✨｜sdxl

[meager canopy]

There are still 1.5 tunes/merges that can produce better images than SDXL

[crisp owl]

1.5 has a lot of utility built around it, and can pump out good quality images quickly. More than enough for casual users.

SDXL will far exceed what 1.5 can do, but doesn't mean 1.5 is gone

and the fact that the more popular platform, a1111, cannot even fully support SDXL yet

[upbeat summit]

yeah it's amazing what the latest 1.5 models can do. fidelity, coherence - on a very high level

[hardy cipher]

that's not really the point. "better" is relative. I prefer xl, but doesn't mean I think 1.5 is suddenly useless now

people still play nintendo

[crisp owl]

damn right

[hardy cipher]

people still buy vintage crt televisions

[meager canopy]

https://tenor.com/view/unbelievable-smh-facepalm-seriously-gif-24551485

[hardy cipher]

and 1.5 still has it's strengths. uses fewer resources. versatile, etc

[thorny frost]

Well, thats true. I still have my vinyls.

[upbeat summit]

how else should I play retro games?

I need my real scanlines

[meager canopy]

Trousers?

[hardy cipher]

sony trinitonics still sell for quite a bit actually

[thorny frost]

But u think SDxl images are consistently better in the same conditions. (Not Controlnet and such)

[hardy cipher]

they have a higher win rate

[upbeat summit]

because they were the pinnacle of CRT TV tech - just the best

[thorny frost]

Both. LoL

[hardy cipher]

with all the tools and accessories surrounding 1.5 I just don't see it going away. if anything I could see it being more integrated into workflows with sdxl as time goes on

[upbeat summit]

yeah. multi-model processing is definitely the way to go imo

I've seen some SDXL upscale workflows using SD 1.5 models - it works quite good

[hardy cipher]

1.5 can create quick rough drafts

[meager canopy]

Like this one? 😄
https://civitai.com/models/119257/gtm-comfyui-workflows-including-sdxl-and-sd15

[upbeat summit]

Exactly 😄

[hardy cipher]

everyone into those low contrast images though. not my preference

[crisp owl]

I love contrasty images, those are my jam

[hardy cipher]

yes, hdr is all good. but when the high and low are so close, that's what I'm not a big fan of

and also that highly blurred background thing

cinematic

not really my thing. it has it's place. but I like everything to pop

[crisp owl]

yeah sdxl pushes the bokeh pretty strong if you let it

[hardy cipher]

people think it's stuck in bokeh just don't know how to prompt though

[hardy cipher]

you have to tell it what to do

[crisp owl]

yup for sure

[meager canopy]

It's still so disappointing at hands though 😦

[hardy cipher]

it recognizes patterns

doesn't understand how the human body works

[ionic dragon]

yep

[hardy cipher]

so if it decides the random noise needs 7 fingers, then you get 7 fingers

[ionic dragon]

my selena, my wish

lol

[meager canopy]

Strange how light always shines through belly buttons 😄

[upbeat summit]

it's just a broken pixel on the latent space cam sensor

[meager canopy]

This image too

[crisp owl]

Probably just a bellybutton ring shining 😆

[hardy cipher]

I find people's acclimation to the incredibly quick progression of text-to-image AI almost as impressive as the progression of the technology itself.

stable diffusion hasn't even been out for a year, lol

15 months ago text to image was weird, uncanny horrors

[crisp owl]

just a bunch of daydream images, or super janky dogs and cats

[hardy cipher]

man, I started on deepdream, lol

and then image on image. you had to earn your wins back then

[crisp owl]

I tried using it, but wasn't a fan of the outputs, so just abandoned trying anything further really.

[molten gull]

is there a way to send a request to my locally running comfyUI server to check if it's still processing something or not ?

[clever verge]

Check if the queue is empty?

[meager canopy]

I started using MJ about 13 months ago and then switched to SD the day it became available. Been prompting pretty much every day since!

[hardy cipher]

you mean if you're not home?

I just have my outputs folder synced with google drive

[high skiff]

Great job civit

[molten gull]

no, i m starting the thing via my python script, but it returns immediately (asynchronous)

and i want to create a second image once the first is done

but i want to wait until the first is done

but i dont know how to figure out if its done

[meager canopy]

My very first AI image, created on MJ.

an epic battle between heaven and hell, cinematic

[unreal plover]

i was amazed by disco diffusion images

for who know what it is

[molten gull]

grats for making your first image galaxy 🙂

have fun

[meager canopy]

That was 13 months ago 😄

[bitter swallow]

Does anyone know what kind of settings do the bots on this discord run?

[meager canopy]

The same prompt from SDXL

[hardy cipher]

let me see what my first ai image was. it was a real winner I'm sure

[unreal plover]

i did this +1year ago with disco diffusion

30 minutes of diffusion

[molten gull]

anybody got an idea regarding my comfyUI problem ? 🙂

[upbeat summit]

yeah. still blows my mind. from 30 minutes for an image, down to a few seconds locally. there aren't a lot of advancements like that

[crisp owl]

Not a clue

[hardy cipher]

first image

doges and cats

[meager canopy]

Gone a bit nuts with the upscaler denoise there! 😄

[hardy cipher]

that was like 2 and a half years ago

[jovial flower]

That looks like something straight up out of an Escher painting.

[hardy cipher]

dude

there weren't options like denoising, LOL

[vocal rapids]

Definitely! lol

[hardy cipher]

how long are these images taking?

my cats and colored pencil scribbles. the image on image was better than deepdream by a lot, but still a chore

[meager canopy]

Some of my old MJ prompts are amazing in SDXL!

[upbeat summit]

nice!

yeah I've been going through my MJ prompts as well 🙂

[hardy cipher]

I need to change to 3x2 or 4x3 ratio

[molten gull]

images take a different amount of time, depending on size for example, the last ones like 28sec

[molten gull]

mount doom?:)

[ionic dragon]

which one?

[molten gull]

neither

[hardy cipher]

malicor, so what is the situation exactly?

[meager canopy]

Supposed to be a Phoenix hovering over an erupting volcano

[molten gull]

i call "python runme.py", which will then call the comfyui server and generate an image

that all works fine

but the "python runme.py" exits immediately (it seems to send the call to the comfyUI server asyncronously)

now i want to do a loop and generate 5 images

but i want to do the second call once the first image is finished processing

question: how do i figure out when the first image is finished processing?

[ionic dragon]

@hardy cipher

[meager canopy]

The first one, because 2nd has slightly distorted eye...but base SDXL creates better versions 😄

[hardy cipher]

2

I guess it's a draw

for now

[molten gull]

cool helm 🙂

[hardy cipher]

I've determined it's not worthwhile to even try loading checkpoints off my external hard drive

[heady vale]

why not request a batch size of 5, then all 5 complete at once

[hardy cipher]

I mean, it's cool for those times I want to wait 5 minutes to load

[ionic dragon]

prompt pls

[ionic dragon]

no, sdxl base isnt creating selena gomez's face

[meager canopy]

#✨｜sdxl message

[hardy cipher]

how many images of ms. gomez ae you using, qwerty?

[ionic dragon]

you think this is great?
she doesnt even look like selena

[ionic dragon]

for training?

[hardy cipher]

are you using them for other things as well?

yeah, just curious how many you're using

[molten gull]

i can look for the prompt later, it's somewhere deep within the code and a pretty complex workflow

and i dont want all images to be sent to the comfyserver at one time, since it could potentially be hundreds

[hardy cipher]

hmm

there should be an easy answer

[inland vessel]

me and my friends generating this in 2020 for like 4 hours and thinking the result was cool af

[hardy cipher]

it's not bad for 2020. I actually didn't see your comment at first and thought it's something someone just made, lol

[inland vessel]

this was the best we could get out of shrek prompt

and we were truly thinking this was amazing (in a funny way)

[meager canopy]

Anyone tried this? It's really easy to setup and use, great for beginners and MJ fans:
https://github.com/lllyasviel/Fooocus

[hardy cipher]

well that's what I'm saying. any of our random generations now would have taken some serious skills even a year ago

[inland vessel]

also the fact that it took hours to create

[meager canopy]

Just created these using Fooocus

[upbeat summit]

These are beautiful!

