kinda delays them

[native moon]

vram is not the proplem

system ram is

[honest flint]

tcmalloc is a pretty badass memory allocator anyways

[lilac raven]

Yes, but the problem with RAM is only triggered with --lowvram

[honest flint]

if you can use it, why not load it up, it will only make things better

[lilac raven]

Other that the --lowvram scenario, I noticed no change

[native moon]

it still uses 20 gb system ram

[lilac raven]

But it stays there

[nimble heart]

saw on the a1111 issues thread somewhere you can manually drop cache using a shady super use command

which frees all the wasted mem

[lilac raven]

On my laptop it's 28 GB with SDXL

[native moon]

how mutch ram do u have

[lilac raven]

The laptop has 48 GB, the desktop 16

[static prawn]

A1111 1 Image -> 20minutes or more, Comfy 1-2 Minutes. Especially on lower End Hardware A1111 is completely unusable atm with sdxl

[native moon]

so is 32 gb ram not engoh?

[lilac raven]

32-28=4, so it should still work

[native moon]

my system uses 4.2

[nimble heart]

a1111 doesn't even support using the refiner anyways, it shouldn't use more than like 16 or so gigs

[native moon]

it works with image to image

[nimble heart]

not really how you're supposed to use it but ok

[thorny hill]

It would be interesting if Stability opened a channel here on the server called ''Workflow-Comfy'' where people could post workflows...🤔

[astral jay]

You guys notice any difference in tagging your dataset as opposed as just using the images by themselves?

[nimble heart]

any images people post will have workflow meta embedded

so you can just drag and drop images into comfyui and it'll rebuild the workflow

[native moon]

i guess i try comfy ui i realy thoght 32gb would be engoh

[orchid oar]

I cant use Sd 1.5 Loras for SDXL, right?

[thorny hill]

I know, but I'm talking something more controlled and organized.

[nimble heart]

it is but a1111 has had those issues for a while, sdxl just makes them way worse cause the models are larger.

I've had a1111 use 60 gigs on sd 1.5 before

[zealous horizon]

there is no reason to not try Comfy. It doesn't take long until you'll learn to love it.

[native moon]

like the model only uses gpu ram so why is system ram a proplem here or am i misuderstanding something

[nimble heart]

models need to be loaded into both

[lilac raven]

Or you can use Comfy on Android and then you may hate it...

[native moon]

yes i am downloding it now

[honest flint]

huh? i do an image using SDXL in 1-2 minutes with A1111

[native moon]

but on system it takes way more space

[honest flint]

something is seriously broken with your system if A1111 is taking 20 mins for an image

[nimble heart]

it's cause the base is still loaded when it switches to the refiner I'd assume

[static prawn]

probably stronger hardware

nope , on sd 1.5 everything is fine

[native moon]

probably. why is it not nloding it first

[midnight shuttle]

A1111 uses more VRAM. If you exceed VRAM and start using system RAM, performance will plummet.

[nimble heart]

make sure these are on 0, beyond that idfk

[nimble heart]

if you exceed VRAM torch will just give up

[static prawn]

dunno i never get low vram messages, it just takes 20minutes

[honest flint]

not with modern nvidia hardware and its unified memory tech

[midnight shuttle]

Not with newer NVIDIA drivers. Instead it will start spilling over into system RAM and get really slow.

[static prawn]

not even using -medvram

[honest flint]

modern nvidia will page out vram to system ram

[static prawn]

but my auto 1111 is running fine

[native moon]

yep it was on 0

[midnight shuttle]

That is why there was a rumor that newer Nvidia drivers caused A1111 to be slow. That wasn't exactly true. It was the memory handling not just a mysterious slowdown.

[static prawn]

no Problems with 1.5 or 2.1

[nimble heart]

600 IQ move. Sell GPUs with shit vram capacities that slow down once you run anything to upsell people to the spendier cards

[weak blaze]

When generating with SDXL in A1111 at 1024 with 20 samples it takes about 6 seconds to get to ~90% then takes another ~20 seconds to finish the generation. Using a 4090. Is this normal behavior?

[urban fjord]

NVidia changed the drivers to avoid OOMs on peak VRAM usage. That's a good thing not a evil plan.

[honest flint]

well sometimes you don't care about speed, unified memory allows people to do things not otherwise possible. so while i love to shit on nvidia, i can't say anything bad about this particular feature.

[nimble heart]

anyone that can sell a 3080 with 10 gigs for $650 is evil, no question

[midnight shuttle]

That sounds like the VAE step. But 4090 should have enough VRAM. I would expect that more on 4070 or less.

[static prawn]

might upgrade soon to a 3060 or 4060 ti 16gb

[midnight shuttle]

Nvidia doesn't want consumers using $650 cards for AI. They want AI done on $50,000 datacenter cards so they make more money.

[nimble heart]

or an 8 gig card for $450 or whatever their mid range 40 series costs now

[honest flint]

if you are upscaling a 1k image to 2k, sdxl can use about 33GB of vram fwiw

[midnight shuttle]

They said at 1024 though.

[nimble heart]

consumer cards are already limited on fp64 compute you don't need to nerf the shit out of the vram too

[azure oxide]

takes me that long to generate a picture and upscale it to 2048 on a 3090

[midnight shuttle]

They do whatever makes them the most money. If nerfing cards makes more money for shareholders they will nerf and we will still buy. That's how it is.

[honest flint]

make sure you don't have hi-res fix enabled

[nimble heart]

one of my racing sims used 20 gigs of VRAM; selling a 12 gig card for $800 is actually evil

[static prawn]

i think i have those on default

[honest flint]

2k images with sdxl takes 33 GB of ram like i said so on nvidia cards its going to page out a good chunk of that to system ram which is alot slower then your vram

[midnight shuttle]

But if they only offer 12 GB people will buy it when the choice is no GPU at all. They know this and so they nerf the cards.

[honest flint]

probably better to just generate at 1k then use an upscaler then realegrsan or swinir and scale it to 2k that way

[static prawn]

what is the benefit on caching vae and model into ram?

[azure oxide]

i agree god corporations are just so evil and scummy god i wish i had money

[nimble heart]

2nd hand 3090's go for like $750 in good condition

[cinder aspen]

for simulating id say go for 64 GB forwards anyway 😄

RAM

[midnight shuttle]

Faster to switch between them.

[nimble heart]

and for gaming the XTX is $950 for essentially a 24 gig 4080

[honest flint]

i just want nvidia to release an 80GB+ card for under $5000, am I asking too much?

[nimble heart]

yes.

[honest flint]

you realize nvidia has been capped at 24GB cards for consumers now for like 6 years or something?

[nimble heart]

the original plan was a 12 gig 4080 so 100% asking too much

[midnight shuttle]

Never going to happen when they can sell that same card to a datacenter for $20,000.

[static prawn]

i just couldnt pay for the power consumption of all those high end cards 😄

[honest flint]

my titan card from like 6 years ago had 24 GB of vram

[zealous horizon]

For the ComfyUI noobs. This is ComfyUI in a nutshell

[midnight shuttle]

Nvidia does not want AI happening on home PCs. They want it in datacenters for many reasons.

