✨｜sdxl

[stone fossil]

Look here again?

U asky me why lora, I show u.

[cyan crown]

bob here they are

[stone fossil]

But as said the van Gogh style must be rly in there to be there.

[pure crystal]

it's like a cinematic illustration style, good shading etc

[cyan crown]

well I think prompt makes some difference because I ask always for deep black and sharp images

[pure crystal]

poor VG, nobody gave a shit about him during his life

he could have used some of that starry night poster and mug money

[pure crystal]

yes nice!

he did all the artwork for Terry Pratchett

[cyan crown]

last one from the one you suggested

[pure crystal]

lol amazing

I'm working on a "impossible halloween costumes" set

(sidewalk level, POV, detailed depth:1.3), cinematic painting by Paul Kidby, mesmerizing, bipedal melty pizza slice humanoid in bank ATM, Chicago suburbs, Halloween, highly detailed, vibrant colors , 8k, sharp, professional, clear, high contrast, high saturated, , vivid deep blacks, crystal clear

horrified onlookers works in prompts which made me laugh

[cyan crown]

btw I made a Lora that reminds me that style....

[icy brook]

Aether Aqua.

[pure crystal]

this is going to be everywhere

[icy brook]

I hope so! 🙂

[velvet pumice]

adetailer - did you finally get it to work again? Or show up on the main screen? Controlnet shows but not adetailer

[loud pumice]

adetailer works for me with vladmandic

[pure crystal]

(close up, POV, propaganda:1.3), mesmerizing cinematic portrait, by Paul Kidby, evil dictator woman wearing Gaultier, center nuclear fallout cloud, futuristic retro future, background city, highly detailed, vibrant colors , 8k, sharp, professional, clear, high contrast, high saturated, , vivid deep blacks, crystal clear

[velvet pumice]

how can i tell what version im working with if its vlad vs another one? Do I find that in the settings? I may be working with something different

[loud pumice]

https://github.com/vladmandic/automatic

Just git clone this one and move your models over if you've been using a different UI

Its pretty much just a straight improvement over automatic's webui

[velvet pumice]

ok will do!

ty

[loud pumice]

np

[neat finch]

Hey guys, we recently trained a text to gif model with SDXL. Check it out and let us know what you think!
https://github.com/hotshotco/Hotshot-XL/
We also open sourced the weights!

[tender timber]

Endless Nodes can now save images and JSONs to separate folders

I've updated my collection of Endless Nodes: https://github.com/tusharbhutt/Endless-Nodes

Changes include:

• Added an Image Saver that can place JSON files in separate folders
• Added nodes to convert from one numeric type to another, and to string. Still in beta

Endless Image Saver

This is why I tried my hand at Python in the first place! There are many, many, many, good image saver nodes out there, so why one more? Well, the default saver does not save to UNC in Windows. Some savers will allow you to save to UNC but have built-in folder formats. Sure, you can can cobble some savers to save an image together with a text file, but the timestamp on the text file tends to be 2-3 seconds off from the image. No saver I know of lets you save the JSON file to a completely different folder

So… this node will allow you to save your image file wherever you want, with full support for standard Python date and time conventions, and you can save the JSON file somewhere else. I have more plans for this, but it’s ready for release now.

I also added some conversion nodes, but they are still in beta. I just felt like giving back to the community, so enjoy!

[upbeat summit]

very nice! I'll will definitely check this out

[tender timber]

Awesome, I remember we were fiddling with something like this a few months back 🙂

Break it as hard as you can, Chat GPT and I have a hate-hate relation for coding this, but I'm finally on vacation after delaying it four times so I'm sorta, kinda, up to the challenge.

[upbeat summit]

yeah. I went through many different setups to get filename management working to my needs in ComfyUI. In the end I hacked the SaveImage node and added some stuff of my own (suffix, own date function), so I gladly give this a try. thanks for sharing it!

[crisp owl]

Interesting, so would this save the timestamp at image completion, or how default works right now when you hit the generate button?

[tender timber]

You're welcome, which node is yours, it in GitHub? I spend from 3:00-6:00 AM and then 10:00-1:00 PM today trying to get the folders to accept the timestamping... and then noticed ChatGPT had happily given me both "img_folder" and "image_folder" in the same damn code

[tender timber]

Hmm, will have to check, should be at image completion

[crisp owl]

That would be beautiful if so

[tender timber]

OK, pressed "generate" at 16:22:15 ... Crap... chose an SDXL model... standby

Oh lord, it's doing 300 steps

[upbeat summit]

I have not released it because it's a dirty hack job :D.

I just added a suffix field, my own datetime function so the file receives the datetime when it is written and removed the counter.

I would like to use a more robust counter solution like in a1111, but the problem is that it's not reliable here because I need my images sorted always in the same order.

And here the counter is always getting out of order at some point. So my base image would be numbered after the upscale etc and this makes it really unusable for me if I need to sort through hundreds of images every day. So my solution isn't ideal but it's currently an acceptable workaround.

[tender timber]

How is the counter getting out of whack?

Also... weird that there are not shit-tons of Steven Seagal images being posted right now. Did that die down?

[upbeat summit]

haha

[crisp owl]

hahaha, it comes in waves

[tender timber]

OK, new test. Pressed the button at 16:25:45 ... and... Booyah!

File(s) saved to: C:\temp\2023-10-03 16-26-41 Endless_0001.png

[crisp owl]

Beautiful, I'll be downloading that now lol

[tender timber]

Next up is to make sidecar text files similar to the one AUTO has, and allow the user to decide what info goes in there. The JSON files are a trash fire when you get large workflows

[upbeat summit]

sometimes files that are clearly generated first get written later. it's a bit more complex and I did a lot of testing on this. l test out your node later and if you want I message you with some feedback.

[tender timber]

Does the default one use the process start as the time stamp? I never used it 'cause it couldn't save to UNC

[crisp owl]

Yeah last I tried doing timestamps in my saved file naming from the default save nodes, I would hit for like 20 to generate, and they would all be within the same couple seconds, which would cause some to be out of order, which was irritating especially when doing like an incremental cfg or something

[tender timber]

Absolutely, or here too. I did a "stress test" of 50 files to a NAS from my potato Athlon-1700 (I am so looking forward to Amazon Prime day next week to pick up a new PC and maybe a 4060 TI which has 16GB of VRAM) and all the JSON and PNG were in sync to the second because the JSON never generates its own timestamp, it takes it from the image

[upbeat summit]

cool!

[tender timber]

Seems good....

[upbeat summit]

yeah I just don't want to spam this channel too much with comfyui node development. I can only say that I struggled with filename handling for many weeks until I had a solution I could work with

[tender timber]

Why not? Better than Steven Seagal. Or all the complainers 🙂

Or is there another channel we're supposed to use? Too bad they don't let us have a central node upload place like with the embeddings

[upbeat summit]

yeah there really should be a comfyui channel

[glad grove]

no badmouthing seagal ( 1 year timeout)

[tender timber]

https://tenor.com/view/seinfeld-no-soup-for-you-gif-3570094

[half cedar]

https://twitter.com/johngreen/status/1708515024275189884?t=NVdnwU1mxSbH_QO9ynM2Mw&s=19. What's your number? I'm a 4. Part of the reason SD is so much fun for me

John Green (@johngreen)

It's baffling to me that some of y'all see stuff in your mind. You SEE it? The way your eyes see? I always thought "visualize" meant thinking of the words/ideas/feelings associated with a thing, not actual visuals.
I am such a total 5 on this scale I didn't know 1-4 existed.

Likes

39178

Retweets

2325

Twitter · Oct 1, 2023, 4:11 PM

[crisp owl]

probably a 1 I'd say

[tender timber]

In my undergrad in Education, we learned that 1 in 5 or 6 doesn't hear the voice in their head when reading, so I'm not surprised the mind's eye is different for different people too

[native knot]

Same. I can visualize things in detail in my mind as well. Not everyone can do that, which is wild. Also, I completely hear the voice in my head when reading.

[crisp owl]

Yeah I always wondered if that's why some people feel the need when they're angry to have to shout into a pillow to release anger or something.

[tender timber]

Probably slightly different 🙂

I'd love to meet someone who's synesthetic though, that would be awesome. For me.

