✨｜sdxl

[crisp owl]

Am I not thinking right here, or does controlnet not like to connect to the Refiner ksamplers?

[azure oxide]

screenshot what the nodes look like

[crisp owl]

haha, that's gonna be fun.

[tardy ore]

Hey guys!

[shy kelp]

nice, does it work well for stylized animation / low poly characters?

[shy kelp]

what's up

[crisp owl]

They connect to the base just fine, if I mute the precon ksampler (using the refiner model), and change the steps to 0 for base, it works just fine. If anything hits the refiner from controlnet, I get the mat1 and mat2 error

[tardy ore]

Nothing much how about you?

[shy kelp]

just runnin some generations

[elfin cobalt]

I've been playing with it for a while. It seems to work well.
Though I haven't done a proper blind test.

[pure hazel]

It cannot. I do not believe it was trained specifically for in-painting.

[azure oxide]

what is inputted into the controlnet conditioning there?

[wet nacelle]

https://tenor.com/view/heart-stroke-chandramohan-heartstroke-varsham-pain-gif-12030118758376637558

[crisp owl]

Tried several iterations and they all just give the same mat1 and mat2 error.

In the screenshot it's the base SDXL text encode conditions, which I expect would not work.
So tried duplicating the apply controlnet node and using the refiner text encode conditional outputs, but also the same mat1 and mat2 error

[crisp owl]

lol i know, this one isn't tidied up yet, was waiting for COntrolnet to get these nodes introduced, then pretty it up 😆

[tardy ore]

Lol awesome

I'm looking for help with an image,

[wet nacelle]

@crisp owlYour PFP looks like a bitmoji and it pisses me off.

https://tenor.com/view/bitmoji-gif-8071394

[tardy ore]

I'm trying to find an ai art maker that I can download to make some nsfw art. If I can I'll send the UI, if anyone could help me that would be great

[crisp owl]

hahaha, not where it came from, someone designed it for me based off my prior image I had

[azure oxide]

In the screenshot it's the base SDXL text encode conditions, which I expect would not work.
this was my suspicion and what you followed it up with was my suggested fix but im unsure why it would also throw up those errors

[wet nacelle]

Makes sense. I can't stand the simple looking shit they have.

[azure oxide]

oh wait
i dont think you would even need to apply controlnet to the refiner

[crisp owl]

I guess I'm just curious if someone even has controlnet working on a refiner.

I'm mostly just testing. The results I got using the preconditioner method, then controlnet base, was pretty shotty.

but mitigating the preconditioner and just starting with the controlnet base, worked well. Just wanted to see how it would work if I had controlnet precon

[azure oxide]

ya the refiner by itself doesnt change the image much except adding details so i think controlnet isnt necessary even if it worked flawlessly together

[crisp owl]

Yeah it's the first ksampler, before base. So it wouldn't be doing much, but keeping it applied gave bad results

because the precon is kinda like laying the foundation, but then the base has to fight against that

as such

[azure oxide]

thats strange, the effects shouldnt be fighting eachother that much. imo the precon's effect on the workflow is essentially equivalent to just another controlnet with guidance start 0.0 and ending at 0.03. in my experience that's always work flawlessly with other controlnets

[crisp owl]

Yeah, I've got more tinkering to try out, just d/l controlnet today, so this is my playing with it inbetween stuff loading during my actual work lol

May end up scrapping the preconditioner method if I plan to use Controlnet, because the disconnect between the two doesn't seem to play nice for that style of image generating

precon refiner left, base pass right with controlnet. Can see what the refiner was laying the foundation for, prominant figure to the left of the frame, but the controlnet then took over and made it where the controlnet should be applying the prominant figure.

If I could get CNet on the precon (left)refiner though, could work. But I don't think it works with the refiner at this point

[azure oxide]

ya the precon method sounds to me like just a workaround of not having controlnet to do guidance start/ends in the early stages of gens. controlnet is much better at doing exactly that, especially since you can slide the strength of the controlnet around and also with the many different controlnets that you can (eventually) use

[crisp owl]

Yup that would make sense. I'll see what comes of it all, probably will just get some sort of logic setup in my nodes which would make the switch off of the precon easy. Will take some fiddling

Because that does make some pretty images

[hoary saddle]

isn't there a command switch to keep models in memory so the load times aren't so long between workflows?

[glad fulcrum]

which model?

[strong field]

He’s referring to AIT and @indigo carbon workflows

[glad fulcrum]

oh

[native knot]

Business cat thinks about the choices he's made that resulted him being where he is at this moment. Suddenly, the decision not to become an ordinary house cat weighs heavily on him.

[queen adder]

does anyone train SDXL with a 3090? im trying to figure out if my speeds are normal or kind of slow and im doing something wrong?

[native knot]

One-shot generation for the Hulk. Nailed it.

[hardy cipher]

nice. that's clean

[strong field]

That lighting is mighty fine

[elfin flare]

I like this one specifically

It wrote out "Draconic Knight"

Or tried to at least

[digital shuttle]

Hi, I just installed SDXL, I thought I could get very nice results like you but I get very ugly results.
I wrote "disney style panda zombie hyperrealistic" and I got this:

I am sorry for my creation but I think it is ugly.
What did I do wrong?

[fresh path]

What resolution is that?

[azure oxide]

I am sorry for my creation but I think it is ugly.
😂 😂 😂

[heady vale]

yep 512 res is too small

[hoary saddle]

there a way to keep models in memory with comfyui? a1111 seems to switch much quicker

[heady vale]

some models/loras should never be combined...lol

[hoary saddle]

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

[delicate kelp]

Hey guys, I'm happy to announce the release of NightVision XL Alpha 0.5.2.0 on CivitAI!

NightVision XL is a lightly trained base SDXL model, refined with community LORAs. It produces touched-up photorealistic portraits, ready-stylized for social media. With nice coherency, it avoids weird body issues and biases, offering rich deep blacks and great evening/night time scenes. Capable of both SFW and NSFW output, NightVision XL prefers simple prompts, letting the model do the heavy lifting for scene building. Easy to use and versatile, it's your new go-to tool for creativity. Check out the Civit page for a full gallery of demo images and styles, including NSFW output. I would love and appreciate you all if you could shower it with your art, fill the gallery with your wonderful creations and give me all the feedback you can, both the good and the bad! It ain't perfect (yet) but I think it's the best photography model you're going to find out there right now!

https://civitai.com/models/128607?modelVersionId=140833

[heady vale]

Looks great SCG. Thank you for making these

[hardy cipher]

I want a 4090, but I'm not good at stuff

model incoming

[delicate kelp]

I suck at art, but I'm good with making tools. it's kinda like my art I guess 🙂

[heady vale]

oh its definitely an art form to craft a model

[delicate kelp]

borrowing a prompt from the DynaVision gallery - this is NightVisionXL 0.5.2

this is from the DynaVision listing, just to compare

[hardy cipher]

any suggestions for sdxl models that would facilitate me putting together beautiful images like this?

[gloomy barn]

Looks like my neighbor and her kid

[hardy cipher]

nice

those bottom two look like the moment a small child was possessed by a demon

[gloomy barn]

demon would be scared away from that

[hardy cipher]

honestly haven't had much luck with the old vintage photo look with sdxl. I'm sure it's possible. wish I could find a lora that did the trick

[strong field]

Been waiting for a SDXL model capable of photo realism, can’t wait to test this out, thanks for the hard work.

[hardy cipher]

indeed. I do appreciate all the work put into them

need to start inbreeding some myself

[strong field]

Have you tried the SDXL prompt styler?

[hardy cipher]

yeah. maybe I need to work on emphasizing it more or something

it definitely changes the image, not denying that. but haven't got that same old analog photo feel

[spiral drift]

Hi Guys I really need a comfy workflow for achieve similar results in fooocus. I have been investigating different workflows for weeks, but none of it can beat fooocus. But fooocus cannot do img2img so we really need it

Some workflows can achieve similar results to fooocus, but those workflows sometime output garbage and I need to pick a lot to get good ones. But fooocus always give super good results. A equivalent workflow can help a lot.

[hardy cipher]

well just figure out how to do this in comfy https://github.com/lllyasviel/Fooocus#list-of-hidden-tricks

[hardy cipher]

I do enjoy the people that hate on workflows, lol

I hate on my own sometimes

[heady vale]

reading the git, it looks like he has combined the base and refiner into one ksampler process

[hardy cipher]

hmm

I skimmed it, really didn't dig too deep. I should look at that again

[crisp owl]

Yeah it's just a simplified method after reading it.
It works, but not utilizing the separate clip function to the full extent also reduces the available latent space you can migrate in.

[spiral drift]

Can someone just translate that to english that I can understand

[hardy cipher]

why don't you?

[crisp owl]

TLDR:
It will make pretty images for someone who doesn't care to really get super specific with their image generations.

[azure oxide]

now translate it into japanese

now chinese

[crisp owl]

Ooh, gotta learn that, one sec

[hardy cipher]

I demand service, cater to my whims

the plug and play pretuned services are great and all, but it almost feels like a driverless car to me

[glad fulcrum]

are the control net models working on auto1111? or only in comfyUI?