[static prawn]

im going to get the cheapest with high vram, so 3060 or 4060 ti

[high skiff]

do the 3060, the 4060ti sucks

[midnight shuttle]

For AI get the 4060 Ti 16 GB. Lots of bad gaming reviews but a good card for AI.

VRAM is king for AI.

[nimble heart]

titan was basically a lightweight data center/AI gpu. It had unlocked FP64 compute iirc

[honest flint]

it was also under $2k

[high skiff]

if all you wanna do is AI on it, then sure

It'll likely be slower than the 3060 tho, so keep that in mind

[orchid oar]

I just tested same prompt, seed, steps, cfs,sampler,same size, same batch etc. and sdxl 1.0 model with a1111 and comfy. a1111 took 38,7s and Compfy 6.81s. (If i missed something pls tell)

[static prawn]

i might get it when its cheaper atm 500€ is way to much for me, anyway i also need a new monitor aswell when i upgrade. my monitor doenst support display port

[nimble heart]

and now they killed the titan branding cause too many college students bought titans instead of quadros lol

[midnight shuttle]

Slower than a regular 3060? Really? Maybe the same as 3060 Ti but I don't know about slower since most computation is done on board.

[fathom geode]

there is something wrong with the way a1111 implemented SDXL, it just runs massivly slower, takes up more vram than comfy, and seems to output garbage half the time

[safe tusk]

@primal vault did you mean 4x-Ultrasharp instead of 2x-Ultrasharp in your instructions?

[honest flint]

nah they just released then as x090 cards (3090/4090/etc)

[high skiff]

the 4060ti has an anemic VRAM bus

Extremely small, and it can fall hard on its face in high bandwidth thing s(like AI and rendering)

[midnight shuttle]

Slower because of VRAM usage and swapping. Garbage if you don't choose a correct resolution.

[nimble heart]

the X90 cards dont have unlocked FP64 compute still

[honest flint]

titan, titan rtx, 3090, 4090 is basically the progression and. they all had 24gb of ram afair

[fathom geode]

my settings between comfy and a1111 were identical and a1111 performed way worse

[midnight shuttle]

It would be interesting to see a real comparison for AI. But hard to find reviews on that.

[orchid oar]

Same Seed and everything i think. Left a1111 and right Comfy (a1111:38.7s,comfy:6.81s)

[primal vault]

yeah, it's a typo

[honest flint]

i don't care about FP64 personally, INT8, FP16, BF16, FP32 are the ones I care about

[midnight shuttle]

Yes, I know. Just telling you why it is like that.

[static prawn]

hate that i need to upgrade my monitor for it, bec i really like my monitor and its hard to find tn panels today :/

[deft stream]

Would a 4090 be a good card? I was looking into builds that use it lately

[nimble heart]

for us it doesn't matter, I'm just saying the Quadro/GTX gap is getting bigger not smaller

or A/RTX

whatever they call them now

[high skiff]

for AI, I would go 3090 instead. Same capabilities, and a lot less money

just not as fast

[midnight shuttle]

Maybe I will look at the A1111 code someday to see why it uses more VRAM. But not on my to-do list right now.

[deft stream]

Noted! Thank you

[cinder aspen]

you will want much of RAM as well at that point

[honest flint]

its a decent card for the price, i know alot of ml nerds use https://www.zotac.com/us/product/graphics_card/zotac-gaming-geforce-rtx-4090-amp-extreme-airo this specific card

[weak blaze]

Not sure what the deal is.. Restarted SD and now it's doing them in 9 seconds but still randomly stalls.

[azure oxide]

idk man ive seen 50 it/s reports on 4090s, thats definitely tempting

[midnight shuttle]

RAM is a factor but for AI VRAM is king.

[sharp robin]

the best card available rn if you have the money

[cinder aspen]

if you really want to get deep into AI, 256 GB+ RAM is the way to go

[nimble heart]

if you do gaming also consider the 7900 XTX. Cheaper than a new 3090 and substantially better for gaming at a moderate cost to AI tooling.

[midnight shuttle]

How does that much RAM help if you only have 8 or 12 GB VRAM?

[fathom geode]

if your really going deep, its probably cheaper to start looking at renting gpu space on a cloud

[azure oxide]

its gotta be a1111, im using comfy with 0 issues. im only reading a ton of bad stuff on a1111

[cinder aspen]

no, that isnt enough

at that point youd go for dedicated AI VRAM

[static prawn]

ill take a look if it has an impact if i set those caching stuff to 0

[honest flint]

best card under $2k... the new 6000 ada card has 48GB of ram so if you are looking for under $10k I would go with the 6000 ada personally

[midnight shuttle]

I asked how it helps.

[deft stream]

Thank you all for the input!

[sharp robin]

oof AMD pass...

[static prawn]

dont know what is the Default value on auto 1111 for the caching

[midnight shuttle]

A1111 is less efficient with VRAM for some reason. No one has figure it out yet.

[nimble heart]

if you're like 95% AI I agree but it's been pretty great at everything else.

[honest flint]

and if money is no option the A100/H100 cards are absolute king of the hill

[nimble heart]

AI is pretty mid. Fast but hogs VRAM and has some weird quirks still like 2k VAE decodes taking forever

[static prawn]

hope intel ark is improving

bet they will outperform anything soon

[nimble heart]

they should release a 24 gig battlemage GPU since their AI stack is already probably better than AMD's lol

[honest flint]

i think apple is going to beat intel ark to the goal line at the rate those two are moving

[static prawn]

in gaming they got way better with the developmrnt of the drivers

[midnight shuttle]

All GPUs are good at AI. But AI libraries mainly support Nvidia. That is the difference.

[nimble heart]

a 24 gig gpu for lets say $600 will sell like hotcakes

[cinder aspen]

they are at at least 1.700€ right now

lets see 2024

i need one as well

[midnight shuttle]

I would buy one right now. But more profit for Nvidia to keep AI in data centers. 😦

[honest flint]

can get a macbook pro m2 with 96GB of unified memory, you can shove huge ml models in there

[native moon]

i boght my 3090 for 650 😈

[sharp robin]

used 3090's go for that sometimes

[native moon]

man i wish gpus had ram slots like cpus

ok with cofyui i dont have the ram proplem

[solemn scroll]

Do you know if haveIbeentrained opt-outs have been eliminated from SDXL training? It would mean we just lost 1.4 billion training images out of 5 billion :/

[echo pine]

Is 1 minute per iteration normal for sdxl lora training with bmaltais/kohya_ss on a 3090 at batch 8?

[nimble heart]

It's improving but it's not there yet. On the AMD side they have a script that runs through C code and just replaces all cuda calls into hip versions and it works well enough for some things. used for pytorch and bitsandbytes. No idea what Intel is doing

[midnight shuttle]

In 10 years either personal AI will be illegal or all computers will have fast AI capabilities.

