✨｜sdxl

[fresh path]

yeah ik was just testing samplers

[uncut fiber]

ok

[south frigate]

A1111, comf looks unnecessarily complicated to me, because in the A1111 obtains exactly what I want. But if the comfy is much faster, then maybe I need to think about a change

I thought A1111 was already fully ready for SDXL after such a long time

[wind wagon]

Which GPU are you using?

[fresh path]

6900xt 16gb

[vale eagle]

If you have limited hardware and waiting so long for sdxl, you might want to try comfyui.

[uncut fiber]

Comfui are using even ppl leaning on A1111 because about SDXL performance. A1111 need desperadely new update to handle SDXL

[south frigate]

To be honest, it's still fast (I have a gf 4080) but I always try to gain time

I guess I'll just wait for the new update to A1111, comfi somehow discourages me with the way it looks - thanks for answers

[uncut fiber]

i have 3,14it/s with 3070 1024x1024 🙂

[south frigate]

but in comfy or a1111

comfy right?

[zinc cargo]

hi, trying to run sdxl on vlad, even with diffusers, with/without refiner and not getting the model to even load...
getting this error...
ERROR Diffusers failed loading model using pipeline:
C:\Users\USER\SD\stable-diffusion-webui\models\Stable-diffusion\sd_xl_base_1.0.safetensors
Stable Diffusion StableDiffusionPipeline.init() got an unexpected keyword argument
'text_encoder_2'

[uncut fiber]

right comfy @south frigate

[zinc cargo]

comfy on the other hand...

[dense chasm]

upsample and downsample because of unet

[soft zealot]

I mean ComfyUI can be made to look reasonably decent IMHO, this is my daily driver view as it loads (and yes I have turned off the spaghetti which helps)

It all depends on what you're looking to achieve. I mean I would be the last person in the world to recommend my workflow to anyone who wants to tinker however if someone just want to have some point & click & select boxes ??

Then maybe ...................

[uncut fiber]

i think see all those @south frigate workflow can make panic in human mind. It is working perfectly with most easy default workflow.

[south frigate]

I did a quick test and it actually renders almost twice as fast as the A1111. I will look on civit for some interesting "bulid" to Comfy and see what comes out of it.

[uncut fiber]

also you can gain 20-50% using AITemplate between model and Kdiffuser only.

[south frigate]

This is the first time I've heard of what you write.

Its for A111 or comfy or both?

[uncut fiber]

AFAIK only for ComfyUI, but can be for A1111 dont know

[south frigate]

This may be a different way of asking. What workflow do you recommend downloading to Comfy for me as a newbie?

[uncut fiber]

I recomment play with most simply default workflow to get familiar with it. For me is it good enough.
If something than probably from github @high skiff 's workflow. But you will have missed some nodes. But it is possible to get them all by using manager.
But as i said, i would play with default.

This is the thing that accelerate it, normaly is load checkpoint model connected to KSampler model.

[indigo carbon]

*on 3000 series GPUs, on 4000 series it's around ~95%-120%

[uncut fiber]

But it is not lossless.

[indigo carbon]

we already confirmed that. however, you might get a bigger speed boost on 3000 series GPUs if you compile the AIT modules yourself- the precompiled ones that are shipped with the node are meant for 4000 series, therefor lower boost on 3000 series

[uncut fiber]

yes i am happy with speed i have 🙂

[south frigate]

can i use those boosts in A1111 somehow? tbh i dont know what is aitempalte and kdiffuser totally yet

[indigo carbon]

only ComfyUI, sorry

[south frigate]

any simple tut how to use it in comfy?

[indigo carbon]

nope, I'm sure there will be in upcoming days though

[upbeat summit]

love the effects!

[uncut fiber]

Just reconnect. But you probably will need install manager. I have manager, dont know how install custom nodes other way.

[rustic surge]

Has anyone used this?

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

amazing

[indigo carbon]

is this allowing you to use ComfyUI as a backend with any configuration?

[rustic surge]

Sorry, I don't understand what you mean, because my native language is not English.

But recently, many new ui interfaces have appeared on the basis of comfyui

[gloomy barn]

yes

[strong field]

-same for me. I started looking at real people distantly. Like actually looking at facial features and studying them, admiring them

[vale eagle]

The stableswarmui is the official one to do that.

[south frigate]

"Just reconnect" what? I have manager atm. Where can I find it (aitemplate/kdiffuser)?

[uncut fiber]

i think loaders AITemplate

no wait a min

@south frigate in manager type AIT and dowlnoad it. Then in right click loaders load AITemplate

[south frigate]

this?

[uncut fiber]

yes

[south frigate]

looks like its here

now i need to connect it like i understand it

somewhere

[uncut fiber]

yes as picture i posted. RMB loaders, load AITemplate

[south frigate]

ah, sec

[uncut fiber]

also dont forgot switch it to disable, and resolution in empty latent space to 1024x1024

[south frigate]

keep_loaded have to be disable?

[uncut fiber]

yes

it is VRAM hungry and no performance hit

[south frigate]

ok, its working

i think

8it/s and 15 sec render

from near 30 sec

what a magic

xD

[uncut fiber]

now you can generate or examine nodes 🙂

[south frigate]

is there anything else I should install to start with? is this all I should know to start with comfyui?

[strong field]

What card you using?

[south frigate]

4080

[strong field]

https://tenor.com/view/you-seeing-what-im-seeing-germur-jorge-garcia-darma-mae-whitman-gif-18324596

@indigo carbon

[south frigate]

thats fast or slow?

[strong field]

Thats right in expected values and further confirms all the latest gibber jabber about AIT workflows and the expected performance increases

AIT was compiled for 4000 series gpu, so thats good for you

[vital ermine]

any link to ait?

[strong field]

https://github.com/FizzleDorf/AIT

[south frigate]

i hope they will fast add it to A1111 then, so good

[vital ermine]

thank you

[strong field]

i want to like a1111 but comfy much faster/efficient

[uncut fiber]

@south frigate also you can drag and drop image generated in Comfui into it, and you have all settings and workflow that was used to create it

[south frigate]

even renders that were not made in comfy?

[vale eagle]

only for comfyui

[uncut fiber]

yes only for comfy

[strong field]

yeah thats one of my fav features lol

[south frigate]

What add-ons do you recommend for beginners? Is there anything worth downloading?

i used adetailer in a1111, but like I see in manager there is only facedetailer similar to it

[vital ermine]

Too bad the AIT is only for Comfy as I can't use that.

[uncut fiber]

@south frigate if you are used to watch preview you can enabled it in manager window, but it cost some performance of course.

[peak dove]

The Undead Awake

[indigo carbon]

currently testing stableswarm, how do I make it see all the stuff in my ComfyUI install? it doesn't see the models

[vale eagle]

#🐝｜swarm-ui

[meager canopy]

My workflow includes a few custom nodes that you'd need to download, but it does then give you controlnet and a face detailer. I'm just working on a new version, that doesn't use the refiner.

https://civitai.com/models/119257/gtm-comfyui-workflows-including-sdxl-and-sd15

[peak dove]

What has Stable Swarm got over ComfyUI at all?

[south frigate]

looks a little simpler than most I've seen, I'll try it soon
where should i add AIT if I want to render faster in ur workflow? (https://github.com/FizzleDorf/AIT)

[strong field]

Symlinks, https://github.com/comfyanonymous/ComfyUI/blob/master/extra_model_paths.yaml.example

[candid walrus]

in comfy_nodes

[shy kelp]

anyone doing portraits getting weird teeth artifacts?