[hardy cipher]

I realize not everyone is going to be into putting it all together themselves, and that's all good. no hate on that

[hardy cipher]

https://tenor.com/view/idontknow-illtellyou-noidea-gif-5915997

[visual glade]

that foocus thing is basically just using a regular workflow with dpmpp_2m_sde + karras

[hardy cipher]

I still haven't worked out exactly how those impact things

[heady vale]

so theres no real secret sauce?

[visual glade]

the secret sauce doesn't really change anything in practice

[hardy cipher]

I still like euler a myself. might not be the best for high quality images though

heun is great if you want it to take twice as long for a boring image

[hoary saddle]

Comfy, is there a hidden debug log feature i can turn on? still getting random system locks after a few dozen renders on 2 ubuntu systems with a 3080 and a 4090 and 32g ram each

[visual glade]

basically he does 30 steps total, 20 steps base, 10 steps refiner

[hardy cipher]

that's a refiner heavy approach

not sure how people pull that off. it's never worked for me

[azure oxide]

i thought he uses only 1 ksampler and thats whats different about it, does he not actually?

[visual glade]

steps=30, cfg=7.0, sampler_name='dpmpp_2m_sde_gpu', scheduler='karras', denoise=1.0

[visual glade]

that doesn't actually change quality in practice

[azure oxide]

yeah i didnt think it would

[hardy cipher]

I mean, if you're running them in succession I don't see how it'd be different

[visual glade]

it's more mathematically "correct" and would be better at very low steps like 4 but at 30 total steps you won't see a difference

though it might even be worse in practice because it's a different model so

[hardy cipher]

masslevel, I made that node moderately useful, lol

it actually gives me all the aspect ratio/ resolutions that the stability website recommends

[upbeat summit]

yes I've seen and I've already dowlnoaded it 😄

[hardy cipher]

but then I can do 40 to 1 AR

did you read my quote from AI enthusiast Mahatma Ghandi?

I don't know, it was good practice. learned a few things.

[upbeat summit]

yeah I have (now) 😄

[visual glade]

my opinion on fooocus is that the sytan workflow is better

[azure oxide]

damn

thats some high praise @high skiff

[hardy cipher]

I just want to know what it would take to implement something like this in comfy https://github.com/mcmonkeyprojects/sd-dynamic-thresholding

[high skiff]

let me guess, winston pinged me? lol

oh

NOT winston lol

[high skiff]

[azure oxide]

Native refiner swap inside one single k-sampler. The advantage is that now the refiner model can reuse the base model's momentum (or ODE's history parameters) collected from k-sampling to achieve more coherent sampling. In Automatic1111's high-res fix and ComfyUI's node system, the base model and refiner use two independent k-samplers, which means the momentum is largely wasted, and the sampling continuity is broken.
does this have any merit to it in terms of optimization @visual glade?

[high skiff]

me when comfy praises my workflow

[upbeat summit]

Hey, comfy. It's this time of the month for me to say thank you again for all your work.

[hardy cipher]

I cannot understand why, but for some reason I can't load your workflow files. it doesn't make sense to me. I can load pretty much every other one I've tried out. it kind of bums me out

[high skiff]

Same

[high skiff]

I am not sure why, I worked hard with the restrictions of just stock nodes for like a month in order to ensure reliable compatibility

[hardy cipher]

it's been a game changer for me. I feel like I know a little less, or a lot less, than a good portion of people here, but I'm learning pretty quickly

[spiral drift]

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

[hardy cipher]

sytan, it's not that it's missing nodes. it just doesn't load. I don't get it at all

[visual glade]

maybe I'll implement it because it's technically more correct but don't expect it to change quality at all

[hardy cipher]

I guess I could manually recreate it, lol

[visual glade]

does it load if you copy paste the contents of the json?

[azure oxide]

right, i just know comfy is literally as optimized as can be compared to a1 so it was surprising to see someone come up with another workflow thats even more optimized

[hardy cipher]

might be a worthwhile endeavor tbh. I create some wacky workflows, but wouldn't necessarily call them top notch

[high skiff]

ok, ok, ok

I am about to drop my boiling hot take on SDXL after some further testing

and expect this to be in the 1.1 release of my workflow

[hardy cipher]

you know, I really should have tried that, lol. ugh

[high skiff]

I see 0 point to use the refiner anymore, and I think its hinders more than it helps in 95% of cases

[hardy cipher]

thanks

[high skiff]

[visual glade]

and you need to scroll up a bit or you might not see his workflow

[high skiff]

all of my excellent realism images are from just base

[visual glade]

oh yeah we want to get rid of the refiner eventually

[spiral drift]

let me try sytan a bit, i think i already have it

[high skiff]

refiner is overfit, and introduces weird blurry artifacts that get exacerbated through high res fix sometimes

[azure oxide]

only 1.1? searge is up to like 3.0 or so now, gotta catch up

[hardy cipher]

I can't decide if I like it or not. and I have no idea what to put in the refiner prompt

[high skiff]

does his have high res fix?

[azure oxide]

kinda funny to see how often he updated his workflow tbh

[hardy cipher]

and I don't think anyone does know

[azure oxide]

i have no idea, i just saw him constantly update it since i saw the 1.0 release

[hardy cipher]

I use the searge nodes a lot

I like his strange pipe like nodes

[high skiff]

I know nothing about him or what he does

[hardy cipher]

wow, I had no idea I could copy the actual text from the json and paste it into comfy like that. sytan workflow achieved

[high skiff]

[hardy cipher]

missing the downscale and contrast fix nodes. I could most likely substitute for those. is there an efficient method of determining where to get missing nodes?

[high skiff]

they are all stock, I am not sure why it doesn't work

when was the last time you updated comfy?

[hardy cipher]

or wait, did you just make them red like that?

[high skiff]

no?

Its an upscale by and an image blend node

both are stock

[hardy cipher]

I updated it a couple days ago, but I do have a few errors when I load

[heady vale]

sounds like your install is incomplete

[hardy cipher]

just get sick of fixing them

bah. I wonder what's causing that

[strong field]

I lower the denoise everyday

[high skiff]

I am at the point where my daily driver for SDXL is refinerless

I really see no reason for it, it messes things up more than it fixes it

and with my newer better realism prompting, as well as my realism LoRA, its completely obsolete for any form of realism gens

[midnight shuttle]

The only time I've liked the refiner output was when I was intentionally going for a very detailed and busy background. But it's such a small change I don't really need to use it.

[hardy cipher]

I guess it's debugging time

de-erroring

well isn't that nice

ERROR diffusion_model.output_blocks.4.1.transformer_blocks.0.attn1.to_v.weight shape '[640, 640]' is invalid for input of size 1638400
ERROR diffusion_model.output_blocks.4.1.transformer_blocks.0.attn1.to_out.0.weight shape '[640, 640]' is invalid for input of size 1638400
ERROR diffusion_model.output_blocks.4.1.transformer_blocks.0.ff.net.0.proj.weight shape '[5120, 640]' is invalid for input of size 13107200
ERROR diffusion_model.output_blocks.4.1.transformer_blocks.0.ff.net.2.weight shape '[640, 2560]' is invalid for input of size 6553600
ERROR diffusion_model.output_blocks.4.1.transformer_blocks.0.attn2.to_q.weight shape '[640, 640]' is invalid for input of size 1638400
ERROR diffusion_model.output_blocks.4.1.transformer_blocks.0.attn2.to_k.weight shape '[640, 2048]' is invalid for input of size 983040
ERROR diffusion_model.output_blocks.4.1.transformer_blocks.0.attn2.to_v.weight shape '[640, 2048]' is invalid for input of size 983040
ERROR diffusion_model.output_blocks.4.1.transformer_blocks.0.attn2.to_out.0.weight shape '[640, 640]' is invalid for input of size 1638400

[high skiff]

my friend is making a massive realism dataset for a finetune of SDXL, and he is gonna be sharing some images forwith me, so I will be contuning my realism Lora with much more data soon

my current tests with just 90 images are already damn promising, which is so good to see

[gloomy barn]

your weight shape is invalid for the input size

[gloomy barn]

what is the Lora?

[high skiff]

its a LoRA for realism. Its not out anywhere, I am still early in dev

unlike a lot of people who rush out incomplete LoRA's, I spend a ridiculous amount of time trying to get the small things right lol

[gloomy barn]

I see

[high skiff]

I need to get a several times bigger dataset

and I need to optimize some settings as well

[hardy cipher]

https://tenor.com/view/cry-sad-gif-25488313

[gloomy barn]

the dataset has to be very large for realism?

[glad fulcrum]

anyone got an open pose workflow for ComfyUI? These nodes are driving me crazy

Also, is there an open pose editor for it?

