✨｜sdxl

[west breach]

Is it fast?

[tepid surge]

i only put on the show so you can see me

[vast galleon]

the color transfer takes 7.53 seconds on a 3060Ti. as far as GroundedSAM or OSX I don't know yet.

[tepid surge]

Okay I done with my soliloquy

[vital ermine]

How does one know if a training is releasable?

[tepid surge]

when... when it dus that good good

[vital ermine]

Problem is we just never know that

[glad fulcrum]

The Lora needs to be loaded in the negative prompt or positive?

[tepid surge]

dat good good

also general awareness as a cat cx

[elfin flare]

true

[hoary saddle]

anyone getting big differences in results when running in comfyUI in web ui vs running with websocket script?

or know reasons?

[wet nacelle]

https://tenor.com/view/ksi-yes-gif-26070625

[hardy cipher]

trying to create beauty

[wet nacelle]

@ me when you get something you want to show. I want to see it in the morning. I gotta go to bed. Night.

[hardy cipher]

indeed. goodnight, sir

[hardy cipher]

lol, wat?

[vital ermine]

LOL

[hardy cipher]

neat

steampunk war blimp?

[vital ermine]

yeah, but sort of looks like a bomb going horizontal

[elfin flare]

Oh you dont get it

That is a steampunk bomb bot

it flies to a destination then nukes the entire area

[hardy cipher]

yes, exactly

[vital ermine]

OH? Like FO but steampunk style?

[elfin flare]

Yeah.

[vital ermine]

big boom boom

[hardy cipher]

man, I broke fallout new vegas. slaughtered an entire settlement for the baddies, while inadvertently clearing it entirely of oversized bees or something. so I was able to be good with both factions at the same time.

thought you guys would want to know about that

[broken hare]

ok, ive clearly done something wrong.....anyone know when I run "Revision" my input pics have no impact and i get just random pictures

[vital ermine]

Only one, besides 76, I never played

[hardy cipher]

it was strange

playing for both sides, lol

[vital ermine]

so far I am liking this locon.

[soft zealot]

looks more like an issue specifically wth Ksampler Advanced.

I see this same issue (shown in Console) with Ksampler advanced configured as shown, if I changes to 200 steps starting at 194 then corrrectly does x/6 rather than x/40

If I trasnsfer the setings to an efficiency NodesKsampler Advanced than at 300 Steps starting at 294 it correctly shows 6/6

@visual glade @ionic gulch (as I justsaw summat else)

[high skiff]

Me?

[hardy cipher]

damn, are those guns shooting into each other?

[vital ermine]

looked like it

[hardy cipher]

how is this donald trump, robert z'dar, and snoop dogg?

ridiculous

[soft zealot]

heres are some bullets colliding
https://www.youtube.com/watch?v=KBdvY_4dAfk

[hardy cipher]

deflection

[high skiff]

there is an even more incredible video where bullets completely stop mid air

[hardy cipher]

didn't they find bullets that hit after some war battle? or maybe that was a bs story I read

[soft zealot]

from the Slo MO Guys? I thinks thats the ones I meant to search for but im onlyjust out of bed lol

[soft zealot]

there was an incident allegedly duringthe US civil war

[hardy cipher]

man, being able to make slowmo videos like that and make all kinds of money from it would be a pretty sweet gig

[high skiff]

https://youtu.be/tcQVrD7RnNI?t=1322

this one

[soft zealot]

Destyn!!

[hardy cipher]

that was brutal war

[high skiff]

they got perfect transfer of energy

[hardy cipher]

think of the odds of tha thappening

[high skiff]

the bullet stops dead in the middle

its one of the most incredible things I have ever seen

it looks soooo fake haha

[hardy cipher]

dude, yeah

[high skiff]

all of the others deflected, but that one stopped dead

[hardy cipher]

very satisfying to watch tht one

[high skiff]

assless chaps lmao

[vital ermine]

haha

[high skiff]

getting ready to ship the beta feedback version of my new high res fix workflow to some peers

[vital ermine]

I didn't even prompt that

[hardy cipher]

one on the right needs to do some squats I think

[vital ermine]

LOL, yeah

saggin has new meaning now

[hardy cipher]

man, this combo is just making black dudes. no trump or z'dar influence at all. even turned snoop down to 60 percent

[vital ermine]

Seriously, I am unsure how to really give my lora a work out.

[hardy cipher]

have an llm write you a long crazy prompt

[hardy cipher]

alright, now it's starting to turn a bit

has anyone else noticed when you double click on a value in comfy to type in a number it doesn't automatically have the text box or whatever you'd call it activated?

[ivory blaze]

Why does SD refuse to generate scuba tanks? I mean.. just saying.

1.5 nor XL will lol

[hardy cipher]

I mean, I don't know that's not a scuba tank

it does say ssuba

[ivory blaze]

yeah that looks more like an AC compression tank tbh with the insulators and

I mean and scuba tanks are standardized world wide, they are very easily recognized lol, it literally WILL not except in very rare and alternate media types (not photographic) might make one that looks okay...

I have 1000 images today with so many prompts just trying to get one photograph of haha, I've become obsessed.

[hardy cipher]

hmm, well seems like it's doable somehow

[ivory blaze]

and this? I asked for A scuba tank... it gave me nine NOT scuba tanks lol it is playing with me now 😦

[hardy cipher]

no scuba tank loras it seems

how weird

[ivory blaze]

I've tried different tank names, styles, naming, different tank types / o2 oxygen acetylene and so on .

[ivory blaze]

well I mean it may not be so common that someone took the time to train one, it's just weird it has no idea what it even is...at all

[soft zealot]

aqualung?

[ionic gulch]

Workflow 4in1 + Rev v3.0 beta - improved Revision mode with two input images
Example 1 - Revision Mode OFF - dog like prompted
Example 2 - Revision Mode ON - no TXT prompt (just mixing the two images)
Example 3 - Revision Mode ON - positive TXT prompt (turning the image to a dog picture)
Example 4 - Revision Mode ON - positive + negative TXT prompt (removing the neon color)
https://i.imgur.com/NzM6cDt.jpg
https://i.imgur.com/28VWm1i.jpg
https://i.imgur.com/3Ce7hG6.jpg
https://i.imgur.com/ycztf5Q.jpg

[hardy cipher]

nah, I was being sarcastic there. but you are right. I wonder if there's another term it'd prefer?

seems you know more about it than I do

[ivory blaze]

yeah uh... no... god no aqualung:

[ivory blaze]

it happened today when I was trying to just whip up a face replace with the kid in a scuba set, but... no tank, ever. then I tried on adults. nope. nothing. and now I''m here... wondering wtf cos it isn't as if it is that uncommon, it is well known activity

[ivory blaze]

yeah so, aqualung, is... I have no idea what the heck is going on here how that is "aqualung" related in any sense...

[hardy cipher]

lol. wtf

[soft zealot]

lol

[vital ermine]

Jethro Tull would even do a wtf with that.

[ivory blaze]

yeah and this is another "aqualung"

[hardy cipher]

how is z,dar, trump, kim jong un a black man?

[vital ermine]

Damn, that is close to one of his album covers from the album with Aqualung.

