✨｜sdxl

[trim orbit]

i hate making even tiny changes to comfyui layouts . it's always this kind of stuff. mb

[soft zealot]

Well it won’t break break it.(I don’t think) just means the seed won’t get written to the file named or into the prompts txt

[trim orbit]

i've got a lot to learn about comfy yet. think i want simpler and sexier layouts. efficiency nodes seem to be my jam

[untold crescent]

comfy would benefit from "collapse" setup where you could just make new functions out of bunch of nodes

[fast vector]

https://github.com/ssitu/ComfyUI_NestedNodeBuilder

[untold crescent]

ohh

[trim orbit]

yes i know it's saturday. saving money requires compromises. also, living on an island makes it easier.

yo we'll see you all late.r i got a work thing called my job to get to. good luck generating generators!

[peak dove]

SDXL makes my 8Gb RTX2070 look underpowered when using A1111

[normal bridge]

Yeah i noticed vram usage is quite insane, even for my 4080 i get it filled quite often 😂😂

[peak dove]

Anybody using Dettweiler's triple refiner-base-refiner setup in ComfyUI?

[soft zealot]

I e tried it But hit and miss Depends a lot on what you’re generating

[peak dove]

I learnt a lot about the workings of ComfyUI, its an interesting insight

[ionic dragon]

What does it mean that SAI made finetuning easier on SDXL?

[slender raft]

What comfy nodes are people using for XY graphs?

[fast vector]

https://github.com/LucianoCirino/efficiency-nodes-comfyui

[soft zealot]

same prompt & seed, one with pre conditiong and one without (see file name).

[soft zealot]

drat forgot to take the 3 out f the no preconditioned

[peak dove]

Yes, its a good setting off point for experimentation

[shy kelp]

Anti AI people want to ban AI images on this sub. i need your vote https://www.reddit.com/r/Avatar/comments/156x3cf/poll_should_this_sub_remove_all_ai_art/

[soft zealot]

this is what I should have added as the Non Preconditiooned however first time I left the start step in First Pass as 3 rather than resetting to 0

[peak dove]

[agile quarry]

is it something he posted or from a YouTube video of his?

[fast vector]

yes

[peak dove]

https://youtu.be/2Xe79Nl_6jA

[soft zealot]

Interestingly for these videos hes not using the triple prompt input method, maybe he's going to bring that in in a future video with an explanation of why/how it works

[fast vector]

Example one also doesn't use triple input.

[peak dove]

The video I posted above, he adds a 3 step refiner to the normal base/refiner setup almost at the very end of the video. 3 step refiner; then base step 3 to 12; then refiner again step 12 to 20...

[soft zealot]

I wasnt referring to that, I was reffering to the fact hes just using a simpple Postive & Negative prompt rather than Linguistic Positive, Supporting Terms & Fundamental Negative

[pure hazel]

Hmm weird, keep getting refiner out of memory errors in 1111 :/

[peak dove]

I'm not sure if it does anything valuable, but it certainly opens up possibilities in the use of ComfyUI ...

[peak dove]

OK, I understand 🙂

L8r

[soft zealot]

Another example of "Non PreConditioned" (on the Left) and "PreConditioned" for the same Prompt & Seed

[shy kelp]

what does this mean? base model genning at 128x128?

[woeful patio]

is there a reason why comfy has bad hands all the time and a1111 doesn't?

[autumn forum]

no, latent images are kind of a different thing. it still generates at 1024.

[shy kelp]

i see lol

[wet rock]

sorry for asking elementary questions but how does one "train" the ai?

[autumn forum]

i believe if you were to pixilate a latent image it would theoretically look like 128x128 random colored pixels. if you were to do a 1024x1024 latent it would either take forever or oom instantly lol

[autumn forum]

you can train it by training the whole model itself or making a lora. Lora's are great because you dont have to have that many images 10-1000+. most people use Kohya GUI to train loras.

[ionic dragon]

@wicked frigate what do you mean that training has been easier on SDXL?

[craggy valve]

wait the latent is 128x128?

thought sd 1.5 had a latent 256x256x4

[woeful patio]

does anyone have a link to the sdxl 1 with the 0.9 vae inbedded?

[autumn forum]

https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/sd_xl_base_1.0_0.9vae.safetensors

[woeful patio]

thanks m8

[autumn forum]

it looks like sd1.5 may use 4x64x64 latent space

but im not sure im still learning about this stuff.

[spark bear]

pretty sure latent is 8x smaller vs base res so 64 for 512 model and 128 for 1024

[craggy valve]

that is impressive

[wet rock]

thank you, question about the training. if i feed it images, does it go to a cloud? or will my model remain local?

[craggy valve]

like

the fact that so much detail is just added by the vae

[woeful patio]

there has to be a way to make pictures come out more rugged, right know the generations look to perfect and I kinda hate it ngl

[spark bear]

yep, probably why sdxl isnt that much better than 1.5 with small details and far away faces, has much higher latent res but vae pretty much same as 1.5 just trained more

[static prawn]

can u give me advice for more complex prompts?

[autumn forum]

kohya_ss is all local on your computer. given if you have a decent gpu with sufficent vram id say 16gb vram is good enough for lora training. yes the files stay on your computer and its up to you to set parameters and such. i would suggest watching a couple youtube videos to get a better idea

[static prawn]

i wanna generate a dragon which skin (body) is out of thousand eyes

but i cant figure out how to do it

[wet rock]

thanks friend. i will definitely watch some videos. i appecriate you.

[autumn forum]

no problem hope i helped a bit!

[autumn forum]

yeah i can see that being difficult. ima give it a shot by trying to weight the eyes part higher than the dragon

[wet rock]

hey @autumn forum what UI are you using? Im currently using auto1111 but i see a lot of people here using comfyui.

[vast galleon]

The refiner prompt helps in allowing hires fix for odd resolutions.

[autumn forum]

comfy, i tried using a1111 but seems lackluster with how well comfy speeds up using sdxl. very low vram usage comparitive to a1111.

[wet rock]

i will make the transition to comfy. thank you!

[woeful patio]

nodes just keep adding up huh

[zealous horizon]

better extension for "link render mode" (shape of the cables)
Now you can change it when you need it (saved to storage)

drop into ComfyUI/web/extensions

https://gist.github.com/pixelass/0518d56d62d7b8ad19281eb032bde330

[vast galleon]

yup

[static prawn]

feel like its super hard generating complex stuff with sdxl, doing basic prompts way better but when u wanna add a lot and doing fine editing its hard

[woeful patio]

did the refiner have the vae problem too?

[spark bear]

using 42 inch oled as monitor is a godsend for comfy, gotta have the extra screen realestate

[half ivy]

🔥

[autumn forum]

im getting.. somewhere? lmao

[static prawn]

i want more like a full shot

[woeful patio]

95 inch projector next lol

[static prawn]

hope u know what i mean, like the body (tail) skin where comes a lot of eyes out 😄

but yeah its getting in a direction 😄

[autumn forum]

i know man its so hard haha. maybe there will be a model for more artistic outlandish concepts one day

[spark bear]

hehe depends on how crazy your workflow is, the comfy wall in your computer room/office

[solemn scroll]

Would you have any idea on how I can replace my ksamplers by advanced one in order to have my first 20 steps with base and the last 5 with refiner? I can't figure out how to gen img2img with this technique, only txt2img. How do I add a certain amount of noise on the original, for example?

[half ivy]

mee...tooo?

[autumn forum]

i give up lmao. its difficult haha

[soft zealot]

Personally I would start with @high skiff 's work and extrapolate your won workflows from there.
Its what Ive done and then added in pinches frm wherever else I need too

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

[fast vector]

use ksampler advanced. For example you can set steps to 25 in your refined one and start from step 20

[solemn scroll]

Yup, It's the one I'm using for txt2img, but I don't get how I could set the denoise for the original image

[static prawn]

🥲

[soft zealot]

it uses latent to latent rather than img2img or are you wanting to add an img2img step after?

[autumn forum]

yet it can do a dragon made of sparkly gems lmao

[static prawn]

🤣

[solemn scroll]

Yes sorry, it's latent to latent you're right. My objective would be VAE encode -> advanced ksampler with a denoise of 0.5 for the 20 fists -> Another one with no denoise for the 5 lasts one -> VAE decode. Maybe there's a node to add "denoise"?

[wicked frigate]

wha

[autumn forum]

ugh.

[brazen patrol]

I use my 65-inch TV

[solemn scroll]

Maybe those two?

[static prawn]

what are the best res for sdxl?

[south horizon]

Is there a spot people are sharing Comfy templates/workflows?

[static prawn]

1024x1024

what for horizontal and vertical stuff?

1024x1280?

[brazen patrol]

I'll be honest, I think Comfy's workflow is far superior

[autumn forum]

1:1 [1024x1024 square] | 8:5 [1216x768 landscape]
4:3 [1152x896 landscape] | 3:2 [1216x832 landscape]
7:5 [1176x840 landscape] | 16:9 [1344x768 landscape]
21:9 [1536x640 landscape] | 19:9 [1472x704 landscape]
3:4 [896x1152 portrait] | 2:3 [832x1216 portrait]
5:7 [840x1176 portrait] | 9:16 [768x1344 portrait]
9:21 [640x1536 portrait] | 5:8 [768x1216 portrait]
9:19 [704x1472 portrait]`

[brazen patrol]

at least in terms of result

[autumn forum]

warning, lots of eyes, i can give you this prompt if you want to try it out. using no negatives.

[soft zealot]

I did say it was like a tyre or oil thread lol

[static prawn]

yeah would like that and maybe try to work with that 😄

[brazen patrol]

I mean, for using the combination of Base and Refiner I find the results better

I hate the endless panning and moving about to render and make changes, but I cannot argue with the results

[autumn forum]

(dragon made of eyeballs:1.3), realistic eyeballs, full body

[static prawn]

ty

[soft zealot]

No Panning around needed here on a daily driver basis

Everything comfortably in one place

admittedly if I want to delve under th hood its different lol

[autumn forum]

your workflows itches a good spot in my brain. love it winston!!

[soft zealot]

I will happily acknowledg that I wouldnt want to try and reverse engineer from it if I was anyone else lol

[solemn scroll]

For curious people, here's what I'm trying. Hopefully it'll work.