[crisp owl]

Amazing how many more senses we have, and how varying they are between each person

[velvet pumice]

when installing i get to the end but its stuck at 17:23:21-461643 INFO Startup time: 146.63s { torch=19.91s gradio=3.80s diffusers=3.06s libraries=5.71s
extensions=89.18s codeformer=2.06s gfpgan=0.09s upscalers=0.09s onchange=0.94s
extra-networks=0.06s ui-txt2img=0.44s ui-img2img=0.37s ui-train=0.07s ui-models=0.13s
ui-settings=1.51s ui-extensions=4.06s ui-defaults=0.95s launch=0.58s api=0.22s
app-started=0.71s checkpoint=12.60s }

[loud pumice]

Doesnt seem stuck to me, just go to 127.0.0.1:7860 in a web browser

[velvet pumice]

kk

[dim steeple]

Does inpaiting with SDXL in auto not work??

[velvet pumice]

AssertionError: extension access disabled because of command line flags

when installing adetailer

[native knot]

This is one hell of an inclusive baseball team from a couple of perspectives.

Like, I don't even know how a couple of those dudes play.

Zion's baseball team.

This one came out surprisingly well. Swedish women's Chef National Team.

[slender coral]

Is the $2 an hour real for a100? https://lambdalabs.com/service/gpu-cloud/pricing

[native knot]

Tatooine national team.

[pure crystal]

Pro Wrestling Team

[native knot]

Nice cat in there.

[pure crystal]

"copious various $whatevers" makes groups more varied

[native knot]

Aren't you a little short to be a cheertrooper?

[pure crystal]

Elf bowling team

[pure crystal]

I tried 'class picture, 'team portrait', what prompt keyword are you using?

(close up, POV, propaganda:1.3), mesmerizing cinematic, (mutant monster football class portrait:1.3), by Paul Kidby, copious various mutant football players, (Gothic architecture football stadium), highly detailed, realistic colors , 8k, sharp, professional, clear, high contrast, high saturated, , vivid deep blacks, crystal clear

[upbeat summit]

maybe group 🙂

[velvet pumice]

We need this as a video game lol

[pure crystal]

I'm an unoriginal POS

[pure crystal]

You must do 1 other-than-car image to every 2 car images tn

[static berry]

https://tenor.com/view/lag-talking-why-are-you-talking-blah-blah-blah-talk-to-the-hand-gif-17718857

sidjhe by sirede

ignore4 double audi logo ai is being siutopid

[pure crystal]

fixed it both ways (haha)

[steady grove]

https://github.com/hotshotco/Hotshot-XL/blob/main/README.md oh wow lookit

[static berry]

lesgoo it generated a current gen bmw

[supple knot]

Blanka's High School Photos?

[crisp owl]

Here's the op #✨｜sdxl message

[steady grove]

nice!

[peak dove]

Weelz are too smorl

I am currently following FollowFox on X/Y Plots @ComfyUI https://followfoxai.substack.com/p/mass-produce-xy-plots-with-comfyui

[uncut fiber]

IF only i know how i did this effect 🙂

[pure crystal]

(chiaroscuro lighting:1.3), dim lighting, couple standing at a beach cafe, night time, (Van Gogh starry night sky:0.3), string party lighting, (POV:1.3),

[uncut fiber]

yes this is o.k. @pure crystal Just over glass is seen gogh starry night. out of glass is sky normal. Or i am confused 😄

[pure crystal]

oh that part, looks amazing, knolling case can help you get there. Great gen

[uncut fiber]

thank you @pure crystal

[ionic dragon]

what do you call this art as?

[uneven gazelle]

I'm not sure, there is a tool in 3D that kinda does the same thing, building lines from motion but i don't remember how it's called

[ionic dragon]

i want to generate them using sd

so i want to know the style

[uneven gazelle]

try to search on youtube/google/chatgpt like 3D animation motion creating lines or something like that, that might help you figure out the style you're looking for

there's a lot of advertisement using that technic, aswell some tutorials

I used that technic in Cinema4D, maybe you can find out

[uncut fiber]

How about slow shutter speed?

@ionic dragon btw your equation must be wrong, but no problem.
(16,000,000)! / (1,048,576)!
because higher resolution, less combination, if you recall it 🙂 But doesnt matter. Just recalled it

[ionic dragon]

No it's correct, you shouldnt calculate combinations, you should calculate the permutations, because changing the position colours would give different images

[uncut fiber]

o.k. just if resolution is 16 000 000 then there is only 1 result.
Anyway i dont want argue, thank you for your help.

[indigo carbon]

JuggernautV2

my last test

I worked on some stuff recently

[peak dove]

Sheep furniture!!!

[indigo carbon]

should I upload this? it seems to be very good at pretty much anything

[ionic dragon]

@uncut fiber for example consider this, if there are 4 ultra large pixels,

1. man
2. katana
3. background
   let the background be the same for both images, in the 1st image a man is holding a katana, in the 2nd image a katana is on the ground and the man is looking at it. the components of both the images are the same, but the placement of them matters, so we have to consider permutations

[uncut fiber]

o.k. i realy dont want argue. Original question was how many unique images is posible at res. 1024*1024 with 16M colors. and i really regret i opened this.
so (16000000!)/(16000000!) = 1
Love and peace! I really appreciate your time you spend with it. ❤️

[uncut fiber]

problem is to control motion of object from left to right. I tried it in night, so playing with lights, still faaar from desired result.

[indigo carbon]

https://civitai.com/models/156522/tdg8uusdxlmixv2

[ionic dragon]

its not (16,000,000)!/(16,000,000)!, its (16,000,000)!/(1,048,576)!. the answer would be near to infinity

[uncut fiber]

i just showed if resolution is way higher.

it is almost impossible to force object moving from one side to another...

[supple knot]

this node stopped in a ipadaper set up in comfyui any idea how to fix it

[indigo carbon]

yoo I also use that node for parts of my IPA workflow

you'd want the BLiP question to be something along the lines of "what does this image consist of?"

also change to caption

[supple knot]

I don't think I did anything different, Just opened a ipa system that I had, that worked. Now its not working

Is there a different" Blip Analyze Node" I can use here

?

[indigo carbon]

set to caption and refine the question

the text widget you have in that node is where you ask the question about the provided image

[supple knot]

does "Interrogate" mean it trys to figure what the image is?

[indigo carbon]

interrogate will say what is in the image, caption will describe the image

trust me, there is a difference, you'd usually want caption

[icy brook]

Aether Aqua.

[west breach]

is it out yet?

[icy brook]

I can’t decide on whether to release Aether Fire or Aqua first 😀

But I was thinking I could release on Friday perhaps.

[west breach]

what happens when you load both? 😄

[indigo vine]

Aether smoke

[icy brook]

No idea 🙂 Lava?

[icy brook]

Yeah. But maybe Aether Cloud can do that already?

Haven’t tried

But if there is a need for it I could surely train that as well

But then, to choose between these two:

[indigo vine]

missed this one. Definitively have to test it!

[west breach]

thought i'd try making something inspired by this SNL sketch

[native knot]

That's a great prompt inspiration.

[west breach]

also great prompt to show how many things can go wrong in one image 😄

[native knot]

Well, clearly our infantile hero is saving that beer from the alien invasion that is occurring by driving away in the only functional vehicle he could muster at the time.

[twilit nexus]

hey can someone recommend me a model for out puts like this

[west breach]

you could try prompting with phrases such as 'character sheet' or 'multiple angles' 'multiple views of same character'

[twilit nexus]

no i meant that kind of style

i have been trying sxz luma

anylora

and some other few models

the one i posted above is the niji model from mid journey

i see on civtai so good looking outputs

but when i put my prompts i get some crap 😦

[rustic garnet]

not "the same style", but I wouldn't say that vanilla SDXL is not able to make high quality character sheets

prompt: "a barbarian wearing fur and leather armor, intricate concept art, clean lineart and color, dnd render of a muscular man, clear outfit design, cutout, game design, character sheet, ultra detailed clothing, white background"

for the barbarian I added "pencil sketch" into the negative prompt, to make it "less handdrawn" as the images you showed above

[twilit nexus]

my potato don't run sdxl 😦

😥

[twilit nexus]

yeah i want something like that

but i cannot run with 4 gb vram sigh

rev aniamted giving this not bad

but most the time i have to put a ref image

i will try just plain text

[uncut fiber]

i would love it but one thing hurts whole image...

[uncut fiber]

need bit more contrast 🙂 @static berry

[twilit nexus]

ok revanimated giving the best results so far

with castlevania lora from civtai

[uncut fiber]

@static berry didnt know you do this things as well

[twilit nexus]

i have to use image to image

[static berry]

?

