✨｜sdxl

[stone fossil]

I do not use A11 tho, only AInodes and Comfy for img gen.

[amber musk]

okay

[stone fossil]

AInodes is powerfull as heck try it.

[amber musk]

okaay

let's see

[stone fossil]

It has many tools not limited to img output.

Vid is supported, roop intergated.

Audio gen is possible.

Its a love letter to cuda if u ask me.

[vital ermine]

I have had signature, text, watermark, etc... in my neg since 1.5 it is always there but in XL it ignores them far too much.

[amber musk]

you speak so smooth....

[stone fossil]

omelette du fromage

https://github.com/XmYx/ainodes-engine

[native knot]

I'll never understand fashion...

[floral island]

#1 og protovision
#2 first decent merge
#3 finetuning block weights on that merge

going to do some more tests

but i'm pretty sure i'll have to drop protovision after i get this merge done properly, i like creative models

[uncut fiber]

just test sorry

[floral island]

icecream cone seats.

[uncut fiber]

🙂 you have nice imagination @floral island

vivid

[floral island]

#3 for sure this time!

so. back to the testing board and have ANOTHER go at merging all the blocks at various weights. again.

perfect block merger for ultimate creativity and fidelty!

[floral island]

it's the result of ... 1.5 days of recollecting ALL of my lyrics prompts in a single master prompt, over 1k images rendered in sdxl

[uncut fiber]

🙂 hard work. But apparently helped with prompting. Ice cream cone seats i couldn't imagine

[floral island]

don't be afraid throwing in some weird stuff at times, or "surreal" or "surrealism"

it makes most models do the most whacky shit, and then is when stuff becomes really cool

[uncut fiber]

will try practice. I like surrealism.

[amber musk]

@stone fossil bruh

unable to figure out

no clues/tutorial

[amber musk]

yeah wait

no tutorial or anything how to use

[wise rampart]

Yamer's Anime Version 3 is finally here:

https://civitai.com/models/76489?modelVersionId=161302

[stone fossil]

Yeah that is a thing.

[amber musk]

where to start from to make a video, i've imported the sd models

[stone fossil]

@sturdy abyss

[native knot]

Time to face the level's boss:

[amber musk]

@sturdy abyss hey

[steady grove]

AI Nodes looks liek it was made by film school students or industry professionals, who are steeped in all of the information they need to understand it all. It's not the type of software that would have a ton of tuts for it.

[uncut fiber]

want to make it other way round. Insect is selling humans... Cant

[stone fossil]

Well the img output ive done wiht it rocks ur socks. 😉

[steady grove]

video editing has so many niche tools that are just undocumented and people know how to use them because of their experience with other tools

[amber musk]

how did you start using it? what to do next

[sturdy abyss]

can't deny, i'm coming from vfx

[amber musk]

where to prompt/which nodes (from the list)to use/complete the generating

[sturdy abyss]

and designed the tool with unreal in mind

[stone fossil]

Hi mixie thanks for joining the chat. 🙂

[sturdy abyss]

it wouldnt make any sense to do classic comfy flows, but it does support using comfy nodes

[amber musk]

@sturdy abyss

[sturdy abyss]

actually the current version includes comfy too

[uncut fiber]

and it includes chaiNNer...

[amber musk]

any basic usage instructions?

[sturdy abyss]

well, some functions of it, via comfy, but yes

normally there are example graphs

but they havent been updated much so they may be a bit broken

[amber musk]

it looks like an all in one application so, very much interested to learn and use

[stone fossil]

Me I am just here sitting back using mix his tools to make some nice output. 😉

[sturdy abyss]

Sd example graph seems fine

deforum kandinsky works for sure

just was playing with faceswapper

[stone fossil]

roop also.

Audio tools also.

[sturdy abyss]

yeah, sd example is fine, just tested

[amber musk]

which one to attach with which one? :3

[sturdy abyss]

left are input and right are outputs

[stone fossil]

Now this is use of cuda.

AND!!!!!!!!!

It supports multiple cuda devices also.

[sturdy abyss]

those two nodes wont do much together, but try opening from the Graphs menu, example graphs, base, stable_diffusion

[stone fossil]

This will allow me to make the best of my GPU's.

[sturdy abyss]

and right click on the first node, press Eval on it

[amber musk]

can i use any deforum to do these or any other tool from the list?

[amber musk]

okaay

[sturdy abyss]

EXEC just makes sure the node connected runs

this is one of the main differences from other node engines / webui's

you can do loops practically

[stone fossil]

AInodes so far does what comfy dont, if usesd correctly.

Also support Kandinsky.

