✨｜sdxl

[upper fog]

/d@cyan crown

[glad lily]

You mean you want to remove the image feed panel that collapses into img icon?

It comes with custom node called Custom Scripts

[native knot]

Leaving the randomizer I built to run last night was great. I did find a few things that needed to be removed as well as I've found it's a bit heavy on a few concepts...Superman, in particular, came up quite a lot. But here are some great examples of what it put out.

[crisp owl]

Spiderman Deer ❤️

[native knot]

If you're interested in seeing all the other images plus other things ones I've intentionally created, I've put them all up for viewing:
https://lychee.soulctcher.net/

[vital ermine]

Not sure I have that as I am currently in Linux so can't check but I will look when I switch over. I wonder why they never gave an option to remove that/disable that? I mean I will gladly remove their entire nodes just to get rid of that one thing.

you know, I am in Linux and it blows me away at how the same fucking EVERYTHING as in windows I go from 1.26s/it to 1.38it/s

for training a Dreambooth

almost half the time

[hoary saddle]

was just looking at this on civi like 2 minutes ago, wierd

does it do well with bearded cats? lol

[native knot]

There's a bearded cat for ya

[hoary saddle]

using shaun?

[native knot]

Just a prompt and a model.

[cyan crown]

Is it my Lora?

[native knot]

No.

No LoRAs.

[hoary saddle]

trying to get consistent grumpy bearded cats, pixel lora does ok but not great

[native knot]

Here's a few more with the same prompt and model, just different seeds.

[hoary saddle]

extreme closeup headshot of a (((grumpy))) white (((old kitty cat character))) with whiskers and a ((massive beard)) (yelling at the camera:1.2) in the voxel art style voxel lora sucks for it

[glad grove]

too many words,remove headshot ,old kitty cat character,yelling at the camera

[vital ermine]

This one?

[vital ermine]

Nope, that wasn't it. I do not have that node installed

[cyan crown]

it's not a Lora but is it sdxl standard model?

[dapper flame]

if i just installed Stable Diffusion A1111 via https://github.com/AUTOMATIC1111/stable-diffusion-webui.git is that sdxl? Where do I look to see what version

[native knot]

I'm using the JuggernautXL model for those.

[glad grove]

this guy looks devious af

[hoary saddle]

seems to be more about the prompt then lol

[cyan crown]

short beard here

[native knot]

P:

 texture, intricately detailed, theatrical lighting, amazing, highly detailed, sharp focus, trending, 8k, 35mm