[ionic gulch]

find yourself a korean wife and find out

[hardy cipher]

https://tenor.com/view/hahaha-dancing-dance-kid-child-gif-5610419

[ionic gulch]

only the pool boy knows :)

[hoary saddle]

cnet decent so far, canny

[high skiff]

I need some community feedback

would you guys rather have a smooth and a high detail high res fix that are fast

Or three options: Smooth, high detail, and a slower texture enhance high res fix? So 3 options in one workflow

[steel mist]

If the quality's better, worth the wait

Assuming it can just be bypassed if you're not looking for final quality

[crisp owl]

Depends if the 3 options are easy enough to figure out even for more basic users or not.
May be better to have a single streamline, but, options are better.
I like options personally

[high skiff]

its just as simple as connecting one node or the other

I am beta testing the workflow across some peers

I only have some screenshots

[hardy cipher]

options are always better in my world

[crisp owl]

Node connecting is simple enough, just as long as it's obvious for those users.
The annoying thing comes when people constantly inquire "how does X work" when it's right in their face

[high skiff]

base 1024x, old (soft) high res fix, fast (detailed) high res fix, textured (slow) upscale

you can really decide the level of texture improvement

Here, let me send a full res comparison

ICAt to the rescue

base 1024x, high res fix 1.0, high res fix 2.0 (fast), high res fix 2.0 (slow)

[steel mist]

slow looks worth it but I also dunno how slow you're talking

[hardy cipher]

nice work

[high skiff]

there is an unfortunate trade off that I am trying to balance

[ivory blaze]

see, scuba tank is meaningless to all things CLIP

[high skiff]

the 2 on the right side greatly enhance details and textures, but they also often leave noise/grit in smooth areas

Not always, but sometimes

[ivory blaze]

Literally... red ball blue box

[high skiff]

I am trying to find a way to fix it

[west breach]

add extra steps to the end stage

[high skiff]

its not something that can be tweaked unfortunately. I am leveraging a really broken and messy process, which works pretty well with how I am doing it, but I need to figure out how to eliminate those small issues

it has to be a specific sampler, scheduler, step count, crossover, everything

change one value and it explodes

thats why the fact I got 2 different ones working is interesting

I have some ideas to fix it

[west breach]

increase the total steps on the second sampler to smooth out residual noise

[high skiff]

I had to test like 300 different sampler/scheduler pairings to get ones that worked haha

[ivory blaze]

Earlier i queried with a1111's clip interrogator with this image and it said "a blue propanent on a white background with a clipping for a clipping for a clipping for a clipping for a clipping for a clipping, Aquirax Uno, blue, a stock photo, plasticien" I mean what is wrong with scuba tanks and AI images?

wasn't much better than your's response

[high skiff]

I guess I could do that, tho its not likely to fix anything based off of how this words

[high skiff]

technically, only the context makes it a suba tank

its just an air/gas tank

[ivory blaze]

the text repetition is like a bad setting in an inference llm...

[sour obsidian]

its just the old clip interigator isnt that great. Sounds like early blip that tended to repeat itself. This took all of 3 seconds to inference the response here

[ivory blaze]

not true, it has different connections, colors and such.

[west breach]

extra step added to total steps on the second sampler

[ivory blaze]

yeah still shows that scuba tank is not realized, And @high skiff I understand that, but even on a diver's back, it does not properly recognize

[high skiff]

I know what you mean, but its not something I can just do here

[ivory blaze]

nor can you get it to generate an image with scube tanks lol, except norman rockwell looking illustrations for some reason

[high skiff]

I have to used a very specific pre-determined split for this to work, which means every step I add is just adding more to a really low resolution area

[sour obsidian]

want an even harder challenge. Try making a crowbar

[ivory blaze]

harder? hrmm, mmaybe equal... also aqualung? Lol

[visual glade]

clip interrogator is using blip first then using clip to figure out what matches best right?

[vital ermine]

Now I get to try and learn this ait

[ivory blaze]

@vital ermine open it in browser :p zoom WAY in

[west breach]

14/15 on stage 1, then 15 to 32 on stage 2

[vital ermine]

early 2000s Anime feel.

[ivory blaze]

they're also around 8k reso

[vital ermine]

I wonder if AIT is worth it? I downloaded it in the manager

[ivory blaze]

what is that

[heady vale]

closest I got to a tank. But that was 1 out of 5 gens

[vital ermine]

AIT is supposed to be super fast

[ivory blaze]

yeah I got things like that occasionally, it looks more like an air compressor you'd find in a gas station lol

And It WILL not put it on a person's back

it does this strange tank a lot, that looks like a gas stank, that would be on the botton of a car.

[ivory blaze]

here is an aircompressor on someone's back lol

[vital ermine]

is there an on and off button for loras in comfyui?

[crisp owl]

comfyroll has one

[ivory blaze]

now if you try to make some norman rockwell illustrations - it WILL totally make a scuba diver

[vital ermine]

what is a cr load?

oh, comfyroll? on it

ty

[ivory blaze]

even though it kind of looks like a baby bottle more than an air tank

another doing similarly, same kind of norma rockwell style affect, with watercolor

from another guy i was discussing this earler

[autumn forum]

be careful when using that. ive made the mistake of turning it off and leaving the strength at weird numbers and it does actually change the image with it off,

[crisp owl]

I hardly ever use any lora's for styles or characters
But, saying that, I do use the offset noise 100% of the time

[ivory blaze]

scuba tank does this too sometimes lol

[fathom hound]

quality 100%

[ivory blaze]

XL makes the best things

[vital ermine]

Yeah, it is weird

[vital ermine]

Dorito Puffs OD. Sad

[ivory blaze]

yeah, wasn't quite what I had in mind, but, close, still it kinda turned out, different

[vital ermine]

trying to use control net in comfyui, so I load it, click queue prompt and an actual crash error !!! Exception during processing !!!

[uncut fiber]

i got just now from nowhere "ZeroDivisionError: float division by zero"

[vital ermine]

ugh

[visual glade]

make sure your controlnet is compatible with the model you are using

[vital ermine]

well, this loaded my screen but the first one is "ComfyUI\execution.py", line 151, in recursive_execute
output_data, output_ui = get_output_data(obj, input_data_all)"

I only have the base model

and those 700mb one from HF

Oh, darn thing it reverted me to 2.1 model

I just spotted it

Thanks @visual glade

[indigo carbon]

does anyone here know how to modify .json files? I finished compiling AIT XL and XLR modules for all gen params (including batch sizes), can't figure out how to get modules.json to see them.

[vital ermine]

Oh, lordie I grabbed the download for ait but I know zip about it

[west breach]

total steps: 30/31/32. Sometimes extra noise works well, other times it doesn't. Let the user decide

[uncut fiber]

Hi comfy, restarted and getting it still. Some mentioning of AIT, but i am absolutely noob. Will try for curiosity changing prompt.

[vital ermine]

which one of these is the merge two images to make a new one? sketch?

revision

[crisp owl]

yah revision

[vital ermine]

that is like the MJ tool

[crisp owl]

ish

[vital ermine]

aye

[indigo carbon]

if anyone here can, DM me. much appreciated

[ivory blaze]

oh yeah, diving suits. But no scuba tank, also... whatever this is ]

[indigo carbon]

this is the only thing holding batch size AIT support back.