[autumn forum]

so you use the refiner for the first few steps? have you noticed a benefit to that at all?

[soft zealot]

its very much a case of YMMV

For more abstract ones such as #✨｜sdxl message IMHO the preconditioned output is much better.

When it comes to photorealism.................... hmm jury is out.

This is Stable Diffusion, This is the way

[south frigate]

Can I just download SDXL and use it as a normal model? Or I need some cfg to A111?

[soft zealot]

@autumn forum another example Of Non PreCondtioned (LHS) v PreCOnditioned (RHS) for same prompt/seed

[autumn forum]

hmm interesting ill have to try that out

[thorny terrace]

anyone knows if you could fuse characters in stable diffusion??

[tight cairn]

Hey, is this discord about the sdxl version on clipdrop?

[thorny terrace]

I think its a general sdxl channel

[spark bear]

preconditioning is pretty interesting, might even be worth while to try finetuning some models just for pre conditioning, maybe pre condition with a line art or monochrome vector art model to get cleaner lines , and finish with a color model, probably need to toy around with the right amount of initial steps

[thorny terrace]

.

[soft zealot]

Non Preconditioned LH Image, Preconditioned RH Image , same prompt & seed

[vast galleon]

how many steps do you use for it? I've found anything over 4 kinda blows it up

[autumn forum]

thats a clear difference!

[soft zealot]

Scott in his vide said to keep it to 3 for the PreConditioning step so I am (and hes a QA guy at SAI so who am I to argue lol)

[vast galleon]

makes sense. it seems 3 is the sweet spot

[spark bear]

thats where comfy really shines, just making weird model daisy chains, especially if you start finetuning custome models for it

[vast galleon]

thats an interesting idea

[soft zealot]

In Principle COmy also allows you to merge models on the fly to use in your flow

[rich coral]

Hmmm, I wonder if in Comfy on my 3090 (system has 32 gigs) if I could have a 1.5 model/SDXL base/SDXL refiner all loaded at the same time and just pipe from one to the other etc without having to load/switch models...?

[tight cairn]

Can you fave your generations and prompts in Clipdrop so you can use them later again?

[soft zealot]

cant see why not, previously on my 1080ti 11Gb I was piping srom SDXL to SD1.5 Img2Img

[spark bear]

hmm thats actually interesting for cooking up model merges..Comfy's kitchen

[rich coral]

I guess I will have to give comfy a try. Feeling cramped in A1111

[soft zealot]

(nb I also have 64Gb system RAM)

[soft zealot]

Comy offloads from VRAM as soon as it can to System RAM so it does increae System Ram usage but then again RAM is cheap and upgadeable unlike VRAM

[rich coral]

yah, but this is a work machine and it is a pain, lol, wish I could just go buy 100$ of ram and toss it in but nope

[soft zealot]

Reqyuest extra RAM for work, say you need it in order to manipulate large Excel documents more efficiently

[rich coral]

haha, oh my job is AI cool so no worries there. but yah I guess I will request a 32g drop soon

I asked for more and they upped me from 16-32... dont think they get it heh

stuff is so damn cheap but many mouths to feed and art is not a high priority

[soft zealot]

Another comparison this time with a more photo feel

Non Pre COnditioned on LH, PreConditioned on RH

[gloomy barn]

anyone have a saved inpainting flow I can use to try to fix bad hands - like load an image and a mask, have it pain only over the masked area inpaint without me having to set up all the nodes? I see images of such a workflow, but can't tell where everything connects

[pure hazel]

It's too bad SDXL is still piss poor at multiple conceps/characters

[rich coral]

great, that gives me a few months to save for retirement

[vast galleon]

same. I wanna know how I can fix these hands

[tender timber]

OOH! Someone who gets that filenames can be have dates and folder locations can be different! Is this the workflow in your PNG images? I tried dragging this image into Comfy but it didn't do anything

[gloomy barn]

The image in the post you replied to looks like a screenshot of the comfy window.

[rich coral]

wow, i get the feeling that comfy is a pain on 1080p

[woeful patio]

zombie vampire?

[stray warren]

does anyone know if there's a node for saving the prompt (in the filename) in comfy?

[vocal rapids]

no lora used but it really look like lora already.....

[tender timber]

Really? On my system, this the default it loads (left) and the SDXL one from the example page on the right

[vast galleon]

I added "crooked teeth"

[gloomy barn]

Wow, I thought it was just a screenshot

[mellow jay]

Bro , which platform are you using to run xdxl ?

[vocal rapids]

colab only with comfyui

[tender timber]

Ha, all good 🙂 Doesn't matter anyway, I got the workflow from one of the PNG he posted, but even using the workflow manager still results in many missing nodes, so I'm gonna hunt them down

[pallid lintel]

Those are all WAS Suite nodes

[craggy basalt]

Negative prompts are still very important in sdxl.

Left is with and right is without

[tender timber]

Yeah, it goofs on installing 😦

[pallid lintel]

You are using comfy portable I guess? What OS?

[tender timber]

Windows 10, so now following the instructions on WAS page instead of going through the install manager

[autumn forum]

i had to give my pc requirements and did somethig with --user idk

[tender timber]

I have full admin rights on the PC, but I'll keep that in mind

[autumn forum]

i do too... sooo yeah lmao, i thought the same thing but still had to do it for whatever reason. but that error doesnt look like the one i had so maybe its fine

[soft zealot]

Yup si I used it as an excuse to add a 32" 1440P monitor to my dual 22" 1080p set up lol

[autumn forum]

may bring out the ol 4k for workflows.. lmao. i use a 1440 for daily and a side 1080 for discord. maybe i should add a 4k uptop hmm

[gloomy barn]

https://www.youtube.com/watch?v=k0AuipqPqsU

[soft zealot]

I have seen one but cant remeber where.

My approach is to write to a text file in the same directory

[autumn forum]

liminal space vibes

[craggy basalt]

Indeed

Imagine how much better it could be with a liminal dreambooth model.

[spark bear]

now winston got me playing with preconditioning...odd results from preconditioning with the same base model, alot of change from just 4 extra initial steps. pre conditioned on the right with 4 steps of base model , same seed

[stray warren]

i think it can be done with a primitive node, i will try it like this and see what happens:

1. convert filename_prefix to input,
2. link it to a primitive node
3. copy the prompt text into the value field like this

[still dove]

I am missing LahMysteriousSDXL v10

[tender timber]

Gah. I tried all three methods (recommended, legacy,and install bat) and still get an error. Oh well, I'll figure out another way to add partial prompt names and dates to the file and output it to a UNC path

[autumn forum]

has comfy responded to you yet?

[normal bridge]

having installed comfy just now as a total noob to it, what is a good point to start off?

[tender timber]

No, it's not the Comfy workflow, nor does the WAS suite belong to Comfy, so nothing to respond to

[shy kelp]

i feel like xl still isnt quite there. idk i was impressed more with the bots and clipdrop than everything ive done or seen post its actual release. feels like a step up from 2.0 era but still not really a game changer idk, wont be a popular opinion here im sure

[autumn forum]

ah okay

[azure oxide]

If it's sdxl you're wanting to use them you can just load this workflow and prompt from there https://github.com/SytanSD/Sytan-SDXL-ComfyUI

[stray warren]

like this (i took the sdxl example workflow to illustrate) @soft zealot
nah, that doesn't make it easier. i hoped i would get a bigger input field, but it's the same as the prefix

[tender timber]

Different strokes for different folks I guess. I can get these quite adequate results from a local install

[azure oxide]

As a base model it's by far better than 1.5's. I'd wait for fine-tuning for actual polished content.

[strange mist]

I'm getting this weird SyntaxError: Unexpected token '<', " in comfyui
I'm using the simple example workflow given in github

[shy kelp]

nice biff

[normal bridge]

do I understand that correctly that I have to adjust start and end steps on the passes if I adjust my overall steps node?

[tender timber]

Thank you, what kind of output quality issues do you get on your local install?

[shy kelp]

its decent, it just doesnt wow me as much as I thought it would. i thought itd be like on MJ level at least with how much some of the stable people were shit talking MJ but I find it pretty weak comparatively

but we'll see how things look in a few months or whatever with some fine tunes, 1.5 base aint all that either

[azure oxide]

Yep it's fairly straightforward enough to use. If you have any issues that pop up then just ask here about them.

[normal bridge]

thanks man appreciate it

[brazen patrol]

https://comfyanonymous.github.io/ComfyUI_examples/

[normal bridge]

favourited ^^

[tender timber]

Yes, people forget how poor SD 1.3, 1.4, and 1.5 could be at the beggining

[tender timber]

This is a good one too... https://github.com/wyrde/wyrde-comfyui-workflows/tree/main

[brazen patrol]

and also

https://github.com/comfyanonymous/ComfyUI_examples/tree/master/sdxl

The second is the one I use for great effect

the second workflow

[ionic gulch]

Why is Scott Detweiler not using the "end_at_step" parameter in his ComfyUI workflow? Doesn't that result in removal of all noise, making the "return_with_leftover_noise" option useless, which would go against SDXL documentation?
https://i.imgur.com/Pkk0uZ7.png
(Source: https://youtu.be/2Xe79Nl_6jA)

[spark bear]

it will end at 20 steps of whatever you set the steps to

[pallid lintel]

@tender timber @autumn forum You might want to ask here: https://github.com/WASasquatch/was-node-suite-comfyui/issues?q=is%3Aissue+is%3Aopen+sort%3Aupdated-desc

I just asked WAS it's not a known issue yet at least apparently

[tender timber]

Thanks, will do!

[ionic gulch]

he sets steps to 12 for base, which from my understanding would result in 12/12 steps with no remaining noise. after that he adds new noise and does another 20 steps with refiner.
all that is not how sdxl should be used by documentations i've seen. refiner should be less steps than base and carry over the noise from base instead of creating new noise.

[soft zealot]

I also wonder why he isnt usng the Positive, supporting, negative triple prompt method however..............

.................I wonder (base don his last 3 videos) if hes going to introduce those other elements.