[strong field]

Just point the path from stable swarm comfy install to old comfy install

[strong field]

Sd1.5?

[shy kelp]

sdxl1.0 0.9vae

i try to put neutral expression

only 1 in 10 images comes out right

[meager canopy]

I've not tried it

[indigo carbon]

huh, awesome! stableswarm is just a fancier ComfyBox! by far my favorite

using my AIT workflow in same speeds as normal ComfyUI and stableswarm, no reason not to use stableswarm

[strong field]

base is kinda weird, crystalclearxl better for me

[mossy canopy]

Need comfyui touch control

[strong field]

i can think of a few

[indigo carbon]

fair enough, but still, it's awesome

[strong field]

hehe, i cant wait till they implement those targets, it will be a great frontend to comfy backen

[mossy canopy]

I might work on it..

[mossy canopy]

Not sure if comfybox is touch responsive

[indigo carbon]

for now I'll stay with my nerdy ComfyUI setup, but stableswarm is way better than I anticipated

[strong field]

thats good to know

[meager canopy]

Just added v4, which now doesn't use the refiner, which means less overhead.

https://civitai.com/models/119257?modelVersionId=141929

[strong field]

i ditched the refiner and have started using other models to refine, initial tests are interesting

would love a technical explanation on right/wrong from someone that knows

[meager canopy]

Same here.

They don't seem to be sure themselves.

I don't think there is a correct way, the answer is always "just experiment".

[strong field]

what i have found really really interesting is that SDXL and SD1.5 work very well together in model stacks

of course you have to upscale/downscale accordingly in the workflow but its wild what it can do

[meager canopy]

Yes, I've been using it for weeks that way.

[shy kelp]

can you stack sdxl with itself?

[meager canopy]

Yes

[shy kelp]

is it any good?

[shy kelp]

i dont know what that is lol

you got poses?

[meager canopy]

The UltimateSDUpscaler, in the middle of that, uses a 1.5 model to upscale, and then another one for face detailing on the right. An SDXL model can be used here, but I've not found it as good yet.

[strong field]

LOL you doing the same shit im doing

[shy kelp]

cool

how can i add an upscaler?

[meager canopy]

In ComfyUI?

You mean an upscaler model?

[shy kelp]

yes

im using the example workflow

[meager canopy]

C:\ComfyUI\models\upscale_models

[shy kelp]

can i just drop the sdxl base there or needs a new one?

[strong field]

different models for upscale

use comfy manager for easy

[strong field]

i could seed a screenshot of my shit but its UGLY lol

maybe one day ill make it pretty

i have two versions, one with regular US and another with USDU

[meager canopy]

Thanks! Mine has the workflow embedded in the image 🙂

[strong field]

will check it out, since you have a similar workflow i might just adapt haha

[meager canopy]

May as well. I'm sure that's how mine started out 😉

[strong field]

why reinvent the whl

[shy kelp]

where can i get one?

[molten gull]

does anyone get considerably better results with upscaling ?

[strong field]

do you have comfy manager?

[meager canopy]

Comfy Manager or https://openmodeldb.info/

[shy kelp]

not sure, i just loaded the sdxl_simple.png workflow and moved it around

[meager canopy]

You NEED this:

https://github.com/ltdrdata/ComfyUI-Manager

[shy kelp]

is that nvidia only? the readme mentions it

[ionic dragon]

SD best works with NVIDIA

[meager canopy]

What? The manager is just for making Comfy easier to use.

GPU makes no difference

[ionic dragon]

the manager works with any

i am saying that SD works better with NVIDIA

not the manager

[shy kelp]

works fine

also these teeth artifacts i mentioned earlier, anyone can help me get clean pictures?

[strong field]

perfect teeth in positive prompt might help

however, this may lead to hilarious expressions depending on the seed

[shy kelp]

i mean i'm trying for a neutral expression

[strong field]

positive: closed lips
negative: no teeth

its going to be a bit of an experiment

[shy kelp]

alright cool

[patent moon]

negative: teeth

putting no teeth is a double-negative

[shy kelp]

yea i tried that

both

still same poor results

[patent moon]

Just inpaint the teeth

[uncut fiber]

teeth 😄

[patent moon]

That's pretty much the way to handle any detail

[shy kelp]

how do i do that?

[patent moon]

You should figure out how to inpaint. It's a useful skill. Find a yt tutorial for whatever you're using.

[ionic dragon]

use refiner

[shy kelp]

what kind of steps yall using?

[meager canopy]

30

[soft zealot]

I prefer to think of it as a a MultiSampler Approach to creating AI generated Art rather than as a (in my case) pre condition, base, refiner pass

Effectivly all you are doing is multiple img2img steps with different models but using the latent space between steps rather than the pixel space

So you could have One sampler sending all the steps through using one model or you could split the creation process using 3 samplers and 3 models if you really wanted too (personally I use the same model dor Step/Sampler 0 & Step/Sampler2 and then the "main" model in Step/Sampler1 )

So for example these 3 images all follow the same format and wer e generated with same seed/prompt etc with only the model changing

LHS :: SDXL1 Base+SDXL Refiner
Centre :: SDXL1 Base + Dynavsion
RHS :: DynaVision+DynaVision

This is Stable Diffusion, ultimately there is no right or wrong way just the way you're happy with.

This is the way

[shy kelp]

with refiner?

[meager canopy]

My workflow doesn't use refiner

[shy kelp]

hm

why not

[meager canopy]

If it's with refiner, I used 26/4

[shy kelp]

cool

[meager canopy]

I get better results without it

[shy kelp]

i tried around 30 and 100 didnt see a lot of improvement

[patent moon]

30+ is decent. I feel like I get improved results up to around 60.

It depends on your sampler as well.

[strong field]

lol good point

[shy kelp]

you think it understands that kind of grammar tho?

[patent moon]

yes

[hoary saddle]

anyone have a good workflow for outpaint, i have a subject but want to change the background

[vagrant wasp]

correct way to use refiner ?

[rustic garnet]

don't do full denoise

the refiner is like an img2img pass and should use 10%-25% denoise strength

also a seed of 0 sounds like a bad idea. Use some random number

usually you use the same text conditioning as for the base model

[devout light]

For my trained model, can I manually input a noise and generate an image when I use it

[broken hare]

hey i was wondering where you went. I went down the same path wondering if it is worth the extra time to upscale images by default in SDXL. As for the current finetunes I'm using, it is just not worth it, and it doesnt improve the image much.....not sure about base

And you know the cool kids from the 2.1 chat are in the SDXL show and tell chat? Come by

[vagrant wasp]

this part

[mossy canopy]

https://huggingface.co/models?pipeline_tag=text-to-image&other=controlnet&sort=modified&search=diffusers/controlnet

[swift acorn]

Is there any option to install a graphical interface on ComfyUI? It seems to me so difficult to use normally that I don't want to look at it.

[mossy canopy]

Yes.. but my pc is wasted right now 😭 using free colab from my phone to run comfyui ... comfyui has no touch control. Barely managing with a halfbaked touch implementation..

[mossy canopy]

Use ComfyBox

[soft zealot]

I 've ended up with a daily view that acts as a pseudo traditional web page view especially with spaghetti turned off

(BTW COmfyUI does had a Graphical Interface, its just node based)

[rustic garnet]

why are all these example images always full of pornography 🤔

[soft zealot]

If you think thats pornography youreally need to get out more

[rustic garnet]

just look at your prompt

[hasty mortar]

First comfyui

[rustic garnet]

I can't even write it here cause its so full of pornographic content

[hasty mortar]

I need to learn how to use the refiner model btw

I'm guessing it helps with distortions 😅

[late marsh]

How is breasts porn?