[noble shoal]

JSON files are plain text. You can just edit them with notepad, if that was the question.

[indigo carbon]

not as simple. it needs clarifications specified I can't figure out how to do. I had to write a new AIT compile script to get it to do BS, it compiled after a few hours.. can't get the node to see it..

[noble shoal]

Ok, then I am out, sorry. Did you asked chatgpt or bing chat by providing the original file? Maybe you find your answer there 🤔

[indigo carbon]

did try, they basically said "skill issue", they didn't give me any explanation except typical "ReStArT Pc" bullsh!t

useless LM models

[noble shoal]

That's a bit harsh of them. If you already edited a json file, you can run it through an online json check (or whatever that's called). You might miss a curly bracket or a comma somewhere.

[ivory blaze]

use a proper editor, gvim or something, so you don't leave hanging opens

[ivory blaze]

really? they've never said that to me. They usually will give me an outline of how to fix it, and then I'll say show me how to fix it, then they do

that step of them giving you instructions or outlines, is normal, but they'll complete what you ask, if they say 3. modify json to be correct

say "could you complete 3 for me"

done done

they've always done that.

[vital ermine]

sounds like AIT is not yet, really, ready for prime time only for the ones in the know and ultra power users.

[noble shoal]

Yeah, I also never had a problem with bing myself. It even figured out how to write complete comfyui nodes.

[indigo carbon]

I already sent the compile BS solution to FizzleDorf, it's a matter of time until it's added to the node

it was able to do BS using test script, and it has a way wide enough range

[vital ermine]

I downloaded that through the manager but if you are going through hell with it then I am going to stay away

[indigo carbon]

I'm going through hell with it because I'm helping improve it. BS1 works flawlessly right away

[ivory blaze]

that worked well...

[vital ermine]

is there a walkthrough page I can set it up with?

BS1 I presume is batch size 1?

[ivory blaze]

...

[noble shoal]

Ohhhhh, I was the whole time under the impression that he was cursing 😅. "BS solution" "get it to do BS". Now that makes more sense now.

[ivory blaze]

also whatever asshat decided backspace should delete nodes needs to be flogged.. .

[vital ermine]

That has bitten me more than once and no undo

[ivory blaze]

yeah it is REALLY stupid.

delete exists, why map to bkspc

especially when there is a lot of text entry in the system

[vital ermine]

at least in davinci resolve I can undo deleteing a node and it reattaches all the node points too

[ivory blaze]

I should join the comfy discord and complain about backspace for an hour and a half

[vital ermine]

Love its node based system which all programs should strive to become.

[ivory blaze]

I use olive, cos I like torture

actually it isn't that bad

OSS too.

[vital ermine]

I use olive oil because it likes the lotion

[ivory blaze]

Olive is an open source node based video editor

it's not, polished well but it works not bad

[vital ermine]

I don't like black magic at all but the devil his due resolve is fantastic

[ivory blaze]

i know, I can tell you enjoy it,

I just use img2img + ffmpeg command line

cos...ease of use

[vital ermine]

What I love is that I can just use the damn thing like an auto1111 or if I need the power I can get into it and use its nodes for a comfyui approach not one or the other but both

[ivory blaze]

and I'm a nerd who knows every command line -switch I will ever need

[vital ermine]

get in, do the deed, get out or get your hands dirty your choice.

[ivory blaze]

yeah,

Olive I can just go a step beyond and fork it and edit it

[vital ermine]

I like fairlight as well but the other one that starts with an F (in its own package) no, it was slow and crashed a lot

[ivory blaze]

and maybe people will like that I added a midi player to it, not for the editing itslef, but just to play one of ten midis I added from the sound track of Quest for Glory 1 and 2, from Sierra online in the mid 90s. So they can be enjoyed while editing nodes

[vital ermine]

everyone complained about it

[ivory blaze]

can't do that with resolve, unfortunately

[vital ermine]

Did you see the new cameras from them? WOW, they have AI in them and will remove the background, etc... on the fly in real time

I bet they are hellishly expensive because they are made for pros. 10k + cameras

[ivory blaze]

I just use a filter app made by some guy in romania that does something like that, you just have to watch an ad for something called wifebeater simulator by popcap games, twice and it'll work for about 20m before another set of ads. Same kind of thing,

[glad grove]

what is a wife beater simmulator

[ivory blaze]

I dunno, I'm'' kidding about the app, but there is annoying phone ad where it''s like some thing and shows decisions you can make, and she says something to the guy and he throws her out in the cold rain, and she becomes homeless. it's some game it kept advertising to me when I let my premium youtube run out on accident

a phone app I think

[glad grove]

oh yea i know those weird edgy phone game ads,theres also one where a guy kicks a pregnant woman in the stomach,

[hardy cipher]

I want to know who came up with that ad format

and why it's used for games that have nothing to do with the ads

[glad grove]

prob to fool ppl into downloading the game,even if they only open it for like 10mins they get money for the ads

[ivory blaze]

comes home, another woman there, gotta go,

[ivory blaze]

kinda wtf... there's another where it had soimething like this, and then it was like "do you choose, take revenge?" ad chose that, and she was in the house with the man and was looking kinda ded in a lazy boy while she was cooking eggs or some such, I was like uhhh wtf

[glad grove]

ai generated ads

[hardy cipher]

I think they're mostly made in china or something. and some of the stuff doesn't translate very well

[indigo carbon]

made some changes to the AIT workflow, which is better? right is new and left is old.

[hardy cipher]

clipvision does strange math

how do those three make the black gentleman?

[glad grove]

maybe put black in the negative

[hardy cipher]

yeah. it's honestly made a black man every single time. different combos of people. I don't get it

I was trying to use the zero out conditioning and only base the results off the input images. but I guess that's not going to work for Kim Jong Trump Z'dar

[glad grove]

try with only zdar and trump

[hardy cipher]

whoa

[glad grove]

real

[bronze arch]

its only been a year with sd release, imagine the kinda finetunes we gonna have in a year with sdxl lol

[hardy cipher]

well I originally had snoop dogg in there instead of kim jong un. I'll do turmp z'dar now

turmp

he has double teeth on the bottom. this is a future human

[hardy cipher]

lol

that doesn't even make any sense

I guess I have to use that negative still. so weird

[vital ermine]

this revision is not acting like I thought it would

[hardy cipher]

I've attempted 12 different combinations of people now, and every result has been a black man

[vital ermine]

woke af, heh

[hardy cipher]

and just using run diffusion with noise offset and photorealistic loras. maybe they're contributing to it, but still

what's it even doing? I'm not upset, just doesn't make sense to me

[vital ermine]

what do you get when two white men marry? A black man. :/

it really doesn't make sense how

[hardy cipher]

put [[[black man]]] in the negative and

that's not what a trump-z'dar love child would look like

maybe in 90 years

[glad grove]

extra white

[vital ermine]

damn

I have not had success with merging with revision

[glad grove]

put old ma,old person in neg

[tepid surge]

imagen being 90 y/o.. wtf

https://tenor.com/view/idubbz-measuring-tape-idubbz-tape-measure-meme-gif-21525107

[hardy cipher]

I'm just crossing my fingers ai or something figures out how to stop aging before I hit that point. or at least do some altered carbon sort of thing

[autumn forum]

i dont think revision is really meant for merging, just variations based off input images.