[static prawn]

still think using artroom ai is a really good option, for sdxl 😊 if people want something easier than comfy, its running really fast

[nimble heart]

they used to

[zealous horizon]

4090 is very good. Haven't had any issues and generations are usually extremely fast. Even upscaling is relatively fast (I've generated several 4k ultrawide wallpapers in under 1min with upscale fix in auto1111)

[nimble heart]

I'm talking new. The A780 had a 16 gig version for pretty cheap once the prices came down.

[native moon]

no

[nimble heart]

yes.

[cinder aspen]

i know only about 80 millions of opted-out images

[zealous horizon]

512 gens on SD 1.5 euler a 7/20 usually take 0.5 sec

[nimble heart]

back in the days of megabytes instead of gigabytes pretty sure some 3d accelerators had expandable memory

[cinder aspen]

did that rise to 1.4 billion?

[native moon]

but not now

[nimble heart]

I know

[solemn scroll]

https://haveibeentrained.com/

[nimble heart]

but you said "no" to me saying "they used to"

So now I'm being petty

[native moon]

oh

[solemn scroll]

Maybe bots, maybe lobbyists abusing a loophole, maybe propaganda, who knows

[nimble heart]

also this just happend in ComfyUI and idk wtf caused it

[cinder aspen]

im not sure on the number tho. Is that official? Because there was a deadline for opting-out as well

[midnight shuttle]

Faulty VRAM? That is a very odd pattern.

[cinder aspen]

he became criminal and is now imprisoned in the AIcatraz

[honest flint]

i started seeing stuff like that right before my last video card died

[native moon]

is it possibel to open the comfui so i can use it in the same wifi on a different device

[inner maple]

What exactly does the refiner model do in sdxl? Is it kind of like highres fix?

[solemn scroll]

I don't know, haven't seen SAI say anything on that subject

[native moon]

it refines the image

[nimble heart]

My card is like 5 months old so hopefully not faulty

Only that one image

[midnight shuttle]

It's a model trained to improve existing images instead of generate new images.

[nimble heart]

also weird it'd be baked into the image. Usually lines are on your monitor post-display

[cinder aspen]

if they took their promise serious then they did it, would be better to do so considering what is going on right now

[inner maple]

Ok thanks. My sdxl results are pretty good so far without it. Hopefully automatic1111 includes support for it soon.

[native moon]

how can i make that i can use comf ui on a differnt pc like in the same network?

[midnight shuttle]

I get pretty good results without it. You can use it in A1111 but it's inconvenient. Send to img2img, change to refiner model, denoise strength 0.25.

[nimble heart]

--listen

[outer oxide]

Chess remains a difficult subject.

[nimble heart]

not even the correct way to use it

gives way worse results

[solemn scroll]

Still, a fifth of the dataset just disappeared, and I can't believe the number went from slowly rising to 80m to 1.4B all of a sudden

[midnight shuttle]

What is the correct way to use it in A1111?

[nimble heart]

There isn't one

[nimble heart]

you have to stop the sampling early and feed the still-noisy image into the refiner, which isn't implemented in any form in a1111

[cinder aspen]

maybe this counts those after the deadline as well

but you dont really need that much of images in dataset for high quality results

[midnight shuttle]

Maybe that's what will kill off A1111 then. But that's too bad because for many workflows such as inpainting it's still better.

[cinder aspen]

well the issue is that the number of images doesnt mean all of those are high quality data for the training

[nimble heart]

I just found this out but Comfy actually does have an inpainting menu, it's just in a weird spot. Not as feature-complete though

[inner maple]

You don't think they'll add it to a1111? I can't see them NOT getting it working....

[nimble heart]

a lot of the samplers don't even work properly so idk

[midnight shuttle]

I found that too but it's missing things like "inpaint masked only". But no SDXL inpainting model yet anyway so doesn't matter so much right now.

[midnight shuttle]

It's spaghetti code and the author writes the code for free. So who knows?

[sharp robin]

there is no 'correct' way of using it, as long as you are happy with the results, the intended method is to stop the gen at about 80% and then finish it off with the refiner in the latent space, BUT if u want to use it in Automatic1111 u can gen an image with the base 1.0 model, then u send it into IMG2IMG with a low denoise and then choose the refiner and run it thru it, not the same results but works and does something

[urban fjord]

We need a controlnet trained to generate the best response to any given chess position.

[inner maple]

Comfy-ui is not spaghet?

[sharp robin]

100% it is spaghetti

[inner maple]

I've only used a1111, so all I know is italian food.

[sharp robin]

a1111 is like lasagna and comfy is spaghetti

[mellow tendon]

It was because of a missing or outdated "SD Ultimate Upscale" Custom node you got that error.

[nimble heart]

Beric's workflows literally look like plates of spaghetti lol

[glad fulcrum]

But better than 1.5. a lora could be made surely

[zealous horizon]

spaghetti is usually less entangled that Comfy

[midnight shuttle]

A1111 is friendly interface, spaghetti code. Comfy is spaghetti interface, friendly code.

[craggy basalt]

hello. Im trying to install SDXL 1.0 locally on Mac. But I got this error in terminal: We do not support vanilla attention in 1.12.1 anymore, as it is too expensive. Please install xformers via e.g. 'pip install xformers==0.0.16'. I read that xformers work only on NVidia cards, so mac is not supported?. Thank you.

[nimble heart]

I'd call this spaghetti

[outer oxide]

a bit better than 1.5, yeah. It sometimes manages to at least get the right number of rows and columns on a chessboard.

[native moon]

here is a good node setup https://pastebin.com/sjhP8Pcj

[zealous horizon]

looks like you spilled too much tomato

[keen phoenix]

ComfyUI is actually pretty neat but for sdxl really miss a well put controlnet to get all the power we love about sd models

[midnight shuttle]

It looks like a lot of repetitive steps in parallel. And I'm sure just like "magic words" of masterpiece and 8k in 1.5, some workflows in Comfy are just nodes people add without understanding.

[nimble heart]

yea I'm not about to install a bunch of custom nodes to see what the spaghetti actually does lol. I just dragged in some image off this discord channel and that's the workflow that was used for it

[native moon]

how can i make teh seed random in comfy ui?

[zealous horizon]

randomize

[nimble heart]

every time you click 'add to queue' it'll randomize the value

[inner maple]

That is horrific

[native moon]

i dont see this in my node setup

[solemn scroll]

March 3rd so we're good, only 80m max. It's still weird how this number went from this to 1.4B in 4 months

[sharp robin]

yeah tell me about it, i usually keep adding nodes then clean them and next thing i know it is full of nodes again

[keen phoenix]

i find that 8 cfg is actually pretty good

[zealous horizon]

it doesn't have to be complicated if all you need is generate images (with LoRA)

[nimble heart]

The really fat node with all the options built in I wrote myself, you can get it here:
https://github.com/Beinsezii/bsz-cui-extras

[high skiff]

Alright, The seller of my first 3090 is starting to give me problems again, and I know how staff members from hardware swap telling me that they think that he's being sketchy, and that he's in the wrong, so I've blocked him and I'm getting ready to take this to full on PayPal court

So tired of all of this

My new 3090 should get here on Monday though, which is reliving

*relieving

[bright sparrow]

A Hong Kong policeman, Korean, handsome, lying in an abandoned factory building, young and handsome, with photos, closed eyes, black military boots on his feet and blood on his body,

[native moon]

witch one

[nimble heart]

ik, my workflow looks like this. the spaghet was from a random image I found in this channel

[lilac raven]

Is touch screen support planned?