[uncut fiber]

hidden portrait

[static berry]

lol

[uncut fiber]

i will not spoil it 😄 i am playing more with text. But in SD, i think in A1111 it is better made

[twilit nexus]

i like A1111

i can't stand comfy spider web of 101010101 nodes

i like buttons not spider nodes

[static berry]

lol

[twilit nexus]

yeah as soon i see that my brain gets very tired

[uncut fiber]

i would love it, but if it is working. AIT which boosted speed by 50% in my case, i have installed but nodes doesnt show in loader tab, not any trace after them. And i got it reinstalled! 😦

[twilit nexus]

in my case it woks

everything is working but darn man i hate nodes lol

[uncut fiber]

AIT? When you tested last time?

[twilit nexus]

i mean the thing works

but what is AIT?

[uncut fiber]

it is for 30 or 40 nvidia gpus, boosting performance. Yes it is working as well for me, but very limited

[twilit nexus]

i got a 3050 gpu rtx

[rustic garnet]

but you write in the SDXL channel 😅

[twilit nexus]

sdchat is kind of dead

no upload image tehre either

[uncut fiber]

i believe 3050 can run sdxl probably, how much VRAM?

[twilit nexus]

4 gb 😢

[uncut fiber]

realized frequently raining inside

[twilit nexus]

i got a bad deal lol

[uncut fiber]

hm, sd is good eighter

[primal vault]

I released update v4.2 for my SDXL ComfyUI workflow on CivitAI and Github - this should solve an issue with inpainting and also adds support for FreeU

[uncut fiber]

prompt: (art by George Romney:1.0) ,art by Laure Albin Guillot, Impeccable Random image, Dark, Circular polarizer, 64K

[crisp owl]

You're supposed to overlay the polarizer over another image

[native knot]

No, this is "art", @crisp owl... 😄 😄
It's just an image of the polarizer.

[muted tartan]

whats the model?

[crisp owl]

haha I know, I was just saying a funny 😆 😅

[native knot]

I know...I was playing off that. 🙂

[crisp owl]

Art can be so complicated 😆

[uncut fiber]

@muted tartan it is nightvision XL but i think that polarizer circle is spottable, just i didnt expected it 🙂

[twilit nexus]

is theer a way to get 3 views always for your character for image generation?

[indigo carbon]

LUIGI
THEY HIT THE PENTAGON

[twilit nexus]

like side view front view and back view

[native knot]

Like nearly all Mario pics I've seen, they're great,...except for the eyes. Most of the time, the pupils on Mario/Luigi are messed up.

[indigo carbon]

seems fine when it's just mario

[native knot]

Some of those are good...though a couple have a double white spot. But often I see messed up pupils in Mario renders.

I'm also not just referring to your renders, but Mario renders from SD in general.

[indigo carbon]

also the new SDXL model experiment I did seems to improve text abilities somehow

[muted tartan]

it looks great man

[uncut fiber]

@muted tartan wrong tag? 😄

[muted tartan]

nope I just love that image, I was just tryna learn which model u used so I can try it out myself 🙂

[uncut fiber]

this image iam not sure about model, but it was made i believe in Comfyui and include all data. Will check it for you!

@muted tartan Various colors of lights are jumping. Captured by static camera. BREAK Slow shutter speed art.
and
Model Jugernaut XL version3

[indigo carbon]

the text is sooo seamless with this

[uncut fiber]

when trying those sparkling lighties things lock in human body it lost all of its motion. Any idea if there is some way to keep its motion?

[steady grove]

how you mean motion in this context?

[uncut fiber]

mmnt will link original images

#✨｜sdxl message
basicaly it is this locked in human body.

[pure crystal]

I was trying last night, I used light painting, long exposure but didnt get closer than this

[steady grove]

i'd say those human bodies have motion, but you dont' want them. negative prompt against a human subject then. "human person posing" in negative might help. maybe just "person"

[uncut fiber]

it can work in glass, i think problem is that light can move in SD but human cant.

There is slow speed shutter as well as long exposition

[pure crystal]

I couldnt think of a good term for that layered look in the original pictures.

[uncut fiber]

@pure crystal what do you mean?

it is in pictures, comfyui, i never strip data. If you mean this.

[pure crystal]

no, from the images the person originally posted

[steady grove]

slow shutter and long exposure are synonymous

[uncut fiber]

it was me i think 🤔

yes i know i need really heavy weight on it 😄

[pure crystal]

the spaghetti look here

[uncut fiber]

oh o.k.
i think it is impossible to do in SD. Because ppl cant do steps or ballet in SD

i mean motion. Light can

[pure crystal]

pure prompting probably not possible, multi controlnet could work

[steady grove]

"i think it is impossible to do in SD." often this isn't the case. Find a photographer who does a style you like prompt for that

prompting is king. if people can't get an image out, they likely are just not casting the right spell

[uncut fiber]

it means simpy interrogate image, will try tomorrow then

[steady grove]

for ballet, controlnet would help a lot. i notice a lot of the people who are like "ITS IMPOSSIBLE" refuse to use controlnet and instead need it all done with one prompt. i don't get it, but it's a powerful option you shouldn't just leave on the table

taking a break from the problem is good but don't declare its impossible. just take another approach

[urban fjord]

Decide, is prompting king or is it not? For controlnet to do wonders you have to do a lot of work for it already.

[uncut fiber]

ballet like animation you mean? Or like boxers punch hand motion? Then it is all possible of course.

[steady grove]

prompting is king. the king isn't the only piece on the board though. everything else works towards the king's goals

[pure crystal]

commisioner meeting

[steady grove]

i'll just block you since you're not here to help people figure things out. you're hear to argue with good advice. why would you try to argue this point?

little astounded right now. who even argues that prompting isn't important?

[urban fjord]

Alright, do what you want. But don't gaslight people. SDXL has limitations and it's foolish to pretend it can do everything.

[steady grove]

gaslighting? okay fine china

[urban fjord]

You are right that Controlnet is powerful, but I find the current ones to be a bit rigid and limited for my liking so I'm training my own.

[crisp owl]

On that front, if something isn't possible, nearly anything I've seen thus far can be with a new trained model/lora/controlnet/etc.

[steady grove]

i'm not sure why you're trying to build bridges after burning them down. oh well

[rustic surge]

I use sdxl to draw pictures, but the effect is always not realistic enough. Can anyone teach me how to write the prompt words in sdxl so that they are realistic, or there is always a vague feeling.

[urban fjord]

But if you need to train a spesific model for everything it isn't really SDXL doing it is it?

[rustic surge]

sad

Why doesn't it look HD at all?

[crisp owl]

trained models is exactly what sdxl is

[urban fjord]

In general, adding photo at the start and some kind of film or camera at the end will help you generate more realistic images if your current results are too cartoony.

But if you're not happy with that portrait you might try one of the many fine-tuned SDXL models which might give you portraits that you enjoy more.

[steady grove]

person trains 12 images on sdxl, suddenly think they made the whole shibang

[noble shoal]

Working on a lora again. Very early stage.

[urban fjord]

Looking good, espeically with the colours.

[pure crystal]

whoa

[uncut fiber]

Good night everybody. Peaceful night!

[noble shoal]

Good night Bernix

[pearl condor]

Hi guys, I'm developing/trying to run a local vue js app for image to image, not seeing any mention in the documentation for input resolution issues. it only seems to take input of 1024x1024, every other input returns an error. Does anyone have an idea how to resolve this?

[noble shoal]

What have i done

[crisp owl]

What have you done, you mean what have Ash and Pikachu done?!

[noble shoal]

Going a lot to the gym, i guess

[indigo carbon]

dragon, fruitcore

[indigo carbon]

"legendary fish"

[uncut fiber]

cant sleep. @noble shoal those are keywords for directors cut.
Inkdrawing, Prompt lora:SDXL_Inkdrawing_Directors_Cut_E:1

[noble shoal]

Yes, but more details are in the LoRa description on civitai.

[uncut fiber]

yes i read it. if you mean by kerres or similar

going to try it

[uncut fiber]

does it look like DC?

[indigo carbon]

tf, it's gone sentient haha, I didn't ask for a sign

[crisp owl]

Well, it's not incorrect

[indigo carbon]

yeah, it's insane how it decided to write and still nailed it

[grave smelt]

That text is clean, damn

[indigo carbon]

it's the new model I made, so far it seems to be incredibly creative

when I made it I calculated what combination of XL models would create a model capable of being great at all styles, and it also got creative in the process somehow

[deft coral]

i give up

i cant train nothing

[uncut gull]

I remember this guy from the other day.