[hardy cipher]

well should I use unclip models then? I thought it was for merging

[vital ermine]

maybe a transfer of our collectiveness into a new ai entity. I could go that

[tepid surge]

i just wanna have like 10 years where i'm like actually alive cx

[hardy cipher]

that's what they were saying? or maybe I misunderstood

[vital ermine]

we go to sleep then wake up with a new android body yet we feel no difference

look in the mirror and do a wtf?

[autumn forum]

well it can merge. im not saying it cant. but it probably would have a hard time merging the likeness of 3 different people lmao

[tepid surge]

eternal robot pain noises cx

[autumn forum]

as for the black man thing. idek lmao

[vital ermine]

No pain which is why I would go for it

[hardy cipher]

well not saying it'd be on point. but doesn't it take the attributes of each image?

[hardy cipher]

yes. this is what we need

wake up as this guy

[vital ermine]

Oh, hell no

[hardy cipher]

look at his double teeth on the bottom

[tepid surge]

live forever as them, but you get to keep the glasses cx

[hardy cipher]

heavy eyeliner on this one

[soft zealot]

Its also no good for those of us on older cards (and no the answer is not "buy a newer gen card", its should be "we the devs will ensure backwards compatability" with NON AIT workflows)

[hardy cipher]

it took twice as long per iteration for me with that ait thing. maybe it's not meant for me

[vital ermine]

Yep. My 1060 is retired now and I am suffering as my bill for this card gets paid on 6th. Wish, I didn't have to do it but it was time and the 1060 was limping about to die. I turn on my system and it would make clack clack clack really fast sound each time. Was a fan near death and third set of fans already in 6 years.

That merge seemed to work...sort of

[zinc cargo]

why so strong?

[soft zealot]

I would prefer Hardware failure than (semi) planned obsolscence due to lack of support ;o)

[hardy cipher]

https://tenor.com/view/flexgirl-muscle-girl-gif-18037996

[vital ermine]

1 str for both

[zinc cargo]

last night i had something similiar, likt was stuck at certian neighberhood in the model and didnt want to go out... @hardy cipher

[vital ermine]

Planned is how Nvidia, or AMD work. Maxwell is now dead for any updates, Kepler before it, and rsn Pascal will get axed.

[zinc cargo]

i got these, even though i was using a lora trained on someone who looks entirely different

[hardy cipher]

it's all black men unless I put that in the negative, then it's outlandishly pale white men

[glad grove]

literally my uncle

[hardy cipher]

indeed

[soft zealot]

Aye and hopefully the devs of the various UIs used in this field we're currently camping in don't bring that date further forwards then Nvidia/AMD do

[zinc cargo]

this is the guy from the original Lora though

[peak dove]

"I just hate hair and six-pack on titties!!!" 😄

[proper depot]

I had the same problem. I really had to fiddle with stuff like 'dark olive tan skin' to get the right complexion

Even then, it's still hit or miss sometimes

[hardy cipher]

well I'm just disappointed it can't be zeroed out conditioning going into the images

[soft zealot]

looks to be of some sort of indian ethnicity to me

whoever he is

[vital ermine]

I fear they may. Already 24GB is barely enough to train on so go buy a 48GB card or pay for the cloud.

[zinc cargo]

we almost all are in a way 🙂

[hardy cipher]

either way, not making a Trump Z'dar I guess

[ivory blaze]

@peak dove will that prompt work?

I just hate hair and six-pack on titties!!

[zinc cargo]

i had the same thing with a few other sdxl loras

[ivory blaze]

I'm going to go with that

[proper depot]

Just FWIW, you can use 'South Asian' to refer to India/Pakistan/Sri Lanka/Bangladesh/Tibet

[vital ermine]

@ivory blaze mixed yours with my cockpit

[zinc cargo]

that's gonna be my new pickup line

[glad grove]

nice cockpit

[soft zealot]

Ok so there are 2 sorts of users of SD.

There are "power users" who like to develop and there are "Joe Public" users who want to simply generate pretty pictures.

Currently the latter can certainly comfortably go back as far as a 1080ti even with SDXL (and I understand that there are others out there using slightly older succesfully).

It's the "Joe Public" audience I want to ensure doesn't get alienated.

Sorry Power Users but I cry for you haveing to struggle to train on 24Gb!!

[soft zealot]

whilst you can I'm sure you also missed several companies out there.

To me the larger gentleman (see I didnt say fat) from the Lora looks of some sort of indian (with a small i not a large I) origin.

[vital ermine]

Without the power users training you would be stuck with just base. You might like just base but most love the models the power user train for them to use.

[soft zealot]

I will concede to that request 🙂

[glad grove]

plus u can make money if u make pretty/beautiful/unique models,just sell them

[zinc cargo]

he's a local kids star from Israel.
(he's a very cool cat with jack black kind of energy)
he's Jewish north African descendent, guessing Iraq or similar - but like i said lately i had a lot of generations being made very "south Asia"...
mostly with nightvision model i believe

[vital ermine]

I wish I could but I suck at such things when money is involved. I always have.

[indigo carbon]

wdym "sell models" never heard of any site selling finetunes

[peak dove]

SDXL "does not know what six-pack is!!!" LOL

[ivory blaze]

lol

it was a joke

was suggesting that telling it you hate it, would make it spit it out

it also has no idea what a scuba tank is

[soft zealot]

and if I saw one I would steer away from it. Same as these people "selling" prompts

[glad grove]

some paid sites have their own private models,they hire ppl who make good models to train a specific style model

[tepid surge]

If models are being breed together that'd be a copy right trial waiting to happen

[soft zealot]

SIGH

[ivory blaze]

I dunno Astra

[glad grove]

you have no power here

[ivory blaze]

Copyright is kinda... ephemeral with AI and training data

[soft zealot]

thats a load of bollocks as the core licence allows sharing and deriratives

[tepid surge]

i'm just gonna post this and sip my tea siip

[ivory blaze]

the byte for byte model may have copyright. but it isn't' a plant

I'm going to post this and see if the guy will sell me his Mogwai

[tepid surge]

mentions a point
chat disliked that cx

[soft zealot]

the models themselves cannot (I believe) be copyrighted. Where thedebate opens up wider is wthere generativtive works made using AI Models can be copyrighted and thats far from a simple answer

[glad grove]

lets talk about cute girls instead

[soft zealot]

well you could aat least have posted a cute girl ;o)

[glad grove]

😔

[ivory blaze]

True, but possible to edit it in a few ways to make it "'your own" under the current copyright laws, as we see with videos of copyrighted material which is discussing it or changing it in some way

[ivory blaze]

I want to bonk the heads of the weirdo artists who think that it is "their" work that made the model anything, no, it's everyting and all of it, remove your little yellow splattrs and two red lines abstract art, and the model will not even change a bit

But it is ASSUREDLY modified to such a degree that their copyright cannot be retained within the model

[soft zealot]

TBF unless you're in a business selling shed loads of generated works etc then on the whole for most "Joe Public" ussers copyright isnt an issue

[soft zealot]

in reality artists /celebs etc would have to make an individual copyright claim for each instance where their copyroght has allegedly been breached so that each case can be determined on its individual merits

[tepid surge]

||controlnet works||

[peak dove]

Pure AI work is treatable as Public Domain/CC0 Copyright. Anyone can use pure AI Generative Art - irrespective of who has generated it! Derivatives of AI Generated Art is more difficult - how far have you added value by modifying the original AI? Modified AI is a minefield when it comes to ownership - but pure AI is free-for-all!

