✨｜sdxl

[lusty token]

I'm on a Mac

[vocal stream]

ram yeh but you can make it consume 0 vram by just disabling hardware acceleration

[crisp owl]

don't think that matters. hold spacebar, move your mouse around

[potent flame]

wait it did use RAM? ok

[halcyon tusk]

can't see, it was deleted.

[potent flame]

someone post a server link to here

seem to be a NSFW server or something but I didnt screenshot

[halcyon tusk]

no, they def. cannot, thanks for the ping

[crisp owl]

snipersamir posted invite to a server called "hub-family" is what it was

[potent flame]

with 3000 members

dont know what is it anyway.

[halcyon tusk]

@smoky patrol dealt with him! thanks everyone

[lusty token]

Must be a limitation of using MagicMouse... I also tried out LiteGraphs demo just rule out any issues with my local install.
https://tamats.com/projects/litegraph/editor/

[boreal bough]

offtopic - but dont accept random invites
that's how you get a lot of private messages about invites once you're on a spam list for crypto people

[vital ermine]

I do just to watch it happen on my lazy days. Just had it happen.

[smoky patrol]

I've got a hammer, and I know how to use it

[soft zealot]

TBH , I never vary it and I rarely alter the steps but th enice thing is If you change the steps then anything else gets recalculated automatically rather than having to work the ratios out

[upbeat summit]

but if they promise a free 4090?

[soft zealot]

11Gb works abosolutely fine 🙂

[crisp owl]

8gb works, I just have to use tiled vae decode for my last 2x upscale to 2048. Otherwise, I haven't hit any roadblocks

[lilac raven]

Has someone ever seen a 4060 Ti with 16 GB anywhere?

[lusty token]

@soft zealot @boreal bough absolutely embarrassed to share this considering all big number vcards floating through here... but I'm happily running SDXL ComfyUI on a MacMini with 16GB of Ram.

I'm actually continually astonished at how well it's running and the quality of the images I'm getting.

For my purposes, which is mainly just toying around, it suits me.

Easily upscaling to 6144x10240.

I'm not obsessed with the Joker... just works as good test image for me 🙂

[boreal bough]

gg. I think you win for lowest nvidia card. with no nvidia card 🤣

[unreal plover]

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

[hoary saddle]

For reals?

[unreal plover]

https://www.reddit.com/r/StableDiffusion/comments/15n6lso/comment/jvki53u/

[lilac raven]

The repository says it's a pickle, wasn't there something pinned before because pickle was dangerous?

[patent badger]

yeah

I would stick to safetensors

[boreal bough]

yep its real and cool

hot damn

Compute
one 8xA100 machine
Batch size
Data parallel with a single gpu batch size of 8 for a total batch size of 64.
Hyper Parameters
Constant learning rate of 1e-4 scaled by batch size for total learning rate of 64e-4
Mixed precision
fp16

for a finetune that is both insane and cool

[patent badger]

I think that one is a virus, the huggingspace user looks sus af

[indigo carbon]

new SDXL AIT workflow:

[boreal bough]

it's real. but its diffusers.

not relevant for A1111 or comfyui

[lilac raven]

That yellow HF STAFF in all caps, scary!

[unreal plover]

I think we can trust "diffusers" repo no ?

https://github.com/huggingface/diffusers/blob/main/examples/controlnet/README_sdxl.md

[upbeat summit]

Somehow I managed to get this working with ComfyUI
https://huggingface.co/diffusers/controlnet-sdxl-1.0/discussions/1#64d48d8fe143badbbafb5b43

[boreal bough]

yeah. I was about to say give it a few days for a conversion... but hot damn not even 24hours

[upbeat summit]

AIT, ControlNet... exciting!

[vale eagle]

Still couldn't get AIT to work. The modules is so confusing. Even I manully download the file and modify the code. I still couldn't get AIT node to work.

[indigo carbon]

if you are on linux it's not supported yet

[vale eagle]

window

[indigo carbon]

there is a new branch for the node that doesn't need complex stuff to set up my workflow

[cold mica]

Join us for comfyui ep 2 today 👀

https://discord.gg/tR2jCs8E?event=1138521432079470595

[indigo carbon]

@vale eagle delete the AIT plugin in ComfyUI\custom_nodes and replace with this: https://github.com/FizzleDorf/AIT/tree/sdxl

[vale eagle]

Let me try again

[vale eagle]

Do you encounter this issue?

[indigo carbon]

it appears you are trying to use live preview. the best live preview that works with AIT is the one in ComfyUI manager.

[vale eagle]

I git clone the repo and switch to sdxl branch

[indigo carbon]

no. you need to start from exactly the SDXL branch. just "download as zip", it will work with my workflow right away- if you use the right live preview or don't use any live preview at all

[vale eagle]

Still the same result using zip

and loaded your latest workflow

[indigo carbon]

odd. are you using live preview?

that's the only issue I can think might cause this

[vale eagle]

What do you mean live preview?

[upbeat summit]

I've seen that error before

[trim orbit]

that tensor is dense

[upbeat summit]

can you go into your ComfyUI folder and than into /models/vae_approx and see if there are any files in there?

[trim orbit]

unrelated, but i've been using that new preview method in automatic's dev branch, maybe main by now. i lik eit

[vale eagle]

yes, taesdxl_encoder and taesdxl_decoder

[ashen oracle]

can you share a workflow you used with this

[boreal bough]

workflow is here

Somehow I managed to get this working with ComfyUI
https://huggingface.co/diffusers/controlnet-sdxl-1.0/discussions/1#64d48d8fe143badbbafb5b43

[trim orbit]

shouldn't be hard right? the other controlnet i just wired up in comfy and it worked, but sucked

oh no theres all this patching and stuff

[ashen oracle]

i'm relatively new to node workflows / Comfy and have learned to at least ask once before I sink 5 hours and end up wiping slate clean to start over 😄

[trim orbit]

same old shit i always see whenever i encounter diffusers

[upbeat summit]

delete taesdxl_decoder and it will work 🙂 the high quality previews are currently not working

[trim orbit]

i still don't understand why diffusers are better from an end user perspective. people talk about them a lot since vladmantic started hyping them for his fork

[vale eagle]

It is the magic

[boreal bough]

ehhh. I mean that image posted at the end, literally has the workflow embedded

drag that into comfyui for the workflow

[ashen oracle]

oh, i have it now

i hadn't opened the link , or had more than a sip of coffee when i asked

[upbeat summit]

I haven't used it yet. I only posted the link.

[trim orbit]

the new TAESD preview. it's quite fast and better looking than approx (imo)

i'm using them right now. on a 4080 though so maybe it's an architectural thing

[upbeat summit]

which one?

[trim orbit]

on automatic1111

[upbeat summit]

ah k

yeah I've been using them in comfyui

[trim orbit]

on comfy however, are those canny nodes, is that part of this? https://github.com/Fannovel16/comfy_controlnet_preprocessors

i can't get that to install on my machine

ohhh hey no it's not. it's just in among the default nodes

someone should just rehost c.safetensors cause i'm not into the whole converting diffusers thing. everytime i've tried going that route, i've had to learn how to code and what diffusers are all about. btu it looks like a c.safetensors exists already and the permissive rail-m license is like "yeah go ahead and rehost that brother!"

So i personally vote that it's a good idea

[elfin flare]

Negative prompt: Mutated. Disfigured. Multiple Limbs. Disfigured weapon/sword.
Steps: 20 | Seed: 468068609 | Sampler: DDIM | CFG scale: 6 | Size: 1080x1080 | Parser: Full parser | Model: sd_xl_base_1.0_0.9vae | Model hash: be9edd61 | VAE: sdxl-vae-fp16-fix | Version: 0a7105d | Pipeline: Diffusers | Operations: txt2img | Lora hashes: "VoxelXL_v1: 796529775279"```

[unreal plover]

reduce controlnet strenght

[modern juniper]

it works so well

sytan workflow with added controlnet stuff

[soft zealot]

theres an option you can add toyour startup (forgotten what it is now) that puts live previews on Ksamplers as they got though the steps, it does eat into VRAM though

[indigo carbon]

we already resolved that issue, he got AIT working on the workflow I made.

[modern juniper]

base, refiner and controlnet fits into my 12gb 3060 so that's good news

[elfin flare]

short toddler knight be lookin

[lilac raven]

4 GB seems to be not enough for ControlNet + SDXL for now

[hoary saddle]

can you stop posting these lovelies until AIT support for linux comes out? 😉

tease

[trim orbit]

https://tenor.com/view/sandlot-forever-gif-5743944

[indigo carbon]

it is patched, should work on linux in the same way me and Masslevel instructed them to set it up

[trim orbit]

weird... usually its the other way around and windows users are left waiting for whats supported on linux already

[elfin flare]

Temples.