(3D voxel style:1.2), pixel style, dynamic```

**N**:
```(worst quality, low quality, normal quality, lowres, bad, oversaturated, undersaturated, overexposed, underexposed, grayscale, bw:1.4), (watermark, autograph, trademark:1.2) (blur, blurry, grainy), morbid, collapsed, cropped, missing, deformed, imperfect, conjoined, deformed, ugly, imperfect, skewed, unnatural, error, mutated, out of frame, cut off, censored, jpeg artifacts, glitch, deformities```

How about papercraft style instead?

[ionic dragon]

How to prompt for these kinda images?

[glad grove]

just dont put anything in prompts and u get that

[hoary saddle]

that's an idea, haven't tried that

[native knot]

Blue_Jizz on Canvas
Anonymous, 2023

$3,050

🤣

[cyan crown]

going to sleep. Good night

[native knot]

Marble sculpture version:

[hoary saddle]

kinda got what i'm after using your prompt idea

[native knot]

Nice...

[hoary saddle]

appreciate ya

[native knot]

Change up the resolution, too.

Like...go for an ultrawide like 1536x640.

[hoary saddle]

thanks!

ugh, damn santa got me 😦

ah, added in some controlnet

[dim steeple]

Does SDXL work with controlnet?

[crisp owl]

yes

[native knot]

aYE

[dim steeple]

Are their XL controtnet models to download?

[native knot]

Aye

[dim steeple]

Or can I use the 1.5 ones

[native knot]

1.5 not compatible

[crisp owl]

https://huggingface.co/stabilityai/control-lora/tree/main/control-LoRAs-rank256

[dim steeple]

So I just put that in the same place as my other controlnet models?

[crisp owl]

Yup

https://huggingface.co/thibaud/controlnet-openpose-sdxl-1.0/tree/main
here's an openpose also

[dim steeple]

cool thank you

I am getting this error:

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

[crisp owl]

Looks like you're in A1111, I know they recently got it working there, but I'm not using that as I've switched to Comfy.
So I cannot verify what works vs what doesn't in A1111 at the moment

[dim steeple]

darn

[crisp owl]

I'd pose the question again in a bit and specify it's for A1111, see if someone else can chime in regarding

[dim steeple]

I also am getting this error in text2img:

RuntimeError: The size of tensor a (64) must match the size of tensor b (96) at non-singleton dimension 3

[glad lily]

No. This one with image feed.
https://github.com/pythongosssss/ComfyUI-Custom-Scripts

[hardy cipher]

What models and vae are you using? That's a size mismatch of some kind

[dim steeple]

that error went away after I restarted auto1111

[hardy cipher]

Ahh

[vital ermine]

That was one of the ones I had disabled too.

[vital ermine]

Still does it

[hardy cipher]

Embosssed

[hardy cipher]

Smooth

[nimble heart]

why are they shrinkwrapped

[hardy cipher]

Thiccc emboss

[glad lily]

Try to uninstall it completely.

[dim steeple]

Do LoRas for 1.5 work for SDXL

[uncut fiber]

no shouldnt

[uncut fiber]

just added evening

[dim steeple]

they won't work?

[uncut fiber]

yes. Afaik, so my answer is basicaly worthless. But i think i read it several times here in channels.

[native knot]

No, they will not work.

[glad grove]

they wont work

[uncut fiber]

for example at LORA with Geishin impact

[upbeat summit]

technically you can send SD 1.5 fine-tunings to the CLIP_L encoder of SDXL, but you will not get the results you are looking for (at least most of time) since all the fine-tunings have been trained on a total different latent space

[dim steeple]

Btw how many images does one need to train a good Lora?

also can I ask what the number parameter does? lora:SDXL-Caricaturized-Lora:1

can it be anywhere from .1-1?

It increases the strength of the lora?

[upbeat summit]

yes - 1.0 is the LoRA at 100%

[dim steeple]

Do I need to add the trigger word outside this part?

or does havinglora:SDXL-Caricaturized-Lora:1 do it

[upbeat summit]

it depends on the LoRA

[dim steeple]

https://civitai.com/models/140430/sdxl-caricaturized-lora

[upbeat summit]

read what it says under the LoRA title 🙂

[dim steeple]

so adding Caricaturized is the trigger word

[upbeat summit]

correct

[dim steeple]

So I need to add Caricaturized and lora:SDXL-Caricaturized-Lora:1?

[upbeat summit]

yes. it makes sense to use trigger words somewhere close to the beginning of your prompt

[dim steeple]

so like this?

Caricaturized donald trump in a suit and tie, full color, cartoon character, lora:SDXL-Caricaturized-Lora:1

or

lora:SDXL-Caricaturized-Lora:1 Caricaturized donald trump in a suit and tie, full color, cartoon character

[upbeat summit]

yes, that's both fine. the placement of the <lora> tag is not important. it will be stripped before the rest of your prompt is being processed

[dim steeple]

okay thanks for the help

Appreciate it

[upbeat summit]

you're welcome! what software do you use to create images? the <lora> syntax does work in automatic1111 but not natively in ComfyUI

[dim steeple]

I use auto

[upbeat summit]

than have fun 🙂

[ionic gulch]

new version 6.2 of my workflow - reduced required 3rd party nodes, cleanup, some smaller fixes and changes
https://github.com/JPS-GER/JPS-ComfyUI-Workflows
https://i.imgur.com/XI4Adpw.jpg

[vital ermine]

That is my next attempt, but it does show it is a naughty node since it is not adhering to me saying disable. Hopefully it works BUT it is one of my most used nodes due to loras having an on/off that is not in the native lora node.

[hardy cipher]

a lora node with an on off switch?

[worthy orbit]

did you img2img @hardy cipher or what 😆

[vital ermine]

yeah, but is a drop down. I did see one with a real switch on it but not sure which one that was

[hardy cipher]

I made lora nodes with switches

[vital ermine]

I don't make nodes it ain't my bag

[heady vale]

comfyroll nodes have a few lora on/off nodes

[vital ermine]

yep

problem is that is the one that I am forced to remove

no way to turn off that queue thing

[hardy cipher]

here's the deluxe version

[vital ermine]

That apparently comes from comfyroll

[hardy cipher]

the one I normally use loads 3

10 is a bit much

I don't know. didn't find any that fit all the criteria. multiple loras in a single node, switches, and normal inputs/outputs

[hardy cipher]

grinds my gears when I find a node that does most of what I want then has weird outputs that don't work with anything else

[hardy cipher]

they look a bit soggy

[glad grove]

she looks greasy

[vital ermine]

Always did

[native knot]

she looks is greasy
FTFY

[native knot]

I just got done w/some workflow adjustments that allowed me to do something like this:

[hardy cipher]

https://tenor.com/view/beauty-painting-beauty-is-everywhere-gif-13407207

[native knot]

All I had to feed it was this:

This is awesome.

[hardy cipher]

oh yeah, your automated system

[native knot]

Did you see the link I posted, @hardy cipher?

[heady vale]

garlic teeth

[hardy cipher]

