✨｜sdxl

[lilac raven]

Warning the user that closing the window will close the running program is a feature that is shocking that Windows terminal doesn't have

[vital wraith]

It likely did, I was just being oblvious, I wouldn't have assumed the settings menu tabbed into the same cmd window

[native knot]

It does if you're closing more than one tab.

[vital wraith]

Found the text buffer options at least, should know if it decides to go off and do it's own thing this next run.

[elfin flare]

ComfyUI got an OpenVINO update?

[high skiff]

Alright, took some time down. Getting ready to crack down on some workflow development for all 4 of my new workflows launching with my 1.1 workflow release

[wet nacelle]

nice

[vital wraith]

It only completed to the second epoch, but it's somewhat functional. I really need to reduce the number of steps.

[high skiff]

this includes:

1. A new base workflow that runs faster, uses less VRAM, and also has a higher quality high res fix that produces better results, alongside the original high res fix.

2. A dedicated img2img workflow (of which I have not started, but I have some ideas for)

3. A workflow for creating frame by frame exported image batches of each step during an images diffusion, complete with variables and LoRA support. Its used to export individual frames to make diffusion gifs

and finally 4. A simplified diet version of my workflow that is optimized to be as light as possible for those with VRAM constraints

[high skiff]

Lovely images!

blueopard

[wet nacelle]

looks great

[vital wraith]

Anyone find or make a good JSON for SDXL and Kohya for character training on 10gigs or less VRAM?

I can do 1.5 pretty well, but SDXL has me scratching my head at times.

[high skiff]

its a big ask, unfortunately

[vital wraith]

I'

[high skiff]

I train occasionally and have a team of trainers, but I have not seen a preset that can do at/below 10GB VRAM

[vital wraith]

I've made a lot of progress so far, and haven't actually hit any Vram issues

[high skiff]

oh?

is it just a settings thing then? not proper training params?

[vital wraith]

Well, it did it's own thing and I lost my epoch progress bar, honestly never ran into anything like it before

It was my first run with SDXL, so perhaps it was something I did wrong.

[glad grove]

we ready for the summer lookin sexy

[wet nacelle]

@glad grove

[high skiff]

@strong field do you happen to be here?

[lilac raven]

I'm noticing something strange with SDXL, rectangular images seem to be more distorted than SD 1.5

[vital wraith]

If I am not using bucketing, do my images need to be the same aspect ratio?

[lilac raven]

The main thing that seems to be wrong are arms and legs

[wet nacelle]

"I become.....robot?"

[native knot]

Are you using refiner and for how many steps? Also, what resolution are you rendering?

[glad grove]

he lookin sexy here

[lilac raven]

The distortion is more noticeable with refiner, the issue starts with 5:4 (as 1280x1024) and from that point it gets worse and worse

[wet nacelle]

https://tenor.com/view/yeah-bwoi-grin-chef-gif-18525191

[native knot]

To reduce anatomy issues, stick with these sizes or darn close: #✨｜sdxl message

Because you're running 1280px width, you'll start to run into more issues at heights above 768px.

It can work, but it's more prone to issues outside of the resolutions on that table.

[wet nacelle]

????

There we go

[lilac raven]

What is the criteria for these resolutions? Is this experimental or documented somewhere?

[wet nacelle]

Would it be simple to add more image nodes for it to read from?

I just use two right now.

[wet nacelle]

clip_vision

[vital wraith]

The aspect ratio info is straight from Stability AI.

There are a lot more listed, but they outlined the best ones.

[high skiff]

generally speaking, you want the product of your height and width to be as close to 1024^2 as possible

So you want your height and width to multiply and be as close to 1,040,000 pixels as possible

[native knot]

I just took the liberty in my post of making a nice table with the decimal and ratio numbers in text.

[high skiff]

even some of the SAI ones don't adhere to that, which I called out not long ago, leading to some aspects having some issues

[elfin cobalt]

https://brage.info/GAN/ganbot2/dc2bb44a-dc3e-4a18-8da0-439d7b2f966d.2.webp

Choo!

[high skiff]

@upbeat summit I see you typing, I have an offer for you

typy typying

[strong field]

Do You want to try out that AIT module?

[high skiff]

yes please!

[strong field]

Alright I’ll dm when I get to desk

[high skiff]

mucho thanks

[upbeat summit]

These are all SDXL resolutions from the official SDXL paper (https://arxiv.org/abs/2307.01952)

I created a preset file for @west breach Mikey Nodes' awesome Ratio Advanced node.
https://github.com/bash-j/mikey_nodes

• the preset file is part of Mikey Nodes and you can find it in the nodes directory: user_ratio_presets.json.example

• I approximated the ARs (aspect ratios) where they would roughly fit

• added where the nearest simplified AR actually is

• added the AR decimals

• calculated all target bucket values (target_width / target_height) proportionally to 4096 being the long side and rounded the height values to multiples of 4

mikey's node also has a swap_axis widget

so you only need one preset to switch between landscape and portrait (x/y axis) mode

[high skiff]

@upbeat summitMind if I hit you up in DMs?

[upbeat summit]

anytime!

the aspect ratios are really only estimates / approximations. I may need to update those. but the resolutions are all from the paper

[vital wraith]

Does this help outline resolutions outside the standard ones I am familuar with? Or is it a tool to easily switch between the best known resolutions?

[upbeat summit]

the list I posted and the preset file contain all standard SDXL resolutions from the official paper

[vital wraith]

I have restricted myself to the common base resolutions: 1024 x 1024, 1152 x 896, 896 x 1152, 1216 x 832, 832 x 1216, 1344 x 768, 768 x 1344, 1536 x 640, 640 x 1536

I am not sure I fully grasp, from what I am gathering, it looks like with your tool I could easily explore others?

[glad grove]

new season of baywach confirmed

[ionic gulch]

cheeseburger scene as cliffhanger of season 1?

[glad grove]

yes he fights against kfc smugglers

[native knot]

The Rock as Kremer and as Jerry.

[lilac raven]

So that is the original source

[glad grove]

i like the giant ear plushie thats floating above the door

[wet nacelle]

https://tenor.com/view/desmondpacito-cooking-chef-gif-12378159

[glad grove]

https://tenor.com/view/neco-arc-gif-25039455

[native knot]

I saw someone's node that had a boolean that acted essentially like an on/off switch and you only had to click it to change between true/false. Does anyone know what node that was?

[native knot]

We need a Seagal/Rock face-off.

[glad grove]

we still waitin for our robert zdar lora

[lilac raven]

I can't paste the bad image here, I wanted to show the comparison

[glad grove]

let him cook 🔥

[lilac raven]

Ok, I think it's going to work now, the bad arms vs. the better arms

[glad grove]

double dragon

[upbeat summit]

Yes, that's what it's exactly for - to easily explore other resolutions and switch between them including the SDXL target bucket values.

It's 2 clicks to select a preset with Mikey Nodes' Preset Ratio Selector or the bigger brother Ratio Advanced after you've set it up, connected the node in your ComfyUI workflow and moved and renamed the example file to your ComfyUI root directory.