[indigo carbon]

I made the model used there just today, so either I used that prompt in the past or you're time traveling

[uncut gull]

Just the dragon made of fruit. It came up when I was claiming SDXL couldn't be forced to make objects out of substances, and you proved me wrong by making a dragon out of fruit, then it took a while for me to replicate the results.

SDXL (technically) animatediff.
animatediff running on realistic vision, then processed through SDXL via mov2mov extension.

[clever verge]

There's some new Porter-Duff stuff in ComfyUI

https://github.com/comfyanonymous/ComfyUI/commit/d06cd2805d86d7a9ed7485b6a0c7e113cff27d8e

[indigo carbon]

ohhh I see. well, the new model I made does all that way easily

I think I managed to nail the creativity while keeping the talents of the merged weights

[pure crystal]

Porter-Duff will be good for compositing depthmap controlnet images https://www.youtube.com/watch?v=yC2i8PUno70

[uncut fiber]

@uncut gull try use riffe on that video. Can make it nicely smooth

[uncut gull]

riffe?

[uncut fiber]

yes video interpolation maybe best but very slow. On those few frames it should be quickly done

you are right RIFE

[uncut gull]

I mean, please explain what "riffe" or "RIFE" is. Internet search does not work these days. It's just results for spearfishing.

[pure crystal]

https://colab.research.google.com/github/HeylonNHP/RIFE-Colab/blob/main/RIFE_Colab.ipynb#scrollTo=Rh6neS2cJFbE
https://github.com/styler00dollar/Colab-RIFE

[uncut fiber]

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

[uncut gull]

I've added it to my future projects bookmarks. Thanks. I have a headache at the moment. I'll make a few more animatediff videos then go to sleep.
AI never gets a headache...

[uncut fiber]

i will try to do it for you tomorrow, if possible and i found proper way in video encoder software i have

[vital ermine]

When loading the graph, the following node types were not found:
AITemplateLoader
AITemplateEmptyLatentImage
Nodes that have failed to load will show as red on the graph.

yet it is installed

yep, I guess comfy killed ait.

[pure crystal]

mantis shrimp beast football player, wearing american football uniform

[twilit nexus]

is it possible to rotate he head at 30 deg angle using AI for and get same image or something?

[velvet pumice]

for Composable Lora. How do you make it work. It shows up on screen / is installed but is there more to it? I have a couple loras in the folder already but how do i use them at this point

[velvet pumice]

DO i just enable and wish for the best fam

[ionic gulch]

the next update of my workflow will allow you to mute nodes for features you don't use - saves resources and may speed up the workflow:

this is how the menu look like if you turn off everything but the standard image:

can't mute the preview images, as they are "end nodes", but the images are cleared and replaced by a single black dot, instead of a red warning border

turning off everything reduces the vram usage on my system from 21-24gb to 9 -12 gb - maybe i can reduce it even more if i also add a switch for control net

[half cedar]

low poly cinestill, nyc roof tops, night time, low poly ninja turtles realism, dimly lit, tragedy, bokeh, photographic, realistic colors, chiaroscaro lighting, blur, long exposure, extremely detailed depth, close up, style of beeple

[uncut fiber]

@uncut gull
it is very problematic because vertical and horizontal lines. Here are two videos. 24fps and 60fps

o.k. i made it mkv forgot it must be mp4....

[uncut gull]

It does look smoother.

[uncut fiber]

yes it does, but it can look even better if not -|- those patterns. It is difficult for that model 🙂

[rain crown]

Anyone use these https://github.com/ltdrdata/ComfyUI-Impact-Pack ? I can't get the face detailer nodes to work correctly, they kinda work but the quality isn't too great

[crisp owl]

Works best after an upscale, then feed the upscaled image there.

This is my default settings that work for me

[rain crown]

Thanks, I'll try out the upscaling before strat

[twilit nexus]

can someone generate 3 views for a model for this image?

as close to it gets

my ram is not good enough to do that

thanks

like fornt and side views

[prime bison]

It still plays if you tap the name, it just doesn't show a preview

[uncut fiber]

o.k. great. Thank you. Result can be better. original is 8fps and results are 24fps and 60fps. problem is vertical bones over horizontal bones. @prime bison

[peak dove]

Using SDXL_Base model and Papercut Style in AUTOMATIC1111

[urban totem]

@uncut fiber I don't know if it's your workflow, prompt style or JuggernautXL v5 (or the combination of them) but I'm get amazing results compared to SeargeXL with SDXL 1.0. I haven't compared directly at the moment but coherency and style is lightyears ahead of what I was getting until then.

[rustic garnet]

???? the whole idea of the face detailer is that it is doing the upscaling automatically. No need to do that manually

[crisp owl]

I'm simply sharing what I found worked best for me.
It was shotty at best trying to utilize it prior to upscaling the whole image.

[supple knot]

Everyone has there favorite Batman , This ones mine

[urban fjord]

Batman!

[supple knot]

Dog looks Comfortable

prompt A man in a wrestling mask with a sign that reads "SDXL can do Text too"

[wet glacier]

now do also Hansel together with Zoolander 😉

[uncut fiber]

@urban totem what workflow do you mean. It is not my i believe. I think i made it according to i think Styler prompter.

[rustic garnet]

I would say you use it wrong... but to be honest, the documentation of the impact pack is really bad

[rustic garnet]

the way the face detailer works is following (you could do that just manually, but of course it's nice to have an automated workflow for that):
1.) find all faces in the image
2.) for each face: take the region around the face, increase it a little bit to also capture the surrounding (this is the bbox_crop_factor)
3.) now upscale this sub-image to the guide_size
4.) make an image2image pass on the upscaled image
5.) downscale the subimage back to its original size
6.) paste the subimage back to the original image. Add some smoothing/interpolation (this is the feather value)

in your workflow for example you use guide_size 256, which is a really bad value for SDXL. As SDXL is best at native resolution of 1024x1024 you want to set guide_size to 1024

(set guide_size_for to "crop_region", such that the whole subimage is upscaled to 1024)

all other parameters are fine. You don't need upscaling beforehand

however, if your face in the image is soo small that you cannot see any details with the low resolution you have, upscaling can help of course. But if you upscale your image, you don't really need the face detailer anymore. The whole idea of the face detailer is to improve quality without increasing the resolution of your image

[peak dove]

Using Face Detailer (ADetailer) in SDXL AUTOMATIC1111 and DynaVisionXL model

[nimble heart]

new default settings

[uncut gull]

Animated cards... Def not usable like this at least.

[uncut gull]

Not going to work. 😕

[uncut fiber]

but it is better, probably one by one?

[urban totem]

I stumbled on one of your answers here through the search and used one of your image. I thought it was yours because of this workflow (https://github.com/wolfden/ComfyUi_PromptStylers/blob/main/examples/workflow.png) but I just noticed there is another one in the same folder 😄

[uncut gull]

I might try animating effects for compositing, like flower petals falling, then layer a character over them?

[uncut fiber]

probably, i cant help you here, you must know 🙂

How you did it?

[pure crystal]

@peak dove ty

[spring fulcrum]

What does IP adapter do?

[half cedar]

Lets you use images as prompt input and will blend concepts and styles into the output.

[urban totem]

Thanks PromptStyler with magic button:

😄

All I want is space marines. By the way I've also noticed that juggernautXL v5 really like to put Trump as the face of my space marines, I wonder if it was done on purpose during training

[vale eagle]

Trying to combine each part in depth map

[half cedar]

It got mixed up and made yoda the xwing

[crisp owl]

https://tenor.com/view/multipass-gif-9536054

[cyan crown]

prompt was : cinematic photo by Luc Besson woman with orange hairs.........

[crisp owl]

Perfect then hahaha

[cyan crown]

last for today

[crisp owl]

that SD's version of the elemental stone caretakers? Or completely unrelated?