[upbeat summit]

So Fooocus uses some post processing stuff in the background

[meager canopy]

Yes, it's listed at the bottom of the readme/main page in the repo

Does something different with samplers

[upbeat summit]

the negative ADM guidance sounds interesting

[ionic dragon]

160

[hardy cipher]

dang, son. did you tag all of them?

[hardy cipher]

I'm looking at the code. not really understanding what they're doing though

[upbeat summit]

and how can I go back to such a minimal interface now that I'm working "under the hood" in ComfyUI

[hardy cipher]

nodes satisfy me

[upbeat summit]

I mean it sounds like that they have some ideas, so let's see

[upbeat summit]

nice hologram star map table 👌

[hardy cipher]

yeah, just doing my normal thing

[vagrant wasp]

is it possible to create a json data of difference traits of images generated

proly some python coding

would be nice to have a node for it

[hardy cipher]

testing out same parameters with different last layer

[hardy cipher]

that does sound reasonably useful

[molten gull]

this seems a bit dangerous to me 🙂

[hardy cipher]

nice t23

[unreal plover]

SDX is really good at low poly

[hardy cipher]

you should try this out https://civitai.com/models/118536/voxel-xl

[unreal plover]

Look the name of the autor of this lora

[hardy cipher]

oh nice

well good job

I use that one too

that's pretty awesome,l ol

how do you train something like that? I'm curious

[ionic dragon]

yeah

[unreal plover]

I make voxel stuff with magicavoxel, so i generated some images and trained on them

[hardy cipher]

how many images did you use?

I'm just curious. doesn't have to be exact numbers

just want to start making things in that realm

[unreal plover]

Come, i am ready

[ionic dragon]

how is this?

[hardy cipher]

but not sure if I need to tag them, or what methods to use

mostly just did loras of people in the past

[late marsh]

Coming!

[hardy cipher]

that's good

[unreal plover]

Lets go

😅 that fail you win

[unreal plover]

I used not many images, the important thing is variation, 2 cars, 2 buildings, 2 persons, 2 animals...

[hardy cipher]

that makes sense

[unreal plover]

But if i found Voxel artists who like AI (here is not many) we can surely make a way better lora working together

[hardy cipher]

what do you use for them?

to make the voxel images

[unreal plover]

magicavoxel

[hardy cipher]

cool

[unreal plover]

but i also have an UE4 voxel project for landscape

using voxel plugin to generate procedural landscape

[hardy cipher]

now how do we make a magicavoxel node?

[late marsh]

I'll catch you 😛

[hardy cipher]

this is a whole new realm for me. didn't even realize it was a thing. just thought that lora looked cool

[unreal plover]

What ? Magicavoxel is an app to make voxel asset, and then you can render it

[hardy cipher]

well yeah, didn't mean it'd literally be the fully functioning program in a node

[hardy cipher]

Sylvester stallone and robert downey jr's love child?

[meager canopy]

I see Tom Hanks in there

[upbeat summit]

extremely art deco a hyperrealism photo portrait of a (upset:1.25) pilot played by a person looking like Tom Hanks waiting for someone at midnight during pre-winter, power auras, sigils, (dark and gritty:1.35), face symmetry, intricate accurate details, cinematic color grading, cinematic, artstation, 8K

using the upcoming NightVision XL model

[hardy cipher]

please make image of wholesome hollywood actor

[late marsh]

@unreal plover Cought ya

[unreal plover]

Are you sure ? You can't beat me, little wolf

[hardy cipher]

I should try out another seed. I just get so attached to them sometimes. and it's not likely we'll meet again by chance

[late marsh]

How about silver bullets?

[hardy cipher]

decent

[late marsh]

Or lasers! 😂

[unreal plover]

i like your prompt style

[late marsh]

Mostly they are just ready made styles or combinations of them. Nothing special 🤷‍♂️

[ionic dragon]

@late marsh do you train loras?

[late marsh]

Nope. I know nothing of training this stuff

[late marsh]

Few days ago

[upbeat summit]

the likeness is pretty good here

[unreal plover]

hands are always the final boss

[soft zealot]

I started looking last night and got annoyed with it (and in the poocess may have annoyed others)

Having come back to it this morning.

Pros:
Its a clean simple interface with very few options even in advanced mode

Cons:
Its a clean simple interface with very few options even in advanced mode

Feedback:
Should make it clear in big letters at the start of the page that this is locked to one model only Should also ideally ask during the install if it's ok to download specific files rather than force downloading them. (NB have added an apology on the issue I riased regarding that point https://github.com/lllyasviel/Fooocus/issues/31)

**Summary:
**Produces nice images and almost gives an MJ like experience running locally , which is fine if thats what you 're after

EDIT Added 11:30::
It does allow use of additional models which makes the forced download of specific models even more annoying however given the target demographic.....

[upbeat summit]

Thanks for the review, Winston!

[meager canopy]

Is it locked to one model? In advanced mode you can choose other models.

[soft zealot]

TBH f that is the case then I am even more annoyed that it foces you to download a specific model to start with.

[meager canopy]

You can copy models into the directory before you start it, like it says in the readme, and then it doesn't download them.

[soft zealot]

you have tpo copy the specific models they list, if those models arent there thenm it force downloads them

[meager canopy]

Oh, ok. Easy workaround with some empty files with the same names...but you shouldn't have to. I think they're just making it easy for beginners, which it does a good job of.

[soft zealot]

see pros & cons and summary

[unreal plover]

by Diego Gisbert Llorens

[ionic dragon]

was that finetuned on SD and MJ images?

[soft zealot]

eh??
Its a UI that uses the base SDXL1 with 0.9 VAE model by default.

THe question seems "odd"

[ionic dragon]

oh, then how is it generating MJ quality images?

[soft zealot]

Iss it ??

Who says it is ?

Wasnt me

[meager canopy]

The experience of using it isn't the same as the quality of the images.

[ionic dragon]

mb

[vale eagle]

Tried and I think it is just a custom workflow of models with selectable style and aspect ratio.

[soft zealot]

hence the "hidden tricks"

which is all stuff the authir has hard coded as it is their opinion these are the optimal settinsgs.

Which is fine if you want the MJ like experience of having little control ;o)

and in case anyone thinks I'm dumping on it i@m not.

This could be the perfect solution to point newcomers at that want to dip their toes in the water of generating images Locally rather then being forced into an online only environment..

[west breach]

I've put this what I hope to be a fairly straightforward workflow together using some custom nodes I made. The sampler does a base -> refiner -> upscale -> base process. There's also a tiled version of the sampler which does the 'hires fix' using a tiled approach

[soft zealot]

Its a real shame you cant define multiple colours in one node (or can you?) then you could colour the Prompt With Style say Grey witht he Positive Prompt Input area Green and the Negative Input area Red

[west breach]

That would be cool

[soft zealot]

actually thats a thought. Can Widgets not be given Titles that display above them?

[west breach]

I think the text input is the only widget without a label

[unreal plover]

very clean

[west breach]

thanks

[vale eagle]

Does anyone try using different prompt for base and refiner?

[iron sonnet]

https://giphy.com/gifs/disney-tarzan-jungle-8AsVmBajCBHPy

[heady vale]

Neanderthal chicken man 😆

[vagrant wasp]

which sampler and scheduler recomended

for mixing

why does my layer turns to cartoon

[heady vale]

add negatives so it cant do cartoon

[soft zealot]

any ancestral based sampler will chnage more than a non ancestral based sampler

[vagrant wasp]

did

[soft zealot]

try using (as a test cos at the end of the day its down to personal preference) DDIM & DDIM-Uniform

[vagrant wasp]

same seed problem ?

[heady vale]

what are you trying to achieve at the end of the sequence? maybe its too high denoise

[upbeat summit]

betrayed by discord's auto gallery crop

[dense chasm]

sdxl controlnet canny is good

[soft zealot]

you get the ides o f the sequence though lol

[upbeat summit]

hehe yep - good sequence 👌

[soft zealot]

I'll be honest, it wasnt planned and that wasnt the order they got generated in but I'll take happy accidents 🙂

[upbeat summit]

well a film isn't shot in order as well, but you still try to tell a coherent story 🙂 it's great

[autumn forum]

That’s a lot of buckets lol. More buckets you have the more vram is used. But I haven’t experimented with that many before.