[modern juniper]

wait does ait change the output in any way? i thought it just improves perf

[trim orbit]

every seed changes the output

[hoary saddle]

the vid @indigo carbon posted last night was incredible for the speed, aye

[trim orbit]

determinism is really just not that important outside of research purposes

[indigo carbon]

actually, good question. I think it makes results slightly sharper and slightly more detailed. not only it's close to tripling the speed, also enhances generations by an unnoticeable amount.

[modern juniper]

sick

[indigo carbon]

refiner is supported!

we did it

workflow with AIT on both base and refiner:

[strong field]

lol i just dug deep into chat to find said video, do you remember which it was?

[trim orbit]

like if this image settings and seed generate differently on another system, i dont really care. This is a good image.

people get so hung up on determinism but i don't think they're really following the thought through to a conclusion, if it matters or not

even then, in most cases where deteminism isn't kept, the changes are very minor

[midnight shuttle]

Determinism only matters when running side by side tests such as understanding the impact of magic words like "masterpiece". Otherwise why try to reproduce someone else's image. The image already exists. If you like it just download it.

[modern juniper]

#✨｜sdxl message

[indigo carbon]

AIT doesn't cause degradation, it's reproducible

[trim orbit]

even in that situation it matters very little. same seed on xformers is fine to get the understanding of how that token affects the generation. xformers is non deterministic, but only barely

[trim orbit]

thats good to know, but also, it doesn't matter

[urban breach]

Keep getting this error? Tried that SDXL branch also, any ideas? Error occurred when executing KSampler: ComfyUI/custom_nodes/AIT/AITemplate/modules/8ceef4c7e2f3884c7ef7e5834bc1bab9324d4b7f842c343b604839d0a68a4b74.so: file too short

[trim orbit]

at the end of the day those pixels are hot 🔥

[modern juniper]

let him cook

[indigo carbon]

Linux?

[urban breach]

Yes

[indigo carbon]

makes sense

wait a few hours, it wasn't entirely patched on linux

[buoyant axle]

seems to work good. getting ~12it/s with it

[urban breach]

ah, ok. the file is actually empty. thanks 😄

[trim orbit]

sure would be cool if someone uploaded this permissively licensed rail-m controlnet as a safetensor for us folks who don't want to create a venv and ducktape together a python script just for one conversion

maybe we could call it c.safetensors because c is for controlnet

[indigo carbon]

although, @urban breach you might be able to resolve this yourself. you can try going to the HF repo with the AIT SDXL modules, then replace them with the broken ones and rename them to what ever the broken ones were named.

[trim orbit]

oh oops. you're not doing a controlnet conversion

i've got tunnelvision

[urban breach]

I tried doing that with the file it was complaining about yesterday and it didn't seem to work, but maybe I need more files

[indigo carbon]

I see. than just wait like a few hours 😄

[boreal bough]

@trim orbit
not sure if you saw it, so I'll link it just in case
#🔧｜finetune message
got a new fun preset going XD

[strong field]

awesome thanks

[indigo carbon]

how many it/s were you getting previously, by how much did it boost for you?

[buoyant axle]

before i had about 6.5s/it. so nearly 100% increase

[indigo carbon]

6.5s/it compared to 12it/s is not a possible difference, do you mean 6.5it/s?

[trim orbit]

nice thanks! i've only got 16 😦 but i'll explore those settings.

i'm also going to train a lora for someone with the token "toyota" now. because that reminds me of an old handle i had "AtoyotasAtoyota" maybe i'll train me as a toyota token

[buoyant axle]

yeah 😆

[indigo carbon]

oh, nice

[trim orbit]

people kept assuming i was japanese and getting really racist towards me so i morphed

[eternal fog]

Sorry if I've missed it, where are these AIT modules/models?

[indigo carbon]

https://github.com/FizzleDorf/AIT/tree/sdxl, put this in custom nodes, load my workflow, boom chakalaka

[trim orbit]

lol i think i just figured out why i'm always lost in comfy ui when i load it default. sytan's workflow is offset above the framed page layout.

[boreal bough]

batch size to 1 should probably work - just didn't want to get peoples hopes up without testing

[indigo carbon]

also as of now it needs windows, if you have Linux wait a few hours

[trim orbit]

theres a joke to be found there about cops and thin blue lines and problems, but i can't suss it out.

[trim orbit]

i'll buy a 24 and test it

[eternal fog]

Cool thanks, and I use windows so it's fine.

[boreal bough]

also, no adamw + full bf16 -> instead adamw8bit might save a bit too.
if it's still over, then text cache will solve it for sure - as it takes away the vram from TE training (so its same requirements as no clip training)

[buoyant axle]

@indigo carbon So looking at your workflow its "just" to put the AITemplate Node between the Load Checkpoint Node and The Samplers (like with loras). I noticed though, that there is also a new AITemplate VAE Decode node, which u are not using. Are we supposed to also use it?

[trim orbit]

I'm curious to see if text encoder training helps significantly

[buoyant axle]

oh and VAEEncode as well

[boreal bough]

it does - but it shouldn't be used randomly as there are a lot of taboo words that instakill your training

[indigo carbon]

no, it's a broken node. also my workflow is set up in a way that benefits as much as possible from AIT, if there will be even more breakthroughs, I'll probably showcase it here.

[smoky patrol]

So, I have been w o r k i n g on this document. S o m e of the links are current...messed up, but they will be fixed: https://docs.google.com/document/d/1BxdWqfBJ3QPggHnBCBx3QIkUpHnBWShd_s3zEt2dTLM/edit?usp=drivesdk As I update it, I'll let ya'll know

Enjoy

[hoary saddle]

#✨｜sdxl message

This one @strong field

[indigo carbon]

I can make a new one, it's faster now

[boreal bough]

rejoice! for as the night fades away, sunny turns it back into day

[modern juniper]

i have to download modules manually right?

[indigo carbon]

no. if you get errors it's because you have an unsupported version of live preview

[trim orbit]

I like the energy in this document

[smoky patrol]

The dawn c o m e s!

Thanks!

[trim orbit]

MWAHAHAHAHA!
its got those dr frankenstein vibes

[modern juniper]

"C:\Users\josef\projects\ComfyUI_windows_portable_nvidia_cu118_or_cpu\ComfyUI_windows_portable\ComfyUI\custom_nodes\AIT\AITemplate\ait\load.py", line 76, in filter_modules
    raise ValueError(f"No modules found for {operating_system} {sd} {cuda} {batch_size} {resolution} {model_type}")
ValueError: No modules found for windows xl sm80 1 1024 unet_control

latest sdxl branch

[smoky patrol]

Well, yeah, that's me, so there you go, LMFAO

[trim orbit]

tyvm!

[indigo carbon]

weird. did you make sure you got the latest version of the SDXL branch?

[smoky patrol]

Np! I will be continuing with updates, so you'll know more SDXL things, and I hope you guys will all enjoy and explore these vectors!

[trim orbit]

OOOOH comfy has a node for loading diffusers controlnets. no safetensors needed

[modern juniper]

pretty sure

[indigo carbon]

you needed the zip file from that branch. also powershell is funky, I personally like to make a bat file to start Comfy

[trim orbit]

i knew that too cause i loaded the other depth model from sarge that way huduhh

[indigo carbon]

OH WAIT i see the issue 😄

you are using Cnet

no Cnet support yet for AIT

[modern juniper]

ooohhh right 😄

[indigo carbon]

I was talking about the AIT repo...

[trim orbit]

oh shit i got tunnelvission again

[soft zealot]

shoul;d have gone to specsavers

[trim orbit]

this has really been the reason i haven't bothered. controlnet will be here any day now

[indigo carbon]

weird also that you used Cnet, I thought you loaded the workflow I built specifically for AIT

[modern juniper]

i loaded this one https://github.com/FizzleDorf/AIT/blob/main/workflows/aitemplate_controlnet.json

[indigo carbon]

nope, that's for 1.5 models. the best current workflow for AIT is likely the one I made specifically for it

[vale eagle]

It should be fine to git clone the repo and switch to sdxl branch

[indigo carbon]

they already got it working (I think) they are trying to use Cnet on it, which wasn't compiled yet

[lilac raven]

After 25 minutes and one failure because of bad nodes, I can confirm the ControlNet works, but for now it's on CPU. I'll make a remote copy to the other computer

[sacred wharf]

during the release there was some talk of some slimmed down Cnet models

is that something that is still coming?

[strong field]

OMG

this is incredible

[indigo carbon]

ah, you got it set up?