[nimble heart]

on my repo it's custom_nodes/bsz-principled-sdxl.py
put it in your custom_nodes folder in comfyui.

[cinder aspen]

apparently those 1.4 billions are for AI organization alltogether, this is not for SD only

[keen phoenix]

thanks man, will be useful

[native moon]

why a py? a json shuld work

[nimble heart]

it's a custom node

with custom code

not a workflow

[solemn scroll]

I think those also helped in the process of banishing images from training, especially given how much shutterstock must weight

[native moon]

oh

[cinder aspen]

yeah

[mellow tendon]

Lots of variety of monsters from that prompt, I like it.

[cinder aspen]

for Shutterstock and their AI it doesnt really matter

just like Getty

[sharp robin]

dis mine

[cinder aspen]

well Getty did sue SAI ofc

[nimble heart]

I think you should add more stages

[cinder aspen]

but they dont have to worry about their AI

and dataset for training of AI

same with Adobe

and Nvidia

[keen phoenix]

do you guys have a good upscale method ?

[sharp robin]

gibs good images and i can upscale p high

[keen phoenix]

i just use ultra sharp works fine but i think it could get better

[nimble heart]

there's two main ways right now

1. render a full image @ normal res, pixel scale, use refiner @ like 0.1 denoise to add details back
2. latent upscale a partially complete image, then do base + refiner @ like 0.65 denoise
   my node can do both, which is preferable depends on what you're doing. Sytan's making a workflow for 1) with some extra goodies using only base nodes, idk when he's gonna release it.

[soft bone]

batch size?

[keen phoenix]

ok i will dive into that

does someone know how far can you go with 3070

[half ivy]

I need more detail in my faces, eyes all seem blurry

[nimble heart]

make sure you use the SDXL versions of the CLIP nodes since you can adjust size targets

[half ivy]

End up taking it Auto1 I paint with face detailer

[nimble heart]

refiner settings?

[half ivy]

20/80

[native moon]

ok i like comfy ui morethen i thoght but its still missing some stuff but its cool

[nimble heart]

20 base 80 refiner?

[inner maple]

What is your prompt for the creature? "Monster" ?

[half ivy]

Like look at his right eye, just doesn’t seem right even from a distance

[half ivy]

Lol other way

[nimble heart]

gotcha gotcah

[half ivy]

My bad 😅

[nimble heart]

what samplers and are you doing noise return or does the refiner add its own noise?

[keen phoenix]

you see it quite obviously when you use too much refiner, you'll get other faces pop on the image

or maybe that's a me problem

[half ivy]

Add noise on base not on ref
EulerA
Return noise on base not on ref

Ill post an img in a bit when I get back, gonna keep trying to figure out this facedetailer node today

[shy kelp]

Question is the refiner absolutely necessary? cause I don't see much of a difference

[astral jay]

imo it's kinda better to just use hi.res fix

[shy kelp]

okay

[heady vale]

in many cases refiner isnt necessary

[astral jay]

Although I guess if you were to train the refiner that'd be interesting

[shy kelp]

ahh training the refiner is an interesting idea

[keen phoenix]

wait that's an interesting idea

[heady vale]

Ive see worse output after refiner

[astral jay]

Too rich for my blood

[nimble heart]

Any examples?

[keen phoenix]

for mine too but someone has to do that

[heady vale]

not on hand. I deleted them

[vast narwhal]

Is @visual glade Online?

[hushed hornet]

♪ "Pizza Horse, where the fun never strays,
Gallop on over for our equine braes.
A mane of cheese, a pepperoni horse course,
At Pizza Horse, we're a different pizza force!" ♪

[shy kelp]

im gonna try merging a xenomorph with a crocodile to see what i get

[half ivy]

Baby zuck nirvana cover 🙈

[inner maple]

I was reading that you can feed sdxl two different prompts and it merges them. is that possible to do in automatic1111?

[keen phoenix]

is it possible to train Loras with old dataset ?

because we can train it with any kind of image ratio now

[deft coral]

lol

[wet raven]

Automatic1111 SD 1.5 I can do 1280x1280 images.

However, SDXL I can only do 512x512

If I go higher, it generates the image, but right before it saves the image, I get an error

OutOfMemoryError: CUDA out of memory. Tried to allocate 1.56 GiB (GPU 0; 11.00 GiB total capacity; 8.68 GiB already allocated; 0 bytes free; 9.31 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

[deft coral]

i tried a nude shot on SDXL it didn't work

[midnight shuttle]

If using A1111, you will need to upgrade to the newer ("slow)") Nvidia driver. A1111 with SDXL requires almost 16 GB VRAM during the VAE step.

[keen phoenix]

i think it's because you don't have enough vram

[inner maple]

I can generate 1280x1280 with a 4090. no errors

It's fairly quick too at 30-40 steps

[midnight shuttle]

You have 24 GB VRAM.

[half ivy]

Im doing 1024 with a 1650ti no prob

[inner maple]

I have tried going higher and then it does generate errors.

[keen phoenix]

1024x1024 under a minute with 3070 8GB

[wet raven]

Mine is 1080 TI FTW 3 (11gb vram) with 64gb ram

[inner maple]

I hope the next nvidia cards have 48GB of ram. I'm ready.

[snow iron]

comfy ui uses way less vram atm atleast

[half ivy]

Just blurry’ish eyes as discussed above 👆🥹

[orchid oar]

im doing 1024 comfyui on 3070 fast

[shy kelp]

Behold, a xenomorph, crocodile hybrid

sdxl 1.0 base model

[wet raven]

hmm.. I'm using a previous version because when i updated, it didn't let me control teh 3 gpu fans. I don't know if the latest one fixed that or if the newest one is still the broken fan controls

[midnight shuttle]

11 GB is not enough for SDXL with A1111 unless you use the new Nvidia driver for unified memory. And maybe that doesn't work on 1080.

ComfyUI uses less VRAM somehow.

[half ivy]

It’s magical

[shy kelp]

I'm using 502.31 driver myself

[wet raven]

I'm using version 531.41 on 3/23/2023
latest driver available 536.67 on 7/18/2023
i'll have to see if there is an in-between version. if so, then i'll try the latest version

[shy kelp]

using the 502.31 driver and the image generation took like 30 seconds

I'm sorry I meant 532.03

any driver 535 and up is having problems with SD.