[half ivy]

The lighting though 🔥

[vital wraith]

My Dude. That is super convient.

[glad grove]

i like this one the brown deformed mickey is my new fav character

[upbeat summit]

@west breach did all the heavy lifting and coding. I contributed some ideas and concepts.

[vital wraith]

Well I will happily piggyback off your labor to save me time and take credit for cool things I am generating. 😉 Seriously though, thanks you two. This is a big time save, I could never do this manually.

[upbeat summit]

have fun! if you have any questions, let me know.

[glad grove]

sexy

[west breach]

@upbeat summit has been a huge help with my custom nodes. Lots of valuable ideas, testing and feedback 🙂

[lilac raven]

I'm going to ask the question: is the AITemplate VAE decode tile broken?

[native knot]

This one?

[strong field]

Pretty sure this one just gives black images at th emoment

[native knot]

Or this one?

[lilac raven]

This one

[lilac raven]

I'm trying to make things go faster. So far, it went from 85 to 68 seconds for 3 1344x768 images

[strong field]

oh wait, are you using fp16 vae?

[native knot]

Hey...NSFW, dude...lol

Too sexy for server

[wet nacelle]

TKFW (TOO KOOL FOR WORK)

[glad grove]

u right im in love already 😳

[lilac raven]

I changed nothing, so it probably is

[strong field]

VAE is compatible if you use the fp16 fixed version

[wet nacelle]

egg head

[lilac raven]

Is this available somewhere?

[wet nacelle]

what?

[strong field]

https://huggingface.co/madebyollin/sdxl-vae-fp16-fix

[native knot]

Burnon balk naces

[wet nacelle]

It's that value that's doing this.

That's fucking funny.

[lilac raven]

Thanks, downloading

[lilac raven]

The first time it runs seem to have a lesser effect, but on the second run the impact is more noticeable. It was taking 85 seconds, now it's 51. This is so good, thank you 👍

[glad grove]

hello ladies 😎

[lilac raven]

And another thing, AITemplate builds for Linux too, very good

[lilac raven]

Is this AITemplate usable on 4 GB VRAM? I'm wondering if it's worth trying to set it up on my laptop

[strong field]

First time it runs it has to load into memory. If your keep it loaded it may be even faster but only if you have 32gb+

[vital ermine]

I need to find a caption tag editor because my old one only works with WD1.4 style tagging

[strong field]

Not sure where the lower limits are but I’m guessing probably not

[lilac raven]

Error occurred when executing KSamplerAdvanced:

No modules found for linux xl sm75 1 1344 unet

RIP ancient GPU

[strong field]

Sm75 not supported by AIT

[glad grove]

then make it supported

[wet nacelle]

He can't forget to wear something black. He needs to look extra kool.

[glad grove]

plus hes a black belt (real)

[wet nacelle]

Plus he can't lose (real)

[strong field]

https://tenor.com/view/ace-ventura-captain-i-dont-have-the-power-cant-do-it-i-just-cant-do-it-captain-gif-15850032

[vital ermine]

Extra hand to help?

[native knot]

That's just Thing giving Wednesday an extra hand.

[hardy cipher]

lovely

[wet nacelle]

These are great

[upbeat summit]

yes they are 🙂

[wet nacelle]

yuppers

[tender timber]

No… I guess not. I believe the article mentions AUTO1111, was hoping it would have appeared in Comfy as well as part of an update… like perhaps it was an Intel module that is pushed out and Comfy doesn’t have to do anything other than update the requirements.txt

[glad grove]

new season of hawaii lookin bloated already

[hardy cipher]

is there any reason al the samplers in the controlnet example workflows are fixed at seed 1?

[hard fractal]

Nope

[hardy cipher]

cool. I guess it was a coincidence that I turned it to random and all the sudden got this and a list of errors

restarted and seems to work fine again

[hard fractal]

gosh...

haha

[hardy cipher]

looking good now

[craggy ibex]

I heard that you can also imitate some artstyle from artist

but how tho?

What prompt I need to type them in?

[hardy cipher]

well, you see, there's the prompt. and you type the stuff in there.

in the style of "artist name"

[craggy ibex]

Just put the artist name in positive prompt as usual???

Damn

[hardy cipher]

a poignant painting of Gary Busey in the style of Hans Memling

I mean, you don't have to put "in the style of" really. say it however you want. that's just how I do it normally

[native knot]

For instance, you could say, "Hans Memling style" and largely get similar output.

[steady grove]

medium in style of artist name. like, painting in style of van gogh, or graffiti in style of banksy

million different ways to do it but adding context helps

[hardy cipher]

yes, I've also just put the artist's name with no context. but problem with that is it sometimes also adds the actual artist to the image

[hardy cipher]

gary busey in the style of banksy?

[native knot]

Gary Busey spraypainted Banksy style on a wall in Rome near a concrete pillar

[hardy cipher]

lovely work

[native knot]

Looks like scenes from an Amnesia game.

[wet nacelle]

https://tenor.com/view/dbav10-vlog-deaf-baby-sucking-thumb-gif-17322649

[hardy cipher]

anyone try controlnet strength 10 yet?

I'm guessing it just makes the image 10 times better

[glad grove]

woah easy tiger

[native knot]

That red hall is begging for a Daredevil fight scene.

[hardy cipher]

base model kept giving me cartoon seagal, but run diffusion did me right

[steady grove]

daredevil has super hearing. he woudln't know how to find the red hall

[native knot]

Bruh. He's got friends that will tell him which hall to go to.

[steady grove]

they gonna lead him to a fight? dick friends

[hardy cipher]

he's a good fighter though

[native knot]

They know what he's capable of.

[wet nacelle]

@hardy cipherSee what this model can do with it if you haven't already deleted it.

https://civitai.com/models/119202/talmendoxl-sdxl-uncensored-full-model

[hardy cipher]

I don't delete models normally. I just leave some of them on my external drive because they fill up my ssd very quickly

but not a worthwhile endeavor to load them from the external. takes about 5 minutes

[wet nacelle]

ah okay

[hardy cipher]

I'm sure you care about all of that. but yeah, I'll try it out

whoa

perfection

that wasn't talmendo though

bumped up controlnet over 2 for that beauty

[glad grove]

step aside guys real men have arrived

[hardy cipher]

oof

[wet nacelle]

I can't post these.

It's too good

[hardy cipher]

what is the deal with civitai? is it just bogged down? are they also installing malware and viruses?

[elfin flare]

Well that's disappointing. They are indeed working on an Openvino WebUI however in the intel arc discord.

[wet nacelle]

I don't have issues right now.

[hardy cipher]

took about 5 clicks on the download button for it to work

[lilac raven]

How do I convert a safetensors to ckpt?