[spiral drift]

I still cannot get as good results as fooocus. My prompts are: Warriors fight monsters in front of a castle in the wilderness. Cinematic and chilling
Negative is: anime, cartoon

I am using sytan. This is the best on after 10 attempts

[strong field]

I’ve been struggling with this too, I didn’t have the wherewithal to remove refiner cause I thought I was just using it wrong. But it’s been lower steps, lower denoise every day for me. Glad I’m getting some confirmation

[high skiff]

not very large, but several hundred images us the ideal, yeah

[gloomy barn]

ah

[spiral drift]

This is fooocus two random results in one click

[upbeat summit]

hooray high quality taesd sampler previews are back when using AIT
src: https://github.com/FizzleDorf/AIT/issues/25

[high skiff]

Caith does LORA's on like 3k+ images

[high skiff]

my current realism LoRA is just 90 images, and it already works so damn good

[spiral drift]

wayy more detailed

[gloomy barn]

I know the guy who makes the crystal clear realistic models uses 300 or so

of course, he also suckers people into helping him, claiming he has a commission and they will get paid, and then when the model is done...nada

[high skiff]

base SDXL vs with my realism LoRA

Identical everything, just without vs with the LoRA

[hardy cipher]

I really want to train a lora on some images I'm gathering but still haven't worked out my approach

[gloomy barn]

the realism lora is on the right?

[high skiff]

without vs with

[high skiff]

yes

[gloomy barn]

odd eyes

[high skiff]

its a small dataset lol

[hardy cipher]

need more vintage loras. old pictures, old maps, etc. but I guess that's not what most people are into so might have to make my own

[gloomy barn]

That one is good

[glad fulcrum]

how can I use open pose in comfyUI

[high skiff]

without vs with

this was with a smaller dataset actually

[gloomy barn]

cant compare because one is black and white

[high skiff]

without vs with

[high skiff]

I asked for black and white specifically, so I hope that gives some idea haha

[gloomy barn]

On the right is Tommy Lee Jones

[high skiff]

also, this is not to say that SDXL base can;t do incredible realism, cause it can

but the point of this LORA is there is no need for realism cramming prompts

like, you can see here

[gloomy barn]

I see

some people put out the unfinished in order to get feedback

get lots of people making images with them

[hardy cipher]

someone needs to create a better refiner. that's the solution. or better yet, two refiners that work in tandem

[high skiff]

the prompt is "a tiger"

No negative

without vs with my LORA

[hardy cipher]

each with their own prompt

[high skiff]

as you can see, my LoRA does definitely push realism haha

[hardy cipher]

very much so

[gloomy barn]

too much pattern in the background of with

[high skiff]

its accurate to what real jungle portraits look like, so I am happy with that

tho, my LORA does support simple background tags

[gloomy barn]

the leaves are in a circle pattern...of different trees

[hardy cipher]

you don't like patterns, bro?

[high skiff]

what? You mean bokeh?

[gloomy barn]

no, looks like they make a circle

and tiger's right ear is messed up

like it's a botched perspective

[high skiff]

I am not sure what you mean about the circle

yeah, the animals are a part I need to improve for sure

I only had 10 images of animals in my dataset

just to see if it could apply what it learned from other portraits, which it could

[gloomy barn]

I know tigers very well...their ears are not flat like that usually...certainly generally not different like that

[high skiff]

it has its ears back in aggression tho

real images of aggressive tigers

granted, the ear placement is a little high, iw ill say

[gloomy barn]

it looks like it's trying to be back, but not placed back, just the shape

[midnight shuttle]

Wait, those 3 are real images?

[high skiff]

yes

[hardy cipher]

alright, having pip scan all the requirements.txt folders.

[midnight shuttle]

Wow. I would have said they all look like AI. It's getting so hard to tell.

[hardy cipher]

it's attempting to install about 5 million things

[high skiff]

best of luck

your install seems royally botched haha

[hardy cipher]

folders, files, you know what I mean

it sort of is

[gloomy barn]

looks like some node types were not found

[hardy cipher]

I trashed the venv folder and started anew a few days ago

[hoary saddle]

and there it went again, 4th lockup today mid render, 2 diff machines 2 diff gpus, everything git pulled and updated

[hardy cipher]

but might have to just start over completely. I don't know

[gloomy barn]

lock up as in won't ever go, or pauses for a while?

[visual glade]

how much ram do you have?

[hoary saddle]

hard lock, gotta hold power button to kill it

[glad fulcrum]

is open pose working? I can't get it work

[hoary saddle]

32g in both systems

[gloomy barn]

ah...sometimes comfy eats up resources and won't give it back. Now and then I get a lock, and I have to restart

[hoary saddle]

32g 3080 10gb / and 32g 4090 24gb

[visual glade]

how old is your ubuntu?

[glad fulcrum]

I downloaded the workflow from the open pose hugging face page.

[hoary saddle]

22.04 fully patched

[gloomy barn]

it could be system ram or vram

[hoary saddle]

doesnt matter if i'm using sytan, searge, my own, just mid render for whatever reason

[gloomy barn]

someimes my system locks after only a few generations. Sometimes it goes for a long time with no lock

[hardy cipher]

last time I did this when I loaded comfy Ievery node was red, lolol

[gloomy barn]

usually the first sign is my browser slows to a crawl.

[hardy cipher]

it's always some aiohttp nonsense

[gloomy barn]

Once you get it working, make a backup without the models

[hoary saddle]

can't find anything in /var/logs

[gloomy barn]

then if something goes wrong, at least you have a cl;ean starting point

[glad fulcrum]

somebody, somebody, somebady, can anybody find meeeeee. some open pose preprocessor to looveee

what does this mean?

is open pose working or not?

[hardy cipher]

when you mix a goat lora with a truck lora

[hoary saddle]

a1111 is running on the 3rd system with a 3070 8gb without sdxl and has been going for several weeks without a crash

[gloomy barn]

watch out, one of these "I have dozens of systems all with better GPU's than most people have on one system" guys

[hoary saddle]

does sdxl care what version of nvidia drivers it uses?

[hardy cipher]

I care

[visual glade]

try installing pytorch nightly

[hoary saddle]

using 525 on all 3

[gloomy barn]

does it change nightly?

[visual glade]

yeah there's a new pytorch nightly build every day

[gloomy barn]

wow

[hardy cipher]

wat?

can't tell if joke or truth, lol

[soft bone]

what are you prompting? I've never seen base XL make that fake of an image unless i put painting in the prompt or something

[hardy cipher]

fake image

[high skiff]

completely bare bones prompts

A pretty woman standing in a forest wearing a black dress

stuff like that

[glad fulcrum]

@visual glade what is this?

When I click on install custom node, or anything it appears that

[gloomy barn]

comparing that simple prompt on SDXL vs with the LORA

[glad fulcrum]

and I can't load a open pose workflow

[visual glade]

just wait for the manager to update

[hardy cipher]

that's the manager I believe?

[hoary saddle]

sudo apt install -c pytorch pytorch-nightly this look right?

[high skiff]

no need to add all the flares of

"A stunning realistic analog photograph by national geoghraphic of a gorgeous young woman in her 20's with a beautiful face standing in a detailed and intricate forest while wearing a stunning designer dress at twilight, textured skin, photorealism, blah blah blah blaahhhhhhh"

[hardy cipher]

my manager never updates. it says it needs to update. but it just teases me and never actually does it

I've accepted it

[glad fulcrum]

Yes, whatever I clickl on the manager that happens

[hardy cipher]

it might be prudent to just start this install over. but hassle\

[glad fulcrum]

but still not working

[hardy cipher]

you just have to wait a little while normally

then it works

[visual glade]

https://pytorch.org/

[hardy cipher]

god damn scipy taking 10 inutes to install

[hoary saddle]

ah, always forget they got that nice install table, thanks

[soft bone]

release when?

[high skiff]

no idea

I need a much bigger dataset, which is gonna take forever to train

[soft bone]

youre not satisfied with it?

[high skiff]

no, it needs a lot more data

[gloomy barn]

in the meantime, release what you have to get lots of people using it

[high skiff]

I am at about 90 images, I want at minimum 400 images

And since I tag them by hand

[hardy cipher]

ugh

  [end of output]

note: This error originates from a subprocess, and is likely not a problem with pip.
error: subprocess-exited-with-error

× Getting requirements to build wheel did not run successfully.
│ exit code: 1
╰─> See above for output.

note: This error originates from a subprocess, and is likely not a problem with pip.
NativeCommandExitException: Program "pip.exe" ended with non-zero exit code: 1.

[gloomy barn]

do you say DAY ta or DAH ta?

[hardy cipher]

game over

[soft bone]

yeah civit has an update system so you can release updates as you make em

[high skiff]

I interchange them

[gloomy barn]

i SEE

[soft bone]

i had that. i had to get rid of nightly

[high skiff]

I have no plans to release it until its done and ready. I don't do incomplete releases

[gloomy barn]

that's called testing, not incomplete

[high skiff]

cause shit is gonna change a lot, and you have to hold this LoRA's hand

[hardy cipher]

it's still running through other things so I'll see what happens. if and when I nuke the install I'm going to be much more mindful about install nodes

[high skiff]

reguardless, I don't want my name on this in the public eye. I have a decent reputation for quality, and I would like to keep it that way 😅

[hardy cipher]

too many dependencies

[gloomy barn]

keep backups when it works

[hardy cipher]

trainwreck,lol

[soft bone]

where can i see your stuff?

[glad fulcrum]

okay I manually updated manager, but still the openPose preprocessor node doesn't load.