[queen adder]

from the tutorial i was following it said you don't need to crop or adjust the resolution because it self sorts into buckets automatically.

but is the issue that the resolution is too high? or that there are too many different resolution ratios?

[tribal jackal]

bruh

[heady vale]

price and name check out

[tribal jackal]

WIld though. 48GB VRAM is three 4080's plus it's DDR6.

[heady vale]

30% faster vram than a 4090... I guess that will go into the rtx 5000 cards

[tribal jackal]

It'd also come out to $2,200 per 4080 to get them in one pcie slot

[thorny frost]

was trying some experiments using SDXL fine tuned as base and SD1 fine tuned as refiner. thehive takes it as midjourney always

[tribal jackal]

Well, per 16GB VRAM

[thorny frost]

I think that leaves little room for doubts about how midjourney works

or people is training with images from midjiourney.

[tribal jackal]

MJ simply can't provide the flexibility of an open source project no matter what they do, so they supplement with convenience.

[dense chasm]

sdxl controlnet make a shock on architecture

[thorny frost]

midjourny has zero interest for me.

[tribal jackal]

It was fun when it was the new kid on the block, but I knew then I wasn't going to keep paying them a monthly fee for limited generation speed and rush hour congestion on requests.

[thorny frost]

Wow didnt know has been released. Downloadaing...

[tribal jackal]

Cries in GPU RMA

[dusk mica]

MJ prompting is much easier imo

[vagrant wasp]

youtube guy doing western to asian face

[heady vale]

ahh Olivio. I see

[queen adder]

aside from training, why would you need this much VRAM?

[tribal jackal]

commercial uses

[dense chasm]

so cool

[tribal jackal]

I like the detail it managed to pull from the eye despite the shaded feathering

[dense chasm]

LoRA with controlnet,then upscaling with refine model,compete with midjourney

[queen adder]

is SDXL refiner model still not working properly in A1111?

[autumn forum]

the resolutions are fine. the amount of buckets will cause a slowdown. are you training a full model im guessing? also could just be the amount of images lol. it takes me 1 hour for about 1500 steps or a little more.

for lora training that is. with adafactor.

[queen adder]

i'm dreambooth training using SDXL as the base model. 750 images because it's a style and not a single character, otherwise Lora probably would have worked. adafactor as the optimizer

[autumn forum]

tbh, styles are very easily done with loras and you could merge that with a model that is close to it. or sdxl base. but straight dreambooth training may make better results. but i havent done a full dreambooth train before. just loras. you have sooo many pics so im guessing 5 repeats? whats is your batch size?

[queen adder]

batch size: 5
training images repeats: 3
regularisation image repeats: 1
epoch: 1

[dense chasm]

https://www.reddit.com/user/FrankChieng/comments/15pzzq0/sdxl10_controlnet_canny_model_is_released/

[autumn forum]

okay maybe its the batch size then? i do max 3 so i can still use my computer. lol. anything more and it maxes out vram and takes forever lol. i hope you can figure it out! im sorry im not much help

[queen adder]

thanks

[thorny frost]

Nope. There is an extension though. Doesnt support controlnet either. A1111 must rise its game or is going to be eaten alive by ComfyUI

i mean sdxl controlnet

[queen adder]

yes im thinking of learning comfyUI. i wonder if the Loras trained on SDXL you are still expected to use the refiner model with it?

[thorny frost]

No problem with sdxl loras afaik

[rustic garnet]

usually not.

[autumn forum]

The refiner is kind of hit or miss. Never expected to use but you can. Works great with base sdxl but sometimes not so much with custom models

[queen adder]

i'm dreambooth training on SDXL model now. from my understanding it will still retain the previous characteristics of the model and wont be a completely new model from scratch?
the reason i ask is because the images i'm training are of people. so if i want to generate a car or a house that will still be possible?

[clever lintel]

It just seems incredible!

https://twitter.com/marcsnv/status/1690733749699297280?s=20

Marc on Twitter

I've just updated the models, and this particular model continues to astonish me. (it is used by +700k people)

Twitter · Aug 13, 2023, 2:35 PM

[autumn forum]

Yes it will still retain characteristics of the model. Your basically just continuing the training I believe. Like adding onto it

Looks anime ish to me, but just not enough steps so it’s a bit noisy

[vagrant wasp]

comfyUI just crash

when clicked Queue

[half ivy]

I am here
I can ask questions
HMU 🤙

[hardy cipher]

what kind of questions?

[late marsh]

Nice

[upbeat summit]

@indigo carbon btw launching ComfyUI using --gpu-only and AIT the LORA lag is gone - instant generation. but after a couple of minutes the performance when using --gpu-only totally breaks down on my computer - like 1 step every 5 seconds with GPU at 100%.

but this also happens without AIT. I need to restart ComfyUI to fix it.

[hardy cipher]

lora lag?

that's one thing I haven't dealt with

[upbeat summit]

if you use a LORA in your chain and AIT it takes up to 4 secs for the generation process to start

every time

[rustic garnet]

the refiner is like an img2img pass with another model. So if you fine-tune the base model and apply the refiner, you might lise details the base model was trained on

[upbeat summit]

and it adds up if you chain multiple loras together

[hardy cipher]

I was saying last night, it's not even worthwhile for me to load checkpoints from my big external drive. I've started to transfer models over to my primary drive and then load them

I wonder if it's an issue of where you have them stored? or maybe I'm totally off base

[vocal rapids]

canny model for SDXL is so creative, lol

[indigo carbon]

my best guess is it's likely because ComfyUI likes to keep LoRA in CPU RAM, --GPU-only forces it to place the LoRA in VRAM, which AIT modules are kept there in an efficient way which the LoRA isn't, so it's likely we need to compile modules for LoRA as well

[upbeat summit]

yeah. just wanted to let you know that --gpu-only fixes it. other people are not having the "performance breakdown" problem as I have.

[bold osprey]

I'm going to try that out because the lag is very noticable

[hardy cipher]

the one place my 64gb of ram actually helps, lol

[upbeat summit]

it's 100% gone for me but I can't leave it running :(. be aware there might be other side effects when you use --gpu-only. I had a Out of Memory message when I switched to many models. looks like with the option more stuff is kept in vram

[upbeat summit]

was that a reply to me, pictures?

[hardy cipher]

yeah

[bold osprey]

does it ever clear the gpu cache?

[upbeat summit]

good question

[hardy cipher]

but sounds like it might be a ram thing then?

[upbeat summit]

I have all my AI stuff on a nvme. so it's not a loading problem I suppose

it's a processing thing especially with AIT

without AIT Loras are much faster

[vagrant wasp]

comfyUI image previews not updating

[indigo carbon]

likely because AIT doesn't support LoRAs yet.. this personally doesn't effect me because I don't use LoRAs

[hardy cipher]

huh?

why would a person not use loras?

[upbeat summit]

yeah I do lots of testing for other fine-tuners and I guess it is really needed for the scene 😉

[indigo carbon]

I'm generating stuff like this in 18-ish seconds without LoRA, I'm happy with the quality of my images, and I never use LoRAs

[hardy cipher]

is speed the deciding factor with this stuff though? are you needing 100s or thousands of images like that?

I mean, not knocking your images. they're pretty on point

[vagrant wasp]

finally it works.. layering

[hardy cipher]

just saying. for me it's not all about speed. maybe when figuring out my approach speed matters more

[indigo carbon]

yes, and the details are already amazing. LoRAs don't improve that much when you use a proper workflow..

[upbeat summit]

no it's not all about speed. I want image fidelity. but having 200% speed is great too 😉

[indigo carbon]

@hardy cipher, however, if you're so arrogant about needing LoRAs you can just wait until we compile AIT modules for SDXL LoRA.

[hardy cipher]

arrogant? lol

you're the one acting like your workflow is top notch, and you post the same 3 images over and over

just saying

I didn't say anything negative about you. I was just curious

but if you want to be rude

[meager canopy]

Settle down guys/gals, we're all just her to have fun. Just stop and post some pics 😄

[indigo carbon]

this is just not true. I never claimed my workflow is "top notch", but speeds are factual; aesthetics are opinional..