[strong field]

yeah still experimenting with quality improvements, looks like a couple tweaks needed

but initial results are phenomenal

[azure oxide]

ya if you think this is good just wait til comfy optimizes it himself

[strong field]

good on you mate

2656 * 1904 is taking no time at all

[indigo carbon]

no need, I think we already did a good enough job to implement it efficiently. heck, I'm making 4k images in 10 seconds..

[azure oxide]

haha thats what we said about optimizations in general
then the optimizations between comfy and a1 were easily made plain

[modern juniper]

yeah it works, but for some reason with AIT for sampling, VAE decode with AIT causes OOM, 12gb is not enough?

[boreal bough]

lol. we can bug him about it live, today in 2 hours

[azure oxide]

i think it was joe who said that ppl suggested things and comfy was like "nah we can do better lmaoo"
the optimization improvements from a1 clearly show that and i think he can definitely do a better job in improving the implementation of AIT

not that this implementation is bad of course

[indigo carbon]

I already told you, it's a broken node

[lusty token]

I did not fully process your instructions earlier..."move your mouse around". I've only been holding space and trying to scroll left and right. I get it now.

[modern juniper]

sorry, i missed that

[indigo carbon]

use the arrangement of nodes and resolutions I provided in the workflow, they are damn close to max efficiency

[vale eagle]

AITemplate reduced 50% time of my workflow.

[upbeat summit]

it's an over 200% performance boost - on all (supported) resolutions

[strong field]

@indigo carbon Whats your 4k workflow, initial latent at 2048?

THIS IS INSANE

[vale eagle]

The bottleneck is vae. If we could use the AITemplateVaeDecode and AITemplateVaeencode, it could even faster

[boreal bough]

not sure of context - but just in case anyone forgot - there's still the fp16 fix vae from 0.9, which will run a lot faster on older cards

[indigo carbon]

the VAE parts of my workflow take a few seconds. we are already getting like 10 seconds per image, speed is what's missing?

[azure oxide]

ya i was just about to mention this

[strong field]

2k images in 5 - 8 seconds what on earth

[azure oxide]

what if we go
EVEN
FASTER

[vale eagle]

I was using my custom workflow built on Sytan workflow. It contains multiple vae encode and decode

[azure oxide]

Ludicrous Speed.

[strong field]

100 - 8k images every second isnt fast enough!

[indigo carbon]

that is inefficient in my opinion. the workflow I built is designed specifically for AIT. if you use AIT that's likely your best bet

[azure oxide]

nothing wrong with implementing AIT into his workflow tho

sounds like he did it as best as he could

also this graph from a while ago looks insane

for a100s

12.8x increases for batchsize 1 training lmaooo

[strong field]

woof

[azure oxide]

jesus they wrote this almost a year ago, i wonder if theres been improvements to it since then https://ai.meta.com/blog/gpu-inference-engine-nvidia-amd-open-source/?utm_source=twitter&utm_medium=organic_social&utm_campaign=blog

[eternal fog]

Hmm @indigo carbon I get this error when using your workflow

Error occurred when executing KSampler:

Input type (torch.FloatTensor) and weight type (torch.cuda.FloatTensor) should be the same or input should be a MKLDNN tensor and weight is a dense tensor

[vale eagle]

It is because you enabled tae preview.

like me

[eternal fog]

You can't use preview?

[indigo carbon]

@vale eagle had the same problem, they fixed it

[eternal fog]

Fixed it when, because I literally just downloaded it

Last update was an hour ago

[vale eagle]

remove files in ComfyUI\models\vae_approx

[eternal fog]

I can't use the preview with this?

[indigo carbon]

you can if you use the preview of the ComfyUI_manager, works FLAWLESSLY

[vale eagle]

You could preview but not that one

[eternal fog]

What preview is this?

[azure oxide]

i remember reading this, there could be a reason for it lol

different context at the time but could be the same reason

[indigo carbon]

I honestly don't know, but it works, and it doesn't effect speed at all

[eternal fog]

ok do not accidentally set Keep Loaded to enable

Seems to have a memory leak issue where it keeps loading it over and over

[indigo carbon]

yes, there is a reason I built my workflow the way I did

[eternal fog]

What's that VAE Decode AITemplate node do

Seems to make it slower for me

[strong field]

should i try a batch of (8 )4k images?

hope i dont burn something

[indigo carbon]

I'm tired of explaining this over and over again.. can anyone else explain?

[eternal fog]

I've not seen it written anywhere and I don't see any documentation unfortunately.

[azure oxide]

your gfx card is designed to do this, youd probly only hurt it if you tampered with it physically or kept using it past the temperatures that its rated for

[strong field]

lol batch is broken

hahaha

[vale eagle]

I think AITemplateVAE is not supported yet for SDXL

[indigo carbon]

it's a broken node, just use the workflow I made in the arrangement I provided it

if there will be more breakthroughs, you can trust I'll update my workflow and share it

[eternal fog]

Yeah I've got that, was just seeing what the other stuff did

[strong field]

without a doubt, the last commit was an hour ago

[modern juniper]

AIT custom nodes break controlnet support even without using AIT nodes anywhere, is it known bug?

[crisp owl]

waits for 2k gpu support

[eternal fog]

14.28 seconds for regular, 9 seconds with AITemplate. Not bad.

[indigo carbon]

I rarely use Cnet, so idk

[strong field]

cnet for sdxl is giving me that trick or treat feeling

[indigo carbon]

the boost should be even more powerful than that

[eternal fog]

That's what I'm getting on a 3080

That's just 30 steps on base

[indigo carbon]

oh, 3080- that makes sense. the newer gen your card's CUDA cores are, the bigger the boost you get from AIT.

[modern juniper]

100%ish improvement on 3060

[indigo carbon]

makes sense =]

[eternal fog]

Interesting, so these modules work for specific image sizes?

[ionic dragon]

https://www.reddit.com/r/StableDiffusion/comments/15nc4fe/sdxl_controlent_is_here/

for real?

[indigo carbon]

yes, what's so special about AIT is that it handles that on its own, and it's not checkpoint specific- it's architecture specific.

[eternal fog]

So if I want a different image size I need to make my own module then

[indigo carbon]

no, it will do that itself

[ionic dragon]

only canny is out?

[eternal fog]

Doesn't seem to

No modules found for windows xl sm80 1 2688 unet

[upbeat summit]

2048x2048 is the current max for a native image - at least it was earlier

[indigo carbon]

this makes sense, so yeah, if you do absurd resolutions like that when you can just do half of that then upscale it x2, so yes- you need to compile yourself.

[eternal fog]

Can't use it with my upscale pipeline then 😦

[upbeat summit]

you can still scale it up after that

[eternal fog]

Yeah but I run a pass over it with the base, which I can't do. I'll have a look at making my own module.

[indigo carbon]

yeah, my workflow already does that on its own =]

[eternal fog]

Yeah I know how to do all that, I just had it setup so it would do a base pass to get rid of certain things

[lilac raven]

The copy of the ControlNet has finished, and having both LoRA and ControlNet seems to work

[sharp jolt]

has anyone run into issues with ComfyUI and Preview Image panels not working? i've spent the last two days doing different installs on Ubuntu in AWS EC2 g4dn.xlarge servers and everything works fine except the Preview Image panel and i can't figure out what to do

also can't find anything on the internet about it

[upbeat summit]

you mean the preview on the samplers?

[strong field]

the speed of generating HQ concepts is uncanny

[sharp jolt]

any preview, ive tried to get samplers working with manager also but that doesn't work either

[gloomy barn]

hq?

[indigo carbon]

high quality i think

[upbeat summit]

have you launched it with python main.py --preview-method auto?

[sharp jolt]

just doesn't show anything

[sharp jolt]

yeah i have tried that

[upbeat summit]

that is only for the sampler previews

[sharp jolt]

the image still saves

[upbeat summit]

the preview image should show up after you have decoded it with the VAE node

[sharp jolt]

yeah i have it tied into the vae decode node, but only save works, not preview

[ionic dragon]

@soft zealot for random prompt, it doesnt log the prompt it fetches?

[crisp owl]

Do you have Manager?

[upbeat summit]

hmm I run it on Windows. the save image node can also show a preview. maybe it's something related to the gpu driver on linux.

[sharp jolt]

yeah i have manager installed and have it set to auto here

[eternal fog]

huh, the speed is exactly the same when I run my full pipeline with the refiner.

It seems like the speed it takes to load and unload the refiner kills any extra speed boost.

[crisp owl]

Do you see the preview at the bottom of your ksampler?

[vale eagle]

Did you implement the AItemplate to refiner?