Nvidia is working on it

[keen phoenix]

when it first came out i was generating in 7 min and after that i updated my driver and it worked fine

[wet raven]

sheesh

[inner maple]

What problems? My driver version is higher than that and it seems fine

I'm on 536.67

[keen phoenix]

same i don't have any issue

[wet raven]

I think they are saying for older cards

[midnight shuttle]

It is not having a problem. It is working as designed with a new feature. But because SD A1111 uses so much VRAM, it causes swapping to RAM.

[nimble heart]

15% refiner vs 0% refiner. I've had almost no images where the refiner made it worse when properly applied.

[inner maple]

Ahhhh yeah I did notice the huge ram use

[shy kelp]

Regardless, Nvidia is working on a fix in an upcoming driver

[hushed hornet]

♪ "Clip clop, trot to Sushi Horse, take the reins,
Feast your eyes on sushi mains!
In every roll, a horse delight,
Savor the gallop, each sushi bite!" ♪

[wet raven]

I wish Nvidia would stop working on video game drivers and just focus on AI drivers, haha.

[midnight shuttle]

It's the same thing.

I don't know how they will "fix" it other than to allow you to turn off RAM swapping. Then everyone will complain about CUDA out of memory errors.

[shy kelp]

Is that the actual prompt used? lol

[wet raven]

I would rather play games in low resolutions with low graphics settings at high framerate and low GPU usage, if i can render AI stuff faster

[wet raven]

i'll have to check the settings. I have comfyui but i like a1111

[midnight shuttle]

You don't have to choose. GPU is GPU.

[wet raven]

i'm just saying, that's my priority. I used to be 100% about gaming graphics cards.
now i'm 99% about AI stuff

[half ivy]

I haven’t touched gaming since SD came out and I used to game daily 😂

[midnight shuttle]

Best way to get a lot of VRAM right now is spend $0.36 to $1.50 per hour to rent a GPU.

[heady vale]

I generated this in A1111 in some of my first image that I remember where I was playing around. This was base only. and I tried refining it and it got worse but I deleted that image.

[nimble heart]

a1111 doesn't properly use refiner

[heady vale]

that could have been the issue. but I think you can agree its pretty good for just base

[orchid oar]

if you want to use refiner in img2img make sure denoising is turned down to 0.25 or maybe even lower

[nimble heart]

yea the base is good

[half ivy]

A switch would be cool for comfy

[midnight shuttle]

You mean a refiner switch?

[half ivy]

I envisage my entire “possible workflow” all connected on one master workflow, and then its just a matter of switching on which work flow you want to work with

[midnight shuttle]

Oh I see what you mean. But that's sort of like saying you want one software program that is all programs and you choose which to use, isn't it?

[half ivy]

Just on off switch

[hushed hornet]

♪ "At Spaghetti Horse, our pasta takes the lead,
A stallion of flavor in every feed!
Gallop on over for a Prego prance,
Or a Carbonara cantor in Italian romance.

Mane courses brimming with parmigiano,
Our marinara races, never slows to a canter.
In this paddock of pasta, you'll find amore,
Spaghetti Horse, our Italian folklore!" ♪

[half ivy]

In a round about way 😂

Can pretty much do the same with the reroute util and just disconnect the model to each path I guess

[orchid oar]

I prefere Ramen

[cursive jolt]

I made a cat... he cute

[midnight shuttle]

That pasta tastes like ChatGPT. 😆

[wet raven]

"These pastas are making me thirsty" - george costansa (sienfeld)

[hushed hornet]

Please never stop with the grocery store frogs. We love them.

[orchid oar]

I won't ever stop

[wet raven]

I'm generated many images of "toad in grocery store"

[shy kelp]

xenocat?

[orchid oar]

lmao, can i see one?

[heavy yarrow]

has anyone figured out how to use a LORA with SDXL in comfyui? i didn't get any luck with the lora example in the comfyui repo

[wet raven]

not SDXL, but ok

[orchid oar]

doesnt matter

[nimble heart]

base only, 15% refiner, pixel upscale + 15% refiner, latent upscale + base denoise + 15% refiner.
surprised how good just a pixel upscale then a smidge extra refining can do

[half ivy]

She’s so cool 😎

[wet raven]

first some toad/frog + turtle hybrids

[orchid oar]

that looks so good

[cursive jolt]

I looked at the first one and was like "this is a real animal what u talkin about"

[hushed hornet]

♪ "Saddle up, folks, to Horsin' Around Fried Chicken,
Where the cluck meets the clop, it's finger lickin'.
Barnyard of flavors, from the coop to the stall,
Galloping goodness that’ll enthrall!

Trot on down, for a bucket that's winner,
In our chicken derby, you're always a grinner.
Hay, there! We're the funniest farm pickin',
Horsin' Around Fried Chicken, it's not just chicken!" ♪

[half ivy]

Frotle

[orchid oar]

yea looks real

[ornate elk]

I am making images of spiders, and discovering that SDXL has an even worse understanding of what spiders look like than 1.4 or 1.5. How would I best go about training it to better understand spiders? I have a 16gb GPU. WARNING: spider--things

[shy kelp]

Xeno-Godzilla attacking a city........

[midnight shuttle]

I've used hypernetworks for that type of thing but it looks like hypernetwork training doesn't work yet (ever?) in SDXL. So maybe a Lora instead.

[heavy yarrow]

hey there, I used that example to use a LORA I trained with SDXL and it is not working. I'm confused, because I confirmed that the LORA worked using the diffusers repo directly. Any Suggestions?

[lilac raven]

The telephones are finally making sense

[ornate elk]

what's a good guide for making a LoRA?

[midnight shuttle]

I haven't made one yet but there are hundreds of guides. I will have to learn since hypernets seem to be dead.

[shy kelp]

Maybe trained models will fix that eventually?

[ornate elk]

yeah that's why I ask. Could pick randomly but looking for the community temperature

[orchid oar]

They look sooo good, well done

The drinking ones are funny

[half ivy]

Why am I into frogs suddenly 😂

[boreal bough]

@wicked frigate
Is there no way to input sampler choice, when I load up any custom comfyui setup?
ksampler nor advanced ksampler pass on the sampler choice 😦
which is a bit of an issue for testing loras

[ornate elk]

got a link to any docs on this?

[half ivy]

The guy on the bike looks so happy

[orchid oar]

big frog

These pics are so creative

He ate too much :(

[shy kelp]

This one looks more like a teenage Godzilla if you ask me.

[shy kelp]

Can I get the prompts to this one please?

[nocturne dove]

I was going to recommend the old videos here. But I see that there is a newer sdxl video on the list already. I haven't watched that one (it's not by the same person as the old ones). But the older videos are how I originally learned to do the whole process correctly. https://github.com/bmaltais/kohya_ss/#tutorials

[orchid oar]

Are u using a1111 or Comfy?

[shy kelp]

a1111

[wet raven]

i've probably generated 10,000+ images

[orchid oar]

u can download the image and open it in png info if you want

[ornate elk]

this happened to me when I forgot to switch back to the main model after using the refiner

[orchid oar]

10000????