[hardy cipher]

yeah, please post super soaker and see through gelatin cube things

[half ivy]

ahh, heres an example, if you que say 20 imgs
and just before teh 2nd image initiates, you change settings, does it change to the new settings on the 2nd img?

[hardy cipher]

I could render 6 step images in a couple seconds

[meager canopy]

No, once it is queued it has all of the settings and prompt determined already.

[indigo carbon]

and I can generate 111 steps upscaled images in the same amount of time..

[hardy cipher]

but no loras because speed?

[half ivy]

thats what i thought, i have moar btw

[hardy cipher]

anyway, I really don't care. I just wanted to hear your reasoning, but seems I've offended you, lol

so forget I asked

[indigo carbon]

no, it's just we haven't compiled LoRA AIT modules yet. will be soon though.

LoRAs won't effect speed after AIT supports it..

[half ivy]

gonna roll one, but im here for q's if ya'll need

[hardy cipher]

I don't think loras have much of an impact on speed anyway

at least not for me

maybe a few percent slower if I'm running several

they take almost no time to load initially

[indigo carbon]

actually, know what? I'm going to compile LoRA AIT modules rn, the compile script is universal anyways atm, so yeah

[hardy cipher]

right on. well go ahead and do that so you can start using loras

[half ivy]

ping me yr best workflows

is CN out?

[hardy cipher]

canny

[half ivy]

eww

[hardy cipher]

my workflows are mostly trash, not gonna lie

[half ivy]

stsrt though

nope

liar

[hardy cipher]

I'm sure someone somewhere has something else going. but I haven't heard about it

[half ivy]

ive been afkbut yr a fav

humble

modest

[hardy cipher]

this one was interesting

[half ivy]

bro

damn

knollingcase pls

straight cat scan vibes

[upbeat summit]

if you are able to do that - AIT speed boost for loras. that would be awesome! 🙂 there are different methods how loras are used. some are loading them in the prompt field and comfyui's default method is using the lora loader and chaining it after the main model.

I tried:

base model > AITemplate Loader > Lora > Lora > Lora... > Sampler
base model > Lora > Lora > Lora... > AITemplate Loader > Sampler

but it didn't made a difference

Not sure if every Lora would needs its own AITemplate Loader?!

[hardy cipher]

[glad fulcrum]

my training on faces don't get good resemblance OR get overtrained. How can avoid this?

[hardy cipher]

i was trying to get it to render subatomic particles, lol

[queen adder]

for captioning images, is there a clean UI setup where you automatically have the image and caption prompt side by side? im tired of having to open the image, then the txt file, and line them up side by side.

[hardy cipher]

that is tedious

[half ivy]

tbh ive tried that prompt alot and whatever model that was did guud i think

[hardy cipher]

prompt, model, a couple styles, juggling loras, skipped a few layers, proper noise offset. it's always a bit more than I can really handle

I just try to make things I've never seen before

btw, if you ever want to see any of my goofy workflows they're normally attached to my images in here

[upbeat summit]

what is your workflow, GTM? so good

[hardy cipher]

almost all of them these days

[hardy cipher]

I just noticed the blue guy in the background

[meager canopy]

Supes 🤷🏻‍♂️

[meager canopy]

I'm using the one here https://civitai.com/models/119257/gtm-comfyui-workflows-including-sdxl-and-sd15

[upbeat summit]

hey I know that person 😄

[hardy cipher]

awesome

[meager canopy]

a scene from "The Boys" netflix series, an angry looking Erin Moriarty as blonde starlight in her white superhero outfit. __location__

Wildcard location at the end there

[upbeat summit]

I'm so stupid - sorry. You have posted it many times :D. It's just too many stuff going on.

[hardy cipher]

did you see last night/this morning when I suggested this lora to fictiverse? https://civitai.com/models/118536/voxel-xl

LOL

anyway, then noticed this

[upbeat summit]

I'll bet he loved it

[hardy cipher]

yeah, I'm glad. it's quality. and it opened up a whole new world to me as I didn't really know what voxel even was

at least not by name

I think anyone that creates these sorts of things should be proud of themselves. whether 5 people like it or 50,000 people use. it. I support anyone putting themselves out there like that

[indigo carbon]

@upbeat summit I looked into it, AIT uses the same modules for base on LoRA, and it works! it's same architecture, so no need to compile for LoRA xD

idk about the delay you mentioned though, didn't experience that

[hardy cipher]

even if it's not my thing, still support it. that's one of the reasons I've stopped looking into the stable diffusion subreddit. so many people hating on others that are putting things out for free

[upbeat summit]

did you read my info here #✨｜sdxl message

[hardy cipher]

I need to read up on the ait thing. I only vaguely know what it is

[indigo carbon]

I can make a LoRA version of my AIT workflow if you guys want, it's very simple to do

[upbeat summit]

I talked to a couple of people who experience the delay...

• you start a queue
• than it loads the AIT module (shows the module guid in the console)
• it hangs for exactly 4 seconds
• than it starts to generate

this is on every image

it doesn't do it if you use comfy with --gpu-only

than it's instant

[indigo carbon]

odd. this sounds like an issue with ComfyUI itself

you can open an issue on the AIT PR though

you will get answered very quickly

back to some gens =]

[eternal fog]

Yeah I don't think that's AIT, it does it normally as well. For me the first time I load up a LoRA its so bad it tries to freeze my PC, but once I've done it once it seems like it's alright. It looks like it tries to load up RAM the first time, hits the maximum and pages. And then next time it doesn't do that for some reason.

[hardy cipher]

so is it maybe a ram issue?

[eternal fog]

I did discover though with AIT you can't blend prompts, like have 1 prompt for 5 steps and then another prompt for the remaining 20.

It just errors if you try that

[hardy cipher]

fancy

[eternal fog]

Maybe, I have 16GB RAM and pretty much all the issues I ever have with SDXL are because it runs out of RAM

[hardy cipher]

yeah, that might be part of the hangup

I'd say 32 is the number

[upbeat summit]

I see. for me it's with every gen. every time before the diffusion process

[eternal fog]

AIT seems good if you have a simple setup and you want more speed. But if you are trying anything fancy, it just causes issues.

[hardy cipher]

I just upgraded from 16 to 64 because it was really a nominal cost difference

32 vs 64

[eternal fog]

How were you loading the LoRA, I was doing Model into LoRA into AITemplate

[indigo carbon]

that's how I did it, works just fine =\

[eternal fog]

It does freeze on lower RAM setups, but so does regular LoRA, so it's expected.

I'd get more RAM but it feels annoying to buy more DDR4, when I'll probably upgrade at some point and then need to get DDR5 anyway

[hardy cipher]

I just like to be able to run ram hog programs like photoshop and still be able to do other things without my computer slowing down

[upbeat summit]

I tried both and it didn't make a difference

base model > AITemplate Loader > Lora > Lora > Lora... > Sampler
base model > Lora > Lora > Lora... > AITemplate Loader > Sampler

[hardy cipher]

although comfy honestly does a lot of the things I do in photoshop in a more efficient way

[eternal fog]

Are you always using multiple LoRAs? Have you tried just 1?

[hardy cipher]

blending and things

[upbeat summit]

of course - same thing 🙂

[eternal fog]

Let me double check what my setup does

[vale eagle]

I also encounter issue using AIT. After a serval gens, I need to restart comfyui to let it work again.

[upbeat summit]

when I was doing SD2 images I used up to 8 - 10 TI embeds in one image and I'm sure I will do the same with Loras at some point

[eternal fog]

Oh yeah a major issue I have with AIT, it's the actual Comfy Module. If it ever crashes, it doesn't unload the model. So then I run out of VRAM if I try render again without restarting.

[hardy cipher]

8-10? that's impressive

[strong field]

I manage to get AIT working with Sytans workflow

[upbeat summit]

yeah. TI embeds were to me like ingredients and I mixed them in the latent space

[eternal fog]

Yeah, it's not difficult to do that. I have a workflow that I made that's very similar to his and it works fine.

It sometimes goes a bit strange on img2img

[strong field]

Steps and denoise are very important AIT

[hardy cipher]

yeah, my issue with the LoRAs after a point is it seems to get exponentially more complicated to get them to all work together properly

[indigo carbon]

yeah, in the workflow I made it's already set on very nice settings.