don't think so

[native knot]

I made a new server to host all my diffusions. The site is here:
https://lychee.soulctcher.net/

If you click on Random Diffusions, that's the output from what ran last night.

[hardy cipher]

ahh yes. that template is quite versatile

[hardy cipher]

man, that's awesome

[native knot]

I plan on running more soon...just am working on stuff for the moment.

[worthy orbit]

yes it is quite fun, always wanted to do cutaway style, mostly buildings

https://civitai.com/models/69211/isometric-cutaway

something like this https://youtu.be/d1RnYfFZK2k?si=S69hS0ozaZ-XNF6y

[hardy cipher]

they're all quality outputs too. no fails in the bunch

I'm all about creating things like that. just haven't really put anything together yet

[native knot]

I found a couple, but they're not terrible and not quite obvious. But yeah, it really pumped out a nice batch. I did have to pull a few nsfw ones. Not that I don't like them, but I've got to keep it family friendly enough.

[wise rampart]

nice

[hardy cipher]

well they can't all be 10s. sometimes you just get a seed that doesn't vibe with the input

[native knot]

Exactly. But considering what ran, it's like 99% good.

But there are some real gems in there... loved the computer monitor setup that displayed an anime wallpaper.

[hardy cipher]

yes. I like these sorts of things.took a few iterations of things to get them where I wanted. started with regular blueprints.

[native knot]

Those remind me of the Metroid ones I did... very nice.

[hardy cipher]

man, I had to keep some of them to myself. they had very phallic appendages, lol. was not going for that at all

and not normal ones either. weird stuff

[native knot]

Pretty funny what this stuff can do sometimes, that's for sure.

Hey...I'm going to create a new version of that ALBUMS picture above...give me 2 totally different styles...one for the background and one for the text itself.

[hardy cipher]

I'm trying to see what sort of things I can mix with the ipadapter. some things definitely don't want to work together

hmm

8 bit background with calligraphy?

I don't know about these things

[native knot]

The font on the image is set...just need some other wild thing to fill it in with.

[hardy cipher]

don't trust em. I know that much

[native knot]

I can do an 8-bit style for the background, though...

[hardy cipher]

8 bit with ussr propaganda style fill

[native knot]

Ok...what should I fill the text in with?

make it different

[hardy cipher]

hmm. ever try pasting lyrics in? you should paste some dr. octagon lyrics in for the text fill

[native knot]

This is the latest one it just rendered. I'll go grab something from Dr. Octagon to kick off the next version.

[hardy cipher]

blue flowers or something. it's nonsensical, but I've made cool things with that sort of approach

[native knot]

Can't wait to see how this ends up.

[hardy cipher]

a bit too deep fried

[native knot]

That's wild.

[worthy orbit]

Very cool, I''ve done some aliens myself today

[hardy cipher]

ahh, going isometric with it. nice

[native knot]

That looks cool.

Good sense of scale.

[hardy cipher]

some of these could be a bit more refined, but I do like the aesthetic.

[native knot]

Stuff like that works really well in tall aspect ratios, too. Like...rerun that, but set it to 768x1344.

[hardy cipher]

that's a good idea, might do that in a couple minutes

cursed

that face in the bottom left

[native knot]

ha...yup

[hardy cipher]

demon

[hardy cipher]

you know, that's quite clean

going weird places

[native knot]

Definitely went for that 8-bit propaganda look

[worthy orbit]

but did you consult jesus?

[heady vale]

Eggplant Man

[hardy cipher]

need to unblurrify this one

alright, I'm going to reuse these. they're almost on point

details are just a bit goofed

[graceful egret]

Is there an inpaint model for sdxl?

[glad grove]

where outpaint tho

[hardy cipher]

the future

[glad grove]

dw openpose where

[worthy orbit]

not yet afaik

[worthy orbit]

I did some inpainting with the SD XL models, it kind of work for changing some details

[glad grove]

bro boarded the wrong plane

[worthy orbit]

Steven doesn't have much chance or

[hardy cipher]

I'm just going to run them back through ipadapter

[crisp owl]

https://huggingface.co/thibaud/controlnet-openpose-sdxl-1.0/tree/main

[hardy cipher]

using this image currently.

those pants

ridin high in the backside, bud

[glad grove]

that aint DW opepose thats just openpose v2🤓 🤓

[vital ermine]

There we go

[crisp owl]

I guess I'm missing context of what DW is 😆

[hardy cipher]

extra deep friend orange man

[glad grove]

a more accurate preprocessor than "openpose_full"

[crisp owl]

aah gotcha

[hardy cipher]

nope

I could make anything, and I choose to make things like this

[vital ermine]

Man, how does someone figure out what epoch is the best one?

[hardy cipher]

well you can save state after each epoch right?

[vital ermine]

Yeah, I have 7 and they all look good except epoch 1

epoch 1 is underdone

[hardy cipher]

there are all sorts of ways to gauge their progress as far as data is concerned

but ultimately I think it's just a judgement call

[vital ermine]

tensorboard is 100% worthless and I ceased to use it after 1.5

[hardy cipher]

beautiful

[glad grove]

he looks squeaky clean

[upper fog]

¯_(ツ)_/¯

[glad grove]

shiny nipples 🤣

[vital ermine]

LOL

[craggy ibex]

What are criteria or recipe to generate excellent images?

Let me guess...

1.Checkpoint
2.Vae
3.Lora (optional)
4.Image Resolution
5.Prompts
Am I right?

[glad grove]

no, prompt is always first if your prompt game sucks your image will suck

[crisp owl]

Prompt #1 for sure.

[craggy ibex]

What

This is nonsense isn't it??

[crisp owl]

Why is that surprising in a text2image product?

[craggy ibex]

Checkpoint is like the beginning of everything yes?

[hardy cipher]

you wouldn't believe the prompt I used for this, guys

complete trash

[glad grove]

steven seagal vs godzilla?

[hardy cipher]

yes

[crisp owl]

Checkpoint I use at stylizing mostly. But just because I use a specific checkpoint does NOT mean my images are going to look amazing

[craggy ibex]

Are there checkpoint even past SD version focus on generate anything?

instead of specific thing

[crisp owl]

huh?

[craggy ibex]

For example

GenerateAnything.ckpt or whatever

[glad grove]

yes but specific style checkpoints are better than all in ones

[crisp owl]

Oh, yeah, what he said.

[craggy ibex]

I wanna try all in one cuz I want to reserve my storage drive lel

[crisp owl]

I mean really, SDXL is pretty damn great at holding it's weight in a ton of different subjects without needing to change checkpoints much

Storage is cheap in the grand scheme of things.
But i get it, money spent is money spent.

[glad grove]

taylor doesnt understand what sensei is talkin 🙏