[gloomy barn]

when you see people's stuff online, it's called "stuffs"

[high skiff]

I don't have much released as of now. I make most things for myself, but my SDXL workflow is my most used/known thing

[hoary saddle]

on a sidenote, the new 4090 24 renders a 1024x1024 with x4 upscale on a @indigo carbon AIT workflow in less than 15 seconds, jaw dropped

[high skiff]

oh man, my stomach hurts

the SDXL indigestion is real lol

[soft bone]

speaking of reputation, i gave up at least on civit. back when i was dreamboothing styles I had insanely good feedback but the moment Loras came out, i got a bunch of bad reviews and ratings because "you didnt make a lora version" 🙄

[gloomy barn]

dont worry about ratings

[soft bone]

manually clone this into your custom nodes folder
https://github.com/Fannovel16/comfy_controlnet_preprocessors

[soft bone]

I don't, and was quickly hired so i stopped releasing anyways. Still it was fun chasing the #1 spot

[hoary saddle]

python3 main.py --listen --port 37860 anything else i should add to startup command?

[gloomy barn]

quickly hired?

[upbeat summit]

--preview-method auto if you want previews on the samplers, --auto-launch if you want your browser to automatically opened when ComfyUI has been started

[glad fulcrum]

thanks. I deleted the installed folder and installed that. I don't have a red message now but still the preprocessor doesn't load. :S

[high skiff]

very nice to see my LoRA is introducing easier support for non standard paired features, like a darker skinned person with blonde hair and freckles

[hardy cipher]

that's a rather unconventional look

[high skiff]

also, my attempt at training in dappled lighting is also seeming tow ork

I need to get more images of it tho for the dataset

[indigo carbon]

Yeah, AIT is actually insane

[high skiff]

the whole goal is to be able to make random mixes of characteristics, cause everybody is unique and diverse

[soft bone]

Most people I was dreamboothing with back before civit existed got hired very quickly for their training skills. Nitrosocke got hired to SAI after making the Elden ring, Modern disney, classic disney, etc. models.
I did some freelance stuff after my Tron, JWST, Cats, and Van Gogh models, but now been developing an app for 6 ish months

[gloomy barn]

I see

[hardy cipher]

I think it's great. keep up the good work, man. do you have a lora tutorial anywhere, or would you be able to suggest one? I know how to use the tools, but not really how to optimize anything

[high skiff]

I am honestly winging it, and my settings are far from good 😅

@boreal bough is a plethora of knowledge tho

he helps me in my manic training sessions lol

[hardy cipher]

I mainly finetuned entire models using dreambooth quite a while back

then extracted loras from those models

[glad fulcrum]

k I give up. I'll wait for auto1111

this comfyUI isn't working

[hardy cipher]

for controlnet?

[glad fulcrum]

for me? yeah

[indigo carbon]

The only downside or rather, unpeculiar property of AIT is that it doesn't decrease VRAM usage as much. If AIT would decrease VRAM usage on the same scale it increases performance, I am confident we would be seeing people running stuff like LLaMa 2 70b on normal 4000 GPUs

[high skiff]

I can run llama 2 70B on my 3090, it just takes a bit

I did falcon 75B at 3 bit IIRC, and it used 38GB memory, but I was still getting like 1.7t/s, which is not much slower than ooba was in its early days for just normal models lmao

[hoary saddle]

@high skiff flow on 4090 vs 3080, only 1 sec faster

[high skiff]

thats peculiar, that makes... no sense

[indigo carbon]

If by a bit you mean 0.08t/s then sure.

[high skiff]

unless you have a very fast 3080 and a very slow 4090

[high skiff]

no, likely about 1.5-2t/s

[hoary saddle]

3080 (not a TI 10gb), 4090 24gb

[high skiff]

which is still usable

Sure, its slower than the like 50t/s I get with wizcuna 13b, but still

[heady vale]

probably CPU bottleneck on the 4090

[hoary saddle]

that definately could be, 4090 in an older machine (card is facking huge and wouldn't fit in the newer case)

[indigo carbon]

I tried this stupid model today on my 4070ti, it loaded the model, pretty fast I might add, but it didn't respond. I left it running for a few minutes and it only wrote 2 words.. pathetic, pathetic "efficient" model

[high skiff]

my 3080 I have is bigger than a lot of 4090s

its the most overbuilt card I have ever seen lol

[high skiff]

thats also a huge issue with the 4070ti and its pathetic VRAM bus unfortunately

fast GDDR6X chips with an anemic 192 bit buss straight off a 2012 midrange card

[indigo carbon]

I'm so angry, I can make high quality SDXL images in about 15-ish seconds, and I can't make a few words?!?! This is so stupid

[high skiff]

take it up with NVIDIA. They are the ones that decided that they would push back the VRAM bus on their cards by nearly a decade cause they felt like it

its not even like they saved really any money doing it

[hardy cipher]

which model?

[soft bone]

looks like @delicate kelp is cranking out some bangers🔥

[hardy cipher]

do you deepspeed?

[indigo carbon]

No, should I? Is this the AIT of LLMs? Heard that name before

[high skiff]

hell, the GTX 460 from 2010 has the same bus as the 4070ti

Its pathetic

[hardy cipher]

it sped things up substantially for me

[delicate kelp]

testing the next update 😉

[indigo carbon]

Huh, I'll try that out. Also EXllaMa, right?

[high skiff]

my favorite bit of info is that the GT 8800 from 2007 has a bigger bus than the 4070ti, and it only has 512MB of VRAM lmao

[indigo carbon]

Man Nvidia really ripped me off with this one, didn't they?

[hardy cipher]

hmm, I haven't tried that one. I used it with oobabooga https://github.com/oobabooga/text-generation-webui/blob/main/docs/DeepSpeed.md

[high skiff]

oh yeah, for sure

They cucked lower end 4000 series so fucking hard lmao

[hardy cipher]

it went from unusable to at least acceptable speeds

[high skiff]

hell, the 4060 has an 8 lane bus

I don't think thats like... ever happened before

[indigo carbon]

4070ti isn't low-end..(?)

[high skiff]

that means if you have an older 3rd gen PCIE mobo (which is likely if you are buying a budget card), you only get 4x gen 4 speeds on your GPU

Its disgusting

[indigo carbon]

I'm generating insane images with AIT on SDXL at speeds close to people with 4090 get to

[high skiff]

its the first one that gets mega shafted by NVIDIA

specifically the Anemic VRAM bus that makes it really shitty at high read/write tasks

[indigo carbon]

Fuck them, I'm probably getting a refund

[high skiff]

if you can, i would

I would just buy a used ampere card

how much did the 4070ti cost you?

[indigo carbon]

The CUDA cores themselves are amazing, no denying that. But man, the VRAM is something I would make out of playdough in 1999

[high skiff]

how much did you pay?

[indigo carbon]

I paid 750$, I was satisfied with it until I tried to use LLMs with over 13B params

[high skiff]

bro, just get a used 3090 for that price

my 3090 was $700, and I have a top of the line EVGA FTW3 Ultra 3090

24GB VRAM with a 2x faster bus

[indigo carbon]

Fuck this

Maybe I can create beautiful images, but I can't run a damn LLaMa model?!?! I am going to smite Nvidia

[high skiff]

welcome to the party lol

[soft bone]

i got that same card for 600 in April😁 love it

[high skiff]

its been a great card so far

I just had to rip it open and clean it out cause got damn was it dirty from the previous owner

[indigo carbon]

headbutts the wall

[high skiff]

honestly tho dude, used 3090 is 100% the way to go for AI

anything the 4090 can do in terms of training, the 3090 can do, just slower

you can give something more time, you can't give something more VRAM

and 2x the VRAM over a 4070ti is also sexy AF

[high skiff]

you can traing LORA's like a beast for SDXL on 24GB VRAM

[indigo carbon]

I'm returning this piece of junk and saving up for when 5000 series comes out

[high skiff]

The state of my card when I got it lol

dirrrtyyyyy

took 2.5 hours of elbow grease but...

Kleeeeeen

i droped, and I kid you not, anywhere from 15-25C depending on what part of the card you are monitoring

[indigo carbon]

I'm so fucking done

How dare Nvidia do this, I will find them

[rustic temple]

you say this like 5000 series won't be similarly predatorily priced and specced as 4000 series

[queen adder]

what is your iteration speed when training SDXL?

[high skiff]

depends, I train a wide range of BS, with TE, without, with buckets, without

I'd say at BS 10 or so, about 6s/it with SDXL at 1024x0124

which means with simple settings and about a 30 image dataset, a LoRA can be done in like 30 minutes?

[queen adder]

ok im definitely doing something wrong then. i used BS 5 and got 60s/it

[high skiff]

my bigger LoRA trained at BS 5, it was about 3.2s/it, and withj 90 images, 30 epochs, and 5 repeats, it took about 2.4 hours

[queen adder]

a lot of the images in my dataset are higher resolution than 1024x1024 and maybe that's why

[high skiff]

oh, are you not using buckets?

[queen adder]

yes, im using buckets.

i copied Aitrepeneurs SDXL json config file so i havent messed with any settings, other than my dataset of images are varied resolutions

[soft bone]

huh? at BS 5, 1 repeat, 120 epochs, 16 images takes a full hour

for me

[high skiff]

let me do some math

I am doing basically 13,500 steps (30 epochs x 5 repeats x 90 images)

and you are doing (120 epochs x 1 repeat x 16 images) which is 1920 steps

[hardy cipher]

that's a bit different

[hollow halo]

You aren't actually using all those images with BS 5 and 16 total images, one is being dropped

[queen adder]

it took me 30 hours or so to do 1500 steps on a rtx3090

[hardy cipher]

50 steps per hour

nice

step per 72 seconds

[high skiff]

ok, something is very much off

are you using gradient accum?