[shy kelp]

it isn't if you don't show nipples right

[rustic garnet]

it's about the prompt. I don't want to blame Winston in particular. It's just something I see very often here

[late marsh]

They are just words.. 🤷‍♂️

[soft zealot]

IMHO you're being very small minded however I do agree that there are some dictionary based words in there that are used in some cpontexts to describe pornographic scenes howevr every word used is "proper english" rather than "vulgar slang" and they are written as "words" not "paragraphs & sentences"

[mossy canopy]

New to danbooru tags?

[swift acorn]

how to "Start the ComfyUI backend with python main.py --enable-cors-header." i dont really know this installation... can u help? (from here:https://github.com/space-nuko/ComfyBox)

[rustic garnet]

yes, I'm still new and disgusted by this.

[crisp owl]

Don't look on CivitAI

[soft zealot]

Dont use the internet!!

[rustic garnet]

I know CivitAI and yes, I find it really sad that such a cool technology like SD is mostly used for porn

[crisp owl]

Or any NSFW Stable Diffusion discord servers

[hoary saddle]

if you have an item that's transparent PNG and you want to add it to a table display, how would you do it? dont really want the original image to change but maybe around the edges so it blends with the ai generated background

[crisp owl]

mostly is a false pretense

[steady grove]

hahaha it's not small minded to feel aversion to pornography. you don't need to be a hedonist to have open minded thoughts

[soft zealot]

again, there is no pron in my screenshot , is there??

[steady grove]

hedonism is such a weird philosophy. it's very incompatible with stoicism

[compact hatch]

one fun thing about comfy is that the workflow is saved on the image, so you can easily load someone else's if you don't want to make yours from scratch

[rustic garnet]

anyways, I never said "block these images to protect our children!!!!1111". I just find it always a bit strange when people post their workflows and then you see stuff like masturb***, submissive women and stuff in their prompts.

[vagrant wasp]

is there lowVRam option like SD

[mossy canopy]

Os windows?

[thorny frost]

hi! have been mostly disconnected for 1 week or so. Any technological revolution lately?

[steady grove]

you're very fine and i see a lot of stoicism in you and that's a-okay. don't let people cut you done because you're not into the same material they are

[rustic garnet]

inpainting?

[soft zealot]

yes and are theose words in themselves pornographic?

[compact hatch]

having fun with the restart sampler, it's only in dev build for a1111 and requires a funky node in comfy but it's nice

[hoary saddle]

used the setup for inpainting on comfy github, not getting great results yet

[swift acorn]

windows 10

[steady grove]

oh it's bdsm. no wonder he's so defensive when people show aversions

[rustic garnet]

I have nothing against pornography in general

[soft zealot]

Most people can imagine what he face loooks like when you describe it as an "orgasm face"

That phrase doesnt mean it pornogrpahic.

Orgasms are a perfectly bnatural occurnce

[mossy canopy]

@rustic garnet join unstable diffusion server.. u will find new meaning in life..

[steady grove]

me neither! i just don't tango with it

[thorny frost]

ty, will take a look!

[soft zealot]

and yet that was yuor immediate reaction, to highlight what you perceived to be pornography (when it isn't)

[rustic garnet]

but in SD its often everywhere and this is something I find just sad. Like even many SD tutorials are using prompts of women in sexy pose. Why?

[soft zealot]

do you know what clickbait is?

[fresh path]

what else can you use sd for

horny furries are the backbone of tech development

[steady grove]

i got kicked off of there because i didn't want to post images in their hardcore channels so i posted in general imagery. a softcore princess peach. something you'd see in maxim maybe. they told me i had ot post in there nude channels. i said "i don't like looking at closeups of buttholes so i don't go in those channels" and they banned me

[mossy canopy]

Do u run using the bat file?

[fresh path]

lol thats funny

[elfin flare]

meow

[rustic garnet]

thats often claimed, but I doubt that the people who develop the methods are the same as the people who use it for jerking off 🤷

[steady grove]

not really. they sort of just pick up trends. if anything they gave us discord. that's the pinacle of furry accomplishment

[fresh path]

I thought you guys would have taken it half seriously

[swift acorn]

yeah, nothing happend prob cuz i dont undesrtand "Start the ComfyUI backend with python main.py --enable-cors-header." - idk what to do with it

[fresh path]

I think the nsfw aspect of generative models still brings a lot of interest in it, if you look at civitai for example

[steady grove]

the whole "pornography drives markets" thing is a myth. vhs never won because of porn. it's because betamax machiens were expensive and had a lot of compatibility issues and in many cases were uglier and bulkier. porn went with vhs because that's where the sales were. not the other way around.

[soft zealot]

you're presuming that all furries are simply in it for the purposes of maturbation? (BTW I find it interesting that earlier you self censored the word "masturbation" but are happy use the phrase "jacking off"

Double standards sir/madam delete as applicable

[mossy canopy]

Open the bat file using notepad and copy the text inside send it to me..

[fresh path]

what a topic

[steady grove]

one context was a prompt to create images. the other was a conversation. different contexts. self censoring is soemthing everyone does and it isn't something we should be lambasting people for

[fresh path]

I think you are right, but its not a very good example

[steady grove]

can we just drop the whole "porn is important and you must like porn" debate. it's never been high level

[crisp owl]

https://tenor.com/view/jason-momoa-chair-popcorn-time-ready-to-party-dis-gon-be-good-gif-24174564

[late marsh]

https://media.discordapp.net/attachments/1131466616526344292/1135471610766315520/popcorn2.gif

[soft zealot]

meanwhile to keep the OP happy here is an image with a large feline in it

(apparently 9u55y is blocked although masturbation & jacking off aren't)

[swift acorn]

its nothing change xD this is inside bat file

[mossy canopy]

Did you get it then..

[mossy canopy]

Let me know if you managed to run ComfyBox

[elfin flare]

Crystal Clear XL + Voxel Style LORA

[swift acorn]

yeah, tryed a lot of thinks, i dont know how to install this

w/e

[crisp owl]

SwarmUI is an option which uses ComfyUI backbone and has a more common UI

[mossy canopy]

It's just two clicks ... you run comfyui in the background with enable cors header argument. And in comfybox just run the run.bat ... that's it

[daring smelt]

+1, please and thank you

feel free to take it to DMs if you'd like to discuss it further, and not place it in the #✨｜sdxl channel

[soft zealot]

actually there's a related question that keeps it on topic.

Ther eare any number of NSFW Specifc Models for SD1.5 such as "Uber Realistic Porn Merge" , has any one found any good equivalents for SDXL yet?