[vital ermine]

Hahahaha

[glad grove]

he got them shiny nipples again 🤣

[hardy cipher]

schnips

man, these images are unsettling

[native knot]

They all just look like overbaked images to me.

[vital ermine]

get new eyes, a better screen, or get off that mobile dinky screen

[native knot]

Nah...they're not good.

I'm not on mobile, I'm sitting in front of a 43" 4k screen.

[vital ermine]

I can't help blindness

[native knot]

If I was blind, I wouldn't be able to say with confidence that this is an image of Monica Lewinski.

But seriously, all that effect does besides make all those images look terrible is maybe make them look like they've been vacuum bagged.

[crisp owl]

Finally put together completely my main workflow. Control switches operate Controlnet/Revision/Ipadapter/etc, and combinations of them together.
Organized by groups to allow for group bypassing to save on unnecessary models/nodes loading in taking up VRAM.
https://civitai.com/models/142535?modelVersionId=158068

[hardy cipher]

that gun is no joke

[heady vale]

thats a cannon

[hardy cipher]

mall cop seagal has some serious forearm girth though. he can handle it

getting weird

[craggy ibex]

Is it possible to prompt a watermark for prevent stealing?

[vital ermine]

https://civitai.com/models/142547?modelVersionId=158085

[crisp owl]

You yourself, not really.
There is a treering watermark tech that can be implemented (to my understanding) by hosting sites for proprietary models

https://arxiv.org/pdf/2305.20030.pdf

[hardy cipher]

there's a way to watermark the images. but I believe that's only to say that it's created with stable diffusion or ai. but don't think it's customizable for the individual

[dim steeple]

Is it possible to use more than 1 img during img2img?

[crisp owl]

probably more of an IPAdapter thing

[heady vale]

https://civitai.com/models/138381?modelVersionId=152897 Try this it works really well

[hardy cipher]

chonker hands

playing with the conditioning pooled output on the last image gave me this

[glad grove]

he looks devious 🤣

[hardy cipher]

he really does

looking intense now

that tuft of chest hair really pulls it together

it's a crazy balance of numbers though. change anything and it breaks, lol

[glad grove]

he looks like an 80s sexy macho badass

[vast galleon]

Finally got the RGB and Alpha Channels all sorted out.

Tomorrow I will work on brush softness and opacity.

[hardy cipher]

pushing the limits with latent Seagal

curious to see if it even works with other seed numbers. 0-1 so far

lol

that image in the back really brings it together

this is the only other one that even rendered. rare Steven

I don't even know if it looks like him anymore

[clever verge]

Anyone know a way to randomize the seed client side without having to do a gen in ComfyUI?

[hardy cipher]

hmm? you mean to cycle through them?

monster hippo seagal looks like he took some front tooth damage

[glad grove]

hippo looks like he had a mask with eye cutouts 🤣

[clever verge]

Seed changes after the generation, it's annoying if you are disabling parts of your flow for speed. You then have to check the metadata of the generated image you like to use that same seed and run the full workflow. If there was a way to change it before the gen or a button that randomizes the seed client side would fix it.

[hardy cipher]

this guy

[glad grove]

https://tenor.com/view/c00kie-get-banned-gif-24947485

[hardy cipher]

what are you using?

[glad grove]

https://tenor.com/view/c00kie-get-banned-gif-24947485

[crisp owl]

hmm they don't allow us to tag a moderator, that's a mistake

[hardy cipher]

you know, the internet has plenty of porn. don't really need to see it here

pretty sweet, dude

[crisp owl]

Must live an exciting life

[clever verge]

Using what? Comfy?

[uncut fiber]

there were some ppl talking to ping them

[glad grove]

https://tenor.com/view/embed-no-no-embed-megamind-megamind-meme-gif-25261934

[clever verge]

Report and block

[uncut fiber]

@smoky patrol

[clever verge]

There seem to be more than me that would like some options controlling the seed randomization. https://github.com/comfyanonymous/ComfyUI/issues/1084

[hardy cipher]

https://tenor.com/view/done-so-done-im-done-monkey-throw-gif-18638117

[hardy cipher]

oh, that wouldn't be too hard to implement

[clever verge]

Never fiddled with the javascripts in Comfy but yeah, should be doable if you know what you are doing. 😄

[craggy ibex]

Does prompt with Color Hex work?

like

#66fdf9 hair color, short hair

uh

[hardy cipher]

well I believe it's python, but close enough. easiest way would probably be to create an external seed randomizer and have it keep a log or something. or create an image loader node that displays the seed number

or I guess you could modify the sampler

wouldn't be that hard to add it to a sampler, but then it'd just be in that one sampler

[ionic dragon]

how to denoise an existing image?

[uncut fiber]

i thought i got magic prompt for architecture, but when removing black magic part, it is even better 😄

[clever verge]

I think it's somewhere else than in the sampler. I'd have a look and see where it's done now.

[hardy cipher]

how do you mean?

[uncut fiber]

@ionic dragon denoise like in photoshop? Or some latent you mean?

[hardy cipher]

the seed number? it's right there in the sampler unless you manually changed it

[ionic dragon]

for example, i have this image, i want to denoise(0.3) this image and generate a new image

[clever verge]

seed generation in the code

[ionic dragon]

latent

[uncut fiber]

i think it is not posible.

not sure

[clever verge]

Sampler just takes seed as a argument to the function so it's generated somewhere else

[vocal stream]

Which CN model do you use for tile upscale? The preprocessor is showing up but I dont see a tile model

[hardy cipher]

def common_ksampler(model, seed, steps, cfg, sampler_name, scheduler, positive, negative, latent, denoise=1.0, disable_noise=False, start_step=None, last_step=None, force_full_denoise=False):

[ionic dragon]

@clever verge can we do something like this?

[hardy cipher]

you just set denoise to 0.3 and do it

[uncut fiber]

in img2img

[ionic dragon]

oh ok

can someone link to any img2img woorkflow

[hardy cipher]

just run the image through a vae encoder

[ionic dragon]

@smoky patrol he is posting NSFW stuff, and using abusive words

[hardy cipher]

then it turn it into a latent

I blocked him a while ago. no more lame porn gifs

[vocal stream]

too bad discord still doesn't do blocking of people properly

[glad grove]

@worldly temple

[vocal stream]

I don't need to see the mentally ill's constant '1 blocked message Show message'

[hardy cipher]

indeed

[clever verge]

Could be in widgets.js:
`function seedWidget(node, inputName, inputData, app) {