[indigo carbon]

SDXL is trained on LAION, which isn't copyrightable. idk why people think they can copywrite AI models

[ivory blaze]

@peak dove it is not a minefield, it assuredly becomes copyrighted with signifigant change to differ it from the original

@indigo carbon it needs no (c) to sell it though

[soft zealot]

as Isaid, average Joe Public user doesnt give a toss

[ivory blaze]

@indigo carbon as well, see if a big company won't fire a legal letter off to joe christmas who is passing around their uncopyrightable AI model

and see how no one cares and even a suit he'd win hands down, he proabbly won't want to pay an attorney and deal, basically a SLAPP but for distribution of open public domain material

I bet we see some of that in the coming years

[peak dove]

e.g. I make a graphic novel. The narrative, typefont, layout of this publication © me - of that there is no doubt! But! The pure AI Generated photos are not copyrightable - and so may be lifted from "my publication" by anybody - and freely used, as I can claim no copyright over those AI Generated pictures!

[ivory blaze]

if only the AI photos are

however, if you adjust them, in a manner which is different from the base you started with,

cos see here's where the thing gets... funny

normally it is the copyright owner who forces the person used their IP in some manner, to prove it is fair use

[peak dove]

Yes, modification brings them closer to being copyrightable - but test that in a court of law - you may be surprised as to what degree of modification is necessary before © you!

[ivory blaze]

however, in this case, it would be the faire use user, going after another party, and there''s no real...case law for reverse defense of fairuse ownership of this sort of thing, so

because it isn't per say fair use

but the same qualifiers to is this "signifigant" enough of a change, but without the need to prove so much as one would need in court against an owner, as none exists here

[ivory blaze]

i think the defendants in such suits will be smmall people, not companies, so the legal system will probablyh not bow to the end user, but the corps

[peak dove]

See Kris Kashtanova and the award of copyright on her Zelda Graphic Novel - which copyright was revoked by the USA Copyright Office - three months after it was awarded!!!

[vital ermine]

he who has the most money always wins

[glad grove]

yea doesnt matter if its fair use,if u are a poor person against a big corporation u already lost

[peak dove]

MidJourney paid for her lawyers to bring her copyright to copyright court!

[glad grove]

unless i punch them in the eye

[crisp owl]

Sleepy time for me.
But here's a woman and her Raven for my departure

[zinc cargo]

safe travels

[vital ermine]

Money shot here, lol

[ivory blaze]

this is how I feel about copyright law..

[peak dove]

Sharkbucks!!!

[ivory blaze]

yes

I do not know why I made these.

i forgot, I did the wrong thiing and searched for png in my output... apparently I have over 102,000 images :\

[vital ermine]

Oh, SB is about to update their logo and one was AI made and the other by a human. The AI one was so much cleaner

[ivory blaze]

so now I am finding all my very... wtf stuf

[tepid surge]

but do we "own" the depth map cx

[ivory blaze]

who needs euclidian space...

[tepid surge]

in any case, based shark memes

[strong kernel]

I like it when looks Ai

[peak dove]

© me later! 🙂

[peak dove]

Summer's day in London 😄

[tepid surge]

will send royalties @peak dove

[tepid surge]

I also came across this earlier today if anyone else wants it

it sd xl ratios/resolutions

[peak dove]

Some ComfyUI this a.m.

[zinc cargo]

that sounds like my last week

[ivory blaze]

my loopback kick is making dump lots of images

[zinc cargo]

image hoarders anonymous ...

[tepid surge]

has anyone gotten facebook's text to music up and running?

[ivory blaze]

Ends up with such awful things as this

[tepid surge]

i figure there's an audio clean up model somewhere too and the two could be used in tandum

[peak dove]

"Is that all?!" 😄

[ivory blaze]

yes

the rest is on google drive :

[ivory blaze]

If public broadcast and community shows were still a thing, I'd get my 30m and play nothing but these awful things I've made and some public domain swing music the entire time.

[zinc cargo]

10\10 would watch this (on drugs)

[tepid surge]

that was 100% a story

[peak dove]

Glittergyno!!! 😄

[ivory blaze]

I totally would if they still had such, I think they have public broadcast still of course, but not the old tv channel where you could just go and be like, I wanna do a show., and then you do

[thorny frost]

It seems Kappa_neuro's false loras have been removed by Civitai

[ivory blaze]

what was a story? the video? lol

[ivory blaze]

So the way this worked is using loopback img2img, generate, using prompt, send image back repeat, with lower noise levels. each cycle is 12, every 2-4 cycles II modify the prompt a wee bit, or with big changes. So you have control over it. now, doing 12 frames and then making over4000, it does take a minute

[ivory blaze]

cos each run is a 12 batched feedback look

[vital ermine]

Look at what is on his right shoulder

[ivory blaze]

but, So here is MY question about this. At the end, when it goes all hyper contrast colors and mayan / egyptian looking artsy, WHERE did that come from, cos... 100% the prompt was "A field of golden sunflowers" noothing more. and it went down that strange road.

here's another for when you find enough drugs @vital ermine

[vital ermine]

shit

[ivory blaze]

its a bit of a longer one.

[vital ermine]

yeah

[peak dove]

The entire input over several hours or days "stays in the background" and can erupt sporadically and in a random fashion?

[ivory blaze]

@peak dove I press generate, for each set of 12... cos I would decide if that set of 12 would be good for the half second of video i was going ot use, if not , delete and retry, or move on, then adjust prompt a littel ever 1-2 seconds

took about 16 hours for all the frames

[ivory blaze]

it was manually interacted with the entire time. So its random output seemed. really strange to me. I could not get it to do anyting BUT that.

[ivory blaze]

@tepid surge this one has more "story" like elements to it, in a way, it is "somehwat" more cohesive with one prompt shift to the next having somme relevance, if but in a David Lynch kind of manner

[peak dove]

I suspect the software owners allow certain "complementing functions" (unwanted I know!) to grace and 'promote' their software!?

[ivory blaze]

What software do you mean?

[peak dove]

What software are you using at all?

[tepid surge]

ah, i just pick up and put stuff togther in story type elements regardless of intention p naturally, i read tarot and stuff and it p much the same vibe

[ivory blaze]

automatic1111 and some customized loopback script edits, and some other little tweaks I did

[hardy cipher]

https://tenor.com/view/hmm-gif-20018917

[ivory blaze]

@tepid surge I mean it had... a purpose of sorts. like the staring skeletons and the monster, which then exploded, and left the land peaceful. It's not really so much story, I guess, just linear cohesion, weak as it may be

[tepid surge]

There's ComfiUI logic x.x https://github.com/theUpsider/ComfyUI-Logic

[peak dove]

A1111 has an incredible memory, if someone uses one of my prompts on another machine entirely, it looks nothing like my own output. A1111 seems to retain a memory of past input over a long period, which informs and colours future output, and is unique from machine to machine.