[hardy cipher]

I also forget what's what with them since they're all named random nonsense

[strong field]

I’ll do more testing with img2img, it’s not typically part of my workflow atm

[strong field]

For sure but I’ve managed to improve it :p

[eternal fog]

So yeah @upbeat summit first run with a LoRA, it maxes out RAM and takes ages

Prompt executed in 76.39 seconds

But if I run it again after

Prompt executed in 17.10 seconds

Slower than without a LoRA, but it doesn't take "ages"

[upbeat summit]

some of my SD2 stuff - all raw txt2img outputs using TI embeds and prompt magic 😉 (a1111 metadata is included)

[hardy cipher]

yeah, those are good

[strong field]

Wow those are great

[eternal fog]

Without a LoRA it's Prompt executed in 9.28 seconds

[hardy cipher]

very diverse

[upbeat summit]

Thank you 🙂

[upbeat summit]

hmm interesting. as said it hangs 4 seconds on every gen for me.

[eternal fog]

Yeah does seem like it's a little slower with a LoRA.

Ran it again without AITemplate with the LoRA and it's 3 seconds faster

Prompt executed in 14.05 seconds

[upbeat summit]

my image gen is not slower with or without lora. it's on the startup of the gen. when it says Using c102dk095kjmfoi29834... in the console / the AIT module guid than it hangs for 4 seconds. after 4 seconds the image is generated and it does it every time

[eternal fog]

Yeah, I mean that accounts for the entire process with AITemplate and a LoRA taking 4 seconds longer.

[strong field]

Are you talking about first green load into ram?

[upbeat summit]

ah I see. if I bypass the Loras there is no delay.

[strong field]

Hmm interesting

[upbeat summit]

nah it's on every run for me

if loras are in the chain it takes an extra 4 seconds for the image gen to start

as said before if I start ComfyUI with --gpu-only the delay is gone when using LORAs, but in this mode ComfyUI doesn't run stable for me

as Tdg8uU said I will debug this a little more and open an issue on github

[strong field]

My current workflow has Loras, two of them in chain but I haven’t noticed slow downs

[eternal fog]

Sounds like maybe the delay is it moving the LoRA from VRAM to RAM or RAM to VRAM

[upbeat summit]

It's hard to say which factors play a role here. could be my hardware, maybe the python version, maybe a node in my workflow

[upbeat summit]

yeah, but moving a 200mb file from ram to vram wouldn't take 4 seconds? well I don't know the process really

also running ComfyUI with --gpu-only gives me almost 2 it/s more when using AIT. that is super great, but it's not stable

[indigo carbon]

you mean 2it/s more than AIT without --gpu-only?

[upbeat summit]

yep

5.08 it/s vs 6.76 it/s

[indigo carbon]

those numbers are half of what you should be getting when using the AIT workflow

[upbeat summit]

It's in my workflow. I can try with yours

but --gpu-only fills the vram at some point and it almost freezes for me after a couple of minutes - so not sure what's happening

[indigo carbon]

sounds like memory leaks, I thought ComfyUI didn't have memory leaks issues?

[upbeat summit]

I talked to others having the same GPU and they didn't have the problem with the slow down, but also Out of Memory errors when using --gpu-only

[indigo carbon]

very odd. this sounds like an issue with ComfyUI rather than AIT, I suppose you should make an issue on ComfyUI repo?

[upbeat summit]

yeah that issue is ComfyUI or related to my setup, because it does happen with or without AIT

the lora thing is AIT related

[indigo carbon]

It appears that there is an issue with LoRAs using AIT- the fix for it being --gpu-only .to fix --gpu-only I suppose you could open an issue on the Comfy repo.

[wet nacelle]

Canny is out for 1.0

[upbeat summit]

just used your workflow, with the water gun. base model looks normal I guess but the refiner runs at 1.33 it/s ?!

base 67/67 [00:05<00:00, 11.75it/s]
refiner 44/44 [00:33<00:00,  1.33it/s]
Prompt executed in 49.45 seconds

guess I do a reboot next 😉

[hardy cipher]

I should see if sdxl canny would even run on my computer

[indigo carbon]

wtf, that doesn't sound right. try to release VRAM

[upbeat summit]

why only the refiner is slow - I don't understand

[wet nacelle]

I'm trying to find the person that made the workflow I'm using to politely ask them to add it in.

[indigo carbon]

close ComfyUI completely, then try again. my bet is this is caused by memory leaks

[indigo carbon]

OH WAIT

[hardy cipher]

which workflow?

[indigo carbon]

is this --gpu-only @upbeat summit

[upbeat summit]

yes

i think that's the problem

I thought I started without it

sorry

[wet nacelle]

@mossy canopy hey do you think you can at some point add the new official canny to your workflow?

[strong field]

On average I think refiner is slower for me, not as slow as @upbeat summit Do you experience this too?

Maybe 1-2it shower

[indigo carbon]

by how much, that's a vague question

how long does it take to execute prompt? @strong field

[strong field]

4096 in 20s

[indigo carbon]

and on my unaltered workflow?

[strong field]

For your resolution or for 4096?

OG file is like 10s

[indigo carbon]

load this

[strong field]

If I recall. Maybe less

Ok give me a minute to get back to desk will test

[wet nacelle]

This might be a good place to start for understanding a workflow idea. https://youtu.be/_f0qrHQs0jk

@mossy canopy

[hardy cipher]

shouldn't be too tough

[upbeat summit]

ok apologies for the confusion. without --gpu-only and your water gun workflow

base  67/67 [00:05<00:00, 11.78it/s]
refiner  44/44 [00:04<00:00, 10.99it/s]
Prompt executed in 14.37 seconds
4090, Windows 11

[indigo carbon]

yeah, this looks right

[strong field]

Yeah that looks right

[indigo carbon]

it's so awesome we got to a point where we make stuff like this in less than 15 seconds

[strong field]

Pioneering

[wet nacelle]

I have a small brain and also currently literally sick.

I need others to help me so that I can have fun.

[hardy cipher]

oh, I'm not knocking you

[wet nacelle]

I get that

[hardy cipher]

I didn't mean to make it sound like that

[wet nacelle]

I'm joking with you.

I am sick though.

[hardy cipher]

I'm looking at it now. I'll see if I can get it to work for myself. if I do and you're still around I'll see if I can maybe help out

bummer you're sick 🫤

[wet nacelle]

I'll send you the other UI I'm using by the other user to see if you want to try and fanagle it into that one as well.

[hardy cipher]

that works. no promises, I'm not the best at this stuff but I'm persistent so should be able to figure it out

as long as my hardware allows it. might be pushing it a bit but we'll see

[wet nacelle]

When I do send it just unwire the upscaler nodes and find a way to route the nodes to it.

The upscaler can take up to a minute plus just due to how he has it set up. Not at all hardware intensive for that portion from what I understand.

I'll send it in soon.

[hardy cipher]

cool, I'll check it out

[strong field]

something seems off for me

thats using your workflow

pretty sure it used to be faster

i have some ideas

[spring fulcrum]

What scheduler should be used with euler_ancestral?

[strong field]

normal or karras work well

[spring fulcrum]

Does anyone have a workflow they are willing to share that works well that uses all the different components? Such as model, refiner, lora, and upscaler?

[strong field]

do you know how to install custom nodes using comfy manager?

[spring fulcrum]

yes

I just don't know what to connect with all the spaghetti

[spring fulcrum]

In other words making my own workflow with an advanced configuration still elludes me

[strong field]

download the json, and hit load in comfy

use manager to install anything missing

[wet nacelle]

@hardy cipherHere it is.

[hardy cipher]

thanks. might be a minute. we'll see. I'll let you know if I figure it out. or if I don't, lol

[wet nacelle]

I get ya.

[ionic gulch]

any idea what i could add to my new node? (things that are always required for SDXL TXT2IMG)
https://i.imgur.com/jlj0vVo.png
(the node calculates end_step, start_step by the percentage value and sets cfg_refiner to cfg_base if "0" is selected), also offers all recommended resolutions for SDXL)

[hardy cipher]

seed number?