Ie is he doing this in a building blocks method rather than a one shor wham bam thank you ma'am

[brazen patrol]

Unless I am not undrestanding, I see it say enable noise at end of the 12 steps, i.e. return with leftover noise

[ionic gulch]

@brazen patrol if you don't use end_at_step, there is no noise left, as shown by the preview image he does from the base result. that's why his refiner node needs to add new noise. that's not required if you use end_at_step.

i would understand that as simplified solution or his personal way, but he says that they use that workflow internally at stability ai.

[pallid grove]

anyone know where I can get a copy of NKMD upscalers? the links dead ?

[brazen patrol]

This is the standard ComfyUI workflow for SDXL:

There is no end at step for Base

There is no option for it even

Mind you, this is not my workflow design. It is Comfy's

[ionic gulch]

your screenshot doesn't show where end_at_step is coming from. at least it seems to be not kept at 10000 as in the video i've linked above.
do you have a link to the complete workflow?

[midnight path]

is there any information anywhere about prompt weighting? I recently heard that you can use a picture of a ((red)) (((cat))) to add weights? or something?

[brazen patrol]

https://github.com/comfyanonymous/ComfyUI_examples/tree/master/sdxl

[ionic gulch]

@brazen patrol thanks for the link. that makes much more sense. they use end_at_step in that example and end at 20 of 25 steps and do the remaining 5 steps with refiner:
https://i.imgur.com/gFcfPKg.png
refiner is used for step 20-25.
so why does Scott Detweiler tell us they use his (totally different) workflow internally as default base sdxl workflow at stability ai?

[stray warren]

@soft zealot so it works, but it's the same as pasting the prompt in thestandard prefix field. no need to use the extra node

[soft zealot]

so this just gives you a stupidly long fine name ?

[brazen patrol]

Yes, sorry if my shot was not clear. I reorgnaized the panes to keep it ll in one screen

[stray warren]

well i pasted the prompt in the field and that;s what i got

[brazen patrol]

More precisely, I organzied them so the ones I typically modify are all there.

[soft zealot]

are you referring to this?

So each time you write a new prompt you need to copy & paste it over to the primitive to include it in the filename?

My method of writng a matching prompt .txt file automates it and adds everything from my orompt inputs automatically

[ionic gulch]

@brazen patrol i do the same on my workflow. would be great to have a way to create a real user interface and hide the nodes in the background.

[stray warren]

how?

[soft zealot]

using a series for text manipulation boxes as shown in the screenshot in this post

#✨｜sdxl message

[brazen patrol]

Agreed, but for now:

[stray warren]

stupid me, i overlooked it thanks! 😅 ❤️

[soft zealot]

I'm also using WASNodes Suite Image Save rather than the default

[clever verge]

Anyone tried ✨ as a prompt? Like ✨Isometric✨

[stray warren]

i tried was nodes, but i got an error
not sure though if it was WAS, i installed a lot at the same time so i'll try it again
i'm using a very simple setup right now
would be nice just to add that one @soft zealot

[clever verge]

Seems to be working, guess it's not a fluke that ✨ is the icon for the channel. 😉

[rich coral]

does comfy have support for post processing a step, like being able to sharpen ala photoshop between an img2img step?

[elder niche]

Unicycling Giraffe was difficult in 1.5, but SDXL seems to grasp the absurdity of things and make better inferences. *A giraffe riding a unicycle, giraffe, ((unicycle)) *

[lethal tusk]

help.

[soft zealot]

look under image of the right click context menu oer just dbl click any where on the canvas to bring up the search box and satrt typing

[rich coral]

Ahh sweet, all right, I have run out of excuses... will install it today hehe, much thanks!

[civic forge]

I think refiner model shouldn't be used for generating in a1111

[soft zealot]

apologies but my workflow is wrtten for my personal use, have no intention of adding to it for other people.

Its in all the images I post if people want to use/adapt it but its not an "offical" supported thing and subject to chnages on a whim

[stray warren]

no it wasn't your workflow, i just downloaded a lot of those custom nodes and tried them on sdxl 0.9 (not recently)

[lethal tusk]

I changed to base model and its still broken

[woeful patio]

do you have the latest version of a1111

[lethal tusk]

yeah i think so, i use a launcher that automatically updates it

[high skiff]

Ok, back to the SDXL 1.0 release workflow grind

[woeful patio]

1.5.1? and maybe set the canvas size to 1024x 1024

[lethal tusk]

still kinda broken

could these be the issue?

[woeful patio]

Maybe, try launching with the xformers check turned off?

[upbeat tundra]

Can you use sdxl model in auto1111 img2img to upscale?

[woeful patio]

yes, ultimate sd upscale extention works as well

[upbeat tundra]

Nice! Do you get higher res results?

[winged tree]

Just the model, check around, its up on the civai...or however you spell it

[woeful patio]

yeah definitely some tweaking to get it to work but it works well

[upbeat tundra]

Cool adding it to auto now...we'll see. Upscale 1.5 gens with the new model seems it would be the way to go until the new model is fine tuned

[woeful patio]

yeah models are pretty crap right now, we just have to be patient. SDXl is going to be great in a couple months

[tawdry oxide]

are there good tutorials for fine tuning SDXL?

[bold osprey]

manually pick the VAE instead of usingthe automatic

[woeful patio]

on further inspection I think it might be your torch version not working properly, try putting --reinstall-torch here

[lethal tusk]

I reinstalled torch and xformers and picked the VAE and while its generating it looks fine but it fucks it up at the end result

[woeful patio]

I think you're torch version should be above 2.0.0, that's what I can think of.

[stray warren]

found it, it's in mikey nodes
you just convert the pos en neg prompt to input and then link them to the text node outputs

[bleak garden]

does anyone know of a way to adjust refiner settings in sdnext? i'm getting pretty good results with dreamshaperxl10+adetailer+refiner

[onyx spear]

So I got SDXL running on the webui and I can generate images, but everytime I try to use the refiner model I run out of VRAM, I'm already using xformers and lowvram, does the img2img just use way more vram or am I doing something wrong?

sorry probably wrong channel

[sleek relic]

did someone tried this depth controlnet model already? https://huggingface.co/SargeZT/controlnet-v1e-sdxl-depth

[bleak garden]

try settings to unload model to ram when switching between base and refiner

[leaden marlin]

I have the same issue and I’ve seen other people also mention this. Sdxl doesn’t work in automatic 1111 very well right now so many people moved to comfy which it work so much better on.

[fast vector]

Honestly, you're not. A1111 is quite memory inefficient.

[brazen patrol]

I tried running 0.9 but keep getting a black image.

[bleak garden]

i moved to sdnext. it's closer to a1111

[brazen patrol]

Can anyone tell me what I am probably doing wrong?

[bleak garden]

comfy is to 'uncomgy' for me to use...

'uncomfy'

[woeful patio]

what are you running? a1111 comfy?

[brazen patrol]

Comfy

[onyx spear]

thanks and also thanks @fast vector, I'll setup comfyUI I guess, doesnt seem too cumbersome

[woeful patio]

can you make a screenshot of your comfy ui?

[visual glade]

are you running it with any options?

[brazen patrol]

I am using your Pytorch build

THough with 3.10.9

since 3.11 was slow as molasses

[bleak garden]

love me some alicia vikander...

[real lagoon]

Best parameters were tested with the bots here right? Were the results ever released?

[brazen patrol]

However, the UI and nodes are pure Comfy SDXL refiner setup

[woeful patio]

why does it take almost 3 mins for an image with 20 steps? what gpu are you running?

[real lagoon]

Or the ideal parameters, if not the full results?

[lethal tusk]

RTX 3070 8GB VRAM

[woeful patio]

yeah I have the same one, that shouldn;t take 3 mins but tabout 25 seconds. so something is really wrong there.

I would try to reinstall a1111 in a seperate folder to see if that solves it. Or does it give something in the command prompt on generation?

[digital shoal]

would anyone here happen to know why my primitive nodes in comfyui doesn't save the custom size I've applied to them? when loading the workflow they just go back to default size

[brazen patrol]

If I swithc to plain 0.9 I get a black pic. If I use the 1.0 it works fine. I have tried wit a separate VAE model and not

[fast vector]

why would you use 0.9?

[brazen patrol]

compare results

variety. Why not?

[fast vector]

I mean have you tried both refiner and base? Or just one?

[brazen patrol]

Both

I did not try the base alone

[tender timber]

I am now 99% of the way there after a clean install... just missing one node. I"ll see if I? can grab it manually somehow

[fast vector]

mind sharing json?

[tender timber]

I can't take credit, it's from @soft zealot and is available from any one of his PNG output files

[upbeat summit]

it's this. save the raw file to your /custom_nodes/ directory:
https://gist.github.com/alkemann/7361b8eb966f29c8238fd323409efb68
make sure its name is .py and not py.txt

[solemn scroll]

Hi, do you think it's worth re-downloading the model to get the 0.9 baked VAE?

[tender timber]

Thank you sir/ma'am!

[fast vector]

ah, I see

[upbeat summit]

greetings all 🙂

[tender timber]

https://cdn.discordapp.com/attachments/1089974139927920741/1134917403856347217/SDXL_20230729193500_318187109389229_000001.png should have it within the PNG

[tender timber]

My cousin's been in Texas for a decade, she now says "all y'all" without irony 🙂

[soft zealot]

That’s seed to text. Link in the credit and notes of mine

[upbeat summit]

yeah just pointed him to it 😉

[high skiff]

We are nearing completion for release :p

Syan's 1.0 SDXL workflow with built in 2048x upscaling

[brazen patrol]

Sounds excessive. I mean I don't really need 2M x 2M pixels

😄

[high skiff]

Just have to check some more things, should be able to submit it to the comfu UI wiki later today after I cut out the upscale nodes (the upscale version will be available on my github)

[soft zealot]

wot no "Pre Conditioning" ;o)

[high skiff]

har har har

[high skiff]

oh wait

[fast vector]

Just disconnect the node that goes to upscaling

[upbeat summit]

half a day away and civit is flooded with SDXL content

[void loom]

Hey, are bracketed weights and :1.3 etc style weights still working with SDXL? i've been trying but don't think they are working, what about prompt timing, is this meant to function as like with old SD or is there a new format? thanks

[high skiff]

I forgot to look into that @soft zealot Would you be willing to help me with that in DM's? Maybe I could include it in the workflow

[soft zealot]

or use a jumper ;o)

[brazen patrol]

I usually upscale with Topaz, so this sort of thing was never on my list anyhow

[upbeat summit]

the jumper idea is a good workaround for not having options or settings

[high skiff]

I have a jumper for the workflow

but Comfy prefers I upload a version with no third party nodes, so I will do that with a link to my github for the one WITH the upscaler

[tender timber]

Success! Thanks!