[zinc cargo]

deforum?

[ivory blaze]

?

[ivory blaze]

I think something gets kinked up in its wonky to this day color correction in img2img for sure.

[zinc cargo]

the animation

[ivory blaze]

what about it, I dunno what deforum is :p

I meant to ask, what is deforum, ?

[ivory blaze]

webui + loopback+ auto img2img move for last output + repeat + a lot of time and goofing with it

some of that, like img2img on last output was just a quickhack I shoved into loopback to push it's final image to img2img so on repeat it would not use the prior and so on

[ivory blaze]

I kept noise down aroound .3-.35 on img to img it kept it from losing pixel cohesian but was slow enough to reduce flicker signifigantly

when it loses cohesiion it becomes ugly af

that's from lizard to bird at .01 noise. until the end when I kicked up to .035

[ivory blaze]

it amuses me to see it barf into CMYK

I wonder if that is because of training images being CMYK or if the noise initialization is CMYK

[peak dove]

This is what electric cars are preventing!!! (_8^(|) d'oh!!!

[tepid surge]

uh oh, clip text encode should be the same resolution as the latent right?

[sour obsidian]

when it first starts breaking down it looks a bit like blown out latents. Likely you are pushing the latent values too high and then its ofc trying to recover but it goes into that color spectrum typically seen with it. just crank up cfg to like 30 and you will likely see similar colors

[ivory blaze]

@sour obsidian it's cfg 7 and noise 0.01 for like 1000 frames

I think 0.01 noise decoheres from the resolution of the image

and it breaks down

[ivory blaze]

cos the noise is mid pixel values probably I dunno

[sour obsidian]

the noise it adds is just pure random gaussian, basically rgb static but in latent space

[ivory blaze]

it is not RGB though THAT is what makes it interesting to me

the decoherence here is CMYK

[sour obsidian]

yea that CMYK color is what you will see often when latent values trend too high going through the VAE

[ivory blaze]

I wondering if it is NOT as it appears but is it trying to actually do its job, and using training data, that's output is indeed, decoherent, and the training data is more CMYK

ha ok

[hardy cipher]

whoa

[ivory blaze]

I mean if I just bring noise up, it'll just turn into a bird, it's just the really low noise value, nothing else adjusted, over about 150frames at 0.01 it does this

[sour obsidian]

could be in the long string of things. Seems across different vae's the same color pattern emerges with blown out latents. I actually do wonder why, since those ranges are outside the training data range

[ivory blaze]

So here. IF you lower noise and loopback the image (output becomes input) over about 150 frames, at 0.01 you see this, it is not a single interation

this is 150 full img2img generations

each fed into the next generation

[sour obsidian]

hmm yea. Could see it being a mismatch between the noise scheduler and the ammount of noise it thinks is actually being added. Then it might be trying to "remove" more noise than its adding at those real low values. That would be my totally naive guess

[ivory blaze]

if you move the noise to 0.03 wavy patterns begin to emerge.

[ivory blaze]

and it has a more fluid movement, even though it rarely will reach prompt

[hardy cipher]

that gets more terrifying the long I look

[ivory blaze]

at 28 I move from .01 to .03 and the clouds of cyan and darker blue, coalesce into wave patterns

[upbeat summit]

yup makes me uncomfortable as well - I love it 😉

[ivory blaze]

does she talk like one of those people with the voice things held to her neck

[hardy cipher]

at first glance I was like "huh, interesting necklace"

but then it got worse

[glad grove]

lady gaga new fashion style

[upbeat summit]

great image! so much to explore

[tepid surge]

im loving this xwx

[hardy cipher]

yeah, I've been using your ferrofluid thing. but with that one I did "ferrofluid, hyperfluid, ferrosmoke, hypersmoke"

kick it up the max

[upbeat summit]

nice!

elements, fluids, textures, materials - all super powerful in SDXL

[hardy cipher]

the words don't have to make real sense as long as the ai finds a correlation in it's training

[upbeat summit]

yeah. a sleek nice looking prompt is cool, but in the end the result is important and if I need nonsensical tokens or a very long prompt for that so be it

[hardy cipher]

people have weird ideas about a lot of these things. at least to me. I've read so many tutorials and I guess opinions on prompts and so many people say short is good, long is bad, etc. but maybe they're going for something different from me.

[upbeat summit]

with this tech - everything goes

there are a couple of good practices for quality and fidelity, but there are no rules

[hardy cipher]

I think that's the thing that irks me a bit

[hardy cipher]

sure, there are definitely things that are usually a bad idea, like turning cfg to 50, or 1 step renders, or things that will in general produce illegible results. but beyond that, people should really step outside of their own views and understand that it's all a matter of your preferences and goals

[upbeat summit]

yeah it can be overwhelming. so it's best to never stop learning and find your own workflows

[ivory blaze]

I have this weird built in fear that a second space after a comma, or a trailing comma, will cause my generation to be bad lol

[tepid surge]

yeees these logic nodes are amazing

[ivory blaze]

I am sure it is just superstition but

[hardy cipher]

I'm working on some new nodes

[ivory blaze]

I don't know, cos it does change the token, which could shift the weight

[hardy cipher]

we'll see how they work out

[feral glade]

sdxl controlnet lora What is the difference between rank128 and rank256 ?

[hardy cipher]

bigger number on the second one

[tepid surge]

size/resolution

[feral glade]

what ?

[ivory blaze]

is control net a lora? how is that?

or is it text to control

[feral glade]

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

[tepid surge]

it's the architecture of it but yee it used diffrently

[tepid surge]

256 is bigger, takes more v memory

[ivory blaze]

woah

[tepid surge]

I was literally just figuring out Controlnet with comfi today cx

[ivory blaze]

....that is.. nifty

had not seen, I abuse control net.

[tepid surge]

https://tenor.com/view/anime-eyebrow-raise-gif-21083414

[ivory blaze]

I tried to make controlnet figures in blender so I could animate, tween, and make the frames the blender controlnet "stickmen" what I wanted to generate in terms of placement

it worked... okay

i dunno how controlnet openpose really works, on the geneeration side, I understand it's modeling but not the way it utilizes it in the actual use

[uncut fiber]

is this image o.k. i mean technical aspect -> contrast and such.

@tepid surge can you post your workflow for CN?

[ivory blaze]

the one guy in the middle there, looks like he smells like 1980s bath toys

the weird chemical this miight not be healthy odor

[ivory blaze]

of rubbery floaties

[hardy cipher]

oddly specific

[ivory blaze]

nah, it is specific, very much so.

[uncut fiber]

😄
Should be plasticine

[ivory blaze]

hahah

it looks like mid 80s flesh colored whatever they used for rubber ducks that had that smell

[hardy cipher]

probably cancerous

[ivory blaze]

cos they painted them, and when their eyes came off, in the water, it got that fleshy color like his

they were 't actually yellow..

maybe I am still here, so

hope not

[ivory blaze]

yeah that flesh faded rubber of his eyes triggered a weird bit of memories from my child hood lol

[feral glade]

How does it affect results?