[hardy cipher]

but then that's happened before where it seems to register the clicks all at once

[wet nacelle]

oh kool

[mild path]

hi

[mild path]

looking for sdxl_depth.fp16.safetensors for a workflow, is this one the same with a different name? diffusion_pytorch_model.fp16.safetensors

[hardy cipher]

well, you could always try it out

but being that it doesn't say sdxl or depth in the name, I'm leaning towards no

[mild path]

it comes from huggingface.co/diffusers/controlnet-depth-sdxl-1.0/

[hardy cipher]

hmm

[mild path]

same with IP-Adapter_sdxl_pytorch_model.bin that does not exist, but i found huggingface.co/h94/IP-Adapter/resolve/main/sdxl_models/ip-adapter_sdxl.bin, would that be it?

[hardy cipher]

https://huggingface.co/stabilityai/control-lora/tree/main

not sure about what you were looking at, but try those

[hardy cipher]

not blaming the model, but keep getting these cartoon images. might be how I'm prompting things

or maybe I'll turn controlnet down

[wet nacelle]

I bet it's the model

[hardy cipher]

it is coming from a drawing I rendered. but seems like canny shouldn't be that worried about all that

just need a little while to get my bearings with the layout and settings. hadn't used controlnet much in comfy before now

[soft bone]

@hard fractal will there be no SAI openpose?

not SAI sadly. that one is trained on a consumer card

im talking hundreds of imgs vs. hundreds of thousands

[glad grove]

someone said free a100 for grabs?

[soft bone]

Try Thibaud's version of 2.1 openpose. It puts hats on every single person and impossible to prompt out. Someone else had to make another one correctly.

I'm very familiar with Thibaud, we used to train together in a small tuning server. And you're right, perhaps you shouldn't have engaged on a message directed to someone else

[azure oxide]

whys everyone so aggressive in this channel lmaoo

chilllllll bros

[glad grove]

years and years of seeing nvidia prices

[soft bone]

being "trained on laion" doesnt mean every image in the datasset. SD was trained on laion too, except with hundreds of thousands scraped from it. It could be 50 images "from laion" lmao. and a single a100 in a home pc is considered consumer card to me.

#🌶｜off-topic

[glad grove]

https://tenor.com/view/jetstream-sam-metal-gear-rising-gif-25742644

[soft bone]

#✍🏼｜rules-and-tos message

[vital ermine]

WHOA

[vital ermine]

Here it is with my lora

[lilac raven]

Managed to convert a safetensors to ckpt. This is such a bad idea

[steady grove]

ckpt aren't dangerous on their own lol. its when you get them from weird places they could be

if you making your own then just ask, do you trust yourself?

[lilac raven]

Right now, I don't trust myself 🤣

[glad grove]

whose dad is this

[steady grove]

https://tenor.com/view/tyler-durden-gif-24812513

[autumn forum]

just published my cinematic model if anyone is interested to try it https://civitai.com/models/131843

[vital ermine]

18 could be anyones

[lilac raven]

It seems it's not possible to compile AITemplate on a GTX 1650 Max-Q

[azure oxide]

if this is true then i wouldnt suggest trying to compile a new AIT model, should probably know what youre doin first

[vital ermine]

No, only Ampere, ADA and possibly Turing but no pascal/16x0 cards, or kepler, or maxwell, etc...

[lilac raven]

The error here said sm75. There is sm75 on AITemplate's source, so I think it should be possible, but probably 4 GB VRAM is not enough
Or maybe there is a way to cross compile

[vital ermine]

remember those cards were utter dog poop as they lack everything. Even a 1060 is better than a 1660 which is mind blowing. Things I never had an issue with on a 1650/60 they had to jump through hoops or get all black generations.

[void zenith]

Not hot ngl

[vital wraith]

It's stupid slow but I finally got Kohya and SDXL to play along with 10 gig Vram.

[noble shoal]

I have a 4070 with 12 gig and it is not much faster 🤷

[vital wraith]

I am barely limping along with a 3070 10gig. I had to basically fight kohya to get it to work lol

I still don’t know if it will complete to be honest

[glad grove]

hottest day in russia

[short marsh]

crammed everything on screen

[autumn forum]

impressive work!

[crisp owl]

There's a little extra empty space around those switches!

[short marsh]

maybe the next ControlNet release can go in there 😆

[crisp owl]

That's what I like to hear 😎

[shy kelp]

such a cool lora

[hardy cipher]

grandpa wizard robin williams

[hardy cipher]

she looks a bit stressed out

[shy kelp]

yea

she got some dishes to wash

[glad grove]

she looks like she gonna wash them using my skin

[copper spire]

Wow well done, anyway to share that workflow ahah look way more tidy than mine 😄

[hardy cipher]

yikes

[vital granite]

.

[proper depot]

Which upscaler model do people like for SDXL? 🤔

[crisp owl]

I use my same main one, NMKD-Siax_200k

[hardy cipher]

gotta controlnet some of these

[uncut fiber]

i am using probably poor swinir 4x, just dont know how to say dont do it 4x do it say 2x only. What node helps with this? Do you know?

[hardy cipher]

there are nodes that do that, forget which node packs, I started an entirely new install so starting over on the nodes. you can also 4x and then downscale to 2. maybe you already know, but if you double click in empty space it pops up a search box, you can look around and see what you might have

[uncut fiber]

yes that i know i can downscale and how, but once i got it 4x why throw it away.
@hardy cipher will do!!! and thank you

[hardy cipher]

personally, I like to upscale to 4x, maybe do some detail fixing, then go back down to 2x

then it's not such a huge file, and when I do the fixes at 4x I think it's just all around better

or might be wrong, lol

wtf. why is there an openai module in the mix when I get errors with controlnet

not the end of the world, only happens once a while, but still. what does openai have to do with this. aren't they doing enough ruining their own ai models, why do they need to mess with us?

[uncut fiber]

I found some bislerp will try it

[hardy cipher]

there's iterative upscale. but sometimes I open nodes like that, see 12 input and 50 widget settings, and I just delete them as quickly as they load, lol

not trying to do all that to make my image bigger

[crisp owl]

yeah I just downscale the 4x

[hardy cipher]

I do appreciate the complex options being there for when I might really want or need them. but I just don't have the patience or want to do that sort of thing on a regular basis

really grinds my gears when I have to manually add paths to the environment variable list

this time around, if a node pack keeps getting errors I don't think are reasonable, Imma zap it

[soft zealot]

if you want to do it the really lazy way with 0 widgets etc just use built in wiondows tools

[hardy cipher]

oh yeah, ctrl+e to get to edit mode, fyi

[hardy cipher]

but I like to watch it all happen, lol

with a1111 and 1.5 I'd do a lot of edits in photoshop too

[glad grove]

bro loves his chicken

[uncut fiber]

by some miracle i got it to work 😄 i mean bislerp.

anybody with good eyes got some pos and neg prompt?

i know but dragging will destroy my genial noodles settings 😄

seems road is bit weid, going high with steps

[glad grove]

little bro needs to upgrade his neck 😭

[uncut fiber]

😄

[woeful patio]

does anyone know where i can download the new controlnet models? I can't seem to figure it out

[soft zealot]

https://huggingface.co/stabilityai/control-lora

[uncut fiber]

I think i will render in native resolution and upscale only those which i will like. It is big performance lost and it also change content. Sooo 🙂

[glad grove]

thats what i do,no point in wasting electricity upscaling all of them

[woeful patio]

thanks, Is there any reason revision is there separately?