[upbeat summit]

Winston is building LHC level machines in comfyui 😉

[upbeat summit]

have fun!

[solemn scroll]

Hi, is the 0.9 VAE model worth it?

[tender timber]

All I really wanted was custom file names and folder paths, but many of the other nodes are lacking or buggy. Will tootle around with this one and maybe grab bits and bobs to make my own Frankenstein version

[zealous horizon]

Small update to my linksRenderMode extension: Just select the mode you need. (saved in browser storage)

For installation guide see comment below the gist

What does this do? It toggles the render mode of the cables/lines mode (straight, slight curve, curve)

https://gist.github.com/pixelass/0518d56d62d7b8ad19281eb032bde330

[azure oxide]

Woah woah help each other out here so everyone else gets to learn, I want in on this too

[fast vector]

put this one into custom nodes?

[civic sigil]

@visual glade Are you still planning on releasing a reference_only node with img2img support? If not I will start trying to hack it in myself but I'm not sure if I will be able to

[lethal tusk]

I reinstalled A1111, eta is still like 2-3 mins, waiting on the result

[normal bridge]

Is there a simple way to have flexible workflows in comfy where I can enable disable certain parts with one click?

[lethal tusk]

images seem to turn out correctly now but it still takes like 2 minutes

[brazen patrol]

So I tested it, and it is the Base 0.9 that doens't work. If I ask it to run the refiner only I get a picture. So I will redownload the Base in case it got corrupted somehow

[woeful patio]

Yeah I’m not quite sure what to make of that, definitely taking to long. You should put your sampling steps around 40/45. Are your drivers of the gpu updated?

[lethal tusk]

drivers are updated yeah

[woeful patio]

Do you get any type of error in your command prompt?

[azure oxide]

Ctrl m mutes selected nodes is all I can suggest

[polar epoch]

Hoi, just a questionary as i'm curious if i should continue using comfy as a AIO click, or manually got fetch, install requirements and such, as a few extensions is missing modules, and i can't even install the modules as it's not cmd'd into it's venv, so it's stuck at missing modules lol

1: Stick to 1 click?
2: All manual, and can effectively add modules :P

[lethal tusk]

nope

[woeful patio]

Settings > System > Disable memmapping for loading .safetensors files. (fixes very slow loading speed in some cases), Maybe turn this on?

or maybe enable low vram, I heard not enabling that can cause issues

[lethal tusk]

how can i enable that

[woeful patio]

during startup of a1111 you can check it

[high skiff]

I am messing with it right now to see if it has any place in my workflow or not

Its veryyyy weird lol

[woeful patio]

@lethal tusk

[lethal tusk]

LETS GO THAT FIXED IT

Thank you so much for helping me

Someone promote this guy

[void skiff]

I hope I ask in right place now, after a stroke I forgot somethings, bur I have the "sd_xl_base_1.0" model and Img2Img I want to use a old 256x256 image and upscale it to 1024x1024, but i get a long error and I hope some can say what I do wrong.

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

I have a Nvidia 3090 that was a gift and I think it will be okay, but do I need to change something in the settings to get this to work or have I just missed some step?

[hasty smelt]

Hi everyone, I want to know if my GPU RTX 2070 Super FTW3 with 8GB of VRAM is capable of running the SDXL model. Complete setup includes 16GB RAM at 3200MHz, AMD Ryzen 5700X, and NVMe Kingston with 7500MB/s read and 6000MB/s write speeds.

[woeful patio]

😃 Awesome, always enable low vram, makes life a lot easier

[woeful patio]

I had the same error happening to me, I restarted a1111 and after a couple times it worked. If you're using the ultimate sd upscaler make sure to leave the tile width on 512.

[fresh path]

yes should be able to, you need -medvram option

[lethal tusk]

i always forget to switch between base and refiner lmao

[autumn forum]

so many character loras are showing up in civit cuz of contest lmao

[void skiff]

Just now I am in a episode where my brain skip logick, I removed all scripts for many did not work with SDXL that I mainly want to use, so I just dropped a 256x256 image in the image tab, added a description and pressed generate and hoped it would give me an image based on the smaller image.
I will look into that script and test it... when I find it 😄

My brain do racover but it work slowly.

[hasty smelt]

Thank you, can I set up the low GPU through Automatic1111 & ComfyUI?

[fresh path]

its a command line argument

but I imagine there is settings in-ui for both

[woeful patio]

I understand, You wanna go to extensions and click available then on the load from button, You want to look for ultimate-upscale-for-automatic1111 and click install. click on installed and then apply and restart ui. after that within img2img under script it should now say ultimate sd upscale. you have the settings for that as follows

@void skiff

[void skiff]

Ty. That is what I needed, I feel a bit stupid when miss simple things like that, ty again.

[woeful patio]

You just have to know where to find it :), It could be you still get the error after I'm not really sure what that means but for me it just started working after a couple of restarts

[leaden marlin]

Btw I was able to use just the refiner alone in auto 1111 for both processes

[brazen patrol]

@visual glade @fast vector So just an update: I narrowed the issue down to th SDXL 0.9 base, and redownloaded it on the long shot it had somehow been corrupted. Long shot was a short shot: the new download fixed it.

[onyx spear]

with 8GB vram?

[leaden marlin]

I have 16 gb but I have xformers and medvram

[keen moon]

is there a website where I can see prompts

gallery

such as https://civitai.com

[void skiff]

Stroke is odd, I wanted to say something, I typed, forgot what and.... BUNNIES!

[somber hill]

what are the prompts for these

[woeful patio]

That sucks man I can only imagine being good at something and then just forgetting it, Take it one step at a time man.

[tropic turret]

It takes my RX 6800 XT roughly 3 minutes per image generated by SDXL 1.0 in 1024x1024 on the official workflow for SDXL 1.0 in ComfyUI. Is that decent?

[bold osprey]

So I noticed that loading one of the latest Loras for SDXL take much longer to be processed in the sampler

[brazen patrol]

How many steps?

[fresh path]

yes for windows

[tropic turret]

25 I think

Yes, on Windows

[brazen patrol]

Is that a very slow card?

[fresh path]

no its just amd perf

[brazen patrol]

or old? I dont know AMD

[tropic turret]

Feels like it lol

[fresh path]

I have 6900xt and get about same speed

[tropic turret]

I wish I had a better GPU and more ram for this :c

[fresh path]

6800xt beats 3070 in raw perf

but windows amd has no rocm, so you need directml, which is like 5-10x slower

[brazen patrol]

Is that the 'simple example?

[polar epoch]

huh.. extension needs bitsandbytes, i install it, but then cuda suddely doesn't exist ,

[tropic turret]

yes I think

[low pasture]

Hi, I'm using Comfy with SD XL 1.0, and I'd like to know how to see the seed of an image, knowing that I've generated 4 here.
But most of the time I generate 20 images, and in that case I'd like to see the seed of the selected image, in my example among the 4, how can I do that please?

[fast vector]

click on one and save it. That should do the trick

[fast vector]

try clicking on one of the images

[low pasture]

I see, however I upscale my images and I'd like to avoid having to download an image that interests me and that I want to upscale, and then go looking in its metadata to find out its Seed.

Isn't there a simple solution on the interface to view the Seed?

[woeful patio]

is there any new news on controlnet?

[brazen patrol]

So I ran it on my laptop here (which is my daily driver anyhow) with a 2080 Super and got 25 seconds

[tropic turret]

😩

I bet the 4090 is blazing fast for Stable Diffusion

https://www.reddit.com/r/StableDiffusion/comments/13hyn0c/selfreported_gpus_and_iterationssecond_based_on/

[brazen patrol]

I have no doubt. 🙂 I bought this for a varety of things, but SD was not one. as it did not exist yet

[tropic turret]

3090 is faster? What?

10x faster than RX 6800 XT for SD haha

[gentle mirage]

Idk what to think of that I can make batch 12 at the same time 896x1152 images on 2070. Literally how :D

[polar epoch]

So annoying that the reroute can only go left, or down

[digital shoal]

is there any custom node/plugin for comfy where you can load loras in the prompt? I think I saw it a few days ago but completely forgot what it was called

[fast vector]

https://github.com/LucianoCirino/efficiency-nodes-comfyui

[gilded plinth]

Hey guys, i can manage txt2img, img2img and loras in the same Workflow?

[polar epoch]

I saw the cool superhero looking "drugs" by midjourney, attempted badly at making one for my adhd meds, i ended up with paracetamol instead

[polar epoch]

Img2img is simply just one ksampler shipping it further to a second one. Or of you wanna upscale, upscale base image, then resample it with more details

[tropic turret]

Does anyone have a workflow for any of these two images, but with the use of Efficiency nodes instead? - https://comfyanonymous.github.io/ComfyUI_examples/sdxl/

[empty grail]

does anyone know what causes this to happen?
Using dreamshaperXL 1.0 on a111 with a 12gb 3060 and 16gbs of ram

am i just runnning out of vram?

[gilded plinth]

Is 1024x1024?

[empty grail]

ya

[polar epoch]

I need to test with other lora's, but seems to work fine. This was with base, refiner and 4x upscale with ariel lora

[tropic turret]

Or even with the vae enabled

[woeful patio]

enable low vram check when starting a1111, it has a big issue with vram and sdxl

[polar epoch]

What's denoise and steps at?

[gilded plinth]

Oh amazing

[polar epoch]

Lemme try ichigo and see how that goes :P

[past raptor]

How do ya'll get the right angle wires in ComfyUI?

[gilded plinth]

Is not 1024

[polar epoch]

Or no. This one, one of my favorite characters in it. Gonna try one with and one without lora

[empty grail]

i did hires from 512 to 1024

and 1024

both end up like the image

[somber hill]

numbers are totally undependable

[empty grail]

ill try with low vram

[polar epoch]

You want base being 1024 if you use SDXL

[woeful patio]

low vram will fix the issue bro someone else came in here with the exact same problem today

[empty grail]

will give that a try

ty!