[ionic gulch]

@hardy cipher will have a look if this is possible with my limited skills :)

another nice node will be able to multiply resolution by a factor (without the int to float and float to int nodes required with available solutions):
https://i.imgur.com/TtJDspZ.png

[hardy cipher]

well I think if you'd got this far you can do it

I need to learn how it all works so I can make a node that it seems no one else is really interested in, lol

[ionic gulch]

@hardy cipher depends if it needs some includes from comfyui base (haven't figured that one out)

[fresh path]

guys I had a dream where my comfyui went to 200 it/s

[hardy cipher]

whoa

[ionic gulch]

next step will be to find out how to make it available on github and comfyui manager

[hardy cipher]

do you have a github account?

[wet nacelle]

I had a dream where I met Kanye and talked about movie ideas with him.

[ionic gulch]

i have one, but never used it for own projects.

[hardy cipher]

I wonder what he'd be like in person. he's obviously has some mental health issues, but might be alright in person

or maybe a huge scumbag

[ionic gulch]

also have a node for just the resolution (it's similar to some existing nodes but doesn't have manual inputs, as basic variation, that takes less space in my ui:
https://i.imgur.com/uaYgBxK.png
cobined with the math node it gives you height, width, height x2, width x2 - the only things you need most of the time

[hardy cipher]

if you would like to be pedantic, to keep the same exact number of pixels as 1024x1024 at 4x3 the dimensions would be 4*(512/(sqrt(3))) X 3*(512/(sqrt(3)))

[wet nacelle]

He's literally someone that is lost in regards to thinking that they need to have a larger than life goal that can barely be met.

[hardy cipher]

it's unfortunate his mom died when she did. I think she probably kept him in check in that regard

[wet nacelle]

That's what most people think too.

[indigo carbon]

that's not right. I remember you showing it being faster?

[strong field]

That’s what I thought too

[hardy cipher]

I need to experiment a bit with this. I have it running, but can't tell if it's having any impact on output

[indigo carbon]

try to reset comfy entirely, you might have messed something up. or even restart PC.

[hardy cipher]

never actually tried to use controlnet in comfy, lol

[glad fulcrum]

this node is default or extension node?

[ionic gulch]

it's a new node i'm working on. just have to make some final adjustments to some other node and find out how to share it on github and comfyui manager.

[strong field]

Yeah been trouble shooting

I think I got it back to normal lol

[indigo carbon]

is the speed back to 15 seconds/gen?

[strong field]

Does it look right for your workflow at 20s. I forgot which card you have

I had done the nodes.py tweak and I think I screwed something up

Reverted the file

[molten gull]

that's some guy 🙂

[torpid garnet]

I'm using ComfyUI with SDXL 1.0 and It is creating black images. I tried using --xformers but that doesn't seem to work with comfyui... Does anyone else know of a fix for this?

[strong field]

What’s your resolution settings

[torpid garnet]

768w x 1024h

[strong field]

Not quite but I’m back to 6.9it/s

[indigo carbon]

how long in total?

[strong field]

Can you screen shot your workflow real quick

[strong field]

20s

[torpid garnet]

Its the one you sent scorp

[indigo carbon]

GPU? that's the exact speed I get on my 4070ti

[strong field]

3090

Went back through history that seems about what I was hitting before

[vagrant wasp]

why pixeleted when zoom

[strong field]

OH FLICKITY F*

[strong field]

I’m running my undervolt profile

[uncut steeple]

Low resolution?

[strong field]

[strong field]

You have any issues downloading the modules

?

X formers works fine with comfy

[vagrant wasp]

1024

[uncut steeple]

You are 193% zoomed in, obviously it will be pixelated

[strong field]

Now it all makes sense, I’m ok sacrificing a little speed for the power reduction but I forgot I had it on lol

[vagrant wasp]

ok

[torpid garnet]

No and it was working before. Then I used the searge 3.4 workflow you sent scorp now I get black images

[indigo carbon]

try it without lower power, I think it should get faster?

[strong field]

Yeah definitely, I’ll double check what my max is

[strong field]

Are you getting errors when you run it or just black images

[wet nacelle]

I understand.

[hardy cipher]

it's weird. it seems to be changing things but not in the way I'd expect

[wet nacelle]

Is it overhauling the image more than it should?

[hardy cipher]

not exactly. just creating seemingly random things. can't really say exactly what's going on. I'll let you know if/when I figure it out. just trying to troubleshoot it atm. I'd like to get it working for myself sa well. it'd help a lot with keeping complicated workflows cohesive

[wet nacelle]

Makes sense.

[strong field]

Eh it’s only like 1-2 seconds faster. Not consistent either

So my others issues were workflow

[indigo carbon]

huh. at least it's still incredibly fast

[strong field]

Seems right now

Great optimization on your workflow, I keep having to come back to it when I break it

8k in 25s is about my best benchmark

[coarse hull]

Hi there, I'm pretty new to the party and having great fun with comfyui and the Searge SDXL v3.4 workflow.
I'm not really understanding too much of it tbh, but so far, I got it running. I still get deformations on hands and sometimes faces, any noob-proof tips?

[torpid garnet]

I'm not sure how to fix this does anyone have any suggestions?

[torpid garnet]

I was getting just black images but not I cant even run comfy 😦

[wet nacelle]

You could just use this too. https://github.com/lllyasviel/Fooocus

[ionic gulch]

seed number was easier than expected:
https://i.imgur.com/AdhD1DP.png

[hardy cipher]

nice work!

I really don't understand what aitemplate has to do with controlnet or why it's causing errors. ugh

[strong field]

What gpu do you have and which PyTorch version

[torpid garnet]

3080ti and currently reinstalling cuda and torch

[strong field]

Yes those are next steps, un install torch and install 118 whl

[torpid garnet]

does this process change if I used venv in my .bat file?

[strong field]

Yeah that venv needs to be redone. Will most likely have a different configuration of requirements

[torpid garnet]

so how do I do that I'm really new to using venv all together

[strong field]

Did it automatically generate the venv last time, if so you can just delete and when you run it will redo

[torpid garnet]

so just delete the venv folder

[strong field]

Are you running windows portable comfy

[torpid garnet]

yes

[strong field]

Yes delete venv

[wet nacelle]

Do you think that you should watch the end part of the video where he goes back and shows the node set up from start to end?

You could possibly deconstruct it and.

https://youtu.be/_f0qrHQs0jk

[strong field]

It will rebuild when you run comfy

[hardy cipher]

maybe, I don't know if it's a memory issue or what. works fine when I switch to 1.5 models. I realize they're different things, but seems like it's the same concept

[torpid garnet]

still getting

raise AssertionError("Torch not compiled with CUDA enabled")
AssertionError: Torch not compiled with CUDA enabled

[hardy cipher]

what's weird is I guess 1.5 controlnet models kind of sort of work with xl?

[wet nacelle]

In a way I think that's true.

[hardy cipher]

never tried it, but they're giving me the general shape of the canny input

[wet nacelle]

It was proven a while back that 1.5 net was working with very specific workflows in 0.9

[hardy cipher]

not sure if it'd make anything pretty, but definitely works

[dense lichen]

what i wanted to upscale vs what i got😂

[wet nacelle]

I'll take a look through the base setup he made.

[dense lichen]

using the 4x-ultrasharp method

[strong field]

run python interpreter on windows ( windows search: python 3.10 app)

import torch

torch.zeros(1).cuda()

then run those two commands, what do you get back?

[arctic bloom]

more lens the better!

[dense lichen]

sure😂

[wet nacelle]

More Asian woman the better too.

[hardy cipher]

are you using tiled upscale?

[dense lichen]

yes

[hardy cipher]

that's it

[dense lichen]

what it

[hardy cipher]

in my experience does weird things with xl. like it renders each tile differently

[dense lichen]

its the normal sd1.5 actually

[hardy cipher]

ahh, well I don't know

[strong field]

deonoise is very sensitive

[arctic bloom]

More anime woman the best!