const seed = ComfyWidgets.INT(node, inputName, inputData, app);
const seedControl = addValueControlWidget(node, seed.widget, "randomize");

seed.widget.linkedWidgets = [seedControl];
return seed;

}`

[hardy cipher]

they should also freeze their account for a while if they get several reports and blocks onb one person

[hardy cipher]

well why you need to mess with all that?

[clever verge]

I'm looking for good gens with less of my workflow enabled (upscaling and other stuff muted) to find a good gen that I then want to upscale and do other fixes on.

What I have to do now is to set seed to fixed and look for the seed in the generated image and paste that back into comfy and enable the full workflow.

[ionic dragon]

@hardy cipher like this?

[hardy cipher]

okay, but the seed number is used by the sampler, so no matter what the sampler knows the seed number. and it could easily be modified to log them all if you wanted to

[hardy cipher]

looks right to me. it'll tell you if something is wrong when you run it

[clever verge]

Like this

[hardy cipher]

oh, well you have an atypical setup there.

but it brings us back to the things I proposed originally. an external random number/seed generator that logged previous seeds

or an image loader that showed you the corresponding seed, or even had it as an output

[clever verge]

Nah, just want to be able to hit a button that updates the seed client side. Like automatic1111.

[hardy cipher]

well good luck with that

[clever verge]

Or if the seed is updated just after you hit generate could probably work too.

[ionic dragon]

@hardy cipher whats wrong in this?
the input image and the ouput images are the same

[hardy cipher]

increase denoising

[ionic dragon]

ok

[hardy cipher]

and that's not the same image you posted earlier. looks like it changed to me

[ionic dragon]

no this is a different base image, this is the base image used

[ionic dragon]

should I also decrease the cfg?

[hardy cipher]

if you want to I guess. see what happens

[ionic dragon]

ok

[hardy cipher]

show me an output image

[crisp owl]

If you wait a few minutes, I'll be uploading a quick update to it. Just tidied up some things.
https://civitai.com/models/123048

[hardy cipher]

you have denoise set at 0.2. increase it if you want to see more change

[clever verge]

Think it's working.

[hardy cipher]

oh, a random seed generator?

[glad grove]

@uncut steeple

[ionic dragon]

I changed it to 0.4, and got minimal change

[hardy cipher]

well then change it more

[uncut fiber]

you can try 0.6 what was default somewhere, or even 0,75

[hardy cipher]

you got that spicy cfg as well

13 is a bit hard to manage without some other things going on

also, you have an entire huge workflow right above yours

[ionic dragon]

0.7

[crisp owl]

just pushed the updated version

[ionic dragon]

yeah, just muted thatt

[hardy cipher]

might have to give it direction or something. and that cfg should be lowered unless you really want your image cooked

[ionic dragon]

cool tq

[crisp owl]

Civit flagged this for review

[glad grove]

🤣

[hardy cipher]

one other thing. you probably want to cut down that image size

[glad grove]

speaking of first picture id like to see a fetus lora for some horror experiments but looks like theres none

[supple knot]

https://www.youtube.com/watch?v=gQPN3UKQM-U

[crisp owl]

Wonder if they'd allow it?

[glad grove]

prob not 😔

[hardy cipher]

so they can have the most degenerate sexual depravity imaginable and that's all fine. but a shiny skull crosses the line

[crisp owl]

hahah ikr

[ionic dragon]

because sd isnt good at 2048x2048?

[ionic dragon]

damn, what are your settings

[hardy cipher]

well because if you don't use 1024x1024 it won't render correctly. well it just has to be 1 megapixel

if you want to use any of my settings just load any of these images

I just added a couple things to your workflow

[hardy cipher]

using some models you might not have, but you could download almost everything from the manager

[crisp owl]

That's a ui error, what'd you do?

[ionic dragon]

is there any node, which downscales the res?

[ionic dragon]

i didnt get you

[crisp owl]

Usually errors like that are because of a backend issue

[hardy cipher]

I added it

[ionic dragon]

ig its not fetchng

[ionic dragon]

we can get a cool animation

[crisp owl]

Your Comfy updated?

[ionic dragon]

i last updated it around 15days back

[crisp owl]

that's a while...

[ionic dragon]

ok

lemme update

@crisp owl which custom node has this?

[crisp owl]

Threw that image into my flow haha

[crisp owl]

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

pretty sure that's this

yah it is

[ionic dragon]

got an error damn

[crisp owl]

missing upscaler model?

[ionic dragon]

how can we addimg2img to sytan's workflow?

i find sytan's easy and simple

[clever verge]

@visual glade Do you have any idea how to best implement this without breaking other stuff? I've modified stuff in widgets.js now.

[crisp owl]

It's just your Efficiency nodes install bonked. Remove it and reinstall.
Use the manager if that's easier

My img2img flow I made the easiest way possible, there is a single node to change, the denoise.
Everythign else is done for you

[ionic dragon]

i used manager to install

[hardy cipher]

lol

[crisp owl]

Then delete the folder and install via the git

[ionic dragon]

yea doing that

[crisp owl]

manager gets wonky with some things

[ionic dragon]

ywa

[ionic dragon]

this image doesnt contain the workflow

[ionic dragon]

same error

[crisp owl]

Hmm.... I'll have to check efficiency pack later then, but I know I recently updated it and it's fine for me.
In the meantime, bedtime for me

[ionic dragon]

ok bye

i;ll figure out the issue

[uncut fiber]

gn

[hardy cipher]

interesting armor

[ionic dragon]

@hardy cipher can you please share your settings?

are you even running them through refiner

[hardy cipher]

you can literally look at the settings if you drg that image into your workspace

[ionic dragon]

the image doesnt have the metadata, i dragged it, but not happened

i think you copied the image and pasted it here

[supple knot]

I dont think you can,discord strips the meta data

[ionic dragon]

if you attach it, you'll have the metadata

[hardy cipher]

discord doesn't strip anything

oh yeah, well I might have done that with the last one actually

[ionic dragon]

now it works, tq

[supple knot]

Are you using Windows or Linux?

[hardy cipher]

me?

[supple knot]

yeah

[hardy cipher]

windows, but I sometimes run things in linux

[ionic dragon]

@hardy cipher do both load images be the same?

[hardy cipher]

load images?

[hardy cipher]

optional

[ionic dragon]

ok

[hardy cipher]

well that's ipadapter

[ionic dragon]

https://huggingface.co/comfyanonymous/clip_vision_g/tree/main

is this the link for clip vision??