[static prawn]

how i can set the steps on sytan workflow?

[polar epoch]

Is there a node to simply turn off/siderail process if i wanna do before and after images without having to replug node threads all the time?

[brazen patrol]

Someone said Ctrl-M

Back in SD 1.4 and 1.5 I asked it to make me a new logo for my channel on YT based on Chess and tech. I was not stuck on perfection, but I couldn't get anything remotely usable. Logos were a dead end, and anything with chess gave me pure gibberish

Now? I first asked it a very offbeat idea:

beautiful company logo of (chess and circuit board), AI, court of Louis XIV, intricate, sophisticated, extremely detailed, 4k

and it gave me:

[boreal bough]

I run a 4090.
Comfyui -> batch 3 is max for vram
takes about 8~14 seconds per image depending on your steps & settings for base+refiner+vae (running in batch cuts time down by around 30%)

Training: batch 10~12 is theoretical max. 8 is good if you still want to use your pc.
If you run an optimzed workflow, training a lora with a dataset of 30~100 images needs about 14 minutes to reach its ideal state. (though I usually go 50~100% over the ideal, so I have more options to choose from after I finish training)

Is it the best value?
No 3090 wins every time in money/performance ratio.
If all you do is inference, or occasional training. 16gb vram will give you access to everything that sdxl has to offer.

(In case you're wondering about finetuning - it's a complicated topic, but even if 24gb vram can do it, you should just do a LoRA instead. If you plan to make something on the level of waifudiffusion XL, rent a proper 8stack A100 on runpod)

[civic sigil]

Used 3090s are cool for now, I will buy new when there's a > 24gb Vram card

[plain maple]

anyone got a 3090? how many it/s do you get with the base model?

[autumn forum]

i get about 3 it/s to 7 it/s

[soft zealot]

I do rather like this 🙂

[brazen patrol]

WHich was kind of cool and well themed with gold Lousi XIV ideas. but I added the word FLAT to get something less picturesque and it again delivered

[plain maple]

hmm I get only 2.3 with base and 4 with refiner

[autumn forum]

it starts out 2.6 then climbs to 7 with my setup. what ui? comfy?

[brazen patrol]

Super impressed

[plain maple]

yep, comfy

[soft zealot]

and "neonpunk" ified

[plain maple]

with this setup

oh, I forgot to ask you the image size 😄

[autumn forum]

yeah sounds about right. i heard you can use --bf16 or something to speed it up. or --gpu-only or something. but im also doing 1024x1024. but i also 16:9 ratios as well

[plain maple]

with 1024x1024 I'm at 1.8 damn

[tropic turret]

These are all handy tips, but I don't even know how to train a lora or make my own custom model (I'm also not sure if the two are the same). You could leave your message up for others though, as it may help. I'm very new to SD.

[boreal bough]

bf16 is in theory out-of-date, due to the automatic fp16 support now

only relevant if you're working with 0.9

[autumn forum]

ah okay.

[autumn forum]

okay just did a test, 2.75 it/s base, 2.45 it/s refiner. dpmpp_2m_sde karras

[plain maple]

hmm

[boreal bough]

2m sde still working for you in 1.0? O:
I had to switch to standard 2m, as the results were otherwise less than ideal in my tests

[nocturne dove]

Yeah it's important to specify the sampler. Although some of samplers take fewer iterations for a good image.

[autumn forum]

doesnt do to bad. should i change to standard 2m? lol

[inner ruin]

When training faces in 1.5, the model usually extrapolated the body form fairly well. For XL it doesn't do so as much. Have you found a solution for that yet? Wonder if you @boreal bough know

[boreal bough]

I'm using the timestep hack. But so far this gives me a hell of a lot more detail on faces again, and reduces the background blur a bit, that's baked into some of the prompts

[primal vault]

It is finally ready! The latest update for my SDXL 1.0 Comfy workflows and extension: https://civitai.com/models/111463
Now with text2image, image2image, and inpainting workflows. All of them using base + refiner!

[autumn forum]

do using that many steps help with faces and eyes? im usiung 34 max on both. seems to be doing fine

[plain maple]

switched to dpmpp_2m and I'm at 3.3 now. I guess the dpm ancestral was slowing it down

[zealous horizon]

Small collection of ComfyUI extensions: https://github.com/failfa-st/comfyui-extensions

I hope these are helpful. Enjoy.

[autumn forum]

yeah samplers make a big difference

[boreal bough]

sadly none of the big fixes I was hoping for, happened in 1.0 😦
faces got the smallest bit easier to train now - but nothing that I would call "easy"

Still trying different methods to find one that is both easy to explain, and doesn't take an eternity to train

[inner ruin]

yeah those 256 dim LoRAs are bullshit haha

[autumn forum]

404

[inner ruin]

like 2gb for a LoRA??!!

[upbeat summit]

page not found

[boreal bough]

might end up making a 'community training data' which is pretrained on a 5k dataset, which you then put into kohya as an unfinished training thingie, and continue from there. Essentially a pre-existing offset. That way I can stick to 43mb + default sdxl base

still hoping I dont have to resolve to that though T.T

[autumn forum]

thats a good idea!

[zealous horizon]

try again (forgot to publish)

[inner ruin]

I'm willing to help. Is the idea to train on 5K images of faces? What's your hypothesis?

[boreal bough]

just faces works, but breaks the existing knowledge of 'crop'

vs uncropped

[polar epoch]

Sub SDXL loras does not work great lol. Trying to generate this guy's lora lol

[upbeat summit]

perfect and thank you for sharing it 🙂

[upbeat summit]

the screenshots still give a AccessDenied

[boreal bough]

so I'll prob need to do one folder all cropped, one folder all uncropped
that or block weight training - just that I haven't tested which blocks now occupy face data, as its a lot more than before

[zealous horizon]

fixed.

[boreal bough]

an additional thing that may end up working as an 'easy' solution, is TI for refiner only.
Essentially unet lora for base + TI for refiner.

Haven't tested it yet, but in theory it should work\

[inner ruin]

and you can't do a TI for base? Because of the 2 text encoders?

I think TI was what understood the body based on the face if I had to guess

for the 1.5

[upbeat summit]

nice QoL features 🙂 thank you. much appreciated

[indigo carbon]

should I downgrade to WIN10 from WIN11? I'm getting compatibility issues sometimes

[boreal bough]

TI for base works - but from tests from others, it usually does too much damage to the clip model to be worth it

[inner ruin]

hmm I wonder why

[plain maple]

for upscaling the image, should I use the base model or the refiner for the second pass?

[nocturne dove]

Well I hope something for the refiner works out. Is there any problem with training a lora for the refiner? Because it's annoying to train a lora for a likeness and have the refiner change lip/nose/eye shapes and ruin the base model's work.

[midnight shuttle]

No it's basically the same OS just with some different appearance. Win11 works properly.

[boreal bough]

clip was trained in a very specific way - unless we perfectly duplicate it, we essentially have to train it for a fair amount of steps to stop it from self destructing. But this inevitably causes overfitting.

[indigo carbon]

I am getting compatibility issues I didn't have on WIN10

[midnight shuttle]

What issues exactly?

[indigo carbon]

environment variables

they don't function the same way as in 10

[midnight shuttle]

Environment variables have existed since Windows 1.0. Windows 11 is not causing that issue.

Yes they do.

So you need to explain your problem exactly so someone can help you.

[indigo carbon]

they don't. it works with a naming system, programs just assume it's a blank name- which doesn't exist in win11

[midnight shuttle]

I can't understand that sentence.

Give an example. Exact example.

[indigo carbon]

when I run programs dependent on environment variables- it doesn't see the library. each time I use a library located in the system's environment variables I have to manually insert the directory, which wasn't an issue with win10

[midnight shuttle]

I can assure you that environment variable behavior is unchanged between Windows 10 and Windows 11. Also Windows 8, 7, Vista, XP, 2000...Something else is causing your problem. You need to give the exact example with exact names.

[static prawn]

i really feel like auto 1111 is handling my prompts for sdxl better than comy, the usage of refiner is kinda useless in auto and anything else is better in comfy aswell

but i feel like in auto i get most of the time more what i want

[midnight shuttle]

Are you using the same sampler, same scheduler, same resolution, same steps?

Also same seed for testing.

[static prawn]

yes

oh i didnt test yet with same seeds

i just generated a lot of images

and overall it felt like

[midnight shuttle]

Be sure to use CPU for seeds not GPU so you can be more reproducible. Results should be the same.

[indigo carbon]

no. in win10 you didn't have to specify the library's name, in 11 you need to do so. programs assume it's just in PATH without any titling. for instance- I ran Nvidia's code that converts checkpoints to TRT, that program relies on environment variables. it didn't see the library. I have set it up on the correct directory.

[static prawn]

especially humans are always super deformed in sdxl for me, dunno why

[midnight shuttle]

If you want to downgrade to Windows 10, you do you. And if you keep asking someone will eventually tell you to do that. But I'm telling you that something else is causing your problem. I can't really understand your sentences so I can't help beyond that. Good luck.

[midnight shuttle]

Are you using same CFG?

[autumn forum]

another picture of my moviestill lora!

[static prawn]

yes

[upbeat summit]

that lora looks like something I totally have to try!

[indigo carbon]

basic English? what's wrong with that?

[static prawn]

dunno i just feel i need all those extra nodes to get really good results, and thats hella overwhelming for me

[midnight shuttle]

Very strange. Everyone else seems to think that Comfy is better than A1111. You probably need to show screenshots of your Comfy workflow and your A1111 settings.

[static prawn]

i really hope sdxl will be better implemented in auto 1111. especially the refiner

[autumn forum]

its hit or miss, lmao, ive realised that 1.0 does a good job already but this adds a special ✨ touch ✨ lol

[boreal bough]

this is default behavior if you go too small in resolution.
You sure your resolution isnt off by 64, and accidentally 'defaulting' to like a 300px small in the background due to that

[midnight shuttle]

Have you tried SD.Next? It's supposed to be like A1111 but supports SDXL better. I haven't tried it yet but I might try it soon.

[upbeat summit]

is it yours or can I test it somewhere?

[indigo carbon]

AUTOMATIC said he doesn't know when he will do that.