[soft zealot]

Ok so presuming that your entire pc is drawing 500W at full load on everything then an hours running is 0.5kw/h

At the rates I pay (29.936 p per kWh) thats around 15p per hour so presuming an upscale takes around 2 minutes on my poor old 1080ti then an upscale costs peanuts

[soft zealot]

no idea, read the model card

[soft zealot]

or to look at it another way running my PC 24/7 costs me around £3.60 per day (if running at 500w draw constantly)

[glad grove]

i still only upscale good ones,otherwise it would take a while to do hiresfix on 500+ images

[uncut fiber]

in my case not much about prices, but about 3x more time.

[soft zealot]

heres a pro tip.

You dont need to sit & watch things happen.

You can walk away and do other things like walk the dog, make a coffee, do some housework

[glad grove]

yea its all about time,im literally rotting and dying here every second

[uncut fiber]

yes for sure, but i am still experimenting

[soft zealot]

most likely because ReVision is not really a ControlNet but who knows, ask one of the SAI guys

[woeful patio]

yeah i think because it's clip vision and not control net, thanks

[soft zealot]

SO I was curious and decided to see what ReVision would spit out gievn 2 photos of myself

So equally weighted I got

[peak dove]

clip_vision_g.safetensors is taking a month of Sundays to d/load!!! 🙂

""You can walk away and do other things like walk the dog, make a coffee, do some housework""

[woeful patio]

stupid question but I cant seem to find the following node: MiDaS-DepthMapPreprocessor
anyone know where to get it?

[soft zealot]

a few more from the same process

[peak dove]

Try Github https://github.com/nbeeeel/Distance-Measurement-Using-MiDaS

[heady vale]

its part of this. you can add via the manager

[woeful patio]

I'm an idiot how do you get there?

[heady vale]

do you have the manager nstalled?

[woeful patio]

I have regular comfyui installed

[heady vale]

bottom of the toolbar is the manager button

or install from here https://github.com/Fannovel16/comfyui_controlnet_aux

[woeful patio]

is it supposed to be in here?

[heady vale]

yep

[heady vale]

install the manager. it makes life easier

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

[woeful patio]

amazing, I'll install the manager

[woeful patio]

It works now, Thanks for the help.
I'm trying to cross over from a1111 to comfy so everything is a little confusing right now lol

[heady vale]

nice one

[craggy ibex]

What checkpoint is recommended for create spells particles like this?

This is from Craiyon btw

But I'm not really sure. what model or checkpoint I can use?

[upbeat summit]

base SDXL is pretty good with spell magic effects

[craggy ibex]

with txt2img?

Searge?

[upbeat summit]

yeah. it's all about the prompting

[upbeat summit]

try something like "casting an ice spell" or "surrounded by magic spells" try to describe the effect you are looking for

[craggy ibex]

alrighty

[woeful patio]

is multy controlnet a thing in comfy?

[hardy cipher]

do you mean running more than one controlnet model/lora in the same workflow?

if so, it's certainly possible, just have to connect the nodes correctly

although it does increase resource usage so unless you have sufficient vram it will slow things down A LOT

[woeful patio]

my 8gb 3070 would like to say that it's not happening then

[hardy cipher]

yeah, I have a 6gb 3060, lol. I tried 2 in a workflow, first one ran at about normal speed, but the second one would have taken way way too long so I just cancelled it. it's more efficient to just run one, then take the result and run it in another

you might not want to use the default workflow unmodified for the second pass though

for the second round what I've been doing is using the original image as the controlnet input, but then using the result of the first round as the latent input for the sampler

[hardy cipher]

cat

[shy kelp]

cattonaut

[hardy cipher]

man, I forgot all about it but I made a bunch of cattonaut images way back with 1.5. yours is much higher quality though

[shy kelp]

sdxl make it more realistic

[craggy ibex]

Something I don't like about SDXL is that

it take so long to generate images

[soft zealot]

not what I expected lol

[hardy cipher]

well, it's rendering an image that's 3 times bigger by default

[soft zealot]

how long do you call long ?

[craggy ibex]

around 1-2 minutes

[hardy cipher]

Sekaiza is quite emo. he's not happy about much it seems

[craggy ibex]

Nah I am happy

But I am impataitn

Impatient

[soft zealot]

thats not long in the grand scheme of things (personal bugbear nobody has any patience these days and wants instant gratification)

If it was taking 20 minutes yes I can see it would be an issue but 1-2 minutes?

purleeeeze!!!

[hardy cipher]

I mean, 3 years ago I would have been amazed with these results if they'd taken an hour per image

or 2 years ago even

[soft zealot]

or 2 months even lol

[hardy cipher]

yeah, I was pretty adept with 1.5. but that was more of a numbers game

render a bunch, find a few good ones, maybe a couple very good ones

[craggy ibex]

with SDXL

Thaumaturge

DONE

GAH DAMN ;-;

I hate waiting

[hardy cipher]

I like that one

[shy kelp]

i still dont know how midjourney create an amazing images, did they have their own model or something?

[craggy ibex]

Why A.I. is good at creating spells?

and particles???

[hardy cipher]

it's magic

[hardy cipher]

yes

[hardy cipher]

I do believe midjourney does a lot of optimizations under the hood

maximum saturation activated

[craggy ibex]

from... AnythingelseV4

Thaumaturgy

[hardy cipher]

looks like a google maps satellite image

[wind zenith]

steam punk shit

[wind zenith]

wtf quantom computer in a volt?

[craggy ibex]

Look like an album cover to me

[hardy cipher]

plugged blueprints into clip vision and made some pretty nutty blueprints, then sent those through controlnet

[craggy ibex]

oooo I love this!

[wind zenith]

arabic art shit

[craggy ibex]

Yeah

Imma use these pics for DND

Dungeon and Dragon Virtual Tabletop

[hardy cipher]

blueprints, lol

[hardy cipher]

try "2.5D" with stuff like that

in the description words

[craggy ibex]

Alright

Which checkpoint I should use?

[hardy cipher]

I don't know much about any of those besides base

[soft zealot]

all about that base, bout that base, no trouble

[hardy cipher]

https://tenor.com/view/ace-of-base-song-always-have-always-will-linn-berggren-album-flowers-sweden-gif-22137854

[wind zenith]

https://tenor.com/view/rock-one-eyebrow-raised-rock-staring-the-rock-gif-22113367

[hardy cipher]

"load controlnet model (diff)"

anyone use that one?

[twilit grove]

hello, what's the difference between these two ?

[hardy cipher]

you want the second one I believe

[wind zenith]

vae and normal model?

[soft zealot]

one is the original realease the other is the follow up release with 0/9 vae baked in

[hardy cipher]

someone needs to make a brian peppers lora

[twilit grove]

so what should i download ?

[hardy cipher]

lol

bro just ignores me

[soft zealot]

most people would probably say the one withthe bakedin 0.9 VAE

[twilit grove]

ok, i use automaic webui so i need to put it in models/stable diffusion or models/VAE ?