Dynavision works and is great if your particular taste veers toweards everything lookin Pixarified

[daring smelt]

Sorry, this is not a NSFW server and we do not encourage conversation about the like

#✍🏼｜rules-and-tos

[soft zealot]

thanks for the link. and I do note there is a difference between posting NSFW content and simply discussing it. None of the former has been posted ;o)

wanders off back to his hole

[swift acorn]

run_nvidia_gpu.bat right?

and when comfy is ON, i trying .bat from the comfybox and its window for 1 sec and insta off, nothing happend

[mossy canopy]

That should work.. maybe try running both as admin

[soft zealot]

Well I did try contacting you directly but it appears I cant

SIGH

[half ivy]

can contact me bruv

[mossy canopy]

Did u git clone comfybox? Then u will have to build it... here is the link for built version https://nightly.link/space-nuko/ComfyBox/workflows/build-and-publish/master/ComfyBox-dist

[dense chasm]

i've just uploaded a new architecture LoRA model on huggingface,enjoy it https://huggingface.co/frank-chieng/sdxl_lora_architecture_siheyuan

[soft zealot]

nah you're fine, it was flower brought it up 🙂

sorry fruit

fruit/flower related

[steady grove]

at first i thought this was a new lora architecture and was getting ready to dig into documentation

[daring smelt]

You can contact using a #1010934719455707218 ticket here on the server-- I don't accept DMs 🙂
but since you aired it out here:
I'm not here to debate you on the server rules, just kindly asked us to move the conversation as we don't facilitate this sort of convo
it's not a debate, argument, or something for me to get granular about with you in what or what not is classified as "nsfw" or "pornographic"
So, let's kindly move on
ty

[steady grove]

still looks cool. thanks for uploading

[pallid path]

https://huggingface.co/diffusers/controlnet-canny-sdxl-1.0-small

ayo what is this

https://huggingface.co/diffusers/controlnet-canny-sdxl-1.0-mid

and this

smaller models!

we need comparisons

[soft zealot]

Ta

/me waits patiently for rule #4 to get updated so that it alkso sayd " No Discussion of NSFW content will be tolerated"

cos at the moment it doesn't ;o)

[daring smelt]

would you like to move on?

[thorny frost]

is there a way for comfyui to update all my node packs automatically at start?

it is tiresome to be git pulling every pack manually

[soft zealot]

Not really but I probably will 🙂

[daring smelt]

Great, thanks!

[half ivy]

opens door

[soft zealot]

NPs, consider the threat understood ;o)

[pallid path]

Gee willy winkers, an online threat!??!

[hasty smelt]

Hey everyone, does anyone have an optimized workflow for a low-resource system to share? Something simple without additional nodes, as I'm not familiar with working with them yet. I just want to generate some cool images, but my 16GB of RAM is running at its maximum capacity. I just encountered a Windows blue screen, and I don't want to stop using SDXL because of this. If anyone has a simple, optimized workflow for a modest PC, please reach out to me.

[soft zealot]

there's probably a rule about that somewhere lol

[pallid path]

amount of VRAM?

[hasty smelt]

I have 8 gb vRam and 16 gb ram

[pallid path]

Yeah with 6GB of vram I could get the base model working, idk if the refiner is really possible with 16GB of RAM :(

maybe if the models are unloaded from ram everytime you switch the models, but that would be dog slow

and I'm assuming you're using comfyui, so idk how much less resource intensive it could go

[soft zealot]

you dould run it GPU only so it doesnt swap back to System RAM but with only 8Gb VRAM thats probably not a wise move.

I would get on amazon and but a couple of cheap sticks of say 32Gb to intsall as a 64Gb pair

[hasty smelt]

I have a 7300MB SSD; I think it can handle this quickly. My issue lies in the VAE pass and also during the Refiner model; they are consuming all of my RAM.

[mossy canopy]

Tiled vae?

[hasty smelt]

yes

[mossy canopy]

Fp16 vae?

[hasty smelt]

Yes, I think I'll need to do that, stop trying to squeeze blood from a stone and get a new RAM kit. Unfortunately, I don't have the money at the moment; that will have to wait. I just wanted to try creating something cool in time to participate in the Civitai contest. It would be nice to have a new GPU as well.

[hasty smelt]

one moment

[mossy canopy]

https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/resolve/main/sdxl_vae.safetensors

[hasty smelt]

This this [is the only workflow that manages to open without having red windows, I adjusted the passage to (Tiled), but even so, there are moments of high RAM consumption peaks, 99% during the VAE passage.

[hasty smelt]

I'll see that, thanks buddy

@mossy canopy should I replace the 6gb model with it?

[mossy canopy]

Models/vae folder

[hasty smelt]

ok

[half ivy]

serious now, what does this model at the top do? like its not a thing like openpose or canny and the like?

[hasty smelt]

?

[late marsh]

It's one of the new and fancy controlnet model for SDXL

[mossy canopy]

Yes..

[half ivy]

but what though? like all of them in one?

[late marsh]

Depends what you downloaded. For some reason they have the same name 😆🤦‍♂️

[mossy canopy]

Run comfyui with this.. --force-fp16 --fp16-vae --dont-upcast-attention
Load vae separately in the workflow

[hasty smelt]

ok, I work on it

[half ivy]

im so lost and i think you've made it worse 😂

[mossy canopy]

Or You could just try my colab notebook

[late marsh]

Dl new ones! They are small and cute https://huggingface.co/models?pipeline_tag=text-to-image&other=controlnet&sort=modified&search=diffusers/controlnet EDIT: adding openpose https://huggingface.co/thibaud/controlnet-openpose-sdxl-1.0

[mossy canopy]

https://colab.research.google.com/github/x-legion/legions-sd-notebooks/blob/main/legions_comfyui_colab.ipynb#scrollTo=HdWvgiQJQ0Ga

[late marsh]

Dl new ones and rename

[hasty smelt]

Everything is new to me, I was used to 1111. I don't understand much about how Colab works, however, I have a premium Google account. I'm thinking of running SDXL in the cloud until I can upgrade my RAM. Let me know if this seems like a good idea?

[mossy canopy]

You could run it on free colab.. no loss there.. per day 2.5-3 hours max on a single account. Never tried colab pro.. so no clue ..

[mossy canopy]

Try running a few colab you will understand how it works..

[hasty smelt]

I'm not sure how it works, I bought cloud storage from Google a while ago. From what I can see here, SDXL is allocated in this space. I'll follow the tutorial and see if the workflow becomes more comfortable. Well, I guess anything is better than blue screens, isn't it?

[hasty mortar]

Guys

Help

Can't find sampling in comfyui

😦

[crisp owl]

sampling as in sampler name and scheduler?

If so, in the ksampler nodes

[vale eagle]

Double click empty space, enter sampler

[strong field]

Is the checkpoint merger in A1111 the best option?

[vocal rapids]

just tried out the depth preprocessor , look good

lol

[vale eagle]

or SuperMerger

[soft zealot]

That would be because @high skiff spent a lot of time ensuring it was written using only standard OOTB nodes included in ComfyUI .

As long as it doesn 't crash with an OOM error I wouldnt worry about it peaking at 99% usage

[high skiff]

peeng

Ah, yeah

It took me more than 2 weeks to figure out how to get a high quality functioning high-res fix equivalent for SDXL using only stock provided nodes to ensure compatibility

[strong field]

stay in latent space?

[high skiff]

No promises on a when, but I I'm currently working on the 1.1 release of my workflow, which should be coming complete with three separate pipelines.

One will be a very light standalone version of SDXL that can run on weaker computers, or just for people who want something light and simple

Another will be a full featured workflow with a secondary high-res fix that I am beta testing right now. From what I've seen, it's better in most ways, but you'll be able to choose between them

And then finally I am also prototyping a dedicated image to image workflow as well

[high skiff]

No, staying in latent space causes a ton of issues

[mossy canopy]

In the meantime i was just using ultimate upscale node

[high skiff]

My workflow is base image, forex pixel upscale, down sample, in code to latent, send to a high resolution sampler, and only resume sampling from the high frequency details, in order to preserve the shapes and underlying textures of the base image

[soft zealot]

especially for @high skiff

[high skiff]

I found that in almost every way comparable, ultimate upscale is unfortunately not as good as my current or even new high-res fix solutions

SDXL is not fine-tuned enough to deliver consistent images across tiles without having to use high denoise values with seams fix

My new high-res fix workflow that I have been working on uses a very stupid trick that shouldn't work, but clearly does lok

[strong field]

"resume sampling from the high frequency details," - explain, are you using step control?