[hollow halo]

It takes me an hour and 35 minutes to go 12400 steps at BS1 for an iA3 on my 3090. 2.16it/s

[high skiff]

if not, thats why

[soft bone]

i am

[high skiff]

I mean for aphellion

Also, what optimizer?

[queen adder]

gradient accumulate steps? it's set to 1 which is the lowest on the slider i believe.

optimizer is Adafactor

[indigo carbon]

Just gave Nvidia a nasty letter

[high skiff]

Your steps may be set to one, but do you have it enabled with the checkbox?

Ohhhh, I think Ada factor explodes with higher BS

Some optimizers use more or less VRAM than others

For example, with the prodigy optimizer, you can only run BS1

BS1 uses about 13.9 GB RAM, but BS2 uses about 39 GB

[frosty pine]

any free SDXL site beside Clipdrop?

[queen adder]

i trained with BS 1 and it was very slow too. oddly enough it was using 23GB/24GB VRAM on BS 1 which didn't make any sense.

[frosty pine]

free forever not 25 images perday stuff

[high skiff]

Nope

Unfortunately, you do have to pay for compute

[frosty pine]

ah to bad, why nobody put the model here https://aqualxx.github.io/stable-ui/ ?

its not out yet for the masses is it?

[high skiff]

Not many sites are keen on throwing money into fire

[frosty pine]

it still Delibarete model on aqualxx

[soft bone]

#🔧｜finetune message

I dont think images get dropped, but the last batch would be only 1 image, so not ideal anyways

[high skiff]

The way I was told by caith and other good fine tuners is that your batch size should always be evenly divisible into your image account

[queen adder]

is it possible the amount of buckets i had was causing the slow down?

[high skiff]

How many images did you have in the data set?

[soft bone]

it definitely should. but i dont think the leftover gets dropped entirely

[queen adder]

500 or so

[high skiff]

Wait a minute, that data set is terrible

Why do you have so many images above and below 1024^2 eq?

All of those low resolution images are really going to throw off your latents

If you are going to train SDXL at a lower than optimal resolution, you should keep the lower than optimal resolution continuous

[indigo carbon]

Should I really get a refund for my 4070ti? Or is it not necessarily an issue of NVIDIA using a decade year old memory bus

[queen adder]

i could never get an answer here on what resolution the data set was supposed to be, so i just winged it.

[hardy cipher]

ooh, that link hsa a furry mix model. just wonderful

[high skiff]

1024x1024, or as close to equivalent as possible

[hollow halo]

Just gonna put this here for convenience

[high skiff]

@queen adder honestly, my guess is that you went severely over your VRAM limit, which led to the additional memory pulling in your system memory, which led to a significantly slower training session

[hardy cipher]

I made an infinite aspect ratio node

[queen adder]

sweet, this is the answer i was looking for lol.

i also saw a tutorial saying varying up the aspect ratio gave better training results, but what they left out is that every resolution used probably added up to 1024x1024^2

[hardy cipher]

well maybe not infinite

[high skiff]

You don't want 1024 X 1024²

You just want 1024²

[queen adder]

yeah my bad

[high skiff]

Or 1,040,000 IIRC

[hardy cipher]

lol 1024^3

[high skiff]

If you set the resolution to 1024 X 1024, and then provide it with images at that resolution or higher, it will figure all the math out for you

Most of my images for my realism data set are actually 4K or higher resolution

Even have some 6k, and maybe even a couple 8k+

[queen adder]

Oh, you mean set the canvas size in photoshop to 1024x1024 and then just process each image in there

[high skiff]

No, I mean if you set the maximum resolution in kohya

[hardy cipher]

what about if an image is too small? upscale or trash it?

[high skiff]

It will still automatically bucket them, but it will bucket them to where they are the maximum resolution that is equal to 1024 X 1024

[indigo carbon]

I am soo angry at Nvidia rn, how DARE they have the AUDACITY to use a decade old memory bus for all cards under the 4080? I think I'm going to get a refund, this is unacceptable

[queen adder]

ohhh, so you think it was just the images that were smaller than 1024 x 1024 was causing the issue?

[high skiff]

yeah!

[soft bone]

yes you should upscale the smaller ones. but only if it comes out looking amazing after

[indigo carbon]

Why the fuck did they think this is a good idea?!?

[high skiff]

if you give is a 2560x1080 image and tell it to bucket to 1024x1024, it will downscale the image to the resolution that is closest to 1204x1024 total pixesl while being the same aspect ratio

[vapid roost]

Do we know much more than what’s on report on how the refiner was trained? What was the dataset that was used and resolutions of images in that dataset? Did they also use the other conditioning parameters mentioned for the base?

[queen adder]

perfect. i will just weed out my smaller resolution images and trash them, or AI upscale the good ones.

[high skiff]

as somebody who is one of the leaders in workflow processing and premade workflows for th emasses, I am ditching the refiner as a whole, as I have found it hinders more than it helps, especially when you get better at prompting fundementally

[hardy cipher]

does it round to 64 pixels?

[high skiff]

awesome, best of luck

[soft bone]

I heard it was trained on openclip only, while base was trained on openclip and openai clip. all i know

[high skiff]

@gloomy barn also, here is the single tiger image in my dataset which it was pulling from hard for that "tiger" image lol

[high skiff]

it rounds to whatever you set it to

[vapid roost]

Maybe we need to get better at implementing refiner too?

[high skiff]

I god popular specifically off that

its just not something that should stick around

its extra issues, extra requirements, and actively damages most images

and it makes using LoRA's a waste of time

[soft bone]

I assume finetunes will make refiner obsolete in every way

[vapid roost]

But I agree at least for now that ditching refiner might make lot of sense especially if we want community trained stuff - LoRas, fine-tunes. Refiner just ruins them

[indigo carbon]

Does the 4090 also have those old memory busses?

[high skiff]

@hardy cipherKohya advanced setting

[high skiff]

no

the 4080 and 4090 are fine

just the 4070ti and below are cucked

or well

actually

the 4080 is kinda cucked too, but its not as bad as below cards

16GB GDDR6x on a 256x bus

the GTX 560ti from 2010 has a 320bit bus

and what... 2 GB VRAM?

GDDR3 IIRC too, so like 1/16th the speed?

[hardy cipher]

ahh, thanks. as soon as I put together a decent set of images I'll try it out. but having a hard time with this one

[high skiff]

best of luck man

[indigo carbon]

Wait what are the specs of the 4070ti's memory bus? Am I yelling at Nvidia for no reason?

[hardy cipher]

thanks. I'll get there. just sifting through a lot of non-win images

[vapid roost]

thx, I wish we knew more

[hasty smelt]

Hi guys! I would like to know if someone could share a Workflow for image-to-image conversion. If anyone could point me to a link with resources on this topic. Currently, I'm using Sytan's SDXL 1.0 Workflow for image generation. Thank you in advance for your assistance!

[high skiff]

can you share what you are training? I might be able to point in the direction of some good image sources

[high skiff]

thats me!

:D

[tribal jackal]

I'll be going for a refund or at least an upgrade when asus rma team eventually gets back to me about my 4080 for sure

[high skiff]

Oh, I guess I should get to work on my img2img workflow as well haha

[heady vale]

its not specifically an old bus, its just smaller due to the size of the gpu chip. older cards had larger buses as the silicon was larger. Its a physical limitation

[vapid roost]

where do I find this famous Sytan's workflows?

[high skiff]

its me

just a sec

[tribal jackal]

4090 minimum, keeping an eye on the new cards they announced but those are so pricey

[high skiff]

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

there you go <3

[crisp owl]

I've seen this node with a "longest" value in the "side" field. Why do I only have "width" and "height"?

[hasty smelt]

Sytan, I'm using your Workflow but I don't have an image to image option, sorry for my lack of knowledge, I'm new to ComfyUI

[hardy cipher]

I'd like to tain a model on things like this. old technical drawings, da vinci sketches. but I'm finding a lot of really basic, faded stuff and then a lot of midjourney generated images for some reason

[high skiff]

no promises on when guys, but I will have 2 new workflows at some point hopefully soon

One for even more efficient SDXL usage as a base

And another dedicated to img2img workflows
both should run more efficient on all GPU's, as well as take less experience to prompt, and also upscale a little easier

Also, full built in LoRA support from text encoder to upscale result

[hardy cipher]

internet archive has a bunch of stuff, but so much of it is just not worthwhile for training

no hate on their sketches, lol

[high skiff]

yeah haha

thats gonna be an interesting LORA

let me see if my go to site has anything

[tribal jackal]

I have been waiting to make that cough drop joke for over a year

[hasty smelt]

Tag me when you do this, I'm learning a lot with your workflow.

[hardy cipher]

I appreciate it.

there are already probably a thousand plus weird nsfw loras for sdxl. but kind of lacking in some of the stuff I want

[high skiff]

hmmm, doesn't seem to have much in the way of technical style drawings specifically unfortunately

[high skiff]

for sure, thats why I am being meticulous for realism

[hardy cipher]

all good. I didn't think they'd be so elusive

[hasty smelt]

Can anyone help me with an image to image workflow?