[sharp jolt]

no, basically no previews work

[crisp owl]

Try restarting your instance if you haven't

[sharp jolt]

i've connected to my ui from my mac on Chrome and Safari and from my desktop on Chrome and preview doesn't work on any browser

i've actually fully uninstalled and reinstalled probably 10 times now with slightly different config

[eternal fog]

Yeah the timings seem a bit variable though, sometimes it's about the same other times it's a few seconds faster

[sharp jolt]

deploying a new instance right now

yeah, totally lost honestly

[vale eagle]

I don't. It worked for me to speed up 2 pass latent with refiner.

[pallid path]

I can't wait to try controlnet

[eternal fog]

I think it might be my 16GB RAM causing issues

[pallid path]

^

it starts to pool into your Swap (or 'pagefile' on windows)

when I use both models I hover around 16.5 or maybe even 17GB RAM usage

[honest marsh]

what is that thing manager

[worthy orbit]

I updated Automatic1111 but I cannot load SD XL model, 0.9, followed some guides

[eternal fog]

Also seems like img2img with the AITemplate doesn't work as well

[vale eagle]

The peak is 17.2 vram in my 2 pass workflow

[crisp owl]

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

[honest marsh]

custom nodes?

[urban fjord]

Does anyone have a benchmark with the AITemplate improvements? Is it really significant or just 10% on a 3060?

[worthy orbit]

Does ComfyUI runs on AMD, RX 6700 XT?

[strong field]

linux?

[worthy orbit]

Windows

But I would love to install in linux

[strong field]

Windows = directML

[worthy orbit]

I have ARch

[eternal fog]

ok yeah it doesn't work very well with img2img it seems

[strong field]

bummer

[vale eagle]

I have 100% speed up in very high steps 2 pass latent workflow using 3090.

[worthy orbit]

Arch Linux

[strong field]

here is a bench on A100

i am seeing 100% improvement on 3090 at least

[worthy orbit]

what is that? care to explain, A100?

[strong field]

a100 is massive 80gb gpu

[urban fjord]

Do I need to do something to activate it in ComfyUI?

[strong field]

#✨｜sdxl message#✨｜sdxl message

[eternal fog]

Hmm actually something odd is happening, sometimes it's fine, othertimes it's not.

[strong field]

i noticed that when trying to gen people

still trying to figure it out, not sure if its prompting or otherwise

[eternal fog]

It was adding this horrible fuzziness to the bottom part of images, but then it suddenly stopped

[strong field]

Yes, the fuzziness was a prob for me on a few gens

[pallid path]

I hope someone prunes the controlnet model for sdxl, I aint downloading a 5 GB file 💀

[vale eagle]

Sometime, comfyui moved the model to ram? and the error prompt said it should in same device. It might due to my workflow

[eternal fog]

This is why SAI didn't release controlnet straight away and are working on something different.

Otherwise every single controlnet model will be 5GB

[pallid path]

why something different, is there something wrong with pruning

[worthy orbit]

It seems Automatic1111 still running on an old Python version, I need to uninstall and install Python again, delete the automatic1111 venv

[patent badger]

Sorry for being the dummy to ask, but what does AIT mean?

[pallid path]

like the original controlnet model here for example takes up 9.6GB, but the pruned version can go down as low as 728 MB!

[strong field]

not going to pretend like in know: https://ai.meta.com/blog/gpu-inference-engine-nvidia-amd-open-source/

[pallid path]

unless they can make a new solution that takes up even less space then idk what's holding them back

[strong field]

the use a neural network to transforms something smart and im lost

[eternal fog]

Yeah it seems more likely to do it on larger images. Left is with AITemplate, Right is without

[strong field]

that was the problem i was having

couldnt figure it out

[soft zealot]

it does if you have the txt output working correctly, if you've fiddled then unfiddle

[strong field]

what upscaler are you using

plan to try an iterative upscale when i get the time to sit down with it and see where it breaks

[eternal fog]

It's a pixel upscaler, so it shouldn't make any difference.

[vale eagle]

I have no issue with this in multiple gen

[strong field]

was goign to try latent upscaling

[eternal fog]

I'm trying a different resolution, but 9:16 it does it every time when I pass an image through IMG2IMG

[strong field]

havent gotten aroudn to it

[eternal fog]

That's always been complete shit for me on SDXL

[vale eagle]

let me try 9:16

[strong field]

yeah but nah

back to sd1.5 days

[worthy orbit]

did that, get this message

This program is tested with 3.10.6 Python, but you have 3.11.4.
If you encounter an error with "RuntimeError: Couldn't install torch." message,
or any other error regarding unsuccessful package (library) installation,
please downgrade (or upgrade) to the latest version of 3.10 Python
and delete current Python and "venv" folder in WebUI's directory.```

but it is installing

[strong field]

gotta use 3.10.6 as far as i know

[worthy orbit]

That would run SD XL models?

[eternal fog]

Tried a 1:1 image AITemplate still does weird stuff. What's interesting though it that the top left corner of the image is fine, but around that gets wobbly

[worthy orbit]

That's what I change it, but it didn't work

[eternal fog]

Left is normal right is AI Template, look how everything has gone all blurred to the right

[ionic dragon]

i had the manual text prompts in the text boxes, so they might have interfered, gotta test by removing them

[strong field]

gotta figure out this people thing asap

the squirrel is a nice touch

[urban fjord]

_lzma.LZMAError: Input format not supported by decoder
Joy... (From running the custome module)

[strong field]

you download the zip of the SDXL branch?

Now that is a happy accident

v

[urban fjord]

AITemplate is a lot faster yes, now I just hope Auto1111 implements it too

[strong field]

glad you got it working

[indigo carbon]

yeah, it's so fast I'm generating images like this under 10 seconds

[trim orbit]

zoooooooom

[indigo carbon]

oh, shit, there was just an earthquake

[strong field]

trying a latent upscale workflow instead of pixel

[crisp owl]

Well when you got speeds like that, what do you expect?

[trim orbit]

i've drilled so hard for this. i want to believe i'd run for cover and stand in a structural supported area or hoddle under a desk. But in reality i'd probably post about it online while it was happening

[crisp owl]

I usually just go, "oh an earthquake".

Then be not surprised why my dogs were acting weird before it

[trim orbit]

all my life we hear about "the big one!" and how geologic records show there is a megathrust event that breaks the scale, and that we're way over due for one. so while they do always happen here smaller ones, they always put me on edge and got me considering "is this one it?.... nah i'm good"

[indigo carbon]

jesus, this was crazy

[crisp owl]

If the "big" one ever happens, ain't nothing much we can do. I'm pretty close to Yellowstone. If that goes, it'll just be a lightshow until poof

[trim orbit]

if yellow stone pops we're all done

you like living in alaska?

[crisp owl]

UT

[trim orbit]

cause everything will be alaska

[crisp owl]

haha

[trim orbit]

impending doom is a fun topic. i got on it the other day acidentally too.

[hot rover]

I live along the San Andreas fault line…every earthquake is an “oh shit” moment ha

[indigo carbon]

yeah, there is a giant volcano there waiting to wake up

[trim orbit]

yeh i'm just a little north on the cascadia subduction zone

[crisp owl]

I used to be there

[upbeat summit]

everything alright?

[crisp owl]

SAF

[trim orbit]

its active it's a wake

she's just hanging out tho

[indigo carbon]

yes, It was just shaking for a few seconds

[upbeat summit]

if I would not be used to it, that would make me feel very uneasy

[indigo carbon]

not really shaking, like, you know when on a roller coaster goes down? similar sensation

[upbeat summit]

low end rumble

[trim orbit]

that sub sound that can't be heard but plates of the earth rubbing on each other can create

[crisp owl]

The rolling ones are pretty wild to see if you're outdoors and can see it coming through

[trim orbit]

your intestines act like an antenna an vibrate with it

[strong field]

science you crazy

[candid walrus]

@indigo carbon thanks so much for posting on the AIT stuff, just cut things in half, what a win

[strong field]

i cant get people i was getting grotesque level with an iterative pixel upscaler

other styles, 2d, vector art, fantasy do SO WELL

[indigo carbon]

I don't have any issues with generating people.. are you using my workflow?

[strong field]

yes, changed widths and heights

[indigo carbon]

ah, makes sense

[strong field]

ill go back to yours to see if i broke something along the way

[vale eagle]

Use 576:1024 for 9:16 and do the upscale

[indigo carbon]

I think you can go a little higher. I just have a rule of thumb that the resolution value should stay the same in total, only changing ratio

[hoary saddle]

i'm at the exact place this guy is, will wait 🙂
https://github.com/FizzleDorf/AIT/issues/12#issuecomment-1672155839

[vale eagle]

I tried 768:1344 but out of context

[strong field]

oh snap, i didnt think about that

[trim orbit]

i am for megapixel on the first pass. if i do a 2x pass setup then i'll sometimes go around 700k pixels

high denoise on the second pass in that case though.