[high skiff]

Yes, I'm using the advanced k sampler from comfy UI in order to resume from a later step rather than starting at the beginning of the diffusion process

It also injects noise before it continues from that later step, which only influences the higher frequency details rather than the lower frequency details

Overall, it takes the general shape of the original image sprays a little bit of light noise onto it, and then applies just a few steps of additional diffusion to bring back those crisp and clear high frequency details

My new high-res fix also works in that way, except I'm using additive/multiplicative noise field which are very finicky, and I'm messing around to find the best value

It's the same speed as the original, but it greatly increases texture preservation, and overall high frequency detail

[steady grove]

what are higher/lower frequency details? i'm not groking this analogy

[high skiff]

its not an analogy, thats actually what they are called

[steady grove]

they're wavelengths?

[high skiff]

low frequency is big shapes, high frequency is fine shapes

[high skiff]

effectively, yes

diffusion is based off of refining low to high frequency noise, which is what we call the "steps"

[steady grove]

that seems like an analogy. nevermind. i dont want to argue i was just trying to understand

[high skiff]

no worries

[steady grove]

i still don't understand why big shapes are a frequency

but not going to argue this

[strong field]

dammit thats the step im missing, i was using base image > upscale > down > send to sampler with lower noise values

but you refined it,

[azure oxide]

I'm using the advanced k sampler from comfy UI in order to resume from a later step rather than starting at the beginning of the diffusion process
any advantages of doing this over just fully denoising and starting a new ksampler with less noise multiplier or is this just a shortcut

[mossy canopy]

https://huggingface.co/faisalhr1997/comfyui/resolve/main/ksampler_sdxl.py

[strong field]

never considered only changing high freq

[high skiff]

That's what I was doing for a long time as well, but I ended up finding out that the advanced k sampler does not work how I initially thought

I thought that because it had no denoise slider, it would either entirely redefuse the image, or continue off of where it left

When instead, it injects noise on the step that it's on, preserving the already diffused details and only refining the finer details above

[strong field]

Id be very curious the answer here

[steady grove]

i guess i gotta study up on fourier transforms. i had no idea it was a big part of the diffusion process

[high skiff]

The benefit to it is that instead of rediffing the image from the lowest steps all the way up to the highest steps, it's able to dedicate that entire diffusion range just to the finer detail steps, and it also prevents the model from changing the underlying shapes or deviating too far from the original image

[strong field]

hmmmmnmm

[mossy canopy]

Try this node with this workflow.
https://huggingface.co/faisalhr1997/comfyui/resolve/main/2workflow.json

[strong field]

ok im going to test your hypoth, see if get something similar

[high skiff]

So, effectively, the way my high-rise fix works is it upscales, downscales, in codes, and then only diffuses the high resolution or small details, preserving the underlying composition much more reliably. And because it only has to do a few steps to achieve the full diffusion spectrum, it's also significantly faster

[strong field]

this is the important bit imo

nd then only diffuses the high resolution or small details,

[high skiff]

So, my high-res fix only needs about three steps to be successful

[strong field]

......i must see this in action

[azure oxide]

interesting, ya i kinda wanna test it out to see the differences myself

[high skiff]

That is the most important part of the workflow, in order to get properly refined high resolution details, you have to turn up the denoise, but when you turn up the denoise without specifying what step to continue on, it starts to change and misshape the underlying composition of the image

[swift acorn]

@mossy canopy comfybox is finally working, can I edit this graph somewhere? idk yet how but I want to add AIT (https://github.com/FizzleDorf/AIT) to faster renders, is it possible?

[strong field]

how do you decide which step?

[high skiff]

I would explain how my new high-res fix works, but I am not too confident I can properly explain it, and it's still not a finalized process

[high skiff]

In my old high-res fix, which is the one that is available to the public, you just treat it as the denoise basically

Whatever step you start on out of the total image is the inverse of the denoise

So if you give it 50 steps to diffuse, but only start at 45, that means 5/50, or 10% denoise

[strong field]

i believe i understand with your explanation already, i can build myself

[high skiff]

Is the new one uses a different process, you don't select which step it continues on, and it has some weird multiplicative multi-sampler noise jitter effects, all I know is that it takes the same amount of time and it produces even better results in my current one

[mossy canopy]

Sure.. why not.. ComfyBox is fully customizable with custom nodes

[high skiff]

I always get these ideas that fundamentally shouldn't work, but end up working lol

[strong field]

this i can deal with "weird multiplicative multi-sampler noise jitter effects,"

[azure oxide]

whatever works, works is how i roll

[strong field]

thats the beauty of comfy

the beaumfy *TM

[high skiff]

That's how I got my original mixed diffusion implementation, which is where the base model doesn't finish defusing before it sends the unfinished still noised latent to the refiner, thus saving time on the diffusion, and also preventing the refiner from damaging the bigger shapes/composition

And then my fractional offset which is still something I'm looking into, that was another little happy accident where I tested something that fundamentally shouldn't work, but introduced some very interesting results, of which are not as important as they were back in 0.9, but still something cool

[steady grove]

isn't that how the original sdxl workflows were made?

[vast narwhal]

About Depth controlnet, i'm doing right?

[mossy canopy]

Add custom nodes in comfyui as usual... it will automatically update in comfybox after restart

[steady grove]

first ones i saw had the refiner chewing on the last steps of the latents

[high skiff]

And then I came up with the first generation high-res fix, which isolated the second diffusion pass using only the base and not the refiner, while also only diffusing high resolution details on top of the original image to preserve quality as well as increased speed

[steady grove]

the SDXL report talks about that process too, the pdf they released during the beta

[high skiff]

And then the new high-res fix is a similar workflow, but fundamentally very different in how it diffuses the last few steps, and it produces even better results while not taking any more resources/time

Here, I have some images comparing the difference between my old and new high-res fix

[strong field]

just out of curiousity are you using this with offset lora?

SDXL has a weird latent shifty goings on

with high denoise it doesnt matter cause it will smash the latent together and redefine the space

[high skiff]

no, I don't use the offset lora at all

[swift acorn]

@mossy canopy Yeah, but... cuz i am newbie here, where should I add this https://github.com/FizzleDorf/AIT I mean, before control net, after etc. i have no idea how to add it well

[strong field]

i was curious if it would help the jitter but i will test

[high skiff]

my new high res fix relies on the jitter

but I am forcing the jitter

again, its a lot to explain

[steady grove]

doesn't work with sdxl controlnets

[shy kelp]

whats a controlnet?

[steady grove]

a gift from prometheus

[strong field]

model loader --> AIT --> next

[high skiff]

base image vs old high res fix vs new high res fix

You can see from the base to old old high res fix, it smooths out and loses a lot of the high frequency detail, and the new version greatly enhances it

base vs old vs new

[azure oxide]

ohhh thats right, new folks entering the scene from the hype of sdxl wont know what controlnet is

gonna be fun to blow their minds

[strong field]

1. middle looks like trash now
2. this has the latent shift i was telling you about from 1/2 to 3, but your method seems to have denoised the artifacts, cycle through 2 and 3 quickly and you will see what i mean

[high skiff]

for some images, the smoother appearance is preferred, but for anything textured or realistic, the new one will be preferred, which is fine cause they will both be included in the new 1.1 release

you can choose between them by simply changing one node connection

it does magical things for realism

mindblowing things

[azure oxide]

scroll down to the examples section https://github.com/lllyasviel/ControlNet-v1-1-nightly, in older SD models we had way more controlnets. some of them are just being released for SDXL now

[strong field]

i can only imagine haha!