where can i get ip adapter?

[hardy cipher]

you want to use "pytorch_model.bin" for sdxl

for clipvision

all those models should be in the manager

you should use the manager to find these things if you aren't sure about them

[uncut steeple]

[ionic dragon]

cant find them in models tab

[hardy cipher]

hmm. well they're certainly all on hugging face. weird they're not in the manager list

[ionic dragon]

i got it

[glad grove]

srry for the ping someone was spammin porn gifs but wyzborrero banned him already 🙏

[hardy cipher]

I like how they were like "yeah, but I'm trolling the server" when I expressed my displeasure with their actions

[hardy cipher]

yes, you'll want to use the IP-adapter models specifically

[ionic dragon]

ok

[hardy cipher]

also, there's most certainly a cleaner implementation of ip-adapter out there if you have trouble with this one

[vocal stream]

is there a tile-upscale for sdxl? I have the new CNs and it has the tile preprocessor but not a tile model

[half ivy]

i am here
what did i miss

[hardy cipher]

nothing

[nocturne trail]

kaj did a backflip

[ionic dragon]

@hardy cipher added refiner, and getting this error

[hardy cipher]

your conditioning inputs are using clip data from the other model

need to give the sampler refiner clip data

[ionic dragon]

oh ok

@hardy cipher i am using this, but I cant find denoise in ksampadv

[clever verge]

I just saw there's a PR for this behavior that seems to still need some work. https://github.com/comfyanonymous/ComfyUI/pull/93

[mossy canopy]

I am new here . Can anyone help me with SDXL on comfyUI. :V

what is this crop_w crop_h do?

[vocal stream]

if you are new I'd recommend just not using comfy but using webui or one of the even easier UIs, comfy is a mess to work with

[soft zealot]

Super Furry

[vocal stream]

but crop w and crop h seem pretty self-explanatory - crops width and height

[mossy canopy]

it doesn't seem to affect my output at all..

crop_w: 400 crop_h:0

crop_w:0 crop_h:400

crop_w:400 crop_h:400

crop_w: 800 crop_h:0

crop_w:1023 crop_h:0

[smoky patrol]

Yeah, shame about that, guys. Thanks for the ping, though I was dead to the world.

C a t s

[mossy canopy]

crop_w:0 crop_h:0

[uncut fiber]

no problem 🙂

[mossy canopy]

The crop conditioning doesn't work very well.

[uncut fiber]

trying to quit smoking. It is difficult.

[mossy canopy]

Crop conditioning is a new feature introduced in SDXL 1.0 that allows the model to take cropping coordinates as conditional input during image generation.

Specifically, crop conditioning provides the cropping coordinates (x1, y1, x2, y2) that were used to crop the original training image. This information is fed as additional conditional input to the model alongside the text prompt.

By providing cropping coordinates, the model is able to generate images that better match the original aspect ratio and framing of the training data. This leads to outputs that are more precisely cropped and framed according to the text prompt.

Some key points about crop conditioning in SDXL 1.0:

• It provides cropping coordinates as additional conditional input to guide image generation.

• The coordinates correspond to the crop applied to the original training image.

• This allows the model to adapt and generate images matching the framing and aspect ratio of the training data.

• Results in outputs that are cropped and framed more precisely according to the textual description.

• Represents an enhancement over previous SD models that did not leverage crop conditioning.

In summary, crop conditioning gives SDXL 1.0 more contextual information to produce outputs tailored to the desired framing and composition indicated by the text prompt. This allows for finer control over the cropping and leads to higher quality results.

[peak dove]

OK, so FollowFox's w/flow is more useful (to me) than I thought! Great to hear.

It has this crop conditioning setup

https://followfoxai.substack.com/subscribe?utm_source=menu&simple=true&next=https%3A%2F%2Ffollowfoxai.substack.com%2F

[ionic dragon]

The logo is perfect

[native knot]

She BEEG. She Queen.
She have army.

The Randomizer run last night gave us this fine image.

[half ivy]

bro 🔥

[native knot]

When you and your homies all max out the same armor and raid together.

Definitely would shop here.

[uncut fiber]

tested A1111 it has nice results, but still slower, on other hand made huge leap in performance.

[native knot]

500 more images, including this fine image of Kirby, from last night's randomizer run added to my gallery: https://lychee.soulctcher.net/

[weary pewter]

#1072220168534642768

[noble shoal]

Yeah, i think have reached peak LoRa Training performance. Made a LoRa that makes every picture look kinda uncanny, unsharp, flat and with a paper texture. Nice 🤣

[mossy canopy]

analog camera??

[noble shoal]

No, my lora that makes things look worse

[uncut fiber]

@noble shoal old newspapers lora 🙂

[noble shoal]

And that's exactly what it was trained on

[zinc cargo]

can i have it?

[noble shoal]

I might upload it later today. Will keep you updated

[indigo carbon]

just made in image input multimodel workflow LOL

[crisp owl]

Oooh, these are cool.
Mobile so can't check if there's Metadata attached, lora effect, or prompted?

[uncut fiber]

Second lora in row i guessed right! 🙂

[indigo carbon]

man, IPA is damn magic

[mossy canopy]

simple prompt.. protovisionxl

[indigo carbon]

McSpider+destroying cyberpunk city= this monstrosity LOL

[mossy canopy]

pcb king

[stone fossil]

1 sec.

[stone fossil]

There we go.

[mossy canopy]

modern pcb

[stone fossil]

More transparant / see trough / rtx.

Nvme waifu.

[mossy canopy]

sure.. wounded modern pcb solider

[stone fossil]

Not many PBC. 😦

Transparant yes.

Cyborg also.

No PCB.

Neon also maybe.

🙂

[uncut fiber]

Thaevil TM

[stone fossil]

To much ice feel. 😛

We can do ice if u will.

[uncut fiber]

sorry my english is too weak. I need easy english 😄

if i offended somebody it wasn't my intention. I hope we all are o.k. 🙂

[fresh path]

Wow I love that one really nice gen

[uncut fiber]

GN ALL

[charred lotus]

does anyone know how to inpaint / outpaint with SDXL? will this require a different model or something similar to 1.5?

[indigo carbon]

LOL

[mossy canopy]

using this lora in the negative will do wonders

[charred lotus]

?

[crisp owl]

Nice, like the outcome 👌

[eternal fog]

It is very cool, unfortunately for me, 2 IPAdapters can sometimes use more than 10GB VRAM, so it slows the generations down.

Have you tried the new version that's out? It uses Vit-h instead of Vit-g. I gave it a go but found it to not be as good at doing what I wanted.

[indigo carbon]

the node? idk about that or are you talking about a new ipa.bin file?

[eternal fog]

Yeah they released a new bin file that's trained on vit-h instead of vit-g