[tepid surge]

id assume resolution, idk how it works technically

[feral glade]

ok thank u

[ivory blaze]

she has an abdominal arm

[tepid surge]

deformation is a siin xwx

[glad grove]

now i wanna eat those mushrooms

[peak dove]

Some more ComfyUI SDXL

[feral glade]

did u use controlnet ?

[peak dove]

No

[feral glade]

okey

[ivory blaze]

@feral glade oh that was you he responded to in the discussion lol

okay, makes sense now

[tepid surge]

fiish xwx

[ivory blaze]

@feral glade I suggest, setting up two control net nodes, two samplers, and just run the same prompt / seed/ all through both , except one have 128 and other have 256 and see how the output differs

[feral glade]

Both give almost the same result and the processing time is the same bro

[sleek relic]

im trying to use the ReVision Feature in ComfyUI but i get some strange results in which the combined image is completly unrelated to the input images - does someone now what could cause this? Im using the clip vision g and sdxl base model

[ivory blaze]

then maybe it is not a big deal, have you tested with heavy handed movements in img2img, eg, making someone looking left, look right, and such, typically that sees huge difference in testing with mutiple models ime

[peak dove]

This is a hybrid photo - the wall pictures are made using ComfyUI SDXL - and the Pool-Table was Generative-Fill in PS, on a Public Domain photo of a room.

[feral glade]

okey bro thank u

[ivory blaze]

what's the clippy output?

those lumpy low poly lookiing putty things, might not be recognized as what they are and the combinatioin of what it thinks it sees, is, that cool race car

or it may think they're scuba tanks...

[uncut fiber]

Anybody knows what is benefit of using -> [Upscale latent by] i have same performance with or without it as well as same output.

[ivory blaze]

ookay....

[sleek relic]

im always getting completly unrelated results here a test with just two portraits - this is the complete node setup im using

[ivory blaze]

ah

[tepid surge]

what even cx faces = vehicle

[ivory blaze]

Hey so

it says G,

your outputs are all running G right, not combine

or L

hold on

I dunno how it worksl ol

[ivory blaze]

oh it may be encoded already

[sleek relic]

yes im using g i tried this one sugested here: https://www.youtube.com/watch?v=1SwDCqgXZ-M&t=110s

[sour obsidian]

why are you doing the conditioning combine?

id remove that and bump both strengths to 1.0

[sleek relic]

thank you will try that!

[hardy cipher]

I've been getting some odd results with that myself

zeroed out conditioning

[sour obsidian]

revision as it stands is a funny one, does some suprising things. I am not suprised it produces some odd results sometimes. Its very visually removed atm working on a conceptual basis mostly

[nimble heart]

geck

[hardy cipher]

it's cool. just trying to understand it. literally only made black males no matter what combination of people I used. and then when I put that in the negative it made ghostly old white men

[sour obsidian]

ok thats pretty interesting haha. No clue why its doing that but alright

[hardy cipher]

I mean, didn't investigate it that much. I realize models and stuff have an impact. but it was run diffusion with the noise offset and photorealism loras. so nothing seemed like it'd really push things in any one direction

[vale eagle]

I have tried to add image as neg condition using revision.You could use anime image as neg to produce more realistic image. Will do more test on it.

[sour obsidian]

yep this works nicely 🙂 good catch

[ivory blaze]

I wonder if you put multiple people would it stop making more than one people lol

probably negate the core, but I suspect if you used grouping like three men, and you wanted one woman, it may work

[tepid surge]

xwx

[delicate grotto]

WTF

hello, me kitty-5001
suit up, we are going to fight scp's

[shy kelp]

what do you guys recommend for sampler

[tepid surge]

dpmpp_2s_ancestral fs fs if you want a swap up from euler_ancestral

i generate several with the diff samplers each run

[west breach]

this has been working well for me

[shy kelp]

wow you really have powerful device to generate all of that

[tepid surge]

that reminds me i still need to figure out what the scheduler does

[stray warren]

try conditioning average

[tepid surge]

just a 3060, it's one after another

[tepid surge]

makes sure i get to explore a prompt kinda stuff

i still need to figure out how AITemplate worrks 😭

appreantly it speeds up generation a bunch xwx

[shy kelp]

what is that

[tepid surge]

https://github.com/facebookincubator/AITemplate

||there's nodes for it too, i just need to spend the time:|| https://github.com/FizzleDorf/AIT

[west breach]

just put it between your checkpoint loader and sampler

[shy kelp]

do comfy ui have like --midvram kinda stuff?

[wicked frigate]

it has various options like that but for the most part you don't need to touch em, it will automatically select the best for your graphics card

[tepid surge]

sd xl runs on my 3gb 1060 with comfi xUx

[shy kelp]

ohh

[uncut steeple]

What does it do exactly?

[shy kelp]

aitemplate makes generation faster?

[tepid surge]

Did it need to got in a new folder in Models?

[west breach]

nup

[west breach]

dunno, it just works

[shy kelp]

[uncut steeple]

Where is the node from and how to set it up? Its new to me

[west breach]

astra linked it just above

[tepid surge]

https://tenor.com/view/weights-lifting-weights-killing-it-girl-heart-gif-5853781

[uncut steeple]

Im blind

[west breach]

install, link, profit

[shy kelp]

its only for xl ? or both

[west breach]

XL only I think

[shy kelp]

when i linked it it shows this

what should i do

[west breach]

what graphics card do you have?

[shy kelp]

3050 laptop

[west breach]

are you using preview?

[shy kelp]

nope

[oblique palm]

guys, help pls

[tepid surge]

wow this thing has a kick to it xwx

yippe!

10 secs off dpmpp_2s_a

[uncut steeple]

I dont have to download anything else besides the node?

[tepid surge]

zoooom

[west breach]

don't think so

[tepid surge]

thank you @west breach ! didnt know all that was needed was linking it up

[oblique palm]

guys

pls

help

[tepid surge]

you ned a model in your model folder

wait model file... a model chekpoint?

[west breach]

the error says it thinks the model is a lora. is it a lora??

[uncut steeple]

Gives me weird artifacts in my setup

and takes triple the time

[oblique palm]

Can you tell me what I need to do, I'm just a beginner and I don't understand at all

[uncut steeple]

#1072220168534642768

[tepid surge]

https://youtu.be/TFqI8XdRjDw

almighty yt

[west breach]

Try asking for help in #🤝｜tech-support

[shy kelp]

i hope i can use it too

[tepid surge]

yeee ate up my v ram though

[soft zealot]

whilst it may be an interesting exercise your card isnt supported

(I have a 1080ti that is also not supported and even tried compling but she say no)

[tepid surge]

the 1060's my old card

i'm on a 3060

def tight on my v ram though with 12gb x.x

[shy kelp]

i have 3050 but laptop

[west breach]

Works on my 3060 laptop

[uncut steeple]

[tepid maple]

between 1344x768 and 1920x1080 therei s huge difference in quality

i wanst expecting it

[west breach]

I don't see any bugs

[tepid maple]

i was expecting same generation in scaling but nope

[tepid surge]

one of these days i need to get myself one of those data center gpu xwx

[tepid maple]

I just use my 3060 btw

[uncut fiber]

yes just 1080 will generate most probably wrong thing, try if native generation 1088

[tepid maple]

scaling 1344x768 with 1.43 scaling perfect

try it

[uncut steeple]

[uncut fiber]

o.k. scaling, i thought generating

latent space..