[cyan crown]

it's just an alien in ancient Egypt by Luc Besson

SDXL base

[crisp owl]

very nice

[cyan crown]

really the last one! Need to stop ::D

[vital bane]

Crystal Clear D

• chose your own character (consistent image generations of characters and clothing)
• accurate camera angles (front, behind, side, closeup, cowboy shot, full body, portrait)
• no shortage of clothes (heels, sweaters, dresses, skirts, bodysuit, shirts, lingerie, pantyhose, stockings)
• various poses (skirt uplift, dress uplift, sitting, lying, crossed legs, spread legs, extended legs, on knees)
• plenty of backgrounds to chose from
• proper looking hands

Satisfy your dopamine boost with unique and accurate combinations such as: from behind, skirt uplift, full body, looking at viewer
Clothes and character look remain consistent when changing camera angle or pose.
Supports up to 1536px resolution without upscale or deformities. No refiner.

and yes it does nsfw too

[somber hill]

please leave a comment : https://twitter.com/GozukaraFurkan/status/1710096529459933573

[fast bobcat]

Are you using controlnet for this results?

[vital ermine]

No.

[nocturne dove]

I liked the expression on this Zelda

[spring fulcrum]

Thank you that was well said

[noble shoal]

I normally see my LoRa from miles away, but this time I am unsure. Did you use a LoRa?

[nocturne dove]

Yeah but also ipadapter with something that looked a lot different than your lora. And technically I used a merge of your two loras from civitai (not for any particular reason though, it's just what I had set).

Pretty consitent in style but most aren't as good as the one I posted first.

[noble shoal]

Looks very nice 👍

[nocturne dove]

As usual, don't count the fingers! And sdxl still has no idea how a person should hold an object in their hands.

[noble shoal]

Having more or less fingers in drawings doesn't look as weird as in photos. Nice work.

[heady vale]

The images Ive posted over the last few days have all used your Lora. I really like it. Nice work

[uncut fiber]

consistency of faces is incredible, i cant get different face 😄

[grim wharf]

is it just me or clip skip doesn’t affect sdxl checkpoint at all, atleast with auto1111 ui

[uncut fiber]

it was said, it isnt needed @grim wharf

[noble shoal]

Have you tried prompting for names instead of genders?

[crisp owl]

Really shouldn't be used with SDXL, but saying that, are you using Comfy or A1111?

[grim wharf]

I use A1111, just want to experiment a bit as usual with 1.5 clip skip did yield some interesting result but SDXL seem to be non affected by clip skip value

[crisp owl]

I don't know for 100%, I know I read a comment from a dev or Comfy himself on the matter of ClipSkip with SDXL, from what I recall (can't find the comment), it was mentioned that due to SDXL being trained with a G+L clip, Clipskip doesn't really do anything.
Not to say it never won't...pretty sure if someone trained a new checkpoint but mostly trained on the Clip L, it would likely affect it?
TLDR:: SD was never really intended to have clips skipped, and even more so with the SDXL model.

[soft zealot]

Update/Tweaks to my Cruiserweight workflow (not yet uploaded as part of a new package to Civitai)

#1 Added as second Upscale sampler so a 1x upscalar such as ArtClarity can be used befor the main upscalar. Zoom on face image is the comparison between my OG on the LH side & the new method on the RH Side)

#2 Replaced standard Ksamplers with those from the Inspire pack and added a "Noise Mode" Selector. This allows you to switch between ComfyUI mode (CPU) and A1111 mode (GPU). Full body is comparison with CPU on left, GPU on right.

All 3 raw images also attached and named (and include workflows, new workflow are images with type2 in filename)

another quick comparison of CPU vs GPU noise mode, all other parameters were identical

[ionic gulch]

preview of the next version of my workflow with real (muting all corresponding nodes, saving resources and time) on/off switches for all major features:
this will reduce VRAM usage to < 10GB VRAM, if you disable some of the more demanding features
simple txt2img with 35 steps and "Euler a" can be reduced to 10 seconds per image on my system (9900k+4090)
my usual setting of 70 steps with "dpm++3m sde gpu karras" takes 17 seconds

[crisp owl]

Well that sounds intriguing!

[uncut fiber]

No, will do!

[high skiff]

@visual glade Are you on by chance? I know I have not been here in a good long time, however, I am running into an extremely infuriating bug with comfyUI I would like to bring to your attention, and that is VRAM management when comparing models

I have tried GPU only, normal VRAM, and Low VRAM, and no matter what, I am hitting VRAM limits on my 3090 generating just a singke image across 5 SDXL pruned models

I press gen, it loads the first model to 8GB VRAM, then it finished and loads the second to 14GB VRAM, then it finished and loads the third to 18GB VRAM, then it finished, and loads to 23GB VRAM, then it finished and pools to 27GB VRAM and the last image takes several minutes

Its inconsistent in how frequent it happens, but I have had it happen several times on all three VRAM settings, and its leading to my tests for a new project being EXTREMELY slow and tedious ._.

[pure crystal]

original, high res

[high skiff]

if it doesn't happen the first time, the VRAM just stays loaded to 23GB every image gen until it eventually forces an overflow into pooled memory. I am not sure why this is happening, but its hindering my tests considerably

[uncut fiber]

@noble shoal your lora making paper box looks real life

[high skiff]

there is no reason for this to be happening when on low VRAM

I am generating single 1024x1024 images

I drop from 2.2it/s down to like 5 seconds/it

[fierce hollow]

did you try starting with --disable-smart-memory

and no normalvram/lowvram

[high skiff]

I guess I can try that. Thats a pretty bad solution if thats the case

I shouldn't be worrying about VRAM constraints when doing single base res image gens on a 3090 lol

I'll try that right now

[fierce hollow]

yeah the smart memory thing isn't that great in my experience at least

actually you may want to try starting with --normalvram too

[high skiff]

same issue

[fierce hollow]

otherwise it will probably default to high

[high skiff]

it does it on lowVRAM, normal, high, and GPU only

[fierce hollow]

well that's weird

[high skiff]

It does it much worse on GPU only tho, cause that unloads nothing

@fierce hollowlooks like you might have fixed it, you mad man

its much slower now, as to be expected, but I guess thats still not as slow as 5s/it

[fierce hollow]

oh that's cool

slower as in swapping between models I guess

[high skiff]

it just fully loads and fully unloads every time

[high skiff]

yeah

screw it, its cold here. I am lifting my power limits :p

I am doing some realism model comparisons, and it generates across 5 models each gen, which has just been terrible for VRAM with these issues

luckily I have enough system RAM to eat up all the models

got a cool 38GB of SDXL models in my RAM ATM lol

I would run some of my comparison tests here for people to vote on, but I am sure they will get drowned out the second people wake up haha

@fierce hollowok, this is a lot faster. Thank you so much. The loading and unloading is still so much faster than dropping to several seconds/it

[half cedar]

Happy to look at an x,y when you have them

[high skiff]

its not really an XYZ, just 5 models next to each other haha

[west breach]

the next problem is running out of CPU RAM

[high skiff]

so far, I am pretty happy with some results

[ionic gulch]

same experience for me. i have to use --disable-smart-memory on my 4090, because smart memory management messes things up (unless it changed with a newer update - haven't gone back to smart memory management for a few weeks now). it keeps so much in vram, that nvidia (old, better driver version) will start to move things to virutal vram. my guess from when i least tested was that smart memory isn't fine tuned well for 24gb and doesn't allow enough buffer.

[high skiff]

I am comparing my super light weight and small realism LoRA against 4 very highly rated SDXL models for realsim, and I feel like mine is matching/winning most of them

[west breach]

just seen all these new sampler nodes. where to even begin!!?!

[high skiff]

like here. Mine is the far left lol

"A portrait of a handsome young European man with light blue eyes and short black hair lit from the side by dappled lighting through slatted blinds"

I might have to resume training on this LoRA when it cools down here

cause its promising

"A portrait of a stuffed elephant plush in a rain coat sitting next to a river"

"A portrait photograph of a black cat dressed in a witch hat sitting next to a carved pumpkin"

@half cedar

if you wanna see some so far

[half cedar]

Cool ty

[west breach]

throw my model in the mix 😛

[high skiff]

I can throw out one of the models in here for it. I know which one is doing the worst

[west breach]

https://civitai.com/models/141592/pixelwave

[high skiff]

screw it, lets do it

I will be sparing no criticisms :p

[pure crystal]

yours looks best, 3 is very good

[high skiff]

thanks! My loRA is extremely small for what I am trying tod o, so I am sure with a much bigger dataset and more training, I can really nail it

my LoRA is only trained on 90 images

[west breach]

I was training loras with 10k images that took 3 days 😄

[high skiff]

@west breachyour model will be replacing juggernaut, cause it has not won a single fime for anybody so far

[high skiff]

my 90 image one took nearly 6 hours, so its a pretty beefy training for the image count

I wanna try to get to at least 500 images

your model is done, testing now

[west breach]

I would usually try to get it to train 30 epochs overnight, then spend the day mixing it in, then gather more images and training again, bloody exhausting

[high skiff]

I need to re-organize the models for bias prevention lol

@west breachalright, its going down

@west breachOk so ummm 😅

I do not think this model is a good fit for a realism test

😅

[high skiff]

it is very 1.5 looking for realism, which isn't a bad thing, but it has THAT look

I'm curious to know what your prompting looks like, cause your model sticks out like a sore thumb for me

lets play a game of "find the outlier" lol

[west breach]

2?