[sturdy abyss]

for deforum, also from example graphs open deforum

[stone fossil]

I love that model lol. 🙂

[amber musk]

https://discord.com/channels/1002292111942635562/1089974139927920741

[sturdy abyss]

xl has some issue, ill check

[amber musk]

something like this

[sturdy abyss]

yeah, has modelscope

and film/rife interpolation

[amber musk]

okay

hey, why don't you make a video course for learning this great pack, will help a lot of beginners

[uncut fiber]

have to check it. Last time i used rife i had only 1070ti now bit better.

[stone fossil]

Yeah there is also default scemes u can choose from the top.

[amber musk]

3060 (laptop gpu 😦

[stone fossil]

Ah yes thats a thing, if you want to chain a stupid amount of things you might want vram.

I have not seen the OOM msg for a long time hehe.

[hardy cipher]

same

^5

thought it was nice when I got it last year

[uncut fiber]

This is probably closer

[amber musk]

Avinash, a seasoned Indian intelligence spy agent, had been captured by the Pakistani military and imprisoned in a notorious high-security facility. He had been there for months, but he had never given up hope of escaping.

One night, Avinash managed to pick the lock on his cell door and sneak out of the prison compound. He made his way to the perimeter fence, but he was spotted by a guard.

The guard opened fire, but Avinash managed to evade the bullets. He ran across a field, but he was cornered by a group of guards.

Just when it seemed like all was lost, a helicopter appeared overhead and lowered a rope ladder. Avinash grabbed the ladder and climbed up, escaping to safety.

But the story doesn't end there.

Avinash soon learned that he had been betrayed by someone within his own organization. The Pakistani military had been tipped off about his escape plan, and they were waiting for him when he reached the helicopter.

Avinash was forced to fight his way out, and he was wounded in the process. He managed to escape again, but he was now on his own and deep in enemy territory.

With his injuries and the Pakistani military on his tail, Avinash knew that he needed to find a way to get back to India. He made his way to a small village, where he was taken in by a kind-hearted family.

Avinash hid in the village for several days, recovering from his injuries. He also began to make plans for his return to India.

One night, Avinash learned that a group of Indian fishermen were planning to cross the border into Pakistan. He convinced them to take him with them, and they managed to evade the Pakistani border guards.

Once back in India, Avinash was reunited with his family and colleagues. He had escaped from Pakistani jail, but he knew that the danger was not over.

The traitor within his organization was still out there, and Avinash was determined to find them. He also vowed to get revenge on the Pakistani military for their treachery.

[amber musk]

will make a video

Avinash's escape from Pakistani jail:

Avinash ran across the field, his heart pounding in his chest. The Pakistani guards were closing in on him, their guns raised.

Avinash knew that he couldn't outrun them, so he ducked behind a large rock. He took a deep breath and peered over the edge of the rock.

The guards were fanning out, searching for him. Avinash needed to find a way to get past them and reach the helicopter.

He spotted a narrow alleyway between two buildings. It was a risky move, but it was his only chance.

Avinash sprinted across the field and into the alleyway. The guards heard him and turned to follow, but it was too late. Avinash had already disappeared into the shadows.

Avinash ran through the alleyway, his footsteps echoing in the silence. He could hear the guards getting closer.

He reached the end of the alleyway and found himself in a courtyard. The helicopter was waiting for him on the other side of the courtyard, but there were two guards standing in his way.

Avinash knew that he couldn't fight his way past them. He needed to find a distraction.

He spotted a pile of garbage in the corner of the courtyard. He grabbed a handful of garbage and threw it at the guards.

The guards were caught off guard. They ducked to avoid the garbage, and Avinash took advantage of the opportunity to run past them.

He sprinted towards the helicopter, the guards hot on his heels. He reached the helicopter and climbed up the rope ladder.

The guards fired their guns, but Avinash was already safely inside the helicopter. The helicopter took off, and Avinash was free.

and guess what, i'm stuck here

[sturdy abyss]

open the stable diffusion example graph to get a grasp of the process

[dim steeple]

How do you guys reccomend making an image look less "burnt out"?

if that makes sense?

[sturdy abyss]

what model, resolution, guidance scale, and steps?

[dim steeple]

sdxl

20steps

1024x1024

7 cfg

using the ip adapter

[sturdy abyss]

then go for cfg scale 3 - 5 ish

[crisp owl]

what weight/s for the ipdapter?

[stone fossil]

1536x1024 nothing more noting less.

[crisp owl]

singl eimage, two image

[dim steeple]

1

[crisp owl]

1 image input or two?