[soft zealot]

models (with or without baked in VAEs) go in stable_siffusion

Standalone VAEs fgo into VAE

[twilit grove]

ok, but i don't have the lora file is that normal ?

i use 1.5.1

[hardy cipher]

https://tenor.com/view/the-invisible-man-hat-raining-clouds-gif-16516946

sometimes these questions really discombobulate me though

[craggy ibex]

Magical Bag...

[hardy cipher]

how do I use sdxl without the lora?

[craggy ibex]

How to make it generate ONLY ONE SINGLE BAG

JESUS CHRISTSLKDGNKLSJFNG

[soft zealot]

dont add the loara

[half ivy]

Just wish there was a SDXL bro that could tell me where im going wrong with my control net

(kicks stone n sulks

[hardy cipher]

I was being sarcastic, lol

[mossy canopy]

no clue why comfybox is so underrated..

[craggy ibex]

Are there any model for generate such items?

I can't

[hardy cipher]

ya got me, I have no idea

[twilit grove]

excuse me but how do i launch automatic1111 ?

[craggy ibex]

I did my best

[craggy ibex]

I even properly typed my prompt

[mossy canopy]

how come

[craggy ibex]

but the model is so fucking stupid

[hardy cipher]

try something out. just start adding whatever is in your head. it can be WAAAAY longer than that

[soft zealot]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-NVidia-GPUs

[shell steeple]

Does anyone know how to change the pose of a character image without changing his/her outlook and style? Tried on Automattic 1111 img2img and controlnet but it doesn't work. It changes completely the outlook

[hardy cipher]

certainly doesn't need to be super long, but you should try that, find a happy medium. short prompts leave the ai less to work with. so it improvizes more

[half ivy]

control net ref and lineart? new open pose?

[hardy cipher]

some people might argue that you don't need a long prompt, or that long prompts overwhelm it and cause it to do random things. but I'm a long prompter myself

[shell steeple]

From the first image I want to have it in the pose of sitting down like the second image. I provide the first one as input on img2img and in controlnet I upload the second image with open pose but it generates something super random

[hardy cipher]

you could probably do something like that with clipvision

[half ivy]

also not sure myself if kyle is sitting or standing, might have to preprocess a different image where sitting is clearer

[hardy cipher]

alright, I'm done with the controlnet sample workflows. nothing against them, but I must forge my own path

[half ivy]

or this

which i dont know

[soft bone]

openpose can't tell he's sitting. use a real photo

[craggy ibex]

I even used diffrent model

It's impossible

Nothing can generate a cool backpack

[hardy cipher]

https://tenor.com/view/emo-emo-to-the-extremo-extremo-gif-21200072

[craggy ibex]

I give up

You can call me that

[hardy cipher]

they're not mistakes or failures, they're just happy accidents

[hardy cipher]

what the heck is controllorasave for?

[half ivy]

just saw that recently, weird

[hardy cipher]

https://tenor.com/view/bob-ross-painting-yas-wisdom-art-gif-4514528

[half ivy]

ahh bob yea lol

[hardy cipher]

https://tenor.com/view/de-jwein-eczemavocht-ikbendejwein-retlaw-jweinross-gif-20070181

https://tenor.com/view/zl-gif-22933534

[shell steeple]

Oh, never heard about clipvision

[hardy cipher]

it's super cool actually. basically seems like unclip to me. but apparently it's better. unclipXL

uses images as prompts

well, you still send a prompt through, but it can be empty. not sure why it's even required

[half ivy]

control net reference would be my go to, with a new open pose

[hardy cipher]

yeah, that's probably better

[half ivy]

but its not SDXL so ewww

[hardy cipher]

lol

[shell steeple]

Can't see clip vision 🤦

[hardy cipher]

oooh, a1111

[half ivy]

join me in the frustration pit, gonna figure this SDXL thing out

[hardy cipher]

yeah, not sure how to use it with that

[craggy ibex]

I SAID

[half ivy]

share your control net settings

[craggy ibex]

GENERATE ME A FUCKING BAG

[craggy ibex]

ARE YOU STUPID?

[hardy cipher]

lolol

[half ivy]

https://tenor.com/view/telmo-coca-harina-raquetaso-esnifar-gif-25660568

[shell steeple]

Here

Not sure why it doesn't display any model....

[craggy ibex]

Why that girl look like she in some shitty condition

[half ivy]

model should be openspose

[craggy ibex]

not tryna be rude but

it really is

[hardy cipher]

you need to downoad models to display them

[half ivy]

why only 1 control net though?

should have 3 bruv

[shell steeple]

Oh... where do I find the models? Sorry if it is a dummy question

[shell steeple]

three pictures to get the pose?

[half ivy]

https://huggingface.co/lllyasviel/ControlNet

[half ivy]

three different control nets all at once... POWER!!!!!!!

[hardy cipher]

or here https://huggingface.co/stabilityai/control-lora/tree/main

the ones I posted are efficient, but no openpose in there

[ionic dragon]

isnt he Emad?

[craggy ibex]

You know what fuck this shit

Imma use this

https://civitai.com/models/121716/helloobjects

Nothing can generate backpack

[hardy cipher]

Emad?

download the backpack lora

[twilit grove]

why do i get thoses buggy pictures ?

[half ivy]

dont see reference though o0

[shell steeple]

Need to figure out how to install it. Have seen they point to a Github repo to install as an extension. When trying it says it is already but I can't select any model on ControlNet

[twilit grove]

i asked for a beach lmao pls someone help

[half ivy]

have to put them in yr models/ control net folder

[hardy cipher]

cfg 22?

[half ivy]

https://huggingface.co/lllyasviel/ControlNet/tree/main/models

[twilit grove]

?

sorry

[half ivy]

download and move to yr folder

[soft bone]

think about what you're asking it. anything about the image that you dont specify in the prompt, the model is free to interpret for itself (which is almost all of it in your case).

picture in your mind the image you want. what is surrounding the bag? empty space? a solid color background? then TELL it that, it can't guess. if it generates more stuff you dont want, put that stuff in the negative prompt. also think about other images that look like the one you're picturing. what do people call images that look like that? my guess is something like "game asset", "object design", "product photography", "single asset", etc. imagine youre trying to google image search for the image youre picturing. obviously "a bag" would get all kinds of unrelated crap. get specific.

"product photography of a blue backback, studio backdrop, plain background, game assett, object, centered"
neg:"anime, person, people, face, animal, action, artistic, abstract, colorful, detailed, portrait, outdoor, busy background, intricate"

and adjust as needed. If you dont get near what you want within 3 images then your prompt needs work. and use someones workflow that is proven to work, your settings could be the problem.

[shell steeple]

Ah, ok. I just have to download them and add them on modles/ControlNet foleder right?

[half ivy]

yessir

[craggy ibex]

I would rather download even more models to solve many peoblem

beside

There are no any models that can generate everything

[half ivy]

and then restart yr .bat file obvsies

[craggy ibex]

They are all but specific purpose model

Anime Model, Object Model

[soft bone]

so you dont want help then. just a defeatist. clearly everyone else can get what they want. do you want me to generate a plain blue backpack to show you its possible?