[hardy cipher]

lol

I'm so thankful this channel isn't flooded with creepy anime images

[dense lichen]

i think that ive done something wrong

[strong field]

so using SeargeSDXL broke comfyui? That is strange to say the least

[hardy cipher]

I'll go over that video. I think I'm just the only person on the internet that prefers to just read this sort of stuff. but everyone making videos so I have to click through hoping to catch a glimpse of the workflow,lol

[strong field]

nah, your not the only

just wish someone would put a picture of their workflow instead of talk about something something for 5 minutes, subscribe this, like that

[hardy cipher]

subscribe to my patreon to see unblurred images of my workflow

[dense lichen]

some people prefer to watch a video u know

[hardy cipher]

you know who makes good videos? dudes that show you how to fix cars. they barely know how to video themselves a lot of times. no creepy uncanny weird face thumbnails. they don't spend 10 minutes explaining a minute worth of things

[strong field]

no doubt

[strong field]

im not knocking the format, to each their own

[hardy cipher]

and those videos are legit worth watching because it's hard to visualize a lot of that stuff

or to understand it from an image

"imagine"

[dense lichen]

u mean like when they show u an almost naked women?

[hardy cipher]

I've never seen that

but maybe one day

[vagrant wasp]

what is up ? it becoming cartoon image

[strong field]

Resolution?

[vagrant wasp]

512 upscale by latent 2

was working earlier

[hardy cipher]

can't tell you much without seeing the actual workflow

[vagrant wasp]

it was working fine earlier

maybe needed restart ?

[hardy cipher]

I guess

[vagrant wasp]

something not right

[hardy cipher]

I'll have to come back to it later. Have other things to do now. Hopefully you find a solution before I do

[hard fractal]

working on a crazy new controlnet 👀

[hardy cipher]

wat?

[hard fractal]

no text prompt.

no OCR

[hardy cipher]

huh. that's impressive. I can't even conceptualize how you'd train for that

[hard fractal]

oooo

[strong field]

What model?

If sdxl is trained at 1024 so 512 will break a lot

[urban fjord]

Is this the new more efficient controlnet that were discussed previously?

[autumn forum]

im curious what the use case of this would be lmao.

[urban fjord]

I guess not everything must have a clear cut usecase at first, but imagine if you had a v2 of this where you used coloured text to colour the different part in the final output.

[autumn forum]

then youd have to make a picture with colored text and what not lmao, that just seems ummm time consuming. but im sure thats not the only use case, maybe its like a 1 size fits all kind of control net.

idk just speculating

[urban fjord]

With a good workflow that's not really an issue.

[hardy cipher]

could maybe be combined with unclip models?

[urban fjord]

Unclip is just a worse version of reference_only "controlnet".

[hardy cipher]

okay

[pure hazel]

Speaking of SDXL does anyone have a good IMG2IMG/refine workflow especially interested in adding detail/fixing faces for SDXL in ComfyUI

[strong field]

https://civitai.com/models/119528?modelVersionId=135409

Maybe this?

Dang discord full-size that’s why don’t you

[inner wyvern]

weird, I could hardly run SDVN6-RealXL with A1111. I ran out of memory before it even started with 12gb of GPU and 16gb of RAM. But with ComfyUI it runs like any other model

[arctic bloom]

weird indeed

[soft zealot]

wonder how many of these colour names Stable Diffusion "recognises" ;o)

https://colornames.org/

[queen adder]

when training SDXL on your own dataset images there is no need to bother grabbing 2k images as the max resolution is 1024,1024 right?

I have a bunch of 1280x720 images and i could get their 1920x1080 but im assuming that would slow down the training with no upside?

[abstract blaze]

what is the best workflow in order to generate high resolution images with A1111 on SDXL? I mean above 2400 pixels

[urban fjord]

Start with 1024x1024 images and then upscale your favourites to 2400x2400 afterwards as this saves you a lot of time by not having to upscale everything.

[hasty smelt]

Hi guys! With just 16 GB of RAM, I've installed SDXL and ComfyUI on my system. During the VAE pass, memory usage hits 99%, causing minor stuttering. I'm wondering if it's possible to limit memory usage to around 90%, even if it slightly extends rendering time. If not feasible, I'm considering using Google Colab, as I have an unused premium account there.
Also, I'd like to know if working on Google Colab is a good idea given my current setup.
My current setup features a 2070 Super FTW3 with 8 GB VRAM, a 5700x processor, 16 GB of RAM, and a 1 TB Kingston SSD 7300mb.
Thank you for your assistance.

[autumn forum]

thought about using tiled vae?

[mellow tendon]

Or buying more RAM...

[abstract blaze]

I want to replace my Hires workflow but in SDXL, so I need the native image resolution above 2K because end result should be above 16Mpixel (I use other AI tool for 2x upscale in an additional step)

[hasty smelt]

I do not know how to do it

[autumn forum]

VAEDecodeTiled is the node name. maybe it will help? takes longer tho

[hasty smelt]

oh, I'll look into that right away, thank you very much. I'll try to find a video explaining how to use it. Thanks buddy

[autumn forum]

hope it helps! i dont usually use it so i dont know the implications exactly but it may work

[hasty smelt]

If it reduces ram usage by 10% or something like that it will help a lot with stuterings.

@autumn forum I found it, but I'm not sure where to place it. Could you assist me? Sorry if this seems simple, I'm new to ComfyUI.

[autumn forum]

so youll need to open the vae decode at the top of that first save image node by clicking the left grey dot, then replace the connections on that with the new vae decode tiled

hope that makes sense

[hasty smelt]

It seems to have worked, I just need to organize things now.

[autumn forum]

nice! below 90% ?

[hasty smelt]

94% is good before i have 99%

Nice

😍

I work more test

[autumn forum]

are you using the upscaler? if not maybe try changing out the other ones too

[hasty smelt]

yes, I will try to organize everything. Should I also do the same in Upscale?

[autumn forum]

wouldnt hurt hah

[hasty smelt]

Thanks for the support Via

[autumn forum]

not a problem.

[autumn forum]

me over here trying to find the best upscaling method with the least artifacting and contrast issues. smahing my head against the wall but i feel im close.

[west breach]

Just ask the operator to patch you through to the best method 🙂

[hasty smelt]

@autumn forum I managed to reserve 1GB with the adjustments, in addition to VAE decode (Tiled), I also replaced VAE encode (Tiled). Before this, memory was consuming the entire system, and there is no more system stuttering. Thank you very much for the valuable tip.

[autumn forum]

nice! glad to see it.

[wet nacelle]

Can someone link me the guide that explains each official node in detail?

[hasty smelt]

that would be good

[wet nacelle]

I think it's a real thing is the issue.

[heady vale]

I havent seen one

[wet nacelle]

https://comfyui.creamlab.net/nodes

@heady vale@hasty smeltIt's always the Japanese that do this crazy great shit.

[hasty smelt]

Thanks

[wet nacelle]

https://tenor.com/view/yes-watermarlon-for-miki-yes-nig-gif-17318458

[heady vale]

the info on the page is just expanding the node title slightly it doesnt really say what they might be used with

[strong field]

Stock dynavision?

[heady vale]

yep

[wet nacelle]

I am disappointed now. I just clicked a link and thought "YIPPIE!"

[heady vale]

gonna need a translator

[wet nacelle]

What are you using for a browser?

[heady vale]

FF

[hasty smelt]

if you click on it you can see an explanation, we'll need a beer to translate that

[wet nacelle]

Same here. Do I maybe have an extension? https://addons.mozilla.org/en-US/firefox/addon/traduzir-paginas-web/

[strong field]

This has some info. https://blenderneko.github.io/ComfyUI-docs/Core Nodes/Advanced/LoadCheckpointWithConfig/

[wet nacelle]

This seems good too actually.

[hasty smelt]

bing have translator, I like

[wet nacelle]

That's rad actually.

[wet nacelle]

Did you mean Edge?

[hasty smelt]

I don't know, where I'm browsing, microsoft always changes things, until today I don't know if it's bing or edge

[heady vale]

bing is a search engine like google but I guess both their own translate page abilities

I never use edge or bing

[wet nacelle]

I LOOOOOOOOOOOOOOOVE EDGE!!!!!!

I will never use it unless I want to try experimental sites.

[hasty smelt]

yes edge, sorry

@autumn forum amazing adjustment, I have 2 GB free

[torpid garnet]

When I generate I get black images and this is what is says...

[steady grove]

sourdough

[steady grove]

i use edge for video streaming since its the only one that is usually fully compatible

[wet nacelle]

BALLS!

[indigo carbon]

Does anyone know some kind of node that allows batch count to create a grid like A1111 does? I kinda miss that feature

[wet nacelle]

Have you tried searching for a node like that?