[high skiff]

base vs old vs new

[strong field]

will build later and see what results i get

super intersted

thats amazing

[high skiff]

and when my high res fix dropped, it was really the best option, while also being super light to run, but now I was able to crank it up to 11 with really no downsides

[swift acorn]

but in comfyui or comfybox?

[high skiff]

its even MORE amazing for textures and skin

base vs old vs new

[strong field]

that is EERIE

[strong field]

comfyui

[high skiff]

base vs old vs new crop's

[steady grove]

have you had success using ait with sdxl controlnet? i thought it still wasn't working

[high skiff]

the new one just holds texture so much better

[steady grove]

maybe i need to update

[strong field]

also, i hate the way the middle looks, thats why i stopped using USDU

[high skiff]

base vs old vs new

it holds onto crunch very aggressively lol

which is why both will be an option

[strong field]

cant you control how much it holds onto tho? very customizable i imagine

[high skiff]

base vs my new workflow

also, the contrast difference is outside of the upscale

I also implemented a small image processing pipeline

[high skiff]

so with the old way, yes, with the new one, no

The trick to hold on to high frequency detail ONLY works at a single step value

Any higher and lower and it nukes

I am still looking into it, so we will see

[strong field]

ooooooo

so latent > sampler 1st step?

[high skiff]

the old one let you choose how faithful it was cause you could change the split

This one is the balancing of 10 plates at the same time aligning the stars to make it work

but it always works at the value I have now

[high skiff]

yes

[strong field]

ah now it makes sense why you said 3 steps

[high skiff]

where did I say 3 steps?

@steady grove

low to high frequency noise/detail

low frequency is big shapes, high is fine details

generally speaking, diffusion figures out the low frequency before the high

so I cut it and restart it again at just the high to preserve the low and medium

this one is more accurate to how image gen works

[high skiff]

ohhhh, no thats for the original

[strong field]

lol

[high skiff]

that was the minimum you could run it at for speed save

[strong field]

ah ok, i was confused

[swift acorn]

so at first I have to put it from ComfyBox into ComfyUI - but saving this file (.json) and trying to launch it in ComfyUI nothing happens, it just doesn't work. why? Everyone else is loading.
i use this "save" to pull this .json

[strong field]

try to make a picture with your workflow and load it in comfyui

i dont use box so im not sure how they wrap the comfy metadata

[icy brook]

👆 Made by @wet glacier using Aether Real SDXL by me

[hasty mortar]

I'm curious too

[strong field]

excellent results, is Aether Real SDXL available now?

[icy brook]

Not yet. Soon 🙂

[high skiff]

base vs old vs new

so much tegsture

[strong field]

looking forward to it

[icy brook]

🤗

[hasty mortar]

K

[south frigate]

where is model loader? i cant find it also

[hasty mortar]

I'm using 10x steps and 5x cfg

Works better

Not realistic though, need to work on that

[strong field]

like this

[indigo carbon]

already made a workflow that incorporates that, if they want they could just load one of my gens

[hasty mortar]

I'm at steps 200 cfg 25

[strong field]

they trying to get comfybox to do it

[steady grove]

these are waves

[hasty mortar]

Lacks realism, but I'm getting closer

[high skiff]

2-200 steps?

bro you need like.. 20

[hasty mortar]

I'ma try at 20

[south frigate]

gimme link to ur with AIT

if u can

[high skiff]

this image is 25 steps

[steady grove]

so samplers do actual wave forms on the schedule? neat

[south frigate]

finally something that doesn't look mega-complicated, thanks, I'm going to test it now

[high skiff]

that one is more accurate to what it looks like, yeah

[steady grove]

i thought it was more gausian random distribution each step

[high skiff]

this is what an incomplete latent looks like

[hasty mortar]

Do I need to run new nodes with python?

Like python main.py ?

[high skiff]

@steady groveThis is what an image looks like as it diffuses

well, thats with the residual noise, at least

[steady grove]

looks like unipc

[high skiff]

actually here, I can make a visual of it, its not that hard

[steady grove]

i've seen step by step progressions

appreciate the examples too

[native knot]

@high skiff - Really want your new upscale steps to improve this image I've been working on today:

[hasty mortar]

Latent image tricks https://youtu.be/OdMtJMzjNLg

[high skiff]

thats old news by now, you can do a lot more than just that

[hasty mortar]

Now would you please help me to install custom nodes? 😭😭😭

[high skiff]

kinda a dumbed down video of how it works, IMO

[strong field]

use comfy manager to install the nodes, thats easiest

no need to run the scripts of a custom node

comfy will pull them in at startup

[strong field]

great stuff

[native knot]

Thanks! I've stumbled upon a few words recently that add some real nice impact to images.

[marble kindle]

can anyone help me with depth controlnet on comfy why is mine like this

[strong field]

try lowering the strength

[high skiff]

fun

[hasty smelt]

Guys, I managed to replicate ComfyUI in Collab, but I'm unsure about how to add upscaler models. How do I install the upscaler models on Collab?

[high skiff]

fun

[vast narwhal]

Works better with base + refiner in my tests

[elfin flare]

LORAs are incredibly powerful tools, man.

[marble kindle]

can soemone please share their depth CN comfy file

[hasty smelt]

Guys, in Collab, is there a way to obtain another upscale model besides the three available for access? I want to know if I can use 4x_NMKD-Siax_200k?

[strong field]

!wget -c <link to any model> -p /models/uspcale_models/

[hasty smelt]

Hooo Thanks buddy

[strong field]

#!wget -c <LINK HERE> -P ./models/upscale_models/

cauition, i missed the period

take out the sharp to uncomment

i mean hashtag, music brain...

[high skiff]

I did it

after much manual labor

fully functioning step by step frame diffusion workflow lol

you can take all of the images and put them together in a gif sequence to see how the image diffuses

[crisp owl]

Good lord lol

[high skiff]

I can make it compatible with LoRA's too, but that will take over 20+ reroutes

so screw it, I'm in

[hasty smelt]

@strong field Does this sound good?

[high skiff]

also, it doesn't save the images just yet, so I have to redo all of the image nodes

they are just previews for now

[strong field]

capitalize the P

you misspelled "upscale"

under the folder name at the end

then should be good

[high skiff]

should I do 1-10 on top and 11-20 on bottom, or should I do 1 top, 2 bottom, 3 top, 4 bottom?

[strong field]

https://tenor.com/view/panda-bear-gif-21452758

[strong field]

makes more sense 1-10 top

[high skiff]

alright, i can do that

[strong field]

so you can naturally see the progression

[soft zealot]

0r
0-9 top row
19-10 bottom row

so it runs clockwise

[high skiff]

1-10
11-20

it now saves the images as well

not just previews anymore

as it is now, the samplers are actually set up in that

1,3,5,7,9
2,4,6,8,10

order, so I need to rearrange the whole thing to not have an aeurism lol

[hard fractal]

...wow!

um...

https://github.com/comfyanonymous/ComfyUI/blob/master/latent_preview.py

You can run comfy with --preview-method auto if you just wanna see it.

If you wanna save it, add a lil' something to the end there.

e.g.

for idx, x0 in enumerate(latent_list):
    preview_image = previewer.decode_latent_to_preview(x0)
    save_path = f"saved_images/image_{idx}.jpg"
    previewer.save_image(preview_image, save_path)

This way, each image will be saved with a unique name in a saved_images directory.

[crisp owl]

I oom if I try to do preview in the ksamplers

[hard fractal]

aha, I see

could also save each steps as a latent

and pull 'em out of the VAE later.

[hasty smelt]

@strong field I did it! Thank you very much for your patience in teaching me. I did it!