[boreal bough]

trained a lora on 50 variations, of the exact same image 🤣

[upbeat summit]

and how versatile is it ?

[trim orbit]

that doesn't look like wallstreet

unless i'm very wrong about how ws works

[boreal bough]

eivor, a girl standing in starbucks, getting a coffee | eivor, a girl in the middle of a crowded metallica heavy metal festival

[strong field]

hahaha

[crisp owl]

I mean, it keeps them sane I suppose

[boreal bough]

eivor, a girl standing at wall street

[trim orbit]

ahhhh thats more accurate.

[trim orbit]

in a street anyways. Cool results!

[boreal bough]

I mean. I'm not sure what I was expecting XD

[trim orbit]

https://tenor.com/view/hulk-the-hulk-bruce-banner-avengers-endgame-gif-14705590

[slender coral]

Really not wanting to post this question since I don't want people thinking I'm into weird fetish stuff. But I'm trying to get a video game Icon for an item of poop.
Not able to get anything, are things of that nature not trained on sdxl?
poop shaped like food. ui asset Trending on artstation, award winning.

[unreal plover]

voxel syle, lora:VoxelXL_v1:0.9, None a couple is playing game console on a sofa, miniature, tilt-shift, film, bokeh, professional, 4k, highly detailed

[trim orbit]

ewww poop !? you're gwoss!! ||poop is a palindrome||

[fresh path]

i dont think poop would be in.. the training data

[trim orbit]

awwww my fun poop fact was blocked

[pallid path]

as disgusting that is, controlnet would help with that maybe, having an outline of a pizza for example and then just typing poop, but really, that's probably censored out of the dataset 😬... try maybe 1.5 or 1.4

[boreal bough]

one variation of the original image - so I wasn't exactly working with prime material here 🤣

[trim orbit]

dangerous fun facts

[hasty mortar]

did you solve it?

[slender coral]

😄 thanks for the suggestion, I think at this point I'm just going to hire an artitst for some things 😄

[hasty mortar]

I'm good with python debugging since i wrote a blender addon, let me have a look

[pallid path]

poor artist

lmao

[hoary saddle]

negative

[hoary saddle]

stuck at the same place too?

[trim orbit]

wonder how many people are saying "poop" outloud rn

[hasty mortar]

didn't even start, will let you know if i solve that

[boreal bough]

eivor, a woman posing for a picture, model photoshoot, studio

[clever verge]

What other parameters did you train it with?

[boreal bough]

#🔧｜finetune message
plug and play of my new faces preset

although, I did add keep n tokens = 1
and made sure my trigger word was the first word on all captions of my images

[clever verge]

Thank you!

[halcyon tusk]

yes, it was auto-flagged and blocked, try to stay away from too explicit language, keep it civil people!

[lilac raven]

I think one of the LoRAs I trained has "learned" something I did not try to train at all, every time I ask for a car, it's an old car
It's because of the clothes that the "default" car is old?

[crisp owl]

sdxl tends towards older cars I've noticed

[boreal bough]

did you train clip?

cause that has a tendency to shift a lot around

[trim orbit]

you could probably convince people these were controlnets they're so consistent

[boreal bough]

when i try to have it make real people, it just does the exact pose XD

so yeah. I'm gonna call this the poor mans high effort controlnet

[unreal plover]

When you dont zoom it look to real

[trim orbit]

it twas minefield territory. will keep it better

[strong field]

not bad

[trim orbit]

sdxl loras really catch poses from the source so well i've noticed

[strong field]

oops i pasted the prompt in twice on that lol

[lilac raven]

I set no option to disable it, so I guess the answer is yes

[trim orbit]

this new canny model from diffusers team is great stuff

[boreal bough]

--network_train_unet_only

[trim orbit]

a lot slower so i can understand why stability is opting to explore lighter weight options

[lilac raven]

This is the thing that was not set

[trim orbit]

sounds lovecraftian when put that way

[lusty token]

On Civi , does the Workflow: 18 Nodes, relate to ComfyUI?

[autumn forum]

Yes

Would be my guess.

[tribal jackal]

My 4080 is dead

Looks like I'm just watching everyone have while I get an RMA issued

[indigo carbon]

this is so sad

[tribal jackal]

Tried reseating all cable connections, different cables, different monitors. Couldn't even get it to detect. Fans weren't spinning. Power light was on at the cable connection site. Replaced with my old 2080FE and here I am.

[hoary saddle]

wow, this controlnet works great!

[strong field]

I TOOK YOUR IDEAS

[tribal jackal]

don't tell me controlnet came out for sdxl the same day my gpu died lol

please

[strong field]

death and stuff

lol that sounds like its bad, its a big skull

[strong field]

you sir are an inspration

[trim orbit]

its un canny.... 😮 see what i did there? cause it was the canny model that came out

[strong field]

ooooo

[indigo carbon]

Thank you, glad you guys enjoy the workflow I built.

[barren olive]

does anyone know if there is a way to get StableSwarmUI to work with an AMD GPU?

Or just any way to run sdxl 1.0 with AMD GPU

[trim orbit]

install arch

[strong field]

prob best way is via linux

i tried and failed many times on windows, wsl, ubuntu, rocm, HIP

[trim orbit]

theres directML but its slow and buggy

[indigo carbon]

Wait, how are you using AIT then?

[strong field]

LOL

i pulled out my 6950XT and bought a 3090 FE

[indigo carbon]

Ah, I see.

I was confused because I know as of now, AIT doesn't like AMD

[strong field]

most AI doesnt like AMD

pytorch especially

[indigo carbon]

It's the other way around.. in terms of hardware - AMD are not that far behind. But man, their software is a nightmare

[barren olive]

thanks for the answer

[strong field]

direct ml might work

it was decent, but there is no comparison to xformers and now AIT

[alpine pine]

has comfyui added AIT or is it a manual process?

[indigo carbon]

Yeah, AIT is now sooo fast I spend most of my time coming out with new prompts. The generation time is as fast as my brain can imagine stuff

[gloomy barn]

I dont think it came out

[trim orbit]

canny is out

[tribal magnet]

Oh my.

I'm running batches of simple prompts with no processing and getting such good results. This one popped up in the middle of basically (very innocent) "elf lady reading a book in by a tree"* and it is so... suggestive. In the best way She's definitely flirting. XD I'm so impressed.

*not the actual prompt

[gloomy barn]

not for sdnext

[strong field]

exactly this, im spending way more time coming up with better prompts

[trim orbit]

that's an sdnext implementation problem.

the model is out

[gloomy barn]

no it's not

[trim orbit]

the weights are published

[gloomy barn]

It's a it will soon be available for SDNext

[trim orbit]

i'm using it in comfy right now

[wet glacier]

is AIT it just for Linux for now or also windows?

[gloomy barn]

right, and will soon be implemented in SDNext, Nowolf

[trim orbit]

doing a 2x pass with it too

[gloomy barn]

let's see a result

[azure oxide]

i dont think comfy added his own implementation of it yet, he mentioned it would be something that you just turn on simply as opposed to using a custom node

[trim orbit]

le dafoe

[indigo carbon]

Just windows, no Linux yet

[trim orbit]

geeze. again. last time controlnet came out you kept badgering me for more hands everyday insisting it didn't ever work

it works

[gloomy barn]

You said it works, so I asked to see a result/example

[wet glacier]

keeps crashing at the Ksampler node (windows)

[hasty mortar]

no here

[gloomy barn]

why is that unreasonable? I am curious to see how it works

[trim orbit]

it's just silly. the model came out and there's people here showing it off all day. and you're like "NO! It's NOT out"

[gloomy barn]

hands are still borked

[trim orbit]

all this can't and speaking with such "i'm right" about it really make it a problem for those of us trying to spread news that isn't so fake

[indigo carbon]

Preview is on, it's a know issue. You need the preview feature of the ComfyUI manager

[gloomy barn]

I didn't claim it's not out. I correctly stated it's not yet implemented in SDNext

then I asked to see a sample so I can see it in action

[trim orbit]

i said its out and you started baiting arguments. come on now

[gloomy barn]

nope

I asked to see it in action

[hoary saddle]

sooo in love with this controlnet, results are incredible

[gloomy barn]

it's so hard to get info here when peole like Flowolf attack every time someone asks a question

[trim orbit]

helps to try to keep up to date a little bit

[azure oxide]

you dont need to keep diggin the hole, we can move past this in the channel lol

[gloomy barn]

Ill ask somewhere else

[hoary saddle]

setup if someone wants it

[wet glacier]

you mean this?