Damn

[orchid oar]

if not i can send you them

[native moon]

yes this happened xD

[wet raven]

I started back when all you could do is go to websites and discords like LAION

and I even started back in Google Deep Dream

[orchid oar]

Damn, your are an expert at creating frogs

[ornate elk]

for sxdl or for all of sd? because I was doing like 100k a week around september last year

[orchid oar]

i dont even know that one

[native moon]

one of the ogs

[wet raven]

i'm only sharing a super ultra tiny fraction of images i made

[orchid oar]

Yea, i think sharing 10000 is a bit complicated, but your pics look awesome

[wet raven]

8 years ago

[orchid oar]

i didnt even have a pc back then

[native moon]

2015 will forever be 5 y ago!!!!

[inner maple]

Scorn?

[static prawn]

@wet raven finally someone uses GTX 1xxx series 😄

[ornate elk]

?

[native moon]

its ispiered by a artist and scoren also was inpiered by the same artist

he is called beksinski

[shy kelp]

Wait seriously the only prompt was "frog at the gym"?

[ornate elk]

oh yeh he's in my prompt

[orchid oar]

Yes lmao

[wet raven]

"toad mascots" with a drink on a golf course

[orchid oar]

I got lucky

or something

[shy kelp]

Watch me try to do something similar and I get chicken scratch

[orchid oar]

Very creative

[static prawn]

@wet raven are u using comfy?

[orchid oar]

i wish you good luck (i use the refiner model)

[shy kelp]

Fine....dog at the art museum. let's see how good 1.0 is

[wet raven]

these images were made with many different things

[static prawn]

but its sdxl? @wet raven

[orchid oar]

Bro doesnt know that the Tv is not even on 😔

[wet raven]

I prefer Automatic1111, but i have ComfyUI, NMKD, InvokeAI and many others

[native moon]

made with jsut the refiner xD

[orchid oar]

lmao

[wet raven]

some might be, i should have posted in another channel, sorry

[midnight shuttle]

Simple prompts often work well. But then people started adding "magic words" such as masterpiece, dslr, 8k, best quality. Without testing to see if those magic words really worked.

[static prawn]

nono dont worry , was just curious and confused

[wet raven]

I get a bit excited about sharing AI images

[ornate elk]

camera words can often protect an image from certain effects

[orchid oar]

"I have no enemies"

[heady vale]

pretty sure SAI said "masterpiece" wasnt required with XL

[midnight shuttle]

Yes but often those words are added without testing. Only added as magic words.

[ornate elk]

i generally use nikon, canon, leica, specific f-stop and lens lengths to help it along, if I am doing something like that

[shy kelp]

Dog at an art museum

[native moon]

what is he doing

[nocturne dove]

Nothing is "required". But you can fix your seed and see what changing a few words around does.

[orchid oar]

😂

[static prawn]

frogs also have their needs @native moon

😂

[native moon]

😭

[heady vale]

reps

[orchid oar]

Wrong one

[shy kelp]

I think the frog has seen his best days behind him

[ornate elk]

more spider

[native moon]

ewwww

[orchid oar]

Looks good

[ornate elk]

spider fren

[shy kelp]

that girl ain't happy. I know i wouldn't be

[hoary saddle]

Wow, that's nice. Thanks for note node too!

[orchid oar]

Im Sorry

[orchid oar]

Froggy doesn't steal

[static prawn]

this was a fail i kinda liked

[native moon]

interesting

[static prawn]

haha

i mainly do horror stuff, but got way better results on sd 1.5

[shy kelp]

Been getting real busy at the office these days.

[orchid oar]

Looking handsome

[shy kelp]

Fido's the man! 😄

[orchid oar]

(He has no license)

[static prawn]

really want to use sdxl too 😦

[native moon]

u can use it on clipdroop

[orchid oar]

Why can't u?

[static prawn]

i know but more in a good innovative way , like i did with 1.5

[native moon]

who wins?

[orchid oar]

zuck

[static prawn]

just reaching my limits with the hardware and dont really like comfy 😅

[orchid oar]

can i send pics with blood with a spoiler warning?

[winged bear]

I cant get sdxl to make anything other than black images in automatic 1111. 1.5 works fine I was hoping someone here might have a solution to my problem. I have tried with my 3060 and 4070ti here is my batch script and all extensions are disabled @echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS= --api --no-half-vae --disable-nan-check --xformers
git pull
set CUDA_VISIBLE_DEVICES=0
call webui.bat

[ornate elk]

idk

[native moon]

comfy is okey and its fast to setup

[orchid oar]

If its bec of comfy just use a example one, for text2img you dont really need to change anything. Hardware wise i think sdxl is better than sd 1.5

[ornate elk]

I just know some people hate spiders and these images are kinda disturbing

[hushed hornet]

♪ "Rise and shine, at Breakfast Horse we're neigh-bors,
Serving sunny-side up, where flavor gallops and roars!
Clop on in for waffles, dressed in mare-ple delight,
Or canter for our 'Horse Oats'—a morning's first light.

Pancakes prance on your plate, as roosters perform their call,
Our scrambled eggs and bacon canter, at the break of stall.
Bridle your hunger with our Morning Mane Muffin,
At Breakfast Horse, you'll always leave with somethin'!" ♪

[midnight shuttle]

Remove --disable-nan-check and see if you get an error about a tensor with alll NaN.

[static prawn]

wanna use a good workflow but i cant even get all those stuff on 1 monitor screen, dunno it feels overwhelming

[shy kelp]

That will happen in time. These are early days for 1.0

[winged bear]

tried that as well same result just black images

[orchid oar]

Understandable, was thinking of sending a picture i created of zombies, but im not sure

[winged bear]

also no errors just black image

[orchid oar]

comfy has examples u can use

[midnight shuttle]

But no error about all NaN? Then it's something different from what most people see.

[half ivy]

Anyone else feel powerless without control net 😅😅😅

[vast thistle]

?

No

[midnight shuttle]

Make sure you are using the right VAE.

[orchid oar]

im still entertained by the sdxl model without

[winged bear]

no errors To create a public link, set share=True in launch().
Startup time: 14.6s (import torch: 3.0s, import gradio: 0.6s, import ldm: 0.4s, other imports: 3.5s, load SD checkpoint: 5.0s, create ui: 1.7s, gradio launch: 0.2s).
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:09<00:00, 2.05it/s]
Total progress: 100%|██████████████████████████████████████████████████████████████████| 20/20 [00:14<00:00, 1.42it/s]
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:08<00:00, 2.32it/s]
Total progress: 100%|██████████████████████████████████████████████████████████████████| 20/20 [00:13<00:00, 1.43it/s]
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:16<00:00, 1.24it/s]
Total progress: 100%|██████████████████████████████████████████████████████████████████| 20/20 [00:21<00:00, 1.10s/it]
Total progress: 100%|██████████████████████████████████████████████████████████████████| 20/20 [00:21<00:00, 1.27it/s]

[sullen linden]

So when I try to load the SDXL model I keep gettin Failed to create model quickly; will retry using slow method. and then my computer turns to glue and I have to kill a1111.... any ideas?

my 1.5 models, even 8 gig models, load fine

[static prawn]

i got a really nice workflow but its just too much on one screen and feels just uncomfortable to use 😅

[winged bear]

yep checked vae set to automatic

[midnight shuttle]

Try None instead.