[hard fractal]

definitely use this:

https://github.com/madebyollin/taesd

[high skiff]

there we go, added LoRA Support

[hard fractal]

lord almighty

[high skiff]

gonna add a primative for sampler as well, this is fun haha

[hard fractal]

https://giphy.com/gifs/fallontonight-reaction-lol-fxkihLhhmp6LHRidye

oh, that's not the right reaction at all

[hard fractal]

woah!

rita?

[high skiff]

is gorjus

[upbeat summit]

yep

[native knot]

ORA! ORA! ORA!

😄

[hard fractal]

Oh, now I gotta spam my Ritas.

[high skiff]

Ah, i had Joe blocked, my bad, I thought you were somebody else @hard fractal

I saw you recommend the viewer

this is for making gifs to share with people

[hard fractal]

[high skiff]

it saves all of the images and you can sequence them

something I wanted to mess with

[native knot]

Nice

[hard fractal]

are you using the LoRA on CivitAI?

[high skiff]

@hard fractalOh, Joe, I made yet another high res fix whish is even better than before. It runs in the same time and outperforms other methods even more, and also still runs stock

and supports LoRAs from end to end

[hard fractal]

I thought you had blocked me!

[upbeat summit]

nope. a new fine-tuned SDXL model by @icy brook not trained on famous people (focusing on skin textures). rest is prompting - post is just upscale and some film grain.

[high skiff]

I did, i am not sure why

[hard fractal]

[high skiff]

I unblocked you, I thought you were somebody else

[hasty smelt]

Guys, I received this error as soon as the image was sent to the Upscaler 4x_NMKD-Siax_200k. What should I do to make it work without errors?

[hard fractal]

Got a workflow?

[strong field]

whats the model brutha?

[elfin flare]

I stand before thee, thy tiny knight.

[high skiff]

Working on 3 new ones for my 1.1 release

New workflow with a tighter pipeline, and 2 different high res fix options

Another one dedicated to just img2img

and a third light and easy to run pipeline with no fluff

but man, the new high res fix way outshines the other two

it preserves textures and enhances clarity much better

doesn't sandblast skin texture

[hard fractal]

Any ControlNet workflows?

[high skiff]

None yet

Haven't needed controlnet yet

I am also working on a realism LoRA

which is in the early steps, but shows a lot of promise

[hard fractal]

Nice!

[elfin flare]

Does controlnet work for SDXL on Vladmantic yet?

🤔

[high skiff]

base SDXL (with my LoRA) vs old high res fix vs new high res fix

[elfin flare]

"new high res fix"

What is this new high res fix you speak of?

[high skiff]

a new process I found yesterday

it shouldn't work, but it does haha

still testing it and learning more about it

the difference is small in that image

here, have this one

[elfin flare]

I would love to generate much higher resolution images. Without coherency lost that is.

[high skiff]

base SDXL vs old high res fix vs new high res fix

new one absolutely kills for realism and texture

[elfin flare]

Yeah, that looks quite good.

[high skiff]

it also does fantastic for textured things like water color

these are crops

base vs old vs new

new one massively enhances texture and tonality, while also picking up more grit like real water color

[elfin flare]

Is this high-res fix of yours working in SD.Next?

Or is it specific to a WebUI?

[strong field]

dumb question but where do i find diffusion_pytorch_model.safetensors

[elfin flare]

Allow me to get it for you.

[high skiff]

its specific to comfy UI

it relies on comfy's special advanced sampler

[elfin flare]

I do not use nor want to use that UI.

[strong field]

been googling for like 10ms

[elfin flare]

🤷‍♂️

[marble kindle]

can someone share their comfy ui depth CN workflow

[elfin flare]

https://github.com/vladmandic/automatic/wiki/SD-XL

[high skiff]

base vs old vs new

you can see how much better it enhances texture and tonality, while not hallucinating duiplicates and such

[hard fractal]

1825x1080?

[high skiff]

they are screenshots

image is actually uhhh, let me find it

[strong field]

thanks landed on the HF page and wasnt sure that was what i needed, appreciate it

[vast narwhal]

https://huggingface.co/diffusers/controlnet-depth-sdxl-1.0

[high skiff]

2704x1600

[hard fractal]

Wanna see a comparison?

This is on the bot:

[high skiff]

weird that the nose, ears, and eyes don't align

[vast narwhal]

Nice

[high skiff]

still much more control than before tho, thats cool to see

[hard fractal]

Ah, just realized I ran that depth map through midas

oops

But still.

[high skiff]

this is using my realism LoRA with my new high res fix as well

[hard fractal]

There it is:

[high skiff]

actually, let me convert that oen so you all can see it at full 2048x

[high skiff]

ahhh, ok, much better haha

I was like, that other one did not match lol

[hard fractal]

double depth.

[hard fractal]

looks super good!

[high skiff]

there we go

fill image converted to JPG

I have not tested this new upscale to 4096x, I should try that

[strong field]

anybody familiar with this error?

when trying to run this

[vast narwhal]

I can't see what is wrong this way

[strong field]

this any better?

[vast narwhal]

Do you have all the nodes? Screenshot the whole thing

[vast narwhal]

Maybe trying to update comfy ui and the nodes, there's not much to go wrong, the midas is working fine

[strong field]

ok will try, thx for help

[rustic surge]

Why can't I run Ultimate SD Upscale

please help me！！

[vast narwhal]

This is the end of the error, let me see the beginning

[rustic surge]

[rustic surge]

i don't know why

[vast narwhal]

You're using a SDXL model, that's the problem

[rustic surge]

Doesn't this node support SDXL?

[short marsh]

Loving ControlNet

[rustic surge]

so how do i get more details

can't I run Ultimate SD Upscale

[vast narwhal]

It works but i get this error trying to use the same prompt encoder.

[hardy cipher]

steampunk dia de los Muertos

[rustic surge]

how do i solve this problem

[vital ermine]

Precisely

🙂

[shy kelp]

Why is controlnet not working for me with the bot?

[vast narwhal]

Make new prompt and prompt encoders, make new everything, you will get more freedom, you will be able to use a different model

[rustic surge]

OK

I try it now, thank you for your help

[shy kelp]

guys i made skibidi toilet

[hardy cipher]

does anyone know what this means?

adm 2816
making attention of type 'vanilla-pytorch' with 512 in_channels
Working with z of shape (1, 4, 32, 32) = 4096 dimensions.
making attention of type 'vanilla-pytorch' with 512 in_channels
missing {'cond_stage_model.clip_g.transformer.text_model.embeddings.position_ids'}
model_type EPS
adm 2560
making attention of type 'vanilla-pytorch' with 512 in_channels
Working with z of shape (1, 4, 32, 32) = 4096 dimensions.
making attention of type 'vanilla-pytorch' with 512 in_channels
missing {'cond_stage_model.clip_g.transformer.text_model.embeddings.position_ids'}```

[hardy cipher]

wow, thanks for the unsettling circa 2004 cgi guy

[gloomy barn]

5-7x smaller controlnet model? Hurrah. Still not supported in SDNext...boo. I'm about to try that comfy front end thing.,...can't remember the name

[rigid adder]

Having fun with cotrolnet running my old Disco Diffusion png's with the same prompt even keeping trending on artstation

[crisp owl]

hoodlums

[unreal plover]

Ruuuun !

[unreal plover]

That is a game i am making, i hope i get this visual result

or this

[vital ermine]

Well, lol

[upbeat summit]

reassembling in progress

[vital ermine]

I wonder how that happened?

[unreal plover]

very high res make this kind of stuff, or refiner

[vital ermine]

No refiner used

[hardy cipher]

weird dimensions?