[half ivy]

Tattoos in SDXL are so much better

Like way better

[static prawn]

which resolutions u recommend

[autumn forum]

its mine. ill be uploading it to civit when i think its good enough. im close. with this round. may try a couple more trainings with different settings

[static prawn]

i feel like 1024x1024 is not the best

[indigo carbon]

he currently doesn't seem to work on anything

[autumn forum]

@upbeat summit heres another pic

[upbeat summit]

nice! I'm looking very forward to it! If you need testers let me know 😉

[boreal bough]

try 1024x1024 just to make sure that things are okay. if people disfigure here there's a problem somewhere else. If this resolution does work, then at the same time you know that your resolution was at fault

[autumn forum]

will do i may take you up on that offer!

[midnight shuttle]

It will be difficult for him to adapt his code for a new workflow. His code is not maintainable.

[static prawn]

this is the res i always use, but people tend to have 3 arms, no fingers etc.

[indigo carbon]

yeah, and Vladmandic's UI is just kind of unstable.

[midnight shuttle]

SDXL is made for 1024x1024. You should use that or one of the other supported resolutions. Any other resolutions will give bad results.

[golden quarry]

@boreal bough I just pushed an update (a few actually) to the UI, to add support for full bf16, and do a bit more error handling

keep in mind full bf16 isn't possible on windows

[boreal bough]

ddim 50 steps, cfg 6
(not ideal, but great for testing if the error was in your sampler/step choice)

[golden quarry]

because bitsandbytes 0.35.0 doesn't work

and any newer version on windows causes issues

[boreal bough]

❤️
also fp16 is now working again - since 1.0

working and recommended

[austere ferry]

I put together a basic beginners guide for both sdxl and comfy ui Incase anyone is interested.

I am still learning myself so I compiled my recent learnings into a guide. If I’m terribly wrong at anything please also let me know.

https://youtu.be/1r6hWhF0UoU

[indigo carbon]

that's also very disappointing. many people are used to his UI so switching to ComfyUI is not going to be easy.

[boreal bough]

isn't vlads close enough to make the jump?

cause diffusers are long term supportable

[static prawn]

i feel like i did way to complex stuff in the past for sdxl, thats why i probably need more trained models and kinda disappointed

[indigo carbon]

Vladmandic's UI is unstable- memory leaks all over it

[midnight shuttle]

ComfyUI is a fundamentally different system and is better for some workflows but much worse for others. It is not really a replacement for A1111 or vice versa. It is like saying that Notepad is a replacement for Word. It's two different things.

[boreal bough]

ah. oh well 😦

[midnight shuttle]

Or saying that Elden Ring and CoD are the same because they are both video games.

[boreal bough]

StableStudio is the A1111 equivalent, which while running comfy in the background, saves you the trouble of actually interacting with it.
Same goes for StableSwarmUI - which just needs a bit more styling, but is functionally already better than any alternative

[midnight shuttle]

My workflows fit well with A1111 and don't fit well with Comfy. I understand Comfy and I know how to use it but it's not efficient for my workflows. Some other people it might be very efficient.

[indigo carbon]

however, if there would be an A1111 extension that replaces all checkpoint instances to Diffusers, then there is hope for A1111 to live on.

[upbeat summit]

mostly seed luck (not stable yet) but I'm always experimenting on film and video prompt builds 😉

[indigo carbon]

in other words- if vlad's ui would be stable, it will be the replacement

[vast narwhal]

Any inpainting method that works? This one doesn't

[indigo carbon]

also the way that it handles prompts.. not even close to A1111.

or A1111 would make a switch to diffusers, IDK

[boreal bough]

ummm... comfy talked about it recently. its already well supported XD but no one wants to go through the trouble of building it, and writing a mini guide for the keyboard + mouse shortcuts that are needed to make (quick and low effort) use of it

[boreal bough]

StableStudio will also add proper inpainting support on top of comfy in the 'near' future. But obviously not a solution right now

[midnight shuttle]

The nodes look right. I've tried a similar workflow and it worked but didn't give good results due to not having SDXL inpainting model yet. If you want a cat face on the woman, just specify that and don't include any other background info in the prompt.

[autumn forum]

@upbeat summit heres a nice before and after, same seed, same prompt

[upbeat summit]

nice hehe. the hair improved 😄

[indigo carbon]

this just looks like the death of A1111 webui, unless he pulls something off

[boreal bough]

everyone knew that the programming debt would catch up eventually.

[upbeat summit]

instant 78/80s film vibes

[boreal bough]

was just a matter of when

[midnight shuttle]

He will need to do a lot of work to make the code work with the new models. And also VRAM usage is inefficient.

[upbeat summit]

yeah. we had those talks when SD2 came out and had the same conclusions

[indigo carbon]

as much as I hate using ComfyUI, it pretty much is the only way. I personally prefer A1111 over ComfyUI, but it seems AUTOMATIC is slowly starting to abandon it, sadly.

[midnight shuttle]

I am looking at the code in another window and I still can't make very much sense of it.

[half ivy]

how to make 1 subject o0

[zealous horizon]

I was very hesitant to use Comfy but went from 90% Auto1111 to 98% Comfy within 2 days

[boreal bough]

his commits are insane. He's definitely working full time on it though. even had the 1.5 release recently, which fixed a lot of old issues that hadn't been addressed.

[midnight shuttle]

I think he's working on it but it will be a lot of work.

I hope he doesn't give up.

[boreal bough]

yeah. it's not a realistic task for a one? man operation

[zealous horizon]

Auto should be moved to an org with a larger number of core devs. With the current userbase and number of issues that's the smartest thing to do.

[boreal bough]

comfy, mcmonkey and multiple solo devs joined SAI. I say that the lack of A1111 says more than anything else.

[indigo carbon]

I don't understand what his priorities are. SDXL has been available for over a month for people like A1111, and still acts like the refiner or diffusers or any new breakthroughs exist.

[pallid sail]

Even though I don't use either comfy or automatic1111, I think I would pick comfy from the fact that the workflows are deterministic and shareable. Don't you have to bounce an image through multiple pages to tweak it in automatic? I would forget everything in about 10 minutes

[pallid sail]

I also love node networks so I'm biased

[plain maple]

anyone got any success with latent upscaling?

[boreal bough]

it was always intended as a backend solution.
And now with the first frontend iterations produced, it shows just how much that payed off

[midnight shuttle]

Comfy workflow is better for creating a workflow to generate one image or one type of image based mainly on prompting. A1111 is better for iterative workflows such as geneting a bunch of images, picking the best, sending to inpaint, and then inpainting small areas many times to fix details.

[zealous horizon]

sure, everything has to be done manually in auto, but it's buttons, simple buttons to click. People love buttons. Clicky clicky
Gotta respect those hard clickers

[indigo carbon]

like, I feel like if A1111 would accept to switch to diffusers all of his struggles with trying to load multiple checkpoints and making the VRAM want to kill itself- the WebUI might return to its former glory.

for now using comfy as a backend and swarm or something like that as a frontend seems to be the way.

[boreal bough]

congrats! that one would pass for a genuine photo XD

no background blur either

[autumn forum]

thats a good photo. but i wish sd could do better on wheels lol

[polar epoch]

I relaly need to mark down what prompts i need for X thing. As i pretty much only use prompt info from civitA.I and generate SDXL from there lol. don't know squat of artstyles, cameras, techniques or anything lol

[timid sonnet]

There are loras for 2.1 that do wheels well

[autumn forum]

oooo.

[midnight shuttle]

I have finally found the actual code that runs the txt2img process in A1111. Trying to figure out why it uses so much VRAM.

[timid sonnet]

Offroad Corvette

[soft zealot]

part of the issues is(unlike comfy) it doesnt release anything from VRAM after its used it

[pallid sail]

Trying to load LoRAs in diffusers but getting key errors atm :/

[midnight shuttle]

The issue I've seen is more related to peak VRAM usage. Especially with the SDXL VAE. SDXL 1024 generation uses only about 9.9 GB but this goes up to 14.3 GB during VAE.

[pallid sail]

Are there any instances of torch.cuda.empty_cache() in the code?

[visual glade]

a1111 has always had issues with memory all SDXL did is make them more apparent

[soft zealot]

and that , it does like to spike at the end during image save

[midnight shuttle]

The code is giving me a major headache. I might give up on this project.

[soft zealot]

no idea, not looked

[midnight shuttle]

It spikes during the VAE step. I just don't know why or if it is normal.

[visual glade]

of course the code is bad, if it was good someone would have fixed it already

[soft zealot]

you get a spike in the VAE step with comfy as well, its just better managed

[primal vault]

you could try my inpainting workflow: https://civitai.com/models/111463

[pallid sail]

That's what I use in my Unreal script to make sure that I've flushed the model before I init a new one. I let diffusers handle loading/offloading the individual model components to and from CPU as needed and I never run out of VRAM, it'll just slow down

[midnight shuttle]

A spike makes sense since it's using another model. I just wanted to see if there was a reasonably simple explanation for the difference (a detailed technical explanation, not just "one code is better").

[static prawn]

i feel like every zombie horror stuff got those white lines in sdxl 😄

white canals

dunno how to describe 😄

i still like it

but im trying to get rid of it 😄

[drowsy coral]

How do you feel about SDXL 1.0? High contrast? High saturation? Is the age settings of the character too high?

[hasty smelt]

Hey everyone, is there any website or document to get the best prompts and appropriate resolutions?

[soft zealot]

@visual glade Question:

Is this pale blue box supposed to indicate where ComfyUI will open up at ?

[visual glade]

yes

[civic sigil]

Is it possible to use diffusers controlnets in comfyui?

[visual glade]

yes, they are supported

use them like regular controlnets

[civic sigil]

I think one I downloaded is glitched then

[soft zealot]

hmm If ive managed to create everything completely outside of it is there an easy way to find it cos if I zoom out its so fine and pale I cant see it lol

[soft zealot]

ok found it

[visual glade]

it's something I added recently so make sure you are updated

[static prawn]

this is like in every horror picture xD

dunno what sdxl likes about those white canals

😄

same here

[civic sigil]

I think that worked, seems to be running 👍

[soft zealot]

fixed that (ok maybe a bit too aggresively lol

[static prawn]

😄

but its weird i got this "style" in every horror picture

[polar epoch]

Add "in focus, hyper detailed" to it, wanna see if it makes your more detailed :P

[static prawn]

could give u like 10 horror pictures, every of them has those white sceleton

canals

i really love this one here 😄

what is the best rez for a full body portrait picture?