[static prawn]

but with the GTX 1070 its anyway a real struggle

[winged bear]

good idea on it

[static prawn]

for 1.5 it was fine

[orchid oar]

understandable

i hope for u that there will be a1111 support fast

[half ivy]

Could be version of PyTorch 🤷‍♂️

[winged bear]

set vae to none still black images

[static prawn]

kinda wish KI would utilize more the CPU, would be more a upgrade i would do than my gpu 😂

[midnight shuttle]

No idea then. Hope someone else can help.

[shy kelp]

I keep getting the fake version of stuff while you all apparently got the real thing. Your pictures seem to have a lot of detail except for me.

[orchid oar]

Blood warning, Cartoon (Dont click if you cant see blood or Zombies)

[half ivy]

Thats not bad

[winged bear]

I might try on another computer and see if its something dumb with this one. I have a 3080m in a laptop that should work

[orchid oar]

ok, i was not sure

[static prawn]

might just get a used 3060 soon, will make everything easier i guess

guess thats the best upgrade for me

[half ivy]

I’ve seen worse from sawX advertising

[orchid oar]

thx for the super reaction <3

[midnight shuttle]

3060 is enough for SDXL on Comfy. But not enough for A1111. It will be slow on VAE step.

[orchid oar]

Yes, my 3070 is very slow on a1111 but on comfy its super fast

[midnight shuttle]

Because of VRAM. A1111 needs 16 GB VRAM for SDXL.

[true hazel]

has much changed from 0.9 to 1.0?

[orchid oar]

Yea sadly

[orchid oar]

i think the training was made easier

or something like that

[shy kelp]

so that's why comfy has suddenly become so popular?

[static prawn]

im pretty sure they might improve it in the future

[midnight shuttle]

I hope so. Also hope SAI will release official FP16 VAE for SDXL.

[orchid oar]

a1111 doesnt have proper support for sdxl yet

[midnight shuttle]

And inpainting SDXL model.

[static prawn]

dunno AI is not my main usecase, and i dont wanna spend to much money for it

[orchid oar]

and now that i used it once its so much faster

[static prawn]

and for gaming my 1070 is totally enough atm

im always in cpu limit in games, so i just wanna do a small upgrade for AI Stuff

[orchid oar]

Tried "Duecki" on the license plate, didnt work quite well

[ornate elk]

I also managed to get a realistic spider crawling on a naked girl, but it took unreasonable effort and inpainting

[wet raven]

Mixing Xenomorph alien with Predator gives some really creative outputs for new ideas for monsters.

[static prawn]

do u think it would be possible in the future KI Image Generation gets more CPU based?

[orchid oar]

I can imagine the work, nice👍

[midnight shuttle]

No, but I think CPU will eventually have AI optimized features.

CPU is good for doing lots of things complex in a row. GPU is good for doing lots of simple things all at once. AI is lots of simple things all at once.

[orchid oar]

did you try running a1111 on cpu (if thats possible)

[midnight shuttle]

SD on CPU is possible. 10 minutes for one image. Same image takes 3 seconds on GPU.

[static prawn]

it takes hours, i was just interested if its possible in the future, its getting faster for cpus

[orchid oar]

I dont think so, your ram will be used and ram is very slow

[static prawn]

A cpu upgrade would be more useful for me in general 😂, than GPU haha.

[orchid oar]

U cant compare the speed of ddr4 ram and vram

[midnight shuttle]

Also because CPU is 4-12 cores and GPU is 3000+ cores.

[wet raven]

I want GPUs to end up being basically their own computers, with upgradable RAM, etc, However, Nvidia would rather you just keep buying new ones.

It's crazy that they take high end cards, and laser cut damage them, so that they are slower, to sell them at a lower cost, because it is cheaper for them to do that, than to have another manufacturing processing line and inventory to actually start with cheaper hardware.

However, since the 20 series cards, they have boosted their price, so maybe they are trying to recover the cost of doing that, or maybe they don't do that anymore.

[shy kelp]

you mean even with 1.5.1 ?

[half ivy]

Noice 🔥

[ornate elk]

yeah it was mostly to see if it was possible. it is, but it doesn't like drawing nipples OR spiders so it needed a lot of help

[winged bear]

technically you can add more ram if your savy and have the right equipment

[midnight shuttle]

AI is multiply weights and add bias millions of times. Easy for GPU. Slow for CPU.

[orchid oar]

The Vram needs to be as near as possible to the chip, im not sure if that will impact performance. But still nividias prices are really bad

[midnight shuttle]

You can add more RAM to make the number bigger on the screen but the GPU can't access the RAM. I saw the same video.

[wet raven]

yeah, but it won't work as intended, I saw a post about someone who did.

[static prawn]

im pretty sure intel will get stronge the next years, im allready really impressed by the arc series

[winged bear]

oh really i must have not finished it looked like he got it working

[orchid oar]

Spider warning

[wet raven]

it "worked" but didn't "work"

[orchid oar]

@ornate elk tried it myself, looks terrifying just as yours

[ornate elk]

😦 yeah, I now have like 1000 failed spider images

like 1 in 50 looks ok

[wet raven]

weird, SD XL is letting me do 1024x1024 now. maybe because I selected "Apply color correction to img2img results to match original colors." ? that wouldn't make sense.
or maybe because i remove restore faces, or restarted 10 times, idk, who knows.

[midnight shuttle]

It was classic Youtube garbage. On the screen it showed 24 GB, but the GPU still can only use 8 GB. It was only clickbait.

[ornate elk]

and it can't do certain angles at all

[static prawn]

ok its 4:16 am here, really need to sleep. wish u guys a good night, see ya

[orchid oar]

Yea thats the shared Vram. Your Gpu can use Ram as Vram but its way slower

[inner maple]

When you use sdxl you really see what's possible. wow. Now we're playing with power.

[wet raven]

I love bugs, so i was generating bugs, then with the spider "spoiler" I remembered that many people do not like bugs.

[orchid oar]

i need to go off sometime soon too, bye

[wet raven]

SDXL 1024x1024

[native moon]

by

[static prawn]

german?

[orchid oar]

Yes indeed

[static prawn]

me too 😅

[winged bear]

so i am getting that error now rebuilt sd in a new folder modules.devices.NansException: A tensor with all NaNs was produced in Unet. This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try setting the "Upcast cross attention layer to float32" option in Settings > Stable Diffusion or using the --no-half commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check.

[midnight shuttle]

No, what I mean is the system info showed the card was 24 GB. But the extra VRAM was not useful to the GPU. It was only a clickbait video.

[orchid oar]

Nice👍

[orchid oar]

Oh okay

[midnight shuttle]

Very strange but try --no-half. You are getting NaN in Unet. Not even in VAE yet.