[indigo carbon]

yes, couldn't find anything

[wet nacelle]

Can you tell me the places you looked? I might know another site if you tell me the sites you searched on already.

[indigo carbon]

HF, Reddit, here, CivitAI, Tensors,

[heady vale]

Efficiency nodes has an XY plot, but Ive never used it

[indigo carbon]

I'm looking for a feature that behaves like how A1111 behaved when doing batch count, the batch size on the other hand does make that grid, is there any way to have batch count make a grid like that?

[strong field]

Wait, which version of searge did you download? And from where. There is an issue with nodes that are updated with in old workflows. You need to make sure you are using the workflow that came with whatever you downloaded

[heady vale]

dynavision + voxel lora

[shy kelp]

anyone know if there's a token limit when running sdxl on local machine?

[strong field]

Efficiency XY plot is the only thing I know

[ionic gulch]

can someone help me with the code of a simple custom node?
code:
https://i.imgur.com/BdcR5vK.png
results in:
https://i.imgur.com/GAtVRCP.png
why does it show "I"," N","T" as return values?
i did much more complex nodes without problems, also tried to build it again from scratch, with the same result

[fresh path]

HORROR

[hardy cipher]

have you tried connecting anything to the outputs? if I were that's what I'd do first. drag from the node output and then let go in empty space. it should list the most common? nodes it can connect to, then you can click "search" to see everything it would connect to

[hasty smelt]

OK, I like this. I confess that now, after 24 hours of using SDXL, it is countless times better than the previous options. I'm loving it!

[fresh path]

yeah if you want to make berries

[native knot]

Or kickass pixel art of kaiju alien tech

[hasty smelt]

I have to start somewhere, don't I? 😎

[azure oxide]

curious, did you have kaiju in the prompt? i did it with an animal and it didn't seem to work but i wonder if sdxl knows it with other subjects

[native knot]

No. But I'm actually trying some iterations of that since I thought of that word when responding.

[azure oxide]

hah gotcha

[fresh path]

Imagine the berry fine tuning

[native knot]

This prompt is just insanely cool.

[ionic gulch]

i solved the node problem (imo that's a bug in comfyui) in a way that doesn't need the additional node.
workflow v13 is working fine in my first tests (img2img and txt2img):
https://i.imgur.com/WJOwCgJ.jpg

[native knot]

That is straight fire right there.

[lilac raven]

I had this idea of having pages that had a "prompt" and images would be generated and the pages would be endless galleries, but I had to give up

[fresh path]

u could do a scale and resize with no filtering to get actual pixel art

[hasty smelt]

Mack, can you send me this prompt? I love vintage art and would like to try something with it.

[wet nacelle]

gothic (demented:1.8) smear dry vintage illustration of knight in bone armor holding spear

[hasty smelt]

Thanks buddy

[wet nacelle]

yea

[fresh path]

and dont forget ur negative embeddings

idk if theres any on sdxl actually

it looks fine

[native knot]

That doesn't "just work" for everything. Pixel art tends to be drawn in certain ways with limited palettes and specific styles.

Lots of things lose important details when you try just scaling/resizing, like eyes, etc.

[fresh path]

I meant as a fix to the model not getting pixels correctly square

but yeah you are right

because the model doesnt exactly put the pixels on a grid

[native knot]

Ah...yes. Actually, I am using that trick. The final step in my workflow is a nearest-neighbor resize.

It corrects the exact problem you're describing.

[fresh path]

yeah thats what I meant

but when you upsize it, you need to use no filtering or else it looks blurry

unless the res is just rly low

[native knot]

Final res is 640x360.

[lilac raven]

On ComfyUI, what is the difference between latent and LATENT?

[native knot]

It's not blurry, tho.

[fresh path]

oh must be my browser

[wet nacelle]

Guys I live here.

[lilac raven]

How do img2img works on Comfy? I have a LoadImage and then a VAEEncode, but I didn't get how do I set the strength, it seems to change everything

[fresh path]

look at the example

I think you are looking for denoise in ksampler

[lilac raven]

So it's not usable with the other ksampler that has the end_at_step for the refiner? Because the ksampler with the denoise parameter don't seem to have the end_at_step

[strong field]

this might work

but its a funky workaround and may cause headaches lol

[ionic gulch]

you can use steps and start_step for sdxl for img 2 img - the higher the start_step for base, the more of the img will be kept

[lilac raven]

I think this is working, let me test a little more

[ionic gulch]

should work without problems - here i start base steps at 35% for example:
https://i.imgur.com/WJOwCgJ.jpg

(right side is the original - left the result)

[lilac raven]

Can confirm it works, thank you

[hardy cipher]

JPS, could you tell me what the deal was with that I N T thing? I'm really curious. I looked into it a bit and couldn't figure out what was going on. t

[strong field]

so the joker is batman now?

[hardy cipher]

obviously it spelled INT, but then I could be integer? N number? T ???

[strong field]

yeah me too, i looked over the code for like 10 minutes an couldnt figure

[hardy cipher]

I actually have no idea, lol

[vast galleon]

you want to do

 return(int(largest), )

[strong field]

bruh was it just the comma?!

[ionic gulch]

for some reason it didn't show the name as it worked for all other custom nodes i made and showed multiple outputs instead of one / also it didn't calculate the values and always gave me the default value - don't understand why, as much more complicated calculations work fine in other nodes i made.

[vast galleon]

looks like it

[strong field]

i checked and double checked all the brackets lol

i was thinking it had to be syntax related but damn, a comma

[ionic gulch]

the comma doesn't make a difference. i tried with and without commas on multiple places - if there is nothing after the comma it just gets ignored

[hardy cipher]

I was trying to figure out if those letters actually meant something, but I barely know enough about python to fix little errors that pop up. that was a bit beyond me

[vast galleon]

you still need it where i had it so it must be another issue.

[ionic gulch]

ok, let me try again

[hardy cipher]

well it had to have a reason to think each of those letters was an output. I want the flawed node so I can see what the outputs would connect to

I guess I could just use that code

[ionic gulch]

so this should work?
https://i.imgur.com/kLtKFeW.png

[hardy cipher]

I hope this isn't a dumb question, but with the fine tuned sdxl models, should I still use the 0.9 vae, or would I be able to use the vae output from the checkpoint loader?

[ionic gulch]

you can copy the code:
lass JPS_Math_LargestInt:

def __init__(self):
    pass

@classmethod
def INPUT_TYPES(s):
    return {
        "required": {
            "value_a": ("INT", {"default": 1,}),
            "value_b": ("INT", {"default": 1,}),
    }}
RETURN_TYPES = ("INT")
RETURN_NAMES = ("largest")
FUNCTION = "get_lrg"

CATEGORY="JPS Nodes/Math"

def get_lrg(self,value_a,value_b):
    largest = int(1024)
    if value_a >= value_b:
        largest = value_a
    if value_b < value_a:
        largest = value_b
        
    return(int(largest),)

[hardy cipher]

I've been using the 0.9 vae, but I don't know if that's correct

[strong field]

think my new prompt generator is smarter than i am now

[hardy cipher]

nice

I've been making some cool comic book stuff

you should download the ussr propaganda lora

[vast galleon]

also try def INPUT_TYPES(cls): instead of def INPUT_TYPES(s):

[hardy cipher]

it adds flair

[ionic gulch]

here is the result with comma:
https://i.imgur.com/kLtKFeW.png
https://i.imgur.com/jDfZfNl.png
for some reason it takes each letter of the "RETURN_NAMES" and creates outputs for the first three letters

[strong field]

they officially released one with .9 so i think you are good

[hardy cipher]

well I'm using dreamshaper, and I've been testing out using the vae output from the checkpoint vs the vae loader

not even sure which is better, they're different though

[hardy cipher]

I've been doing fine just using the regular vae though so I'll stick to that

[strong field]

if it aint broke.

[vast galleon]

also try:

largest = int(largest)
return(largest,)

[hardy cipher]

btw, I don't just ask these questions when they pop in my head. it's just hard to find this information online

like trying to find a definitive explanation on how to construct the refiner prompt

[strong field]

scott detweiler is laughing at his cocktail party righ tnow