[soft zealot]

also remeber that it is as much about luck as prompt skill.

This isa harsh reality many people are unable to accept 🙂

[craggy ibex]

"All-Purpose" model doesn't exist

[hardy cipher]

I mean

[craggy ibex]

show me and I'll stop complain about this

but thx for advice btw

[hardy cipher]

try the base model to start

[craggy ibex]

I am just really angry

[hardy cipher]

are you saying the base model isn't all purpose?

[soft bone]

hmm idk. as i get better at prompting, i get the image im looking for in fewer and fewer tries. often i get exactly what i pictured in the very first gen

[ionic dragon]

yeah,, he looks like emad

[twilit grove]

i just downloaded sdxl and webui 1.5.1 and it's all buggy output

[heady vale]

1024x1024. 512 is too small

[twilit grove]

ok

[fathom hound]

Uhh 512x512

Kek

[craggy ibex]

You mean like these?

[half ivy]

look at extensions under settings and add openpose editor, think its there, lets you edit any pose's you make directly

[mossy canopy]

comfybox make it so easy to handle long complicated workflows..

[twilit grove]

i luve u tyyyyy

[hardy cipher]

ahh. I guess I'm out of the loop. not sure who that is

[tribal lantern]

like this?

[hardy cipher]

that's unpossible

[shell steeple]

You mean to look in extensionis on the Available tab for openpose editor? I can't find it

[half ivy]

🤔

[hardy cipher]

latent size?

[soft bone]

literally my very first and only output of the day

also the one above

[half ivy]

1024x1024

[craggy ibex]

Damn..

[half ivy]

https://github.com/huchenlei/sd-webui-openpose-editor

[hardy cipher]

try lowering cfg

[half ivy]

sorry maybe i was wrong

[half ivy]

kkkkk

[hardy cipher]

or, double check to make sure you're not using a 1.5 lora by accident

[half ivy]

no loras

[tribal lantern]

You're really strugglin with backpacks?? try 'product photography, blue backpack'

[hardy cipher]

sekaiza is a peculiar lad

[ionic dragon]

wait, 1.5 loras dont work with sdxl?
i ran some 1.5 loras and they were good

[hardy cipher]

I mean, they work. but they don't work correctly

[soft bone]

second output of the day. luck is not required

[hardy cipher]

I'd guess for particular specific aesthetics a 1.5 lora might turn out to be preferable. but that's not going to happen much

[craggy ibex]

I did my best but it keep generated a human

[hardy cipher]

look at that, fortnite backpack

[craggy ibex]

I used SDXL

base 1.0

Ngl I think this is passed

[half ivy]

bad seeds 😅

[soft bone]

literally put human into negative prompt bro

not bad seeds. i can get a plain backback 5000 generations in a row

[hardy cipher]

I've been telling him stuff like this for days

[craggy ibex]

I already put human in negative prompt bro

but I am very unlucky as hell

[hardy cipher]

ugh

[craggy ibex]

my seed is randomize everytime when I click "Queue Prompt"

Idk why

[soft bone]

download my image and put it into comfy. if you get humans, i will give you $100 cash

[half ivy]

twas a joke ser

wonder if its true though o0

[craggy ibex]

I'm serious, it is randomizing

[hardy cipher]

most people just don't post every single time they get an image they're not happy with

[glad grove]

steven seagal is not in a bikini , emergency meeting

[hardy cipher]

this woman has 6 toes, why am I even alive?

[craggy ibex]

Oh

[soft bone]

guys SDXL cant read my mind how do i sue?

[craggy ibex]

I thought you all are always lucky when you use Stable Diffusion

[half ivy]

run batches my guy, choose yr fav

[craggy ibex]

very very lucky

[hardy cipher]

perseverance, not luck

[half ivy]

might use 1 of every 20 generated, unless yr mack

[hardy cipher]

mack takes the stream of consciousness approach

dude, elon musk is the only person I put into the shitty fan art lora that didn't turn out shitty. and it really grinds my gears

I tried like 4 times

[soft zealot]

this dude has a backpack

[glad grove]

now we only need the Robert Zdar lora to make him company

[hardy cipher]

wtf is going on with elon?\

[soft zealot]

the prompt that produced it was

[shell steeple]

Can't figure out how from an img2img adding an image I jussst want to change the pose. All the time I try it generates a completely different character that has nothing to do with the original image. I just want to create the same character but with a different gesture. Is that even possible? Appreciate

[hardy cipher]

that's what I was saying earlier

[glad grove]

🙇‍♂️

[craggy ibex]

with helloobjects

hmm..

I wish I could make it fantasy style

[hardy cipher]

man, this whoopi goldberg one though. it makes me feel uncomfortable

[soft bone]

why are you telling us? tell the prompt that

[glad grove]

mouth too small for Whoopi imo

[hardy cipher]

she's looked the same for 40 years

[soft zealot]

wonder what ClipVISION isseeing in this combo It likes to add a camera?
Thats 2 in a row lol

[soft bone]

does revision not take a prompt input?

havent tried yet

[hardy cipher]

maybe your shirt?

[soft zealot]

Ignores the prompt input if you have the conditioningseroout in place

[soft zealot]

could be the seagull I suppose

[hardy cipher]

peak performance

[soft zealot]

#3 no camera lol

[hardy cipher]

you know, I've been using that conditioning zero node, but wasn't actually sure if it did that. just assumed it did

[peak dove]

Using Sytan's w/flow

[proper depot]

Does ComfyUI have simple switch nodes? It'd be nice if I could press a button and swap from using img2img to txt2img, or swap from using a LoRA to not. 🤔

[half ivy]

would like that too

[hardy cipher]

I don't think it'd be that hard to make a node that did that

if one doesn't exist

[woeful patio]

anyone have a comfyui workflow that uses the lora i downloaded from civit?

[half ivy]

not that i use my comfy anyways 😂

[hardy cipher]

https://tenor.com/view/jesus-take-wheel-lord-omg-cheers-gif-9932767

[woeful patio]

nani?

[half ivy]

what lora bro??

also i dont use loras so prolly shouldnt get involved

[soft zealot]

Ok So this is with the Prompt Left Blank and Zeroing enabled (ignore Stage 0 thats not in play the way I have my workdflow built)

[craggy ibex]

This is cool!

[hardy cipher]

nice

[soft zealot]

Ok so if I add a small amount of prompt it does seem to get added but.............

That is going to be at the Stage 2 step as ReVision is only applied to STage 1

[hardy cipher]

so would supporting terms be clip l?

I think you can also just leave the prompt boxes blank?

[soft zealot]

Yes but Text G is fed to Step 0 (precon) & Step 2 (refiner) in my implemetation

[soft zealot]

I did leabve them blank in the first example I posted

[hardy cipher]

very precise

[ionic dragon]

can someone share SDXL prompting cheat sheet

[hardy cipher]

I'm trying to understand the purpose of conditioning zero out having an input

[soft zealot]

is it the right way ? I don't know.