[high skiff]

I could throw one together really fast, but the problem is it needs a lot of math

[hardy cipher]

I'm just wondering how many images I should use

I guess it depends on the method

[high skiff]

honestly, I think a few dozen should work, at least as a proof of concept

[hasty smelt]

I don't want to take up your time, I just wanted something to help me understand the knots, everything is new to me.

[hardy cipher]

yeah, I might have 10 or 12 that I think are passable.it's going to be a chore tbh

[high skiff]

maybe start with 24? BS 6, 6 repeats, 10 epochs?

see how that goes, then see if you need to change anything?

thats um... an image

[hardy cipher]

and then images like this that are cool, but midjourney

[high skiff]

wow lmao

that looks pretty sick

but getting SDXL to grasp that will be hard

[hardy cipher]

I could probably incorporate it tbh

[high skiff]

what you really need is a single subject in that style

[hardy cipher]

that's what I was wondering about

well I wasn't originally going to span so many areas

but it's surprisingly hard to find these images. I mean, in my head I thought they'd be in abundance

[high skiff]

explain what the subjects are

It figures out what it has to do to its fundamental understanding of said object to get it that way, then it should be able to apply it pretty well to other things

[hasty smelt]

This is crosshatching style draw

[high skiff]

yeah, you might be able to augment your dataset with some cross hatch stills

[heady vale]

Nightvision alpha - no refiner

[high skiff]

||looks straight out of every 1.5 model to me ||

[hardy cipher]

yeah, I'll definitely try it out. do you tag manually or do you use any kind of model for that?

[high skiff]

it changes from run to run at the moment

you could also look for concept sketches

[arctic ocean]

i really need help trying to come up with an AI image it is so specific in detail

[high skiff]

like these

[soft bone]

@high skiff what if I challenged you to replicate what fooocus (or whatever) has done with the sampler mix

[high skiff]

not even sure what you are referring to

[hardy cipher]

maximum croshatching.

[high skiff]

definitely not pointilism lol

[hardy cipher]

well my original idea was using vintage technical/mechanical drawings, but then adding weirdness to them

[hasty smelt]

This sounds good

[hardy cipher]

digital crosshatching

[soft bone]

the way they combined the refiner into the same sampler as base because "a lot of latent gets dropped when transferring between samplers" or something

[high skiff]

eh, i wouldn't waste the time. Its pretty agreed upon in the training and high level prompting scenes that the refiner is kinda dead weight

all of my fantastic images said bye bye to the refiner

[arctic ocean]

can i get some help with making an AI image?

[hasty smelt]

I think I got a good result at some point with the bot, let me see that.

[hardy cipher]

want to make stuff like this, but mayb ea bit more vintagy

[soft bone]

but thats part of it, if the refiner performs way better the way they're using it then its less of a dead weight

[arctic ocean]

i was told looking into sdxl since i had something so specific

[hardy cipher]

I guess that's not a technical drawing, lol

subatomic sketch

[high skiff]

all base SDXL realim with the refiner kicked to the curb

[soft bone]

use "schematic"

[hardy cipher]

hmm, good idea

that one is one of my favorites from that batch of renderings

[soft bone]

plague inc

[high skiff]

I think I might have found some help @hardy cipher

[hardy cipher]

but definitely would be cool to have something a bit less futuristic

[high skiff]

"Design draft product sketches"

[hardy cipher]

nice. howd you come up with that?

[high skiff]

Looked for even one good image relating to what you mentioned, looked through the result, found words it used, strung them together

[hardy cipher]

btw, both schematics and "design draft product sketches" are giving me good results

[hardy cipher]

that's some smart thinking, lol

[hasty smelt]

Certainly, I've corrected and clarified the text for you:

Using the term "Crosshatching," which is the correct terminology, the bot wasn't yielding very promising results. However, when I used the term "engraving," I think it worked to generate something similar to crosshatching

[hardy cipher]

that's a nice one

[high skiff]

some fun results

[hardy cipher]

but I made it, lol

[soft bone]

your workflow is completely outside the blue box 🤣 opened it and saw nothing

[high skiff]

@hardy cipherproduct design concept sketches is also dope

[hardy cipher]

oh that workflow is rather unconventional. but what's the blue box? lol. is there is actual center to these things?

[high skiff]

lots of gold here

[soft bone]

you cant see it? blue outline around center

[hardy cipher]

tbh I never noticed,l ol

[high skiff]

lots of good shoe schematics

[hardy cipher]

nice. I have all these search terms loaded. I'll surely have sufficient data here shortly

[high skiff]

community sourcing lol

I come here every so often to community source some needed materials for my workflows

[hardy cipher]

you know, now that I think about it I was wondering why some of my workflows were out of view when I loaded them

[high skiff]

specifically pixel upscalers

[hardy cipher]

just working outsid ethe box

[high skiff]

ohhh, industrial design concept sketches is pretty dope

[hardy cipher]

man, I was kind of at a standstill there. I would have figured it out eventually, but thisvery much expedited things

[vale eagle]

I have an idea a while ago. Use the model to generate sketches. Use the sketches for img2img to generate the realistic image.

[high skiff]

I would be scummy if I didn't help other people lol

[hardy cipher]

it's much appreciated. I have been doing all of this stuff solo until recently. and sometimes it'd be a grind

[heady vale]

Ive used colouring in book line drawings in controlnet. works really well. same applies to drawings

[high skiff]

for sure

[hardy cipher]

maybe include the bean comic?

[high skiff]

I had to rely on my research partners for good image sourcing sites for my images

took me 3 days on google to put together a pretty ass 24 image dataset cause google image search has been so massively neutered

[hardy cipher]

that's what it is. every image I was finding was either ai, low resolution, or watermarked

[high skiff]

yessir

google has monumentally stepped down its image search engine to hold a big lead in data sourcing

[soft bone]

this discord taught most of the training pros the majority of what they know, prompters as well probably. the open source energy of sharing info is what propels SD at light-speed. never stop sharing info and never stop asking questions.

[heady vale]

websites might be adding google image scrape to their block lists

[hardy cipher]

I guess that might be true

man, I evolved in relative isolation so I have some strange approaches. but tbh I just wanted to feel like I sort of knew what I was doing before I started trading ideas with people

[high skiff]

I doubt that free stock image sites are actively trying to not be accessed, as they get money from ads and clickthroughs

even unsplash results are getting rare

and even then, google only embeds their high res images at like 600p

[soft bone]

i wonder if other browsers have more now

[indigo carbon]

how come I can generate stuff like this in ~17 seconds and I can't generate more than TWO WORDS with LLaMa?!?!?!!11!

[azure oxide]

llms are pretty complex stuff tbh

[hardy cipher]

need more like this, lol

well not more cars per se

but that's the kind of structure I'm looking for idealy

[indigo carbon]

there is no way, when I run a 30b LLaMa model I get 0.034T/s and with SDXL I make awesome images in under 17 seconds

fuck that, "efficient language model", so stupid

[high skiff]

ok, 30B is too big for 12GB VRAM

30b is reserved for 24GB VRAM only

[hardy cipher]

wow. you should check out "vintage organ schematics"

[high skiff]

jesus lol

[heady vale]

cutaway 3d blueprint diagram of a race car

[hardy cipher]

that cold be an entire lora by itself. why so many images?

[indigo carbon]

really? with 12B I get ~41.3T/s, why is it unusably slow when I go 30B? is this because Nvidia are fat fucks?

[high skiff]

no... its cause 30B models need at least 18GB VRAM

but your very cucked VRAM doesn't help with pooling

[hasty smelt]

I'm looking for some image to image workflow in Civitai, but everything I see there seems complicated, can anyone recommend a noob-friendly workflow?

[crisp owl]

https://civitai.com/models/123048?modal=commentThread&commentId=220968

[high skiff]

like here, on my 10 GB 3080 with a 30B LLM, i would still get like 8t/s

[indigo carbon]

4070ti should be a little stronger than a 3080, and it has more VRAM, why the fuck is this happening?

[hasty smelt]

I will try this one, thanks buddy

[high skiff]

likely VRAM bus being shitty

[uncut fiber]

4070 has bad concept, yeah

[high skiff]

the VRAM bus on the 4070ti is nearly half the bandwidth on a GPU that is decently faster GPU

[indigo carbon]

jesus

[high skiff]

and once you get past the minimum required compute time in latency, it slows down like crazy

[heady vale]

do ADA cards need specific drivers or dlls - similar to A1111 etc?

[indigo carbon]

wait, so why do I get more than 40T/s on 13b models?

[high skiff]

cause a 13B model can fit inside 12GB VRAM

the issue shows up much worse when you need to load from outside VRAM

so cause a 13B model fits all on the card, it doesn't bog down the anemic VRAM bus as much

But then a 30B model is way too big for 12GB, so it caches like 6-8GB in system RAM, which is much slower, then it has to send that info into the GPU while echanging out info it doesn't need at the time, which is a huge VRAM bus hit, and then additionally it also has to read and write to the VRAM in real time for every token as well

So its a massive amount of reads and writes happening per token, which chokes the hell out of the 192 bit bus on the 4070ti, subsequently making it fall behind its clocks and basically computing, then waiting for a long time for all the right info, then computing, and waiting again, so on and so forth

on my 3090, its a fairly small hit for LLM's

Like I ran a 3 bit 70B model on my 3090 using about 36-38GB GPU memory, so 24 on GPU, and about 12-14 GB in system pool. Even then, I was still getting like 1.7t/s I think?