1024x1280?

[vast narwhal]

I need a reroute node that does that

[civic sigil]

Branching is illegal sir

[upbeat summit]

hey great SDXL output . yeah that would be good

[static prawn]

easy eat stuff directly right in your tummy? just get a zombie with a extra mouth

[midnight shuttle]

Looking at the A1111 code at least showed me how --medvram works and I realized it would speed up SDXL on my 12 GB card by getting stuff out of VRAM to make room for the VAE. 1024 generations went from about 40 seconds to about 25 seconds (20 steps).

[past raptor]

@soft zealot How did you get the 90 degree corners on your wires?

[hasty smelt]

Someone posted a document with a list of model prompts for XL, but I can't find it. Can someone repost it? Thanks.
the model have styles pressets

[cedar tusk]

Any guidance on what the height/width and target height/width for the base and refiner clips should be in Comfy SDXL? Should I match the ratio of the target image? Is there a best ratio between the height/width and target height/width for the base clip? Does it matter?

[upbeat summit]

you mean this? #✨｜sdxl message

[hasty smelt]

oh Thank you buddy

[cedar tusk]

Wait hold up how do you route the paths to perpendicular bends??? My workflow looks like a bad day in an Italian kitchen

[autumn forum]

put this in comfyui/web/extensions folder

[cedar tusk]

Thanks

[glad fulcrum]

Is there a guide to train lora on kohya ? I have 12GB of VRAM and it says it can do it.

But I don't even have it installed. I would love a tutorial for training sdxl Lora

[heady vale]

that really cleans up the spagetttt. nice one!

[hardy cipher]

man, I sometimes make some really wacky iflows as far as editting and finetuning the image between renderings. it's pretty much impossible to make it look good. lol. well as far as I know

[cedar tusk]

Should the height/width for the SDXL CLIP be a square ratio?

[autumn forum]

yes

width and height refiner and base i have at 4096. the target width and height i have set to 1024

[cedar tusk]

what about for the refiner? nevermind, misread what you wrote

[autumn forum]

lol

[hasty smelt]

Guys, my 8GB RTX 2070 Super wasn't able to run the SDXL model. There are freezes when loading the model, and even the minimum resolution of 1024 takes 29 minutes to render a single image. Fortunately, it works very quickly with other models. Let me know what the ideal Vram is to run the XL model.

[cedar tusk]

Mine works on Comfy but not on Auto1111.

[midnight shuttle]

For A1111 at least 16 GB. But try --medvram or --lowvram.

[cedar tusk]

Is there ever a reason to change it from those?

[midnight shuttle]

I have determined by reading code and by testing that --medvram increases SDXL generation speed by 25-35% on A1111 due to less swapping to system RAM during VAE step.

[autumn forum]

ive never had a reason. even when changin the resolution of the image.

[midnight shuttle]

Without --medvram 40 seconds, with --medvram 30 seconds per generation.

[hasty smelt]

I think I'll stick with other models until I upgrade my GPU someday. I don't have the patience to wait for long minutes, but what bothers me the most is that the model freezes my GPU for a few seconds during the load. I believe there's not enough memory left for the rest of the system during those moments. It's better to wait for a future upgrade. For now, I'll work with other models.

[civic sigil]

How much normal Ram do you have??

[hasty smelt]

I have 8 gb in total 2070 Super

[autumn forum]

if you want the best possible chance of using sdxl. use comfyui and do a simple node setup with just the base

[civic sigil]

How much system ram

[hasty smelt]

16 gb

[midnight shuttle]

I wonder of ComfyUI unloads more things or moves more things to CPU when not in use compared with A1111. Maybe that's why it uses VRAM better.

[civic sigil]

Yeah Comfy should be able to handle those specs, idk about A1111

[heady vale]

looks like it mostly sits in system RAM then is queued to the GPU when needed. Ive noticed it unloads VRAM after its done. A1111 would just keep it in VRAM

[midnight shuttle]

That could explain it. A1111 moves stuff around more with --medvram and that makes SDXL faster on 12 GB card even though it seems like it would be slower.

Due to less RAM swapping during inference.

[static prawn]

i went from 20 minutes to 2 minutes with -medvram

[midnight shuttle]

I didn't think it would help until I read the code and then it made sense.

The reason for speedup made sense I mean. Not the A1111 code. A1111 code makes no sense!

[tropic turret]

Nice!

[glad fulcrum]

I'm installing koyha... what should I do here please?

well I pressed 1

🤞

[cedar tusk]

So, how do you stop the refiner pass in Comfy from taking the work a LoRA is doing and hucking it in the trash?

[autumn forum]

i believe thats what you do, if it doesnt work try the others lmao

[autumn forum]

i dont think anyone knows yet. if we could train the lora for both that would be awesome but i dont think anyone has found out how to train refiner yet

[cedar tusk]

Hmm

[boreal bough]

you dont

[cedar tusk]

*queue Invincible meme*

[boreal bough]

A.) dont use refiner
B.) give refiner different prompt, to not touch some areas
C.) research for more options that we haven't discovered yet. (lora training refiner is currently not possible)

[cedar tusk]

I was able to 'cheat' it by basically giving the refiner 2 steps only and then set denoising to a very low number (modified the KSamplerAdvanced to make that visible)

[boreal bough]

if its a face lora, then A is your only option

[deft coral]

D - sell caith to me

[boreal bough]

🤣

I'm currently writing a comprehensive guide to lora training - after that I can simply link it everytime 😄

[deft coral]

i can't wait for full checkpoint training

DB is working on it

[west breach]

has someone tried training the refiner on 512x resolution?

[lean knot]

Do we have any recommendation for Lora training that isn't faces?

[boreal bough]

'full'. pruned, right? cause we officially dont have access to the full 1.0 weights

[boreal bough]

?

[deft coral]

will they ever get released

the weights

[lean knot]

Just asking if there's already a few examples/guides on the optimizer, LR, and such

the most popular ones are for faces right now

[deft coral]

ya the pruned model like 2.1 and 1.5 etc

[cedar tusk]

With an 8GB 2070 super, will I be able to do LoRA training for SDXL 1.0? I keep seeing recommendations for 12GB, but some have said 8GB...

[deft coral]

12GB will be for full checkpoint training i think

and 8 for lora

[boreal bough]

is what I'm wondering as well. even when pinged, sai staff ignores that question for now. might be a legal thing as well - essentially we just dont know, and they have no obligation to provide us with knowledge. might show up randomly in the following weeks, might be partner program only 🤷‍♂️

[deft coral]

wait

so this is a legal issues

dam

[boreal bough]

12 is the lowest we got working with 0.9 - which wasn't optimized.
in theory, 8gb should barely be possible now -but will definitely take a lot of tinkering.
I'd rather recommend google colab + kohya, as that will work without any issues, nor do you need any weird hacks that may impact final quality

[ionic dragon]

SAI said that training on SDXL is easier. So in what terms is it easy?

[deft coral]

they say its easier but no one can train it yet

[west breach]

I've had pretty good results from just 1 epoch when looking at the sample images

for a lora

[boreal bough]

to answer in his place - yes its become a lot easier.
just a very few aspects became like 10 times harder, but in return 90% of loras are now 10 times easier

[heady vale]

I think they meant you dont need to fight the model to get results

[boreal bough]

the fact that faces are one of those aspects is just unlucky XD

style loras are essentially fool proof to train now - even bad datasets + bad captioning + bad training settings will give an average quality lora - when compared to what 1.5 loras provided in terms of results

[deft coral]

@boreal bough wanna see a cursed image

[boreal bough]

concept loras are a lot easier to train now - only captioning stays equally important - so if you mess that up, its still a garbage lora

[lean knot]

From examples on CivitAI, there are already some pretty good loras for faces though

glad to know styles are doing better, as that's what interests me

[west breach]

I'm going to try a style lora. Just taking a while to get decent images from the 1.5 model I'm using to create the training data

[boreal bough]

faces/anatomy lora got 10x harder. definitely doable if you have a lot of experience, but dont do this as a first lora.
^ while its easy to 'get results' with sdxl, you are essentially damaging the model significantly if you make it badly. Many tutorials going around that do just that sadly.

[boreal bough]

overfitting is how most do it - it's also the only option I'll include in my guide for now - as its beginner friendly. but as you see in the picture above, that is the cost of doing it this way\

[autumn forum]

that donut looks good. fruit loops donut yummm

[static prawn]

i still dont get how u all get so great results, my results are "ok"

[boreal bough]

try way longer prompts, but written with natural language

[lean knot]

Fair enough. I feel that we had to take that overfitting with a lot of character/styles loras of 1.5 too

[boreal bough]

less commas = good for generally good images, even if not very accurate to the prompt

[deft coral]

[lean knot]

still hoping to get a good anime finetune

[lean knot]

but I want to start trying to train anyways

[midnight shuttle]

Many tutorials going around that do just that sadly.
That's true about much more than SDXL training. Too much Dunning-Kruger in the world of tutorials.

[west breach]

here are some pics from my 'grumpy style' lora, it makes people grumpy 😄

[boreal bough]

never gets boring seeing the grumpy faces

[deft coral]

@boreal bough

i did a bay movie with elon musk

[west breach]

needs more lens flare 😄

[deft coral]

this is cursed

[static prawn]

i mean the concept was ok, im just missing details, hard to describe them in natural language 😄

eyes look generally kinda bad

[autumn forum]

yall look at the bottom right hand corner haha

[glad fulcrum]

Now its asking me again:

[deft coral]

[glad fulcrum]

should I do any optional steps?

[midnight shuttle]

Uh, 5?

Looks like it already did those steps just from the small logs you showed.

[glad fulcrum]

3 and 4 not necesary?

[midnight shuttle]

Look in logs and see what it already did.

[deft coral]

elon and mark

[glad fulcrum]

oh yes you are right

[deft coral]

i can't

[glad fulcrum]

thanks

It seems it worked

[deft coral]

why am i not getting photo realistic images

[high skiff]

@visual glade So sorry to bother, just wanted to ask if you could help me with something.

I am wondering how to remove images that have been uploaded to the image load node, cause its getting really full of things

found the folder, nevermind!