[gloomy barn]

There is no hole.

[trim orbit]

i've helped you for hours and it's always an argument about can't. stop mischaracterizing things

[gloomy barn]

You've never helped me one iota

[azure oxide]

damn flowwolf, youve blocked people for less

[trim orbit]

he spreads misinformation OFTEN

i have to counter

[indigo carbon]

Yeah

[gloomy barn]

please stop harassing me. Block me so you don't see my comments. Let me chat in peace.

[azure oxide]

the man with the cape that we dont deserve

🙇‍♂️

[wet glacier]

I've going to have to read this whole thread, so I can understand what I am doing wrong

[trim orbit]

still insists that hands absolutleyl can't be done right. when we have 100s of examples even before controlnet came out. he went on quite a crusade about how bad hands are and made sure to answer everyone looking for help to let them know how impossible they are

[hoary saddle]

for what it's worth, i didnt convert the pickle to safetensors, it just worked as is

[gloomy barn]

Nope, never said that

[trim orbit]

as a tutor, he drags people back

[gloomy barn]

I correctly state that hands are often bad in generations

[azure oxide]

the preview image node is incapatible with the workflow that tdg8 gave us is really all there is to it, he suggests using the comfyui manager to get around it

[trim orbit]

often jumps at the chance to help those who are looking for help, but douses them with just bad misinformation and a lot of "you can't do this" mentality

[gloomy barn]

nope...I help people with good info, constantly

[trim orbit]

it's a toxicity that immunity must be taught towards

[gloomy barn]

there is no toxicity.

[trim orbit]

the key is for people to become aware of the "can't" nature. awareness of it is 90% of the immunity

[gloomy barn]

there is not a way to generate good hands every time.

[indigo carbon]

I get like bad hands every 20 or so gens..

[gloomy barn]

people must be aware of this attack people who ask questions nature. It must be combated and stopped

[hoary saddle]

i started puttin gloves on every prompt lol, helps some

[daring smelt]

no reason to argue and air this out here-- maybe block one another if you cant see one another's messages without spinning arguments or taking it to DMs if you really would like to. Let's be productive and kind and move about our days.
@gloomy barn @trim orbit

[gloomy barn]

WHAT

I get 9 out of 10 bad hands

[strong field]

depends on workflow ALOT

[indigo carbon]

I am magic =]

[hoary saddle]

https://tenor.com/view/shin-jimin-jimin-whip-whiplash-whipping-gif-12463359

[gloomy barn]

I try, but every time I try to ask a question or discuss, flo buries my comments in onstant attacks, constant criticism

[daring smelt]

block

[azure oxide]

hey fruit while youre here, completely unrelated to the drama above, is there a new method of reporting messages to the mods? the apps selection menu hasnt been working for a while

[trim orbit]

seriously. People often have way more success than he's alluding, but he pretends it doesn't happen because of his personal experience. I tried to help. so honestly tried to help him. at this point i feel like it's a bit he's committed to.

/final thoughts

[gloomy barn]

It doesn't matter if I block. My comment/question gets scrolled off the screen

[daring smelt]

have you tried restarting discord? It's usually on discord's end, and pops back up once fully restarted

[azure oxide]

does restarting my computer count

here's what i see

[daring smelt]

restart discord client itself

[gloomy barn]

in this case, I mentioned that controlnet is not yet available - only one model is for one ui. Then when flo said he was generating, I asked to see one to see how it looks. I didn't keep claiming it's not available.

[azure oxide]

same thing unfortunately

just did it

its not working for anyone else is it?

[daring smelt]

Okay, you two can come back once you're past it and not too emotionally invested

also, @azure oxide

we've been getting reports today!

[azure oxide]

interesting, so im the outlier 😭

[daring smelt]

so i'm not sure.. most times it's fixed when restarting the discord client

in the bot channels it doesnt appear?

[azure oxide]

let me check

ohh i see. was it simply for the bot channels? it works for me there

I must've incorrectly assumed that it was a feature for all channels

[daring smelt]

some channels it can be a bit wonky-- but should always work in bot chans!

[azure oxide]

gotcha 👍👍

[lilac raven]

Is --medvram necessary for 12 GB on A1111?

[slender hinge]

Does anyone have any good information or resources on instance & class tokens for training SDXL? I have seen things such as "ohwx" recomended because its unique & 1 token? but also seen people say with SDXL just use the name of the person or something else? really trying to understand better.

[vale eagle]

#🔧｜finetune message

You could find more information in finetune channel

[slender hinge]

thanks i will ask there

[urban breach]

Elden Ring returns

[vale eagle]

Very appreciate your work on prompt engineering. But most links of the document direct to discord attachments which I couldn't open.

[smoky patrol]

As I said, I'll be fixing those--they're just broken for the second, but they all actually work. It's just because I copied some links wrong but I'll be adjusting them.

[vale eagle]

Thanks for reply. That would be amazing

[smoky patrol]

It won't take me too long--I just gotta go through and find which ones aren't the correctly copied images. (I have all of them.)

[vale eagle]

All links direct to discord attechments is failed to load. I could only open DS examples

Oh wait. The ds example also is dicord attechments

There are some broken before page 13.

I able to open after that

[smoky patrol]

Strange. There are many that work before that, so I can't tell you why they don't work for you.

But you can see what links I'm working on, but I'll remove the warning when I'm done 🙂

[vale eagle]

@smoky patrol Roughly checked. Page 6, 7, 10(v4) 11, 12, 15, 18, 19, 20, 22, 23, 24, 26, 28, 30, 31(Marker lineart: 2), 33. This pages have broken links.

[smoky patrol]

Okies! Yeah, it's cuz I accidentally had copied the bot links instead of the DM'd ones, hahahaha.

Ty!

(I'm on page 6 rn; pages before that should, in theory, work)

[vale eagle]

Your document is a great work. I could learn a lot from it.

[smoky patrol]

Thanks! There's more to come, hahaha.

Always more.

[sinful falcon]

Hey all, I switched from A1111 to Comfy last night for SDXL and it's just amazing. I've searched here and Google and I can't seem to find if there's a way to interrupt Comfy during sampling. If so, it seems like maybe it's not a great idea? I'm doing 12 steps at 1024x1024, then sending to latent upscale by 2x and then sample for 12 more steps. If I see in the Image Preview after the first sampling session that it's way off base, I'd like to stop it if possible. Do I need to just find my zen and wait it out?

[arctic bloom]

there should be menu on right with button "view queue.

After pressing it option to cancle generation would be available

[sinful falcon]

Sheesh. 👴 Thanks!

[vestal goblet]

I am new to all of this, just installed Comfy with SDXL 1.0 but I have no idea what workflow to use. The SDXL Examples (bottle) doesn't give me very good photo realistic images. Any pointers?

[urban breach]

yeah see the images people are posting, if it's a png in many cases you can drag and drop right into Comfy

like that hazmat image

except you'll have to install a bunch of addons from the manager menu, not sure if you have that yet

[nocturne dove]

Yes, this. Keep in mind that some people have weird workflows that are not immediately runnable, like Flagg implied.

[vestal goblet]

I have manager yes. Thanks, I will try that!

[nocturne dove]

That made it sound like I was saying that you (Flagg) implied that they were immediately runnable. I meant the opposite.

[urban breach]

yeah for the most part, if you update comfy and fetch updates with that manager, then install the "missing modules". a lot of the workflows will work. what's the popular one called?

Zartan or something

here it is, this one's really popular: https://github.com/SytanSD/Sytan-SDXL-ComfyUI

[hardy cipher]

if you set things up right you'll be making photorealistic images like this in no time

[steel mist]

Not sure if this is the right place or if it’s tech support, but will I know (via console output or other warnings) if I’m running into memory pressure issues, or will things break “silently?”(ie, slowdowns or lower quality output). The TLDR here is I’m trying to decide if I want to seriously consider replacing my GPU or building a new system but told myself I’m not allowed to unless I run into specific issues

[unreal plover]

First AI generated images was like that, i remember, deep dream

[crisp owl]

If you run out of vram space to process the image in one go, it will automatically switch to a tile vae process

[unreal plover]

is it a real image ?

[glad grove]

real

[unreal plover]

nop

[steel mist]

What would that look like in terms of console output?

[glad grove]

real

[unreal plover]

nop

[glad grove]

real recognizes real

[crisp owl]

It'll say something along the lines of, "unable to process, trying tiled vae"

[steel mist]

Awesome Ty Ty

[crisp owl]

I know I hit it on my final upscaled decode, so I just put a tile vae in place instead of waiting for the console to recognize I need it.

gives me a few seconds back

[crisp owl]

like right now it's using it haha

When you can't help yourself and do an img2img of a guy on a zoom call....