Do I care ?

No

lol

[ionic dragon]

someone share some 30-40 docs, few days ago

[hardy cipher]

https://aituts.com/sdxl-prompts/

[hardy cipher]

that's how I roll

[half ivy]

This is sick 🔥

[soft zealot]

cheers,

Its another ReVisioned generation 🙂

[hardy cipher]

step 1: send your clip vision image data through the negative conditioning hole and produce 2 images
step 2: take resulting images and use as negative conditioning with your original images now used as positive conditioning
step 3: success

[soft zealot]

if you say so 🙂

[hardy cipher]

I've never actually tried it. just wanted to see if anyone else would so I could see what happens

[ionic dragon]

not this, ig sunny or someone shared a docs link

[hardy cipher]

well, I guess you need to find the link

[soft zealot]

I do like the "realism" that I'm seeing in many of these ReVisioned generations

[half ivy]

Finally

https://www.reddit.com/r/StableDiffusion/comments/15wgbwr/new_soap_photo_enhancement_lora_bye_bye_blurry/?utm_source=share&utm_medium=ios_app&utm_name=ioscss&utm_content=1&utm_term=1

[soft zealot]

although thi swas an interesting one lol

[half ivy]

frustration nation

[ionic dragon]

lol, wtf

[soft zealot]

It also generated this, damn those hands are very nearly pefect !!

[hardy cipher]

what's your setup?

[ionic dragon]

dayyum

[hard fractal]

Let's see!

Without SOAP LoRA:

With SOAP:

[hardy cipher]

https://tenor.com/view/hmm-dot-dot-dot-stick-figure-intresting-gif-23376515

[half ivy]

so over the bokeh now

[hardy cipher]

I've never been about the bokeh

[half ivy]

this is slightly better but not quite the space ship i wanted with a full 10 CN

[soft zealot]

If I may be critical:

These are veering towards looking like an image of a person has been superimposed onto a background in The Gimp.

Somethings just a little off about the edging IMHO

[upbeat summit]

https://cdn.discordapp.com/attachments/1139433039169069077/1142922705398665336/masslevel-sdxl-20230820035014-1021388970548490-upscale-post-nm.png

[hard fractal]

i see what you're saying

[soft zealot]

dame I awas replying and you dleted lol

[half ivy]

now i cant unsee it LOL

[soft zealot]

sowwy

[hard fractal]

I didn't realize you were responding to the SOAP images.

[hard fractal]

I agree.

That was at 100% strength. I'm sure that you can dial it back and it'll look better.

[ionic dragon]

@hard fractal is it possible to train loras using inswapper?

[half ivy]

@hardy cipher this is my setup

[hard fractal]

doesn't that use CodeFormer?

If so, not us.

That requires a license.

[half ivy]

a control net that is yet to start controlling or netting decently

[hardy cipher]

ahh, lots of nodes I don't have currently

[soft zealot]

be interested to see 🙂

[hardy cipher]

dial it up. level 10

[half ivy]

apparently i cant math and that can be a thing 😂

[half ivy]

already full boss

[hardy cipher]

incorrect resolution?

[half ivy]

1024 on prepro after upscale and 1024 latent

[hardy cipher]

tis a mystery

[half ivy]

tis me but its getting better, cfg hint was a win

im out for a bit but will be praying to the diffusion gods for a light going forward so should be good

[hardy cipher]

you could also experiment with the tone map and cfg rescale here https://github.com/comfyanonymous/ComfyUI_experiments

[half ivy]

will check that out cheers

[hardy cipher]

https://tenor.com/view/cheers-gif-batman-drink-gif-25961951

[hard fractal]

ah, that SOAP isn't from us

was just testing it out for fun

[soft zealot]

I knew that 🙂

[hard fractal]

ah, got it

[soft zealot]

On the other hand (as you may have gathered) I'm really liking the whole ReVision thing and have incorporated it into my daiuly multistage workflow 🙂

[hard fractal]

❤️

[hard fractal]

glad you're using it!

what are the stages?

[hardy cipher]

yeah, the clip vision stuff is pretty interesting

[upbeat summit]

base SDXL

[soft zealot]

theyre like big wooden boxes puched together that peopl can perfom on inf ront of an audience

[hard fractal]

https://tenor.com/view/enter-the-stage-stephen-gilchrist-glover-steveo-hello-everybody-energetic-gif-25951149

[upbeat summit]

base SDXL model
Beautiful Woman with jewels build with ferrofluid 200 ghz translucent

• negative

that's it!

[hardy cipher]

that's a pretty eclectic prompt

[upbeat summit]

you have no idea - this is super powerful

[soft zealot]

@hard fractal
Stage 0 (or Pre COnditioning 3 steps using Model 2 (refiner)) is being ignored when I enable ReVision
Stage 1 (or Base using Model 1) is being fed from the Revision Process
Stage 2 (or Refiner using Model 2) is fed direct from Prompting without going via ReVision Process (so allowas little bit of style to be applied Post ReVision)

[hardy cipher]

whoa. I hadn't thought about using ferrofluid in the prompt

[upbeat summit]

you just need to be creative with the elements, textures etc

opened up a whole new latent space area for me 😄

[glad grove]

someone here should try fractal maybe it gives goods results like in 1.5

[twilit grove]

what's the best sampling method for sdxl 1.0 ?

[soft zealot]

@hard fractal
I just think "Multi Stage Prompting" is a more descriptive and flexible term than "Base+Refiner" As I regeularly now use Different mOdels in Different stages to achive different effects, almost like on "on the fly poor mans Model Merge"

[hardy cipher]

lol. hmm

might need to fine tune that one a bit

[hardy cipher]

that's awesome

[upbeat summit]

it's organic alright 🙂

[proper depot]

Do I need anything special to use a Lycoris with SDXL and/or Comfy? 🤔

[twilit grove]

what's the best sampling method for sdxl 1.0 ?

[proper depot]

I mostly hear good things about DPM++ 2m

[hardy cipher]

choose the one with the most words

and if there's a tie, use the one with more letters or higher numbers

[twilit grove]

ok ok ty

[upbeat summit]

I'm using dpmpp_2m karras but right now I'm on the new dpmpp_3m_sde_gpu karras, 55 steps, cfg 4

[proper depot]

Or, on a related note, are there any issues using Loras from SD 1.5?

[proper depot]

55 steps? Damn

[twilit grove]

and how does it takes for 1 pic ?

[upbeat summit]

it depends on your hardware of course

[twilit grove]

i have an r5 5600x and a rtx 3060 12GB

[upbeat summit]

I can't really say. could be 28-35sec? also depending of course on the resolution you will make your images in. dpmpp_2m is really fast

[wet nacelle]

They're coming to rob ya

[twilit grove]

with your parameter it took me 45sec

[upbeat summit]

yeah sounds reasonable. not the fastest but I think the output fidelity is quite good

[wet nacelle]

okay. Do you guys remember when civit had the Creation tool for like an hour?