[still dove]

guys, where do I get these nodes

[crisp owl]

impact pack

[still dove]

installed it but no luck

[hardy cipher]

hmm, I think I could modify that node I made a bit to quickly resize all these images to the proper resolution

[crisp owl]

make sure it's updated. at least for ultralytics

[hardy cipher]

decent

[uncut fiber]

you downloaded missing nodes?
it may need restart.

[hardy cipher]

this is true

[high skiff]

some nodes you can refresh, some you need to restart

[hardy cipher]

I always just restart either way

since it takes like 30 seconds

[high skiff]

yeah, its way faster than auto

[still dove]

I installed latest build of impactpack, how can I update it for ultralytics?

[indigo carbon]

anyways, my 4070ti seems to run SDXL extremely fast when with AIT. So I guess this doesn't screw me up that bad because I'm not that into LLMs

[hardy cipher]

dude, I nuked a couple auto installs when it would take 5-10 minutes

sytan, what were you saying resized the images? does the kohya program itself do that?

[high skiff]

the batching? yeah

sorry

bucketing

[vale eagle]

Is it a generated image?

[crisp owl]

Here's his github. In a recent update he added the ultralytics node which will replace the MMDET nodehttps://github.com/ltdrdata/ComfyUI-Impact-Pack

[hardy cipher]

I kept reading bucketing and just didn't know what it meant. and didn't really want to ask since I could look it up

[hardy cipher]

it's possible. but the website gave no indication of that

we're at the point where pretty much anything could be ai generated at this point

I said point a lot

[soft bone]

why does my comfyui suddenly have a chicken on it

[hardy cipher]

I don't see a chicken

[soft bone]

im guessing searges nodes?

[hardy cipher]

that's a bit odd

[soft bone]

tis chikin

oh thats a fuckin furry girl

[crisp owl]

that's an anime girl

[soft bone]

delete

[heady vale]

https://tenor.com/view/look-at-all-those-chickens-gif-20113673

[hardy cipher]

you know, I didn't even notice that was a chicken or an anime girl

[soft bone]

must undo this

[hardy cipher]

I can't even think of what it was before. I saw that and just assumed I hadn't noticed it before or something

[soft bone]

back to normal

[hardy cipher]

I"mfinally just noticing the blue box over here. I feel like it could perhaps be a bit easer to see.

[soft bone]

true is very faint

[hardy cipher]

I wonder where that setting is hidden. probably pretty easy to alter if I could find it

add a little brightness to it

[hardy cipher]

that boy ain't right

[heady vale]

cyberpsycho

they arent right. at all

[indigo carbon]

man, I like how easy it is to do text in images now

[hollow halo]

Anyone else no longer generate sample images when training?

[soft bone]

@hardy cipher bit like that?

[soft bone]

mine stops randomly, and randomly starts again later

[hollow halo]

I don't mean as an issue, I meant like it just not being useful for determining when its done. I found the graphs tell me a lot more and I just save more often, then test when I think it isn't going to get better

[hardy cipher]

indeed. it's sort of deviated a bit from my original idea, or expanded rather

[soft bone]

oh yes i agree, mostly useless. no epochs show likeness and then i use the exact same prompt in comfy and get perfect likeness

[hollow halo]

Yeah that's what I concluded, though sometimes its nice to see something more visual as it goes, but I haven't generated samples for my past week of training

[soft bone]

I only did for the very first time last week

[hollow halo]

Has anyone tried a multi-concept LoRA that uses the same class token for all folders?

[soft bone]

they were talking about that here

#🔧｜finetune message

[hollow halo]

Thanks, I'll check it out

[queen adder]

@high skiff This is strange, i upscaled all my images but when i look at the buckets readout the resolutions look wrong?

[hollow halo]

It downscales them again

To fit in the range you set

which is probably 256 or 512 to 2048

[high skiff]

the range I use is 640-1636 IIRC

[queen adder]

you mean this range?

[hollow halo]

Yeah, 4096 is really high btw, and it tries to fit them around the 1024x1024 you set above that

[queen adder]

what do you reccomend minimum and maximum set to when training SDXL?

[hollow halo]

Sytan's range of 640-1636 isn't bad, the default is 256 to 2048 in Kohya_SS or 512 to 2048 in SDXL branch of Darrien's easy training scripts

I leave mine at 512 to 2048 usually

[high skiff]

@queen adder

[queen adder]

thanks, i will run it again with these settings. maybe the speed will change. I was getting 3it/s on BS 1

the buckets still look exactly the same but whatever. 2.3s/iteration

[hollow halo]

You were technically already fit into that range by coincidence lol

[queen adder]

it's strange to me that 19GB/24GB VRAM is used up training with batch size 1. how the heck are people using batch size 3, 4 or 5?

[hollow halo]

What optimizer are you using?

[queen adder]

Adafactor

[hollow halo]

That one uses the least iirc, but I haven't tried anything over 4

I usually use Prodigy so am stuck on BS 1

[high skiff]

I use adamw8bit of the time

Sometimes I can get up to BS12 on 24GB VRAM, but it depends on what my dataset is like

[hollow halo]

I sit at 18GB with Prodigy on BS1 but anything more will OOM for me

No checkpointing though

[queen adder]

adamw8bit doesn't run for me, gets some kind of python error when i try to run training

[hollow halo]

I used it on 0.9, so not sure what that would be about without seeing the error

[hollow halo]

It doesn't take Scale_parameter and relative_step

not sure about warmup_init

[queen adder]

ohhhh, i should wipe all those extra arguments then.

[vagrant wasp]

comyUI official discord ?

[soft zealot]

SDXL OOM issue with COntrolNet.

Well scratched my head last night and I'm still none the wiser.

Card is a 1080ti with 11GB Vram
ComfyUI is running at standard OOTB settings and has never OOM'd on any other workflow I run on it (execept to switch to tiled VAE on Hi ResFix)

Trying to run the accontrolnet workflow in the screenshot however get the dreaded red wall as sonn as it starts using the Base Ksampler.

Also included on the picture i s a meory graph showing a standard 3 psdd tun (Pre Con-Base-Refiner)

*torch.cuda.OutOfMemoryError: Allocation on device 0 would exceed allowed memory. (out of memory)
Currently allocated : 9.60 GiB
Requested : 80.00 MiB
Device limit : 11.00 GiB
Free (according to CUDA): 0 bytes
PyTorch limit (set by user-supplied memory fraction)
: 17179869184.00 GiB

Prompt executed in 31.26 seconds*

[uncut fiber]

Hey, anybody using AITemplateVAEDecode in ComfyUI with SDXL with 8GB of VRAM? I mean does it working? Getting black images probably have not enough memory, but tried on 256x256 as well and black as well.
Thank you!

[crisp owl]

I have a 2060S with 8gb, but that series isn't supported yet, hopefully it will be eventually

[uncut fiber]

i have 3070, probably same issue. Thank you @crisp owl

[crisp owl]

I believe the 3k series is supported

[uncut fiber]

AITemplate loader is working, not perfect, but working. Just VAE is too slow. And node giving black picture

[crisp owl]

ah I gotcha. I wish I could fiddle around with that, but alas, my hardware situation won't allow lol

[uncut fiber]

i believe official support will make it for all.

[crisp owl]

Yeah, just more "hurry up and wait" situation
I guess it's more impatience because this AI stuff is moving swiftly

[queen adder]

just read this youtube comment SDXL 1.0: it is recommended to use smaller network ranks, such as 16 or 8 to reduce file sizes
does this seem right? i'm using network rank 256 on Lora training and maybe that's too high

[uncut fiber]

@crisp owl if you put it that way, you get error or is it terribly slow?

[crisp owl]

I can't use AIT at all

[uncut fiber]

o.k. Because in my case first run is horibly slow sometimes it can take up to 3 mins. Then it starts working.
Will have to next time set first run steps to very low, to get rid of this pass.
Just asked.

[heady vale]

change keep loaded to :disabled

[crisp owl]

was just gonna mention that, I usualaly see people with that disabled, not enabled

[heady vale]

if enabled it hogs vram

[uncut fiber]

o.k. thank you, but i think it didint work. Going to test thank you @heady vale

[vagrant wasp]

would it be possible to export the prompts in json format later down the chain .. this could facilitate in creating NFT

need a scripting node

[uncut fiber]

Hey working great and VAE is quick as thunder

[azure oxide]

256 is kinda high

that's not to mention this

[azure oxide]

meant to respond to this, dunno why it didnt 😒

the creator of LORA on DIM from https://www.reddit.com/r/StableDiffusion/comments/1223y27/im_the_creator_of_lora_how_can_i_make_it_better/jdqv9ik/ :

It's possible we don't need to use dim=128 and adapt all attn layers. I suspect that we can reduce the size by quite a bit if we are careful about which layers to adapt.