[high skiff]

yes

if you would be willing to share your prompting examples, I would be willing to modify the test depending on how that changes the other models

cause your images look phenomenal

[west breach]

I tried to get it to not look like every other model out there

[high skiff]

see, and at least in this test I ran with my prompting, it looks like every 1.5 model out there

But I really do like the results you got yourself

[west breach]

I used this negative makeup, 3d, b&w, bad anatomy, bad art, blur, blurry, cartoon, childish, close up, deformed, disconnected limbs, disfigured, disgusting, extra limb, extra limbs, floating limbs, grain, illustration, kitsch, long body, long neck, low quality, low-res, malformed hands, mangled, missing limb, mutated, mutation, mutilated, noisy, old, out of focus, over saturation, oversaturated, poorly drawn, poorly drawn face, poorly drawn hands, render, surreal, ugly, weird colors

which I stole from the clip interogator repo

[high skiff]

oh wow

that just made my stomach flutter with excitement, cause I use basically no negative 😅

[west breach]

left with negative, right without negative

[high skiff]

see, it didn't make hardly any difference for me

your image still looks drastically out of place

I wonder if its your gen settings

would you be willing to share more in DM's? I wanna give your model a fair shot

[west breach]

yeah

[high skiff]

cause as of now, its not winning any realism contexts anytime soon 😅

although, I do like that negative!

I will play with it some for the future, thanks for it!

[west breach]

@icy brook was generating pics with random prompts earlier and this one came out, made me think of you 😄

kinda glitchy, realistic

[icy brook]

glitchi glitchi ya ya da da

[west breach]

front shot, 1 man glitching, splitted personality, shattering skin, computer glitch effect, portrait photo of a 35 y.o man, looks away, natural skin, skin moles, (cinematic, film grain:1.1), schizophrenia

[soft zealot]

No LORAs , just using realisticstockphoto model

One image generated using Nosie Mode = CPU (ie comfyuo default) second image generated using Nosie Mode =GPU (ie A1111 mode)

All other parameters equal

[high skiff]

welp, I found the area where my realism LoRA absolutely murders all the realism finetunes for SDXL lmao

forest portraits with detailed foliage

[azure current]

chinese style

[high skiff]

left is mine in all of these

This LoRA does so much better than I remember

[pure crystal]

(close up, POV, extremely detailed depth, intricate detail,photoreal DMT, dappled speckled light:1.3), cinematic painting by Thomas Kinkade, mesmerizing surreal majestic expressive chameleon, low angle jungle, hanging vines, (stargazer lily:1.1), moon lit sparkling god rays, psychedelic spiral distortion, highly detailed, shining dew, black linework, vibrant colors , 8k sharp,professional, clear,high contrast, high saturated,vivid deep blacks, crystal clear, sacred geometry intricate scales, natural textures

[azure current]

who have some lora about man?

[high skiff]

damn, my LoRA also beats them in low light images too

I really need to coninue this LoRA man

[azure current]

how can i get it?

[azure current]

it's amazing u can output text

@azure current

[west breach]

I'm just playing with ya 😉 it only makes good text once in a while, not every image

[azure current]

which lora do u used?

[west breach]

no lora, just my model pixelwave

[azure current]

ooo,thanks

[ionic gulch]

i'm missing a movie picture with "lora rennt" (or the english title "run lora run"):
https://m.media-amazon.com/images/I/51k6iRX6obL._SX300_SY300_QL70_ML2_.jpg

[west breach]

movie poster for ...

prompt is pretty simple

let sdxl decide what the movie is about

[ionic gulch]

guess a little img2img percentage would help to get the original style

[west breach]

of course, the longer the title, the less likely you are to get a good result 🙂

[west breach]

lisa frank and Keith Haring make a good combination. neon colours from lisa and thick black lines from keith

[icy brook]

"This is fine!"

[west breach]

lisa, you're tearing me apart

[upbeat summit]

it can't pop off any more than this 😄

[west breach]

it's pretty intense. never heard of lisa frank before. it's cool discovering artists this way

apparently she was huge with stickers 20 years ago

the dogs barking i circled reminds me of something that used to be on a lot of clothes in the 90s, but I can't remember exactly what it was. maybe Billabong?

[analog roost]

Pity it doesn't fit with the theme of our event, but damn this image is beautiful

[west breach]

I remember now, it was Mambo clothing 😄

[native knot]

Those days when you stare up into the cloudy sky and you see things formed in the clouds...

[indigo carbon]

where do I get eye bleach?

[native knot]

I think I stared too long...

[indigo carbon]

got the holy water

"jar of magic eye bleach"

[icy brook]

Aether FIRE is live!

https://civitai.com/models/157594/aether-fire-lora-for-sdxl

[visual glade]

make sure you don't have any custom nodes

[uncut fiber]

Time to dusting Christmas things 🙂

[native knot]

I've also had crazy inconsistent performance lately. Like, I'll be able to run several generations just fine and suddenly things will just crawl. I'll go from like 1.5s/it to 15s/it with literally nothing changing except a seed value.

[uncut fiber]

aren't custom nodes treasure of ComfyUI?

[noble shoal]

Will be hard to comped against this with my 1900 Baseballcard LoRA

[icy brook]

Haha. What’s the competition? 🙂 What you made there is something totally different.

[native knot]

https://tenor.com/bbtBf.gif

[noble shoal]

That's very true. I wonder if its possible to combine those. Let me give your lora a try.

Close 😅

[uncut fiber]

windows screen saver

[icy brook]

It’s fire 🙂

[noble shoal]

Your LoRA is approved.

[crisp owl]

yes but often they cause stuff to stay loaded. So the more "default" stuff you can get away with, generally the better

[uncut fiber]

i understand, but having not functional AIT, at least for me, simply is essential for me.
Does it work for you?

[crisp owl]

Nah because AIT doesn't have anything compiled for 2k series cards

I just wait my couple minutes for a full process if I like the first base result

[upbeat summit]

are you on the latest comfyui? this fork does work with it. you can use it temporarily until it is merged in the main AIT repo:
git clone -b hacks https://github.com/asagi4/AIT

• delete your old AIT folder and clone this.
• you could backup the modules folder if you don't want to download them again

[uncut fiber]

Thank you, will definitely try @upbeat summit

jump between dimensions
And animals heaven 🙂

@upbeat summit is it posible throught manager? Goint to test it.

Done will test later. Thank you

[upbeat summit]

let me know if you have any questions

[noble shoal]

🤣

[native knot]

I have questions.

how is prangent formed

🤣

[upbeat summit]

I would suggest a dictionary first

[noble shoal]

Now i am curios too

[native knot]

Can u down a 20 foot waterslide pegnat?

God, I love the internet sometimes....

[upbeat summit]

time for an oracle LM

[native knot]

Actually, what we need to do is create a model that is trained only on memes.

[upbeat summit]

dall-e 3 is pretty good with meme imagery

[native knot]

I saw...but we need that for SDXL.

[icy brook]

I think my dog is pregernet???

[native knot]

PREGANANANT!

My friends did a luigi board.. and it mentioned me! PLEAS HELP!?

[icy brook]

https://tenor.com/view/thanks-appreciate-lubalin-gif-21323706

[native knot]

How to use a wedgie board and is it real?

[upbeat summit]

ahm... yeah. brain farts

[native knot]

I'm almost afraid of what a "wedgie board" would be.

[crisp owl]

Did someone say wedgie board?

[upbeat summit]

do you sell them?

[crisp owl]

No but I'm all in

[native knot]

Can you burn a Luigi board?

[upbeat summit]

Aether Fire

[noble shoal]

That sounds like a wonderful plan. When do you start training?

[native knot]

Tomorrow. Always tomorrow.