[tepid maple]

results are completely different Ithink scaling is not just scaling

it is generation by itself

and it takes double the time

[uncut steeple]

If you generate at a different resolution you get a different result

[tepid maple]

it is just normal generation almost

what scaler keeps the details 100% ?

[uncut steeple]

I use Ultrasharp but no upscaler is perfect

[tepid maple]

so none of them keeps the detail

[uncut steeple]

not all of it usually

[west breach]

I like lollypop

[tepid surge]

i was thinking A series but apparently the 4090 beats it out cx

[tepid maple]

I am just using latent nearest

[ivory blaze]

yeah that revision model doesn't seem to work well

[uncut steeple]

Im doing upscale with model for the first pass and then with ultimate sd upscale

[tepid maple]

hmm

[uncut steeple]

but my workflow only really is made for the 4090, cause it takes 80s for a 2048² image

[tepid maple]

cant go higher than 1344x768 thats the most stable one

tried alot of res

and it gives 1080 almost 1920x1099

[uncut steeple]

those are the best res for sdxl i think

[tepid maple]

yep finetuned ones are these

did you tested this ultimate upscale with scheduler ?

does it work stable ?

[uncut steeple]

For me it does, but takes some time and tweaking for good results

[ivory blaze]

definitely not working right...

[uncut steeple]

As i said I do pixel upscale by 4x with ultrasharp, then downscale into refiner and then ultimate upscale

[tepid maple]

too long for me 😄

generating hunderds of image same time at the moment.

with agent scheduler

[uncut steeple]

You can try my simpler workflow if you want

[tepid maple]

it is simple I write prompt send to Que thats it 😄

[ivory blaze]

what you never wasted 6 hours of your life for an 8k image? that's like 1.33k per hour!

[tepid surge]

such bootiful noodles 🥹

[tepid maple]

I will look at ultime upscaler later noted

[tepid maple]

damn 😄

tried generating 4k ones and I was like

helllll nooo

😄

[ivory blaze]

open in browser... and zoom way in on the little house on the bottom right

lol

[tepid surge]

math checks out cx

[ivory blaze]

there's really no need for that kind of resolution but still it was neat

[tepid maple]

I think 1080 is must if you want to share online

was going to make all 1344 but nope

too low res for people to use

[tepid surge]

that insect face is still imprinted on my brain

it's eating my soul xnx

cx

[tepid maple]

[strange mist]

that is one gorgeous woman ❤️‍🔥

[ivory blaze]

is there a way to see what data is getting pushed on a line in comfy?

specifically from clipvision's out

[ivory blaze]

I don't think this model is working right at all I think nothing is coming from those tubes...

I think the weird gens from the earlier guy who could not get it to work were just cfg

on empty

[tepid surge]

try searching for "print" or "debug" there might be node depending on what you're trying to look at

[uncut fiber]

i need a help. I am trying dynamic prompt, but i dont know, where to connect it, it doesnt work with any box. Anybody can help?

[shy kelp]

what node is this

[west breach]

seems to work okay

[uncut steeple]

ComfyRoll Aspect Ratio

[west breach]

also try with just one image and a prompt

[ivory blaze]

it is not passing anything.

same seed = same output if I unhook image to clipvision

so clearly something bork

maybe need some updates somehwere

[shy kelp]

what's the different weight & height vs aspect ratio, are they related

[ivory blaze]

which mmodel y9ou using

[uncut steeple]

So you can set it to 768 and use it on 2.1 for example

[west breach]

night vision xl

[uncut steeple]

aspect ratio uses your set width and height to calculate

[shy kelp]

so the output will be the size of aspect ratio?

[ivory blaze]

I meant the clip model

[ionic gulch]

with some finetuning you can make revision and positive text prompt work, like in this example. but it's not very useful. i get better results only using negative prompts in combination with revision:
https://i.imgur.com/3Ce7hG6.jpg

[uncut steeple]

yep

[west breach]

clip_vision_g

[ivory blaze]

https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k#how-to-get-started-with-the-model this one?

oh it's a whole different model

lemme grab that one hold

[west breach]

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

[fierce compass]

My image is all garbled when using dpmpp_3m_sde sampler in comfyui. What can be the problem?

[ivory blaze]

well that chokes

the shape in current model is torch.Size([1280]).
size mismatch for visual_projection.weight: copying a param with shape torch.Size([1280, 1664]) from checkpoint, the shape in current model is torch.Size([1024, 1280]).

[tepid surge]

i've had weird stuff with the 3m stuff too previous @fierce compass

[fierce compass]

dpmpp_2m is fine

[uncut steeple]

I use 2s ancestral mostly

[ivory blaze]

i only use ddim

[ivory blaze]

never anything else.

[short marsh]

use exponential scheduler with dpmpp_3m

[shy kelp]

weird aspect ratio but works

[fierce compass]

ah! TIL. I was wondering where the karras samplers were (coming from a1111)

[tepid surge]

ahhh it was all my tabs open that was casuing my vram to spill over with the AIT

worked and then didnt worked then didnt cx

[tepid surge]

cuts a whole 10 sec from my generations

time to hook it into my animation graph cx

[shy kelp]

how do you get an idea what to prompt

[uncut steeple]

#🌠｜show-and-tell message

[tepid surge]

I just association when i'm searching for something specifc, or i use:

for when i'm just wanting random stuffs

like i'll start with "forest" then go "Forest, day time" then "forest daytime, lightshafts, dynamic lighting, vibrent" etc

[uncut steeple]

willy yoda

[shy kelp]

where does it linked?

[tepid surge]

https://tenor.com/view/velociraptor-push-cat-love-you-gif-12642577

[tepid surge]

insert your ui prompt wild cards

[peak dove]

Gone to watch BECK on TV - check out these ComfyUI beauties!

[tepid surge]

💯

[ivory blaze]

so Clipvision model, something was off with it, I grabbed it from another source, and it works fine. before I never saw the encoder nodes activate, but now they are . Weird. no errors or anything just did not work

well still doesn't work lol

[ionic dragon]

whats the use of adding this clip?

[west breach]

something about extracting concepts from the image provided and using it as input

[ivory blaze]

not that one

that one is a bit goofy

[ivory blaze]

or at least the one I downloaded had a bunch of torch vomit and size issues, I just used manager to grab it, and it works bettere

[ionic dragon]

so need for txt2img?

[west breach]

I think it still helps to provide a text prompt

[shy kelp]

its like take the image as reference

not the ear

[ionic dragon]

does it improve the output, if I just generate images using text?

[ivory blaze]

it does nothing

[west breach]

you need to provide an input image

[ivory blaze]

then what is wrong with mine?

[uncut fiber]

dynamic prompt, just got some question, if somebody using it, let me know please.

[ionic dragon]

oh ok, then i dont need it

[hardy cipher]

those are just photographs, bud

[ivory blaze]

lol

also welcome to hell

[upbeat summit]

very nice

[shy kelp]

canny

[ionic dragon]

are you on windows?

[shy kelp]

yep

[zenith hollow]

Does anyone know any way that I can train a face on SDXL with colab?

[upbeat summit]

these are amazing!

[ivory blaze]

I'm just hungryl...