https://github.com/tencent-ailab/IP-Adapter#sdxl_10

[indigo carbon]

oh, yeah, I see. is it really better?

as far as I know, ViT-G is bigger, so IDK if it's a downgrade lol

heh, it doesn't even work in ComfyUI lol

bruh

I'll just stick to what I used before, this doesn't even work

[idle zenith]

is this supposed to be usual time taken even with xformers used?

[indigo carbon]

DUDE

[hardy cipher]

Vit-g is bigger, but apparently vit-h gives similar results. But seems like it'd have to be a zeroed out conditioning thing because it can't deal with the clip-g input from the text prompt. I got the same errors when I tried to use it. Need to mess with it more though

[hardy cipher]

You're probably exceeding what your gpu/vram can handle which vastly increases render time. Even then that's still a ridiculous amount of time

[eternal fog]

It does, you need to use their SD1.5 clip vision file.

You also can't mix conditioning from the IPAdapters with the SDXL text encoder

[floral island]

uhh...?

that's not a ... floral girl

[indigo carbon]

SDXL>1.5, I'm probably not going to use that

[stone fossil]

What is love?

[shy kelp]

https://aituts.com/comfyui/ I can't get it to work

[stone fossil]

Gief me more love SD. 🙂

[wet nacelle]

Ask a better question then.

Jesus

[stone fossil]

@mack gief some SD love.

[wet nacelle]

I'm taking a shit right now. sowy.

[stone fossil]

Again ok.

[wet nacelle]

second time today.

[stone fossil]

😄

Hope the toitlet is up to par.

[native knot]

Anyone here have a good comfy upscaling workflow I can see/use/adapt?

I know how to do basic upscaling...looking for something with better results.

[stone fossil]

Upscaling you should never do to in a workflow.

Imo, here I only upsacle if I get something good and use chainer instead.

[native knot]

Why wouldn't I simply be able to have all of the upscaling steps in a group that can be turned on when I need it?

[stone fossil]

Imo you should only upscale the wicked output, my 2 cent.

Because not every image is wroth a upscale.

Thats why imo, my 2 cent again.

[native knot]

Hence the "can be turned on when I need it".

Same idea as the other parts of my workflow.

I just turn the group off when I'm not using them.

[stone fossil]

When I upscale.

I upscale trough the roof.

And do this

[stone fossil]

Else no go.

😉

[mossy canopy]

upload one of your upscaled image

[stone fossil]

Ok.

[native knot]

Is the upscaling adding any detail or just smoothing out whatever you're upscaling?

[mossy canopy]

depends on what kind of upscaling workflow you are using

[native knot]

I'm asking Thaevil1, specifically.

[stone fossil]

https://we.tl/t-Xpqb6oaifM

[native knot]

I'm trying to get a good workflow that increases fidelity/detail.

[stone fossil]

https://wetransfer.com/downloads/f883e8e546e0dc4505f30ab656d0199d20230903080945/8be2d7

[mossy canopy]

still downloading lmao..

[stone fossil]

😉

There is a 2nd one plz zoom.

[native knot]

That chromatic aberration on the lips, tho.

[stone fossil]

Got more if u like.

[native knot]

I think that's the effect from that bad VAE.

[stone fossil]

Fucking disc wont allow file size. 😦

[native knot]

lul

[mossy canopy]

too much artifacts..

[mossy canopy]

even the eyeball isn't accurate

:V

[stone fossil]

Im pretty happy with it.

[native knot]

I'd like to take something like this and make it a bit larger, more detailed. It's great, but it could be a bit more.

[stone fossil]

Like said I only upscale those I think I are good

Using a upscale in the end of a workflow is ass to me.

Why? You will not know what the image will become. So do it after hand. 🙂

And also why? I do not need nothing from the un-upscale imaget to do it.

[native knot]

If you're in Comfy and you change nothing prior to a step, then enable that step, it will pick up from there. I can just keep that group bypassed until I want to upscale something.

[mossy canopy]

i do 10 to 20 batches .. only upscale the ones i like.. I only upscale upto 4k res.. not more ... i never do just general upscale always latent upscale..

[novel birch]

does comfy remember your pending queue if you close it?

I need to stop generating but I have a bunch queued up still

[stone fossil]

I always upscale later.

All my images posted are non upscale unlesss told so.

🙂

But it is what float ur boat, what may float for me might not flow for others. 🙂

[floral island]

@stone fossil

[stone fossil]

Drop one of ur upscales then? 🙂

[mossy canopy]

#🏆｜winner-gallery message

[stone fossil]

Many dupes.

🙂

I dislike dupes.

[stone fossil]

OK ish.

Drop the healer style. 🙂

Full on evil.

[wet crane]

Hey hoping to get some help ? How the heck to I solve this issue I can't seem to be able to get these nodes working

[floral island]

is not healer. is rococo

[stone fossil]

A raccoon?

[floral island]

rococo 🙂 an actual thing

[stone fossil]

😄

[uncut fiber]

Comfy can i have question, suggestion? That if AITemplate detected in workflow, it when fixed seed, do not generate just only one image, as each image using AITemplate is original? I mean each image generated with AITemplate is similar but different. And allow this, it can be interesting doing those small differences animation.

[stone fossil]

I wish comfy would allow for cuda select per node.

[mossy canopy]

who is duping whom?

[uncut fiber]

And second thing make default disable in AITemplate keep in memory 🙂

[stone fossil]

Dupes on the image.

Replication.

[mossy canopy]

like?

[stone fossil]

SHow me one imgs have been deleted.

[mossy canopy]

what are you saying?! I am having hard time understanding you.. 😅

[stone fossil]

Then lets keep it that way. 😄

[mossy canopy]

sure.. why not.. 😕

[stone fossil]

Lets burn some VRAM.

[mossy canopy]

It still looks good .. it's just putting a analog filter on..

[noble shoal]

That might have been cherrypicked. Wait until you see faces

[wet nacelle]

@noble shoalHave you shared your Loras and whatever else you have on CivitAI?

[noble shoal]

Not now. But i will release this soon. Just picking the best....or might i say worst rigth now

[wet nacelle]

Do you have an account I can check up on?

[noble shoal]

Oh, that will be my first release. I will post it here when i have uploaded it. I might need a name for this abomination.

[wet nacelle]

I mean a CitvitAI account that I can check up every once in a while.

[noble shoal]

Should be aimingfail on CivitAi.

[wet nacelle]

wicked

[uncut fiber]

teeth in shape of tooth 😄

[uncut fiber]

reminds me starry night

[stone fossil]

Tadadada.

[tight crown]

Hey - were you able to get the new vit-h model for ip-adapter working? I get dict size mismatch errors

[ionic dragon]

i tried negative prompt as the prompt, and the fingers are perfect

[eternal fog]