[glad grove]

we truly live in a society

[unreal plover]

Heun sampler give interesting realistic textures

[glad grove]

now i get why ppl said the new version of the little mermaid was bad

[crisp owl]

Truly terrifying

[unreal plover]

Loch Ness

[crisp owl]

I KNEW IT

Real af

[unreal plover]

Look fake af

Even AI know it is fake

[glad grove]

real

[urban breach]

cloudy with a chance of nuclear fallout

[modern juniper]

AIT VAE decode now doesn't crash but produces black image

[slender coral]

Lads, is there an easy way in compy to save a filename as uuid?

Or would I have to code that?

[unreal plover]

uuid ? What is that ?

[slender coral]

Basically random string

[unreal plover]

dont know,

[gloomy barn]

that's opposite

[gloomy barn]

The eyes are odd.

[unreal plover]

perfect hand, it is a real image

Time to sleep, GN

after launch

[hoary saddle]

Use the comfy websocket example on GitHub and use python to name the output with uuid string

[slender coral]

So you're just suggesting:

def get_images(ws, prompt):
    prompt_id = queue_prompt(prompt)['prompt_id']
    output_images = {}
    while True:
        out = ws.recv()
        if isinstance(out, str):
            message = json.loads(out)
            if message['type'] == 'executing':
                data = message['data']
                if data['node'] is None and data['prompt_id'] == prompt_id:
                    break #Execution is done
        else:
            continue #previews are binary data

    history = get_history(prompt_id)[prompt_id]
    for o in history['outputs']:
        for node_id in history['outputs']:
            node_output = history['outputs'][node_id]
            if 'images' in node_output:
                images_output = []
                for image in node_output['images']:
                    image_data = get_image(image['filename'], image['subfolder'], image['type'])
                    images_output.append(image_data)
            output_images[node_id] = images_output