[vital ermine]

1920x1080

[hardy cipher]

well that's basically double the pixels

that are recommended

[vital ermine]

Worked with 2.1 so did we take a huge step backwards?

[elfin flare]

So Controlnet works with SDXL now huh?

That's good.

[hardy cipher]

everyone so quick to blame the models. I don't know. depends on the prompt, seed, etc, I'd imagine. I was getting all sorts of 2 headed abominations with the beloved 1.5

[vital ermine]

Never liked 1.5 due to that

I was a 2.1 user as soon as 2.0 dropped I ditch 1.5

[upbeat summit]

this looks awesome - if you need a soundtrack let me know 😄

[soft bone]

producer??

[upbeat summit]

yeah - musician and producing

[hardy cipher]

well fact is all the models are trained on specific pixel counts. so if we deviate from that specifically then they can do weird things. I don't know the science

I have music things myself

[soft bone]

excellent, same! genre/instrument?

[hardy cipher]

I have keyboards that make various sounds. and things with light up buttons that make sounds too

[upbeat summit]

learned the piano - very mixed genres. from blues to ambient, from hybrid orchestral to electronic genres. the latest stuff is mostly demos for virtual instruments by my buddies and used to promote their products: https://soundcloud.com/masslevel

[hardy cipher]

https://tenor.com/view/cat-gato-funny-dj-festa-gif-15160628

[upbeat summit]

https://tenor.com/view/cat-jam-cat-jam-head-bounce-kitty-gif-17946989

[hardy cipher]

https://tenor.com/view/cat-dance-dancing-epic-animal-gif-17713979

[hardy cipher]

I'm pretty big time. let a guy online use one of my songs in the opening sequence of some corporate instructional video he was making.

[soft bone]

sweet! those are nicely mixed

[hardy cipher]

yeah, those are pretty high quality tbh

[soft bone]

I really like Magic Castle 😄

[crisp owl]

prompt: Jiving to retro synthwave produced by Masslevel

[upbeat summit]

Thank you 🙂 Glad you like it

[hardy cipher]

that frog lives a cooler life than I do

[vital ermine]

Hahahaha

[tender timber]

Haven't tried, it but this is supposed to reroute "for real"

https://github.com/rgthree/rgthree-comfy

And some other cool stuff that would be helpful if they work as advertised

[unreal plover]

Here is a variable based reroot system ? Like setting a variable, then using this variable as an input of an other node ?

[hardy cipher]

I kind of like what this does for "analog" images

might need some adjustments

[polar epoch]

Is there a better way to upscale than this way with SDXL? Like throw upscale image to set res, then Ksampler to take last gen's latent to continue generating the higher res where the lower res left off.

[vital ermine]

I need to figure out how to get llama-13b. 7b is pretty cool

[hardy cipher]

how to get it working?

[vital ermine]

Not sure about any of it tbh. 30b too

[hardy cipher]

well what do you use to run them?

[vital ermine]

7b was 80-90% accrate for captioning my dataset

blew blip2 in to the garbage pail

[unreal plover]

we dont need telescope anymore

[short marsh]

happy with my controlnet interface

[elfin flare]

I'm starting to realize more potentials with SDXL.

Using Img2Img to correct some minor issues at higher resolutions is useful.

Starting 1280x720 then upscaled to 1920x1080

[polar epoch]

Sorry lads, too large for discord

Just kidding, here you go lol

[vital ermine]

So long, and thanks for all the fish.

[hardy cipher]

400 mb image

jiminy crickets

[hardy cipher]

good news everyone. for what I believe has been several days now, I've been using the wrong noise offset lora. it's been giving me some pretty subpar results. but I just assumed I'd been using it wrong. which I was, since it's for 1.5. I have the right one and I've had it the entire time. I just, at some point, started defaulting to the wrong one

https://tenor.com/view/think-smart-gif-22411928

[hardy cipher]

and now I know why I kept getting tensor size errors, lolol

[elfin flare]

Well at least you know why now.

[hardy cipher]

indeed. tbh it really didn't break images like some of the loras probably would. but definitely didn't do me any favors

the lack of naming standardization in general as far as loras concerned is extremely frustrating

I'm at the point where I might just purge about 75 percent of them since it's often a huge chore to figure out what they even are

[heady vale]

If you know what each is made for, create sub folders in the lora folder. You can do the same with models

[hardy cipher]

oh, I did for sdxl models. but didn't really seperate things much beyond that. problem is, for whatever reason, I just started using the wrong noise offset lora. it's not in the sdxl folder and I should have known better. but here we are

[hardy cipher]

https://tenor.com/view/halo-master-chief-nod-nodding-pablo-schreiber-gif-26177401

[elfin flare]

What LORA?

Or is it a lora.

big thonkl

[heady vale]

voxel xl

[elfin flare]

Ah okay, I'm also using it

It did very well with master chief

[heady vale]

yeah some errors with the suit but pretty good

[elfin flare]

He lost a wing. RIP.

[heady vale]

SD very rarely draws both wings. sad

[hardy cipher]

I suggested fictiverse check out the voxel lora

and then he told me to look at who made it

[heady vale]

loool

[elfin flare]

Lol

[elfin flare]

Wha

Who is that

[vital ermine]

I dunno

[elfin flare]

He looks right out of an 80's film lmao

[vital ermine]

Yeah, LOL

[hardy cipher]

I don't mean to judge, but he might have fetal alcohol syndrome

[elfin flare]

Possibly.

He could also be...

[vital ermine]

That nose, though

[elfin flare]

https://tenor.com/view/monke-gif-20514159

doing this.

[hardy cipher]

it's not a bd idea

[vital ermine]

DUDE

same seed but a woman

[hardy cipher]

siblings

what model are you using. these are top notch

[vital ermine]

Is that his wife or sister? Yes.

[elfin flare]

The top of their head

Is flat

LOL

[vital ermine]

A lora I just trained without captions and just base

I need to test more

[hardy cipher]

beautiful

[elfin flare]

SDXL is so...refreshing.

Being able to just generate nice 1080P images.

A welcome change.

1080P seems to be the most reasonable

4K is a letdown

[vital ermine]

I have no idea but I think there is the back of his head

[elfin flare]

We now know

that it is not human

I am afraid.

[vital ermine]

LOL

[hardy cipher]

are you trying to render in straight 4k or upscaling/

?

[elfin flare]

Are these the aliens that guy mentioned in the court case?

[elfin flare]

1280x720 img2img upscaled to 1920x1080

Maintaining the original 1024x1024 pixel count for coherency

[vital ermine]

I dunno but this did not turn out as I expected

I need to find a llama-13b

[elfin flare]

I can get you one

A very good one indeed

[vital ermine]

I was using 7b for captioning and damn, 80-90% hit rate

[elfin flare]

GGML or GPTQ?

[vital ermine]

Minigpt

[elfin flare]

MiniGPT. Interesting.

[vital ermine]

7b with 50 beams rocks and I was told 13b is even better

[elfin flare]

I would've passed you https://huggingface.co/TheBloke/Stable-Platypus2-13B-GGML

[vital ermine]

I looked at his

[elfin flare]

It's what I use in Kobold.

For LLM gens.

[vital ermine]

but I was in a rush this morning

minigpt4 only has 7b

[elfin flare]

When you say captioning

I assume you just mean interrogation

CLIP interro?

[vital ermine]

aye

[elfin flare]

BingGPT can accept images now

🤷‍♂️

[vital ermine]

don't like it and no batch

[elfin flare]

But local use is still better imo

Ye.