[twilit grove]

yea thx

[upbeat summit]

also try dpmpp_2m. it might be faster for you. I ran it at 70 steps, but it does also make great images with lower steps

[twilit grove]

ok ok

[wet nacelle]

@upbeat summit do you?

[upbeat summit]

If you don't already have and use ComfyUI you should try and set-up AITemplate with your GPU: https://github.com/FizzleDorf/AIT. It can give you a 50 - 100% speed boost when generating images.

[upbeat summit]

I heard that something was there, but I missed it. But I would think they are probably cooking something up for some time - they will launch an image making service at some point

[wet nacelle]

It was such a chaotic launch too. Too bad that I never got to actually see my outputs.

[upbeat summit]

when was that?

[wet nacelle]

Like two to three days ago

[pine nimbus]

can someone walk me through on how to disable nsfw filters? or whatever else makes my images look black almost all the time

whoops nvm

[twilit grove]

ok

[craggy ibex]

For DND

[hard fractal]

Testing out some controlnets and need some prompts!

[half ivy]

How every read me should start

[hard fractal]

Here's detailed vector art of a beautiful african american woman in times square

With the smallest possible legend on the side.

A photographic prompt here:

[half ivy]

Ubiq Ninja sitting on roof, cyberpunk, futuristic, neon lights, cityscape, emerald green and black

[hard fractal]

You got it.

Got an init?

[half ivy]

Negative

[hard fractal]

There here it is

[half ivy]

Hmm, Maybe add “a close up shot of” to the front

[hard fractal]

oh, swarmui has been unleashed already

Ubiq Ninja sitting on roof, cyberpunk, futuristic, neon lights, cityscape, emerald green and black

I did use a different init.

[half ivy]

Hold up

[hard fractal]

Holding.

[half ivy]

Will delete this next message 🫶

[hard fractal]

👀

Yeah, overall... finding that the smallest controlnet HuggingFace released is the best one.

(and that they're all not as good as the ones from Stability)

💅

Wonder why that would be...

Maybe the big ones were overtrained.

[uncut fiber]

played with graphs in comfyUI, and more confused than when i started 😄

[polar epoch]

Huh, post a image of the graphs, kinda intrigued :P

[polar epoch]

Huh, there's SDXL models for controlnet finally?

[uncut fiber]

very easy graph, just used SDXL text prompts with 2 resolutions + latent resolution and somehow puzzled

[polar epoch]

Ah, so you mean workflow, thought there was a node with a graph you tuned outcome as if it was a EQ for audio

[hardy cipher]

bernix needs a couple 10 stacks in his workflow

[polar epoch]

@uncut fiberThis is my current workflow. Base sample image, then latent upscale, resample, then to refiner, then to upscale image with model

The result with a pikachu that went a little too much "pet" lol

[uncut fiber]

@polar epoch looks nice. i have to do some more complicated workflow as well.

[polar epoch]

And as confusing as comfyUI can get, i love it how it can scale infinitely, but what i need the most is a rerout module that has a on/off switch that disables an entire group of nodes. Like upscaler all in one, upscale resampler all in one and so on

[uncut fiber]

Would love if it has CTRL+Z like undo and redo

[polar epoch]

Though i need to learn how these properly works, as it seems to split up every small bit of process the sdxl generates at

[polar epoch]

Indeed

[proper depot]

Real quick question: can anyone tell me the difference between these controlnet files? 🤔

[polar epoch]

Just remembered a crucial function i'd want in comfy, and that's a ksampler that shows preview of the image i'm making, so that i can cancel if it's too bad and not have to wait like 40 sec if i use 100 steps at 1024 or larger lol

[hardy cipher]

top one says 16 at the end

[polar epoch]

Nice! Only off putting that i noticed right away, was this scribble, otherwise beautiful!

[uncut fiber]

@polar epoch if you have manager you can enable preview here, but you think for sure something different

[hardy cipher]

clip vision

[proper depot]

It's also half the size

[hardy cipher]

took two images, sent them through the negative and created two other images with zero prompt conditioning as far as words go. then used the first two images as a negative clip vision prompt

so basically, created the opposite of what I wanted?

[polar epoch]

There was 3 of those. Encode, loader, and input switch

[hardy cipher]

then used that as my negative

using those two images as input

which are controlnet results from blueprints

blueprints I made with clip vision from other blueprints, lol

https://tenor.com/view/alecpeters-axanar-deeper-gif-18041081

[polar epoch]

Oh nice!

[shy kelp]

idk what to prompt

idea?

[hardy cipher]

ask the ai

[primal vault]

They should be straight forward, the top 2 nodes and the bottom node on that screenshot are probably most useful outside of my own workflow, the other nodes are mostly supporting them

[polar epoch]

Optimus prime riding jormungandr to battle against kaiju at zeus's palace, rest of the prompts depending on how you want the image to look. I usually just look for images on civitai with prompts to achieve what i want.

[hoary saddle]

what's this mean? first time i've seen it in console

ah only happens when using the control canny bin file, guess i can delete that now

[shy kelp]

lol

all combined

[half ivy]

Salute 🖖

[half ivy]

eta would be fancy 💅

[polar epoch]

Or do a hybrid, "combine jormungandir with optimus prime" or "jormungandir mechanical transformer". Want wackier? Crank up CFG, and upp steps for better quality

[woeful patio]

is there any news on when controlnet for sdxl is coming to a1111, comfy just doesnt work for me

[shy kelp]

with higher cfg

[half ivy]

I used to be like you, its true ser, let us move on from the glory days, for better fortunes follow if we can push through the stress of not generating forever while we learn

[uncut fiber]

it seems i am obsesed by weddings 😄

[half ivy]

Also Auto with some SDXL buttons would be dope af

[half ivy]

Less bokeh 🧐

[mossy canopy]

more bokeh

[mossy canopy]

that reminds me sd2.0 was good for one thing .. doing realistic cat images

[mossy canopy]

cnet testing?

[soft zealot]

nope thease are all ReVision generations using

Just cherry picking ones that mach a sort of theme

[fair stag]

Not seeing any commits related to that yet on the dev branch. https://github.com/AUTOMATIC1111/stable-diffusion-webui/commits/dev

[soft zealot]

seem to be either coming through like that or like these

[half ivy]

Wonder what ash would look like

Irl

[mossy canopy]

civitdown?

[crisp owl]

seems so

[mossy canopy]

itsback

[crisp owl]

or just super slow. Refreshing mine still load wheels

[mossy canopy]

giveme a prompt that wont generate bokeh in sdxl

[fair stag]

Someone get this man an extra helping of bokeh

[vale eagle]

#✨｜sdxl message

[crisp owl]

Why so much hate on bokeh though?

[mossy canopy]

why is it so large??

[mossy canopy]

that lora size

nobokeh

[uncut fiber]

cats are nice, that one on last one has nice cute paws 🙂 Can you tell me, if it is noticeable on first sight it is latent upscaled?

[mossy canopy]

what is odd in this pic

[half ivy]

Its all you get when signed up for the average deal

Should be requested imho

Not forced

[half ivy]

Can we marry them, so cute awww ☺️