[soft bone]

when the only good sdxl lora tutorial is 3hr training time 😔

[inner maple]

I can't wait until a few months civit.ai is flooded with sdxl models and loras.

[winged bear]

its super weird makes black images on the other install and this on a fresh install on a new folder in a new drive

[inner maple]

Do you guys use the civit.ai extension for automatic1111?

[midnight shuttle]

Fresh install with new venv and no extensions? Very strange. Try --no-half.

[winged bear]

trying that now

[orchid oar]

Im not sure what you mean, but SD cant use the shared Vram i think, so if you have 24gb vram together and only 8gb dedicated, sd can only use 8gb

[orchid oar]

if i use a1111 then yes

Its pretty handy for the Activation Words

[hearty ginkgo]

are you guys using the lora from sdxl 1 to?

[orchid oar]

Im not

[midnight shuttle]

All Nvidia cards can use shared RAM with the latest driver and SD can use it. But if you solder more RAM on a GPU like in the video it will show 24 GB in Windows but remains only 8 GB usable. Clickbait video.

[wet raven]

Does anyone understand why I can render an image in 1280x1280 but then 1 second before it saves the image, it gives and error about "not enough VRAM" and deletes the image?

Is there a way to save the 99.99% image?

[midnight shuttle]

Solder VRAM on GPU video = clickbait. Simple as that.

[shy kelp]

what happens if you go lower than 1024x1024? like 512x768

[native moon]

it looks very bad

[orchid oar]

Ahhh i didnt know it was about soldering vram lol, my bad

[midnight shuttle]

No because the image doesn't exist until it is complete. It is running out of VRAM when converting latent space to RGB with the VAE.

[winged bear]

same error no extensions

[wet raven]

but the preview exists

[soft bone]

So apparently training lora with batch size 1 gives the highest quality but then you're leaving a lot of overhead vram and taking way longer than necessary. I wonder what will come of that

[orchid oar]

512 vs 1024

[midnight shuttle]

The preview you see is an image produced approximately by a neural net based on the latent space.

[winged bear]

return func(*args, **kwargs)
File "D:\SDXL\stable-diffusion-webui\repositories\k-diffusion\k_diffusion\sampling.py", line 145, in sample_euler_ancestral
denoised = model(x, sigmas[i] * s_in, **extra_args)
File "D:\SDXL\stable-diffusion-webui\venv\lib\site-packages\torch\nn\modules\module.py", line 1501, in _call_impl
return forward_call(*args, **kwargs)
File "D:\SDXL\stable-diffusion-webui\modules\sd_samplers_kdiffusion.py", line 215, in forward
devices.test_for_nans(x_out, "unet")
File "D:\SDXL\stable-diffusion-webui\modules\devices.py", line 155, in test_for_nans
raise NansException(message)
modules.devices.NansException: A tensor with all NaNs was produced in Unet. Use --disable-nan-check commandline argument to disable this check.

[crystal bone]

Is there a FAQ somewhere for SDXL troubleshooting ? Having the py.torch error that a lot of people seem to have.

[midnight shuttle]

No idea. I hope someone else can help.

[native moon]

it was trained on 1024 so it looks better in 1024

[winged bear]

yeah me 2 this is weird

[orchid oar]

yes

[midnight shuttle]

Post the exact error.

[shy kelp]

confirmed, stuff does look worse at lower resolutions and it seems to need 1024. But gonna try 1024x1536

[orchid oar]

sorry i answerd to the wrong message, meant to answer to olternaut

[midnight shuttle]

Use one of the supported SDXL resolutions. There are about 10 of them.

[orchid oar]

For SDXL, it is recommended to use trained values listed below:

• 1024 x 1024
• 1152 x 896
• 896 x 1152
• 1216 x 832
• 832 x 1216
• 1344 x 768
• 768 x 1344
• 1536 x 640
• 640 x 1536

These are recommended resolutions

[astral jay]

Thanks, you beautiful bastard!

[crystal bone]

@midnight shuttle

[shy kelp]

i need to write those down

[orchid oar]

Copied them from comfy

[midnight shuttle]

I've never seen that one before. What version of xformers, pytorch, A1111? (Answer all)

[crystal bone]

How would I check that (sorry for ignorance) ?

this is at the bottom of my a1111

[midnight shuttle]

Version is much too old for SDXL. Upgrade A1111.

Might be easier to do a fresh install because it is so old.

[crystal bone]

I thought the git pull would do it on its own

[orchid oar]

open the folder in explorer, enter cmd at the top where your path is and type in cmd git pull

[orchid oar]

yes it should, where did you put the command?

[crystal bone]

@echo off

set PYTHON="C:\Users\Heidi\AppData\Local\Programs\Python\Python310\python.exe"
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=--xformers
git pull
call webui.bat

in the webui-user.bat

[orchid oar]

thats right

[crystal bone]

I followed some tutorials few months back but havent touched anything since.

[wet raven]

I had to go to releases and overwrite all the files to get SD 1.5.1 to work.
Git Pull didn't work, I tried all kinds of extra commands, like stash and pop and stuff.
I also have to use a 1.x Torch and 1.17 dev version of xformers.
If i go to torch 2.0 and xformers 2.0 it doesn't work, i get a popup error.

[crystal bone]

so i should delete everything and redo?

[midnight shuttle]

Might also need to do --reinstall-torch and --reinstall-xformers (one time only).

[orchid oar]

Can you copy your cmd at the start when you open a1111?

Is there some kind of error

[midnight shuttle]

Easier to start with a new copy because that one is very old.

[crystal bone]

yeah there is actually

that must be the issue

[shy kelp]

well I just did a separate install of a1111 1.5.1 made sense really

[native moon]

maby keep the downloded model weights so u dont have to redownlode tehm

[crystal bone]

open the folder in explorer, enter cmd at the top where your path is and type in cmd git pull

do this?

[orchid oar]

yes, but i think that wont help

its the same command, mb

[wet raven]

this is what i have to use, if that is helpful to anyone.
I can't use Torch 2.x or xformers 0.0.20

[crystal bone]

yeah doing same

[orchid oar]

i will google, wait a sec (sorry)

[crystal bone]

[midnight shuttle]

Why did you edit those files?

[crystal bone]

I did not

Or not that my knowledge at least 😄

[midnight shuttle]

Someone did. Easier to just start with a fresh install at this point.

[midnight shuttle]

Too much work to figure out what is going on with it.

[orchid oar]

try in cmd git stash push --include-untracked

[midnight shuttle]

Did you find that by google? It is not the best solution here. The installatin appears corrupted. Best to start fresh.

[orchid oar]

yes on stackoverflow, it seems like it will remove those changes so you can do git pull

[orchid oar]

If that doesnt work, fresh install

[crystal bone]

Will do

[orchid oar]

did you try git pull after that?

[crystal bone]

just did now

doesnt give me error

[orchid oar]

So it does work now?

[crystal bone]

seems to have updated at least

will restart the webui thing

[orchid oar]

ok

i wish u best luck

[midnight shuttle]

Sometimes that you find on google is not the best solution. Best to be careful with such.