    return output_images

``` and save the image myself? 

I'll do it but is there no quicker way?

I'd have to write a propter as well, you can't subscribe to the portal if I query my own prompts.

[crisp owl]

I know this is their save file manual, that's about it on that front
https://blenderneko.github.io/ComfyUI-docs/Interface/SaveFileFormatting/

[hoary saddle]

That's how I'm doing it for my web app scripts that use the comfy API. If you're have some python knowledge you could probably write a custom node yourself or wait and I'm sure someone else will.

[slender coral]

Yeah I can write it but I don't fully know the system, wouldn't I just be able to update the save? Or does everyting have to be nodes?

Sending requests to the server doesn't give me the portal, so saving a file that I prompt with the ws that way woud defeat using the ui.

filename = new UUID4() that's the code 😄

[soft zealot]

yes and according to WAS this is a bug !!!!

Cpomys OOTB S& R only works with ComfyUIs image save (as far as I can tell) and WAS's Tokens only work with WAS nodes

Its frustrating

https://github.com/comfyanonymous/ComfyUI/issues/1057

https://github.com/WASasquatch/was-node-suite-comfyui/discussions/129

[crisp owl]

Yeah that is a frustrating thing, ran into that as well

[soft zealot]

@visual glade soz to put you on the spot but any comment onthe above please?

[upbeat summit]

hey I've seen that before heh

exposing S&R to all nodes would be so helpful for data processing

[hoary saddle]

Could probably just add a new variable for it in the existing save node, but everytime you git pull it'll probably get overwritten since it's a default comfy node. Could duplicate it and call it something else

[smoky patrol]

Links should all be fixed now 👍

[soft zealot]

my initial thought was a parser for WAS Modes that could read the S&T and pass them into a WAS Token as thats already quite a wide ecosphere howere WASasquatch is clainming th ehwole S&R thing is a but so he wont hence tagging @visual glade

This is the fiurtsrating part of opensource and github stuff.

Its both a blessing and a bloody curse

[strange mist]

thank you for sharing that. I'll have a proper go at it tomorrow 👍

[hoary saddle]

Gotta read more to understand the issue. Hope it's not too bad

[soft zealot]

All the info is in the two discussions I linked too. That’s all I’m aware of.

[soft bone]

why don't my loras work in auto? it spits out thousands of lines in console and then ignores the lora and generates without it

this is my first time trying auto since xl0.9

[gloomy barn]

with sdxl 1.0?

[trim orbit]

i looked quick at uuid in python. import uuid
str(uuid)

pop that into an example node an you're golden. fire that into save prefix and tada.

[smoky patrol]

Np!

If you have suggestions/things/styles you would be interested on hearing on, let me know

[heady vale]

started playing with AIT. It seems to love Euler sampler, but DPM is half speed or less of Euler

[upbeat summit]

yeah, I've not done any detailed runs but I need to test different samplers with AIT for sure

[heady vale]

its still fsater than no AIT. its amazingly good

[upbeat summit]

from my first tests I get a 100%+ speed improvement with dpmpp_sde_gpu

[heady vale]

2m looks like a winner

[upbeat summit]

it's fast - that's for sure

[heady vale]

dpm gpu isnt as fast for some reason

[azure oxide]

Sde is essentially double the steps

That could explain it

[upbeat summit]

yeah that's normal

[soft zealot]

Is that in relation to S&R or something else?

[upbeat summit]

@soft zealot send you a msg

[trim orbit]

no i dont think so? earlier was looking for a uuid in comfy. useful tool but i guess it's not a node yet.

[hardy cipher]

I can't figure out if I like all these prompt prettifiers and things

or if they just make wacky stuff

I told it to make me a meme and it made this

tbh pretty meta because that's how I felt when I saw it

[glad grove]

my honest reaction

[hardy cipher]

I might max out the aesthetic score setting to see how that works otu

[trim orbit]

have barely ever made anything in python, let alone a comfy node.

[hardy cipher]

nice

[trim orbit]

its pretty straight forward tbh

[hardy cipher]

the node building?

[trim orbit]

yeh

[hardy cipher]

I was looking at the sample node

[trim orbit]

i cobbled it from that file

[hardy cipher]

thinking of making a pipe that better suits my needs

[upbeat summit]

I'm currently using refiner pos ascore 10 and refiner neg ascore 3 - getting got results but as usual - it's a work in progress

[hardy cipher]

and/or samplers that can be daisy-chained

seems inefficient to run 10 foot long noodles when I don't have to

[hardy cipher]

how high does it go? I want it so beautiful it sucks

[upbeat summit]

I think 2500

[hardy cipher]

aesthetic score 2500, steps 10,000

[upbeat summit]

I forgot - it caps out at least in the node at some point - or was it 9999?

[hardy cipher]

this is the way

[upbeat summit]

but I could see differences between 6 - 10. so I've been since at 10 pos ascore

[hardy cipher]

supposedly you're not supposed to go over 6 if you want ideal photo-realism

[upbeat summit]

supposedly

I switch between 8 and 10

negative ascore is at 3 for now in my workflows

[hardy cipher]

I feel like the AI is on a different meme plane than I am

[upbeat summit]

ah a sports car spotted in the wild

so lush

[hardy cipher]

it could be performing on some sort of rain man level meta meme plane that I can't even wrap my head around

those typos. anyway

[upbeat summit]

the latent space is a strange place

[hardy cipher]

I'm also doing weird things with the prompts. no idea how that's impacting things.

[upbeat summit]

that's how it should be. you might find more coherence, push fidelity or break it completely hehe

[hardy cipher]

send pos neg prompts to a styler, have that sent to G, have the G prompt sent to another styler, have that output to L, then back in and styled again for refiner

[hardy cipher]

it probably makes no real sense to do this

[upbeat summit]

love those setups 🙂 it needs to be super simple or I process stuff with very elaborate (and convoluted) pipelines

[hardy cipher]

I just need to find something that trims them when they get too bulky

I like that ape man. what checkpoint/loras did you use?

woo

[upbeat summit]

nice water interactions

[trim orbit]

melympia

[upbeat summit]

there's a string trim node in the WAS package

[hardy cipher]

I saw that, wasn't sure how it works exactly. I'll build huge elaborate things and then I'll knock them down like sandcastles and start over

[upbeat summit]

also look at those. you can make a lot of string replacements with regex
https://github.com/pythongosssss/ComfyUI-Custom-Scripts#string-function

[hardy cipher]

had a setup with one prompt thing running into another prompt enhancer and all sorts of other things. but it felt like too much so it's gone now

well it's still in the images I made

[midnight shuttle]

No matter how much is done in the workflow, everything eventually turns into weights for the model. The question is if there is any logical connection between the prompt and the weights after all of the steps are done.

[hardy cipher]

also tried to work out some sort of cfg scheduling setup, but lots of fail there. I need to figure out what "mimic cfg scale" means

[hardy cipher]

yeah. this is true. haven't ever used that stuff before, so thought I'd mess with it. I have had gpt-4 or short bus bard write me prompts before

[upbeat summit]

exactly and you can also choose from different types of weighting methods. there a couple of text encoders that offer stuff like comfy, a1111, comfy++ etc

[midnight shuttle]

I'm surprised no one has introduced a random weights extension yet. That could generate images that would otherwise be impossible with normal prompting.

[hardy cipher]

yeah, I thought about that

I wonder if any smart people are doing this sort of thing with sdxl https://engineering.purdue.edu/ChanGroup/ECE595/files/chapter3.pdf

they use the term "attack" but really just seems like you're pushing the model to weird places

[midnight shuttle]

I've seen that paper. It refers to tricking a classifier into misclassifying an image. Not really SD related.

[hardy cipher]

well yeah, but couldn't it just be done in reverse to a degree?

misunderstanding input. it can be done with LLMs

[midnight shuttle]

What would be the reverse of tricking a classifier? Making a classifier more accurate?

[hardy cipher]

don't know if it can be done with normal words

[midnight shuttle]

Do you mean finding prompts that generate completely unrelated images?

[hardy cipher]

well classifiers identify images right? isn't that just text to image in reverse to a degree?

image to text, text to image

[midnight shuttle]

Not really. The way they operate is very different.

The closest analogy would probably be finding some type of prompt that sounds right but generates a totally unexpected image.

[hardy cipher]

well how about the visual transformer model/models? they're involved in both processes right? I mean, I can't say I know the specifics of how it'd work

and it's a thing https://arxiv.org/abs/2208.04135

not sure what you could gain from it

[midnight shuttle]

Yes, that's an example of generating images based on nonsense words. I wouldn't really call it an attack though. It's just a function of how language tokenization works.

[hardy cipher]

I don't know why any of them are called attacks tbh

[midnight shuttle]

You could give a prompt of sdmlefrjweoijsdfokjoks and you would get some type of image.

[hardy cipher]

unless you're literally trying to do something malicious and damage things

my cat slept on my keyboard once and put in a nice string of randomness. the image it made wasn't bad tbh

[midnight shuttle]

Tricking a classifier can be an attack if you can trick the Tesla into thinking that a wall is open road or a car is a driveway.

The second one isn't an attack. They just call it that to get more attention for their paper.

[hardy cipher]

yeah. I'm just curious about the liminal space so to speak. the out of bounds areas in these models that exist but really aren't accessible with normal approaches

like clipping in a video game

[midnight shuttle]

I'd have to look at the raw output of the encoder and figure out how to produce random values instead.

[hardy cipher]

I've messed with LLMs a bit and figured out how to bypass a lot of the guardrails and things. but not sure there's an equivalent with image generation models. and it's really not all that exciting

[midnight shuttle]

Bypassing the guardrails is more a function of tricking the business logic. There aren't any guardrails in local SD.

[hardy cipher]

it's really not as interesting as I thought it would be. I got bing to talk about sentience. and tell me how to hotwire a car. that was about peak excitement in that realm

[midnight shuttle]

And to be clear, not training the model on NSFW content (SD 2.1) isn't a guardrail. A guardrail is business logic that checks and filters the output to avoid getting the maker of the model in huge trouble.

[hardy cipher]

yeah, omission vs censorship essentially?

problem with those guardrails is they're overarching and essentially make the models dumber. not sure how that's going to work out in the long run

well I guess the models aren't dumber. they just have to play dumber

[midnight shuttle]

Some call it censorship. Some call it safety. But the difference is between what's in the training data and what's done with the output after it's generated.

The big companies will have models without guardrails, or with their own guardrails. ChatGPT was just a tech demo. It was never intended to be anything more than a toy.

[glad grove]

what if we add a monitor software to force all of the gpus to always connect to nvidia db and check if you are not doing anything illegal

[elfin cobalt]

https://brage.info/GAN/ganbot2/ec099909-d163-4de8-91d8-82b39952a7a1.1.webp

Would you play this guitar?

[midnight shuttle]

Show average people the output from SD 1.5 and they just might vote for that.

[hardy cipher]

is that real, baughn?

[glad grove]

they might vote but it has to pass 2 chambers

just like the tiktok ban

[elfin cobalt]

Consider the channel I'm posting it in. 😅

[midnight shuttle]

Show average lawmakers the output of SD 1.5 and they just might vote for that.

[glad grove]

they might or maybe they wont ill ask schummer tonight

[midnight shuttle]

Most people don't even know SD exists and they definitely don't know it can be run in an unsupervised unaccountable environment.

A lot of people wouldn't like that idea. They want accountability.

[hardy cipher]

yes, hello, I'm corporate person

[glad grove]

id vote for jack johnson over john jackson any day

[hardy cipher]

smile

[glad grove]

look at that clean bathroom

[gloomy barn]

Are those the good hands I was told about?

[hardy cipher]

lol

I like how they meld together

[glad grove]

look at legs of the statue

[hardy cipher]

true chad legs

[gloomy barn]

the statue is a wreck

[hardy cipher]

I mean, depends on your perspective

these food images are really getting me hungry

[smoky patrol]

That color palette, yet...that f o r k

[gloomy barn]

how did you ge that color?

the dark red?

[hardy cipher]

using "sdxl prompt styler" forget which node package it's from. anyway, has all sorts of style alterations it can make to a prompt

I think that one was food photography/dystopian/enhance

[glad grove]

it sucks 😔

[hardy cipher]

I don't know what is real anymore

[glad grove]

🍽️

[glad grove]

sensei 🙇

[native knot]

Iron Chef Steven Seagal?

[native knot]

Amazing.

[upbeat summit]

Any great warrior is also a scholar, and a poet, and an artist
- Steven Seagal

[hoary saddle]

Stud!

[hardy cipher]

I didn't realize steven seagal was also an accomplished musician

songs from the crystal cave

one of his albums

[native knot]

SDXL really is awesome.

[hardy cipher]

I feel like the people that hate on it are the people that hate on everything

[trim orbit]

whatever their youtuber tells them to

[native knot]

It's early. The only reason 1.5 is what it is has to do with how much time and effort everyone invested to improve the capabilities.

The myriad of LORAs, etc...

[trim orbit]

controlnet

[native knot]

Give SDXL time and it'll surpass.

[vast galleon]

Alright! Coding is tested and ready for release. 4 custom nodes for inpainting, automatic face fixing, and color transfer

[hoary saddle]

Here's my controlnet if anyone wants it

[hardy cipher]

indeed. it's like getting an untuned supercar or something. but what I really don't get is if people are so disappointed in it, why not just keep using 1.5 and remind yourself it's all free for us anyway.

[trim orbit]

i think its clear and it does surpass now already

[hardy cipher]

yes

[trim orbit]

people getting bad results are cherry picking them to say "see!"

[glad grove]

they are mad because big majority cant run it

[trim orbit]

i guess that's anyone's but their own fault

[hardy cipher]

I'm running on a laptop, just barely, lol

[glad grove]

like most things in life its all about money

[hardy cipher]

if I couldn't use it I'd probably just not use it. or just use colab or something

as I do anyway for things

[trim orbit]

it feels more like a massive digger to me

or a big rig

[hardy cipher]

something they haven't even built the parts for yet

to fine tune

[trim orbit]

https://youtu.be/LJDyr3xiDrE for reference

[hardy cipher]

I think I saw that movie

[gloomy barn]

If you're going to use an image to show off a model, you'd think you'd fix it up, make sure it was good.

[glad grove]

this good

[upbeat summit]

perfection 👌

[gloomy barn]

I want to see the rest of that face

[gloomy barn]

It looks like Segal crossed with Tarrantino

[hardy cipher]

it makes him extra asian in all the pics I've done

[gloomy barn]

try "asian" in the negative

[hardy cipher]

yeah, problem is I just set it to run a bunch and went to do other things. so came back to asian style seagal

[gloomy barn]

maybe it's the model

[hardy cipher]

that's most likely the case

[gloomy barn]

Ive been going through images I made last night. as always, oy those hands

[hardy cipher]

I saw that controlnet is up and running to a degree

for xl

probably lots of vram though

[gloomy barn]

that's exciting news

on comfy right now. The SDNext folks say they will have it working soon

[trim orbit]

just a canny model but its working well

vram i can't confirm anything. it is slower to use

[gloomy barn]

is it canny or cammy?

[hardy cipher]

it's whatever you want to call it

but I know it as canny

[gloomy barn]

what does canny mean?