[clever verge]

Recommendations for a node that can iterate through 100 different loras?

[native knot]

EfficiencyNodes has an XY Input: LoRA Plot node that you can point at a folder, I believe.

[noble shoal]

Wait a sec. Can't you make the LoRa an input, then connect a primitive and set it to increment?

[clever verge]

That sounds good if it works!

[native knot]

But if you do that, how will you know which LoRA without looking at the meta? The XY plot will put the name of the LoRA there for you.

[noble shoal]

But that wasn't part of the inquiry

[native knot]

lol

Trying to be proactive. My bad.

[clever verge]

both in metadata and they will be in order

[noble shoal]

And I am trying to be effective

[upbeat summit]

use a node that can call lora:1 from your prompt and use a __wildcard.txt__ file

[native knot]

I figured out the wedgie board thing.

[noble shoal]

Smart

[native knot]

The beauty of Comfy is that we can find like 6 different ways to do tons of stuff.

[noble shoal]

How about merging all the LoRas into a single one and calling them by their keywords. 🤔

[clever verge]

Worked perfectly, I don't know why I didn't think of it! Thanks! 😄

It's the same LoRA only different checkpoints.

[noble shoal]

Ha! Take this, mister Proactive. @native knot 😅

[native knot]

https://tenor.com/bI9cF.gif

https://tenor.com/bHuma.gif

[icy brook]

My little guy is saying: “F this shit!”

[native knot]

Are you burning my Groot from the poster earlier? How dare you!

[icy brook]

What? 😀

[native knot]

#✨｜sdxl message

😄

[icy brook]

Wow! That’s similar. Should try and prompt for it with Aether Fire 🙂

It should work well bc of how the latent space connects wood and fire

[native knot]

It was similar...that's what made me think of it. lol

[icy brook]

Exactly 🙂

[native knot]

Totally unrelated and random thing, but it looks like our avatars are talking to each other.

[native knot]

[clever verge]

Mix it as an image

[icy brook]

IPAdapter priorities.

[native knot]

This is my source image.

[icy brook]

I should make twelve loras of this concept.

[crisp owl]

hahaha
likely my last two images before heading out for a long weekend

[native knot]

Love it, @crisp owl

Have yourself a great weekend.

[crisp owl]

https://tenor.com/view/old-faithful-geyser-water-yellowstone-steam-gif-4936853

[crisp owl]

his food didn't arrive in a timely manner, and that's the origin of his hatred for the overpopulated universe

[icy brook]

I have to try and set him on fire.

From the love of my heart ❤️‍🔥

[uncut fiber]

Tiger tamer

[noble shoal]

Collect your favorite baseball players, like "The Fridge", "Toilet", "Sandwich" and "Hobo in a Park", soon, with my new SDXL 1900s Baseballcard LoRA. Releasing probably today.

[icy brook]

Nah, it's too far away. You could come closer.

[icy brook]

this is funny 😄

[native knot]

Because that image is too 🔥

[icy brook]

out of all the images I posted THIS gets flagged for a review 😄

and only this

[icy brook]

maybe it's the connection between something kids related + fire?

[native knot]

Maybe.

Civit gets weird.

[noble shoal]

Probably because of Intellectual Property. Detected Pokemon

[native knot]

Pfft...there's a ton of IP images on there.

They probably just wanted to catch this shiny.

[icy brook]

that's possible. but why did Iron Man slip through then?

[noble shoal]

Maybe they don't fear Marvel that much

[native knot]

Closer, you say?

[icy brook]

way too far off

😄

[noble shoal]

Introducing "Aether Air". Prompt: A photo of air that looks like a man, tree and gardens in the background

[native knot]

heh

[native knot]

Closer, you say?

[icy brook]

what is that even? 😄

[native knot]

x-ray of an eyeball

Well...one of SDXL's infinite interpretations of one.

This one is probably closer to real.

[icy brook]

ugh it's getting gross 😄

[noble shoal]

Gross? I have a Lora for that

[native knot]

Let's just zoom out, then....

[pure crystal]

I trained an SDXL lora person, same dataset I used for 1.5 dreambooth, the SDXL only learned the person's lips and that they're bald.

[native knot]

lol

[pure crystal]

it's pretty funny, the lips are perfect in every gen but that's all

[native knot]

@upbeat summit, you did space in an eyeball, so I did the eyeball in space.

[cyan crown]

I trained a Lora of myself for sdxl and it works

[pure crystal]

I will try again, this was first pass

[upbeat summit]

perfect

[icy brook]

"A Whole New Wooooorld!"

[upbeat summit]

one of the most iconic scenes ever - and traumatizing to some 😄

[icy brook]

which one?

[native knot]

T2

[icy brook]

yeah this works well!

a photo of engulfing fire that looks like thanos, cinematic, intricate, heavy grain

[native knot]

Modified your prompt.

Swapped in scarecrow.

[stone fossil]

VanGoghBrushwork SD XL 1.0: https://civitai.com/models/157780/vangoghbrushwork-sd-xl-10

VanGoghBrushwork - Transforming Images into Van Gogh's Painted World

Description:
Introducing VanGoghBrushwork, an innovative LoRa (Text to Image) model that transports your ordinary images into the enchanting world of Vincent van Gogh's paintings. With VanGoghBrushwork, you can effortlessly infuse your visuals with the timeless brushstrokes and vibrant colors of van Gogh, creating a captivating artistic experience that echoes the genius of this master artist.

Use trigger words: v0ng44g, p14nt1ng

Trained on 3500 steps from a highly detailed by hand captioned large dataset.

• Special thanks the people who help me and who I care a lot about:
  @MarkOREZ
  @upbeat summit
  @osiworx
  @mix
  @Thibaud
  @Kamikaze(Elon Musk)

VanGoghPortraiture SD XL 1.0: https://civitai.com/models/157794/vangoghportraiture-sd-xl-10

VanGoghPortraiture - Transforming Images into Iconic Van Gogh Portraits

Description:
Introducing VanGoghPortraiture, an advanced LoRa (Text to Image) model that brings the enchantment of Vincent van Gogh's portrait style to your images. With VanGoghPortraiture, you can effortlessly transform ordinary visuals into timeless portrait-style artworks inspired by van Gogh's signature brushwork and artistic brilliance.

Use trigger words: v0ng44g, p0rtr14t

Trained on 3000 steps from a highly detailed by hand captioned large dataset.

• Special thanks the people who help me and who I care a lot about:
  @MarkOREZ
  @upbeat summit
  @osiworx
  @mix
  @Thibaud
  @Kamikaze(Elon Musk)

VanGoghSketcher SD XL 1.0: https://civitai.com/models/157801/vangoghsketcher-sd-xl-10

VanGoghSketcher - Transforming Images into Timeless Sketches

Description:
Introducing VanGoghSketcher, an innovative LoRa (Text to Image) model designed to reimagine your ordinary images as captivating sketches reminiscent of the iconic style of Vincent van Gogh. With VanGoghSketcher, you can effortlessly transport your visuals into the world of artistry, creating timeless and evocative sketches that echo the genius of van Gogh himself.

Use trigger words: v0ng44g, sk3tch

Trained on 3000 steps from a highly detailed by hand captioned large dataset.

• Special thanks the people who help me and who I care a lot about:
  @MarkOREZ
  @upbeat summit
  @osiworx
  @mix
  @Thibaud
  @Kamikaze(Elon Musk)

[cyan crown]

will use them for my next YT video. .Thanks for the work

[icy brook]

that's perfect!

[icy brook]

I especially love this one

to me it looks like the early 20th century sketch art we have here on the westcoast of sweden

[stone fossil]

Yes me 2 it makes something diffirent then your every day SD output. 🙂

[icy brook]

what's your channel?

[cyan crown]

https://www.youtube.com/@DigitalArtGuide-bz3gn/playlists

[high skiff]

I don't have any custom nodes, at least not enabled.

The problem ended up being solved by turning off the smart memory or whatever it is. The thing that allows the memory to pool and then offloads it.

It made it a bit slower by having to load and unload every single model with every single gen, but the extra few seconds added on is nothing compared to the sometimes several minutes per Jen that was happening from VRAM overflow

[icy brook]

nice! 🙂

[cyan crown]

thanks.. it's a mix of tutorials and showcases

[nimble heart]

I disabled smart memory too lol

[icy brook]

great!