Yeah, but they've moved the files around on their hugging face, so let me make sure the information I'm going to give you is correct.

[hardy cipher]

got the vit-h ipadapters to work

well the ipadapter model and clipvision model

[frigid halo]

is revision better than reference only?

[eternal fog]

How do you find them compared to the vit-g ones?

[tight crown]

@hardy cipher @eternal fog
Old version working:
sdxl_models/ip-adapter_sdxl.bin is working with sdxl_models/image_encoder/(model.safetensors or pytorch_model.bin)

New version error:
sdxl_models/ip-adapter_sdxl_hit-h.bin not working with either sdxl_models/image_encoder

[eternal fog]

Yeah they've not explained you need to use the SD1.5 Image_encoder

[tight crown]

fr?

[eternal fog]

Yeah because it's using vit-h which is the SD1.5 Text Encoder

So this needs to go into your clip vision loader - https://huggingface.co/h94/IP-Adapter/blob/main/models/image_encoder/model.safetensors

[tight crown]

sweet, will try thanks

[eternal fog]

And if you are putting the IPAdapter through an unCLIPConditioning that is coming from an SDXL Text Encode node, you need to remove that.

[tight crown]

doesn't that mean it's a sdxl/sd1.5 hybrid?

[eternal fog]

No it just means that it's only using the vit-h text encoder and not the vit-g

SDXL uses both, or it can use both

[tight crown]

oh got it

[tight crown]

I wonder if they finetuned the image encoder that SDXL or SD1.5 used or made it from scratch

[eternal fog]

I had more of a play with it last night and I have to modify my workflow to make it work, but this is the difference I see from 1 image.

This is with the newer vit-h only

And this is the older vit-g

And this was the input

[hardy cipher]

what setup did you use? this is what I put together. pretty barebones

[hardy cipher]

key is you can't send vit-g clip data through it so not sure you can even use the ipadapter node most people seem to be using. the all in one thing

[eternal fog]

You can use that node, as I'm using it.

You just can't control conditioning through it

[hardy cipher]

you might be able to concatenate the two

[eternal fog]

My Original Workflow I had like this

[eternal fog]

You can't I tried

At least not with the stock concatenate node. Because the sizes of both the conditionings are different.

[eternal fog]

But I've just had to remove that unClip conditioning for it to work. Because I'm still using both vit-g and vit-h text encoders in my conditioning

[hardy cipher]

ahh. I wonder if my conditioning normalizing node would fix that

[eternal fog]

Although let me try a text prompt and see if it breaks it

[noble shoal]

@wet nacelle https://civitai.com/models/142931

[hardy cipher]

I should add those nodes to my github. I just didn't want to put anything up until I knew it was working as it should

at first I goofed and had it modifying cached conditioning, lol

that goes awry quickly with the multiplier node

[eternal fog]

Ok you can still use vit-g conditioning with it, you just can't sent the IPAdapter into an unclip that has vit-g conditioning going through it, which makes sense I guess.

But you could just send that same image into Revision and it would basically do the same thing alongside it.

[hardy cipher]

I don't even know what these terms mean tbh

revision, etc

I know what the nodes do, and the models

[eternal fog]

Revision is the thing where you can provide an image for the conditioning

[wet nacelle]

Wicked. Thanks man.

[eternal fog]

So I have a workaround for my previous workflow, I just have to decide how I'm going to fit it in. And it's a bit annoying because I have to load both Clip Vision G and H models, which starts to max out my VRAM.

[hardy cipher]

ahh, just never really spend much time looking at the example workflows. maybe I should

[eternal fog]

@hardy cipher Have you seen a benefit to using the vit-h one at all?

To me it seems to be a little bit worse for a few things, but similar for most. However the thing they mentioned about potentially using less memory doesn't see to happen for me, it uses the same amount of memory as when I use the Clip Vision G ones.

[hardy cipher]

h on the left, g on the right. identical parameters otherwise

[eternal fog]

Yeah sometimes they are similar, which is fine. Other times, with the model I'm using and with people it seems to go a bit funny sometimes.

But without anything else effecting it, like no lower memory usage, I might just stick to the Vit-G one.

[hardy cipher]

yeah. not sure what the advantage for H is exactly. but I like messing with new things

[tight crown]

@eternal fog what does the unCLIP you mentioned do?

[eternal fog]

It's what's used for Revision. You feed an image into it and it takes that image concepts and uses it for conditioning.

[tight crown]

Is that the control net revision that SAI released?

[stone fossil]

gvd waar is mn vuurtje.

[eternal fog]

It seems like the vit-h and vit-g ones both go towards different things sometimes.

In this example the input is an image of tracer. vit-g creates a real person and vit-h creates a more 3d model type look (you can get both to do it at different strengths, but this is at the same strength)

[eternal fog]

Yeah

[frigid halo]

is there a smaller file for the revision control net? my kaggle breaks above 13 gb ram

[eternal fog]

Don't believe so

[tight crown]

oh yeah what weights and cfg do you recomend for the ip-adapter

[stone fossil]

Spanks @floral island

[eternal fog]

Play with the weights, their recommendations were if you are just trying to make something that looks similar to the target image, then use a weight of 1.

If you are trying to make something slightly different and your adding prompts in then use a weight below 0.5.

Weights of 1 sometimes go a bit nutty, so if it looks weird turn it down

[drowsy tangle]

I have buid a prompt that improve our prompts âne get Best image that you can use with gpt and bard

[tight crown]

@eternal fog using their example:

[stone fossil]

If your need to express your image it shit.

My 2 cent.

[eternal fog]

Yeah, I see no overall benefit between the 2, so I might stick with the old one as it works with my workflow better.

[stone fossil]

Sue me.

[tight crown]

do you have a preferred control net

diffusers/sai/other?

[uncut fiber]

Just did some stress test with random images. Not sure i can run it all night.

[stone fossil]

Whiooa.

[stone fossil]

Noice slmots.

[eternal fog]

I've not been using it a lot, the T2I Adapters are nice because they don't slow down generation time.

[vital ermine]

Well, my NSFW on civit just failed

[stone fossil]

Skate or die.

[elder rose]

Random question... Do you(Or anybody) happen to have a copy of of Hlkys repo with the 0.3.33 changes referenced here.... https://github.com/FizzleDorf/AIT/commit/33ad2d7b7227338593c95489ac593b04853deff9

[stone fossil]

Free.