[azure oxide]

pinged then found it literally 2 seconds later 😅

[high skiff]

yup, thats how I roll lol

when typing out questions, it can help to figure out solutions lol

[nimble heart]

where

[high skiff]

comfy UI, inputs

input folder

[nimble heart]

holy shit

doesn't work for inputs, right?

[dapper current]

nm sorry misread

[nimble heart]

be cool if there was a --dont-copy-inputs or something

[boreal bough]

ummm... I think there was?\

something soemthing right click + keyboard shortcut

someone ask comfy when he's on again

[ionic dragon]

Which aspects?
SDXL better understands a subject?

[serene sleet]

1. is it possible to train on 10GB of vram?
2. what batch size is possible on a 4090?

[deft coral]

7-8 for lora

with a 4090

@serene sleet

[boreal bough]

#✨｜sdxl message

[serene sleet]

did you change anything to achieve that? my 4090 doesnt go above 1 batch without ooming

[deft coral]

thats odd

it should do more if your lora training

[boreal bough]

training a car for example, now takes around 14~21 min to get a damn good lora

[serene sleet]

I tried both LETS and bmaltais kohya, both oom above that, are you using the studio drivers or game ready?

[deft coral]

game

[serene sleet]

welp

so am I

[ionic dragon]

So it takes less time to train too?
But the min vram is 24 😢

[rose vigil]

What does conditioning the noise do? (as described in https://youtu.be/2Xe79Nl_6jA )

[deft coral]

@serene sleet

[serene sleet]

I mean lora training, not generation

[deft coral]

i have't trained lora for SDXL

[serene sleet]

fair enough, sorry forgot to mention it haha

[boreal bough]

16gb vram for super quick training
12gb vram for no-compromise training

below that should be possible now thanks to fp16 support - but haven't had the time to test it out yet

[deft coral]

it should handel more than 1 tho

[ionic dragon]

Oh ok

[boreal bough]

training at lower resolution is possible - but don't. it does more damage than good if you train anything that isn't a style lora

[deft coral]

not bad

8 images at once on 10245

1024

[high skiff]

been messing with this for potential release alongside my 1.0 workflow drop, but I found its extremely unreliable and leads to tons of issues

[ionic dragon]

Yeah, we would need 1024x1024 images or atleast the dimensions specified

The supported res

[rose vigil]

What’s it supposed to achieve?

[high skiff]

for example, on one seed it looks like this

[nimble heart]

☠️

[high skiff]

on another it looks like this

its unreliable and often leads to some very messy and artifact filled results

@soft zealot and I were messing with it in DM's

[deft coral]

[rose vigil]

Thanks!

[boreal bough]

supported res is built in. just use bucket 512 ~ 2048 (same file "resolution" as 1024x1024, just other aspect ratios)

[nimble heart]

what res/batch is using 20 gigs?

[deft coral]

8

with 2014

1024

[nimble heart]

?

what card

[deft coral]

4090

[boreal bough]

yeah, but that's pre vae decode XD

[nimble heart]

huh. you have --high-vram?

[deft coral]

yep

[nimble heart]

Ah.

[ionic dragon]

Is multi res supported?
For 1.5 and 2.1 supports multi-res through EveryDream2

[boreal bough]

assuming you ignore vae decode, you can go up to batch 49 on the rtx4090

[nimble heart]

for 1024?

[boreal bough]

yep

[nimble heart]

damn

[deft coral]

how can i ignore vae decode

[boreal bough]

doesnt help you though, since vae gonna get stuck

[nimble heart]

I'm @ like 18 gigs batch 9 using base+refiner on my 7900 XTX

[boreal bough]

tiled vae - but you do so at cost of color accuracy

[ionic dragon]

@high skiff any more key aspects to make outputs better?

[nimble heart]

took 90 seconds for 30 steps, so 10 secs an image

[high skiff]

jesus the VRAM usage on AMD is high

[boreal bough]

Original Post:

So how efficient are batch sizes? using @nimble heart .json

Here's the table that absolutely no one asked for! 😄

RTX4090 running full model + tiled vae + cli argument --dont-upcast-attention

batch  1 = 10.9s per image
batch  2 = 7.7s per image | 09.6 GB vram + overhead |  15.44s
batch  3 = 7.2s per image | 10.0 GB vram + overhead |  21.50s
batch  4 = 6.9s per image | 10.4 GB vram + overhead |  27.39s
batch  8 = 6.3s per image | 11.9 GB vram + overhead |  50.41s
batch 16 = 5.9s per image | 12.0 GB vram + overhead |  95.32s
batch 25 = 6.0s per image | 14.9 GB vram + overhead | 151.00s
batch 32 = 6.0s per image | 17.4 GB vram + overhead | 192.95s
batch 36 = 6.0s per image | 19.9 GB vram + overhead | 215.71s
batch 49 = x.xs per image | 22.1 GB vram + overhead | xxx.xxs

squared numbers provide peak efficiency - but I also included some off-peak examples for science!
(I just barely ran out of memory for 49, since I have too much overhead on my win11 machine)

[nimble heart]

batch 1 is 13 secs/image for the same settings

[boreal bough]

just saw, I even did it with the workflow you posted 🤣

[nimble heart]

yup. don't think it unloaded anything though

[deft coral]

[nimble heart]

if you unload base/refiner when you're not using them it'd probably go down

[deft coral]

[high skiff]

what in the hell did you mess up lol

[boreal bough]

he did batch8 without tiled vae

I've been there 🤣

[nimble heart]

which .json

[high skiff]

he would have had to do a lot more than that

[nimble heart]

i've got like a dozen floating around now

[high skiff]

BS8 with SDXL runs fine on my 10GB 3080

[boreal bough]

I think it was the first one you ever posted

[nimble heart]

The one you used for your 2b lora?

[boreal bough]

not that it's a relevant test - I just wanted to see number go up

[boreal bough]

yep!

[nimble heart]

imma try 16 see if my DE crashes

gonna take a hot minute for the kernel to compile a 16k square lol

[boreal bough]

🤣

I should be able to do up to 64 now with fp16 + all memory optimizations we have

[nimble heart]

fp16 on the attention or the vae

[boreal bough]

since both are pruned, both? I think that's by default now though - if you use the 1.0 model

[queen ether]

Sdxl loras don't use the text file??? For training??

[boreal bough]

if you're using the 0.9 vae, not the 0.9 built in, then use the 0.9vae with the community fp16 fix

[boreal bough]

bad tutorials. yes they do, and caption files are very important

[nimble heart]

seems stable @ 20 gigs batch 16. I saw it hit 23.9 before dropping so it might've unloaded the refiner model while it's doing the base pass

[nimble heart]

both pruned? officially?

[boreal bough]

it's why the new 1.0 checkpoints are so much smaller, and use less vram

[nimble heart]

is that new? I'm still using the 1.0's from first release day

[boreal bough]

as a downside, it's why 'full' finetuning is kinda not possible XD

[boreal bough]

check size. they're not 13gb, right?

[boreal bough]

also get the new one, with the vae fix embedded. then you dont have to load the (big) vae separately

[glad flint]

I went to bed with my generation time optimized down to 30secs (which was fun)
I woke up and it's 170 secs again, hanging at the vae decode and before starting on the first steps with the k-sampler
same settings, same comfyui w/ nightly pytorch, and even less RAM used by system than yesterday
AHHHHHH
for reasons I don't know it seems to be getting RAM lockup again and I've spent hours trying to figure it out, I've even tested crazy small things to try and replicate the smoothness of how it was running yesterday.

[nimble heart]

base 6.5gb refiner 5.7gb
0.9 vae 320mb

so guess im gucci

[boreal bough]

dont forget windows uses vram too. every additional monitor uses vram, chrome uses vram if hardware acceleration is on, adobe apps use vram

[glad flint]

yeah i signed out of steam and I'm using a tiny monitor and I close chrome when I'm generating
my windows desktop has gone on a whole diet 😂

[boreal bough]

rip ❤️

[nimble heart]

On rocm 5.5 I was running out of vram decoding a 1536^2 image.
5.6 reduced it enough I can at least do that lol. Still gotta tile for 2048 tho

[boreal bough]

Vit-H thinks that rivet from Ratchet and Clank is a cat 🤣

[vale eagle]

@boreal bough I trained a character lora but when I used the workflow like base+lora->refiner, the refiner made the output less similar to the lora. Do you have any idea to avoid this? I could only use base+lora but it kinda makes the refiner useless.

[wicked frigate]

that's close enough

[boreal bough]

yes. exactly

[wicked frigate]

it figured out ratchet&clank too tho which is kinda impressive

[indigo carbon]

these are attempts to make the backrooms

[boreal bough]

its from the flavor chain. It might figure out rivet as well if I add a custom flavor chain

[wicked frigate]

the what

[boreal bough]

only first sentence counts, as that is from Vit-H directly

[nimble heart]

batch 16 in 152 seconds, < 10 secs an image now for 26 base and 4 refiner steps.

maybe when the winter rocm release comes and my gpu is "officially" supported I'll finally beat those 3090 bois 🔪

[boreal bough]

the first part, comes from Vit-H - which is the caption that reads like normal text.
Remaining tags come from a custom text file called 'flavor' - where it compares all words, and sees which 'fit' the image best.

But interrogator in A1111 comes presupplied with a custom flavor text of questionable quality.
If you add your own flavor text, then it uses the words that you yourself supply

[heady vale]

what do those images look like? Id rather have a bit more quality over the fastest gen

[nimble heart]

just using my principled sdxl node's default settings

I like the extra quality of 30 total steps over 20

I have a fast enough gpu it doesn't matter too much until you get to like 3 megapixels

[glad fulcrum]

does anyone got a 1024 regularization images of woman?

[west breach]

For some reason the interrogator extension in A1111 was ignoring my settings not to put artists in the caption

[boreal bough]

artist comes from flavor chain - which is 1 single txt file, so it 'can't' tell them apart from all the other captions

[west breach]

it seems to separate them properly on the analyze tab

[nimble heart]

I've found somewhere around 30-40 is the "safe" zone where a lot images start to become somewhat convergent. 20 -> 30 steps over half of your images get like major composition changes

[west breach]

training a new lora now, fingers crossed 🤞