[soft zealot]

Interesting.

If I run the same flow as standalone (minus the Condition Switchers) It completes although noted it does use all of the memory with no overhead

[molten gull]

does anyone know if it's possible now with a1111 to train textual embeddings?

[soft zealot]

Ok so adding --lowvram to the comfyui "run_nvidia_gpu.bat" helps

(first chunk was standalone with --normalvram then theres a succesful combined run with --lowvram enabled)

[molten gull]

and another question: is it possible to remove the comfyUI metadata somehow again from an image?

[plush gale]

is there any good anime models for sdxl

[dense chasm]

huggingface have some excellent LoRA trained anime style models

[soft zealot]

if you use the WAS Node Image Save you have an option to not wirte the metadata to the file on save or if you prefer you can use something like ExifCleaner to strip the data later

https://exifcleaner.com/

[half ivy]

i am here

any bro's wanna check my messy control net setup, feel like ive done it wrong o0

[meager canopy]

You can compare it to mine, that works. https://civitai.com/models/119257/gtm-comfyui-workflows-including-sdxl-and-sd15

The controlnet part is just the light blue area, bottom left of the flow.

[half ivy]

Cheers man

[charred carbon]

Tyrannosaurus rex

[plush gale]

got a link?

[hollow hare]

hi guys, does the existing textual inversion from 1.5 can be used with SDXL?

[trim orbit]

nope

[dense chasm]

Min-SNR-γ 3.4 times faster convergence

[rustic garnet]

you can train TI with kohya

[molten gull]

i want to use the A1111 process, since i want that "save every trainign step" set to 1 and save every image

i think this cant be done with kohya

[rustic garnet]

it can

[molten gull]

in the same way like a1111 did?

[rustic garnet]

--save_every_n_steps

and sample

yes

[molten gull]

is there some easy way to set kohiya up ? 🙂

[rustic garnet]

I'm currently not on my computer, but you can more or less copy& paste the train command

the train setup itself is easy: make a directory with the name "1_keyword" where keyword is what you want to train on

the rest ist just the parameters. You can basically use same parameters as with lora training

[molten gull]

can i stop training in the middle, change training images, then continue ?

[rustic garnet]

I think there is a checkpointing option to stop and resume training

but it's easier to just stop and restart training

you can give a parameter --weights that points to a pretrained embedding

[dense chasm]

kohya training arguments set save_state=true and resume=true

you'd better run on highly RAM,otherwise it will be OOM

it will not offload any occupied RAM usage

[molten gull]

how can i create a LORA that effects skin-texture ? i need to get better/realistic skin on my portraits
but i dont know ( a ) what kind of pictures to put in the training picture set for the LORA, and ( b ) how to describe the pictures so the LORA will learn only the skin texture
anybody got ideas here ?

[lilac wren]

❤️
https://huggingface.co/thibaud/controlnet-openpose-sdxl-1.0/tree/main

[strong field]

how much work do you want to put into this?

15 pictures, or 150?

[molten gull]

as much as needed, its gotta be good

but rather 30 pictures 🙂

hey, if 150 are needed, i ll do 150 though

[strong field]

@high skiff is working on this very thing at the moment, he has some examples in chat but he thinks it will be a while before its ready/out

you could try to do a "quick and dirty" lora using 30 ish of photos you think look good with texture, it might work decently at lower lora strengths

Do you have your photos already?

Once you get your photos start with BLIP captioning to get you a decent head start

The captioning might take a while

You will need to caption everything that is NOT what you want

[molten gull]

my first question is: what kind of images do i need for training

and how exactly would i have to tag them?

describe just ANYTHING but the skin ?

[strong field]

pretty much

i suppose you are doing close up portraits?

gotta find something you like in the ether

[molten gull]

rather closeup than not, yes

would you have a good example portrait picture ?

[strong field]

actually i take back what i said, 30 isnt enough, refer to this guide, it will give you substantial detail on how to accomplish your goal

#🔧｜finetune message

[molten gull]

this may be a good training picture? and thanks for the link

[strong field]

i hope you have 24gb

[strong field]

this looks decent however the quality of the picture will be important

[molten gull]

yes, got 24gb

[dense chasm]

when you finish lora training with base model,you can render more details with refine model and upscaling nodes

[ionic gulch]

why do so many people dislike Refiner? it helps in many cases imo - without refiner i get a 35 yo woman, if i ask for 50 yo for example:
https://i.imgur.com/f0MvLDT.jpg

[high skiff]

I would say this is not a good training imsge

Obviously I don't wish to give away all of my secrets, as I am already training what you are trying, but I would not use that image

My current dataset it's 90 images, and while I get some good results, it's inconsistent

My goal, if I can get there, is at least 500 images

That's really all I will say, everything else I would respectfully say is up for you to figure out in order to compete honestly

[molten gull]

Any hint why it’s not a good image ?

[high skiff]

I felt the same way before, but I now realize it was more of an inconsistent and unreliable crutch to lean on for sub optimal prompting, reslly

[high skiff]

Too close and not really high quality skin detail in general

[ionic gulch]

the main problem are the eyes imo - at least at 80/20 split they often come out bad with refiner

[high skiff]

Yeah, the refiner trashes eyes

I suggest ditching the refiner and messing around with prompts until you get good results

When you do, they will be far better than the results you would normally get even with refiner

[ionic gulch]

i made a workflow that generates 4 versions: 80%, 87%, 93%, 100% base. so i can pick the best one - and it doesn't take that much longer, because i don't generate all of them from 0% and continue from the last step instead. so only the last 20% have to be done twice and the vae decoding 4x.

https://i.imgur.com/WzwRYik.jpg

[high skiff]

I will really honestly just stick to my assertion that the refiner is more of a negative crutch for bad prompting than it is any form of reliably better result

Especially for photo realism and painterly images, as the refiner kind of mangles fine details associated with brush strokes and skin texture

[glad fulcrum]

Okay I am trying yet to use control net open pose... but the workflow says I am missing the preprocessor, then with the manager I click on get missing nodes and it leaves me to this:

[high skiff]

My 1.1 release of my workflow will be ditching the refiner entirely in favor of just using the base, which is faster, more efficient, requires a few less steps, maintains consistency and compatibility with LoRA's from the first step to the very last step of high-res fix, uses considerably less VRAM and is more accessible for people on lower end GPUs, and overall improves pixel level quality considerably

[glad fulcrum]

this is stressful

[ionic gulch]

i like the concept of expert models and splitting models at x%, but other aspects have to be worked on. refiner is problematic if things like loras don't fully support it.

[high skiff]

I am sure people's dependency on the refiner stems from my earlier releases of my workflow, where I sung my split technique as being optimal, which four a workflow that it is using the refiner, it is optimal

But I have now found that the refiner is more deadweight than useful weight

[ionic gulch]

means the missing node is not registered at comfy manager i guess

[glad fulcrum]

okay after 1000 tries it loaded, wth

I don't know why when comfyUI loads it opens up my audio driver though 😛

[soft zealot]

see generally speaking I'm happy with using refiner although I would agree there is no magic bullet and prompting helps.
Personally I use a 3-5 step flow with the last 2 being optional

Precondition
Base
Refiner
Upscale
Face Detail

NB I 've also tweaked my flow so I can use it with SDXL or SD1521 Models

Mind you this is the joy of it all, there is no right or wrong way, its what you;re happy with IMHO and yes YMMV

This is STable Diffusion, This is The Way

🙂

[high skiff]

I will likely have to provide proof of my assertions that the refiner does more damage than it does good, and I'm prepared to release example images along my workflow, as the faster we can remove people's reliance on the refiner, the sooner we can get people who know better how to prompt, and subsequently rely more on the base which is more compatible with additional networks

That's my view on the matter at least

[soft zealot]

In Fariness many of the Custome XL models do say not to use the re SDXL refiner although all I sdo there is use the same model in the refiner step as I do in the base step.

(minor note I now refer to them in my workflow as Step 0,Step 1 & Step 2 rather than Precon.Base,Refiner as I ve found its fun to mix and match models at different stages 🙂

[high skiff]

That's right, I forgot you still do preconditioning

Preconditioning is pretty neat, but I found that it has an absolute mind of its own, and often explodes some subjects lol

But when you do get something interesting with preconditioning, it an be pretty cool

[soft zealot]

Aye, I just like it 🙂

But anyways as I say in my own mind (and in my workflow) I have now moved away from the terminology I was using to a (in my mind) more accurate method of describing it it

I use multiple Ksamplers in a sequence which are rcounted starting at 0

Hmm should I rename Upscaled as Stage 3 & Face Detailed as Stage 5 I wonder lol

[soft zealot]

these are the 3 outputted images for completeness 🙂

[dense chasm]

u can control and easily generating darker or light images by offset the noise when fine-tuning the model,it's very interesting when you set noise_offset_num with different values

[ionic gulch]

i prefer pre-conditioning with img2img at a very low percentage - this gives you some control over colors and composition.

[soft zealot]

and same prompt & seed through the DynaVision lens 🙂

[high skiff]

that image looks aggressively 1.5

[soft zealot]

in the foreground possibly, less so in the background

Its this one

https://civitai.com/models/122606/dynavision-xl-all-in-one-stylized-3