[high skiff]

Yeah, it just absolutely goes stupid on 24 GB GPUs. Even on low VRAM, it would weigh overfill my VRAM on my 3090 just doing single image gens

[nimble heart]

On my end smart memory would slowly increase my VRAM usage until it was barely at 100% then my desktop system would just brick cause things like firefox and my compositor would randomly fail to allocate

[visual glade]

remove all your custom nodes then you won't have to disable it

you have to remove some not just not use them because some people decided it was a good idea to hook in comfyui code

[nimble heart]

Also Comfy the latest pytorch stable supports GFX11 cards. Might be worth using over nightly

[zinc cargo]

some spells dont work out as planned...

[visual glade]

but nightly has ROCm 5.7

[nimble heart]

You dont use nightly for new cuda versions do you? I thought nightly was just so the 7000 cards worked

[visual glade]

I use nightly myself

[nimble heart]

Rocm 5.7 gains like 2% in SD 1.5 but effectively 0 in anything higher res from my testing

In 1.5 I go from 19.25 to 19.5

[high skiff]

Thats kinda what it was doing for me as well, thought I could assume it was even worse on an AMD card, as they don't have memory pooling, sos hit would just start to break for you

For me, it would do this:

"I press gen, it loads the first model to 8GB VRAM, then it finishes and loads the second to 14GB VRAM, then it finishes and loads the third to 18GB VRAM, then it finishes, and loads to 23GB VRAM, then it finishes and pools to 27GB VRAM and the last image takes several minutes"

it was just constantly edging and going over my VRAM, and my it/s would turn into s/it lol

[nimble heart]

yea idk. I have the AIT nodes installed cause I was trying to make them work a couple weeks ago but I guess I could remove them and try. Else the only nodes I have are my own which dont do any funky hooking.

[high skiff]

but just disabling smart mem and loading every model seems to be faster in the long run. I am lucky though, as I have enough VRAM to just eat the memory usage

[nimble heart]

I'll enable smart memory and remove all nodes but my own and if it still causes a GPU reset imma just assume it's fucked up

[high skiff]

I use no special nodes as of now, so I am not sure. If Comfy means just not even having them installed, then thats something I am not willing to do for the smart memory to return. The non smart offloading with memory is honestly just fine for me in order to have additional nodes as versatility 😅

[nimble heart]

Its usually during the VAE decode I noticed.

the vae decode will spike just a bit higher and it'll unload exactly enough to fit it then my youtube video turns into green barf then my compositor OOMs and triggers a reset

[nimble heart]

yea some nodes are super invasive and still run code just being installed

[icy brook]

underwater fire

[zinc cargo]

i wanted to burn jim carrey's pants, so he'll be a lier lier pants on fire...

when it didnt work, i went to play with the other toys @icy brook made for us

[icy brook]

I must admit that v1 of Aether Fire has some issues putting clothes on fire

but you could try and up the weight as high as 1.4

or experiment with fire that looks like x or x made of fire

[zinc cargo]

you shouldnt have thoguht me how to do it...

#sorry

[native knot]

Groda on fire.

[zinc cargo]

groot i am

[half cedar]

Super clean!!!

[ionic dragon]

hamsters fingers are better than human fingers

[half cedar]

Because they don't do much. (Always that shape vs human hands)

[native knot]

There's a psychological expectation that we have as humans when we see other humans' physical attributes like hands and eyes. Imperfections stand out more with human hands and eyes to us than others do. It's also part of the reason why fictional non-human or humanoid, but not human, creatures depicted in CGI tend to be easier to process/more believable and don't trigger as much of the uncanny valley effect.

[mellow tendon]

Yeah that hamster has a thumb/5th digit on one hand and not one the other.

[native knot]

It's still rather fun to depict hamsters in various scenes. I also really love the Jedi hamsters. Though I think one of these guys is about to get whooped when the others notice.

[upbeat summit]

does anyone hear their GPU making noises in the computer case when it's calculating an image? the noise (chirping?!) has the same rhythm as the progress bar moves for example in ComfyUI.

[native knot]

Yes. All the time.

[upbeat summit]

I don't mind it. just when I don't wear headphones I keep noticing it 😄

[native knot]

It's soothing. 🤣

[upbeat summit]

yes

haha

[mellow tendon]

Its called GPU coil whine, it is so bad on my RTX 3090.

[mellow tendon]

When I roop a face it makes a really distinctive sound.

[upbeat summit]

I always thought coil whine is more constant and not linked to operations, but yeah that's probably it

since I mostly wear headphones I don't really care, but if you want a very silent environment it can be very noise polluting

[native knot]

Coil whine tends to be a constant high-pitched sound. The sounds that I think masslevel and I are talking about are a bit different.

[mellow tendon]

Maybe it's the AI's sentience screaming out for help!

[upbeat summit]

I had some hardware with coil whine over the years and it was different - very annoying high pitch sound. this is more like *chirping * in rhythm of the progress bar when some compute heavy stuff is being calculated and it stops immediately when it's done with the task.

[half cedar]

Bing.

[native knot]

Which is what I hear.

[icy brook]

yes, I thought it was me putting together my pc in a bad way 😄

[upbeat summit]

good to hear I'm not the only one 😉

[icy brook]

it's a really weird thing though

but is it really the gpu?

[native knot]

I'm pretty sure it is.

[icy brook]

weird

[opaque rampart]

https://civitai.com/models/157787/pentellion-xl

[zinc cargo]

i cannot hear my gpu as the fans are too loud

play safe kiddos!

[grave smelt]

When browsing community feeds, i see a lot of different ways of writing prompts. Specifically how people seperate the commands. Some just write (thing, thing, thing, ...) and some do (thing. thing .thing. ...). Is there an advantage of using period over comma or vice versa? Is it same?

[icy brook]

nice stuff!

[zinc cargo]

thanks @icy brook I've added a few to civit

if only smoeone would just made a civit node for comfy- where i can just click and upload that image 🙂

[icy brook]

hehe. this is why I have only posted kids made of water so far 😄

scary kid 😄

[zinc cargo]

burn victim Lora is not a good idea i guess...

[noble shoal]

🤔

[upbeat summit]

@native knot I pointed my dynamic mic at my computer and captured the GPU noise while generating an image in ComfyUI 😄

I pulled up the volume a bit (-12db). time to make some music 😉

[native knot]

Yup...pretty much that, though when I hear it from outside the case, the highest tones aren't there. But same same.

[upbeat summit]

yeah - that is really upclose to a case opening

[zinc cargo]

sending this to GPT, asking to decifer the noise and steal that prompt

[noble shoal]

Is that the beginning of "future breed machine" from Meshuggah?

[upbeat summit]

I mean stranger stuff has been captured from different electrical signals

[nimble heart]

Additionally I've found different samplers produce different frequencies. In the llama world, AutoGPTQ produces a sorta squealing, while ExLlama is just a very soft thrumming as the tokens stream.

[upbeat summit]

that is a 50 steps latent upscale using dpmpp_3m_sde_gpu

[noble shoal]

Use this as prompt on stable audio

[nimble heart]

if you play games you can change the pitch by adjusting your FPS cap too

[upbeat summit]

yeah I noticed this as well. my first 30 step pass almost makes no sound. but the 50 steps...

[nimble heart]

it gets higher pitch the higher your FPS

[upbeat summit]

yup

[zinc cargo]

@upbeat summit can you peak at your DM for sec please?

[nimble heart]

I typically don't hear it at all until like 150 or so, but when a game doesn't cap frames on the main menu and it spikes to like 1000 then it's just WHEEEEEEEE

[noble shoal]

Hey if you have a tip for him to get rid of the sound, i am interested too

[native knot]

Uh-oh...someone's slidin'!

https://tenor.com/w26E.gif

[upbeat summit]

that's a secret - sorry

[noble shoal]

Just use another outlet for your speakers. That might solve it

[upbeat summit]

the sound is coming out the computer

[noble shoal]

The Fuck?

[upbeat summit]

this is not analog noise as you might expect 😄

it sounds like the audio signal flow isn't isolated correctly

but it's not

it's the gpu itself - the hardware makes that sound

[noble shoal]

Is is your powersupply?

[native knot]

The call is coming from inside the house!

[upbeat summit]

hahaha

[nimble heart]

High power GPUs can literally resonate and produce audible noise

[native knot]

It's the GPU for sure.

[noble shoal]

I think it is saying: Drain Drain Drain Drain Drain Drain

[upbeat summit]

yeah. my ears and my dynamic mic says it's the GPU hardware