[dim steeple]

one

[wide crane]

look at the colors

what the hell

[sturdy abyss]

hmm, i might make an ip adapter

looks simple

[dim steeple]

this looks kinda terrible

I hate how burnt out it looks

it keeps doing it too

[crisp owl]

Same results with different checkpoints?

reduce the ipadapter weight

etc

[sturdy abyss]

lowered guidance scale too?

[wide crane]

very sick

[stone fossil]

Bows.

[dim steeple]

Are there any good negatives to make an image look less burnt out?

is that even the correct word for it?

[wide crane]

negatives dont even do anything im convinced

[crisp owl]

negatives do a lot, but they're not super necessary for sdxl

I've seen people do "deep-fried", but I've never had a burnt image issue myself

[stone fossil]

Good yes they depend per prompt you see.

You should always tailor your + vs your -

It is no use to dump a stuff ton of negs into a prompt.

[wide crane]

the eyes

[stone fossil]

If it does not make any sense.

[hoary saddle]

anyone know the best model/lora to use for vikings and spacesuits? lol

[stone fossil]

So you gen like this (here) 1536x1024 3x and if the image output is not good it is either the prompt or the settings like GFC.

Also always rendere in:
40/50
50/70
70/90.

Where the base is the base and plz use a fine tuned model AS refiner.

😉

[floral island]

same as before
#1 og model
#2 first acceptable mix
#3 first attempt weight mix
#4 retrying all block weights and finetuning on #3.

that number 3 image on the final merge. chefs kiss

[dim steeple]

not the regular SDXL refiner?

[dim steeple]

is 1024x1024 worse to use?

[floral island]

at 1535x1024 clones CAN come into existance

1024x1024 is nearly always safe

[stone fossil]

Like stack 2x fine tune.

As base and refiner.

[dim steeple]

I use auto

[stone fossil]

Auto is dead.

[dim steeple]

😦

[floral island]

not dead

[wide crane]

automatic is infinitely more pleasant to use

very simple not mind destroying ui

[floral island]

unless you can make sure i can get in my retart 16k token prompt plus wildcards AND dynamic prmopts. AND output it on console without issue

[wide crane]

but slower

[floral island]

it is slower, the lack of ait hurts 😦

i mean, 100% more generation speed is nothing to scoff at, at all

[wide crane]

yea

[floral island]

however, losing my prompting style and methods, that is something too big to give up

[wide crane]

do you use a1111 or comfy

[floral island]

a1111

[wide crane]

yea ik what u mean

[floral island]

comfy doesn't support BREAK

[wide crane]

idk what break is

lol

[floral island]

the original SD spec was: 75 tokens max

but soon they already figured they can go beyond, just send in a weighted average of multiple groups of 75

so, you either just keep writing and it'll split it up in groups of 75

or you use BREAK, and you can force an early split

[dim steeple]

what is Stability going to release next?

3.0?

[wide crane]

what does that accomplish

[floral island]

so "girl BREAK anime" would cause it to send 150 tokens..."

in my case, i have my verser lines

basicaly, song lyrics broken down by verse, in which i BREAK on each verse

so all verses get a similar weight, instead of one getting broken in the middle

[dim steeple]

Can the preprocessor for IP in XL be more than 512?

[wide crane]

ah

well

i never used anything like that on a1111

all i want on comfy is booru prompting guide

😢

[floral island]

well, 148 tokens it are padding

booru is the worst kind of prompting, imho

[wide crane]

its what all anime models use

which is what i primarily do

so

[floral island]

it restricts the model creativity so hard -> you cannot ever get any interesting output

[wide crane]

i disagree with no interesting output

ive had plenty of crazy awesome images while also using booru prompts

[steady grove]

i've only toyed with anime models a few times, but booru tags really didn't restrict it to mode collapse at all

[floral island]

A1111 prompting style (+ dynamic prompts/wildcards):

BREAK official art, promotional art, surrealism, inspired by __artists__```

[steady grove]

"less creativity" in ML is "mode collapse"

[floral island]

i dare you recreate this in booru language

[steady grove]

thats not anime. i wouldn't use booru for that

[floral island]

want me to make it anime using [anime cell shading:0.2]?

[wide crane]

i actually did use wildcards aswell

idk if they are onnn

comfy

[floral island]

they aren, and not really that great to work with

[steady grove]

no what i mean, is booru is tied to anime data sets

[vital ermine]

I have just about given up on a refiner as I see so little difference. Yeah, it refines but not always.

[cold mica]

Stable Stage is starting soon! 🔴

#1029055412764422214

https://discord.gg/kqNVNhCe?event=1149807096997621870

[steady grove]

thousands of people use booru tagging. NAI invested a lot of money into training them into 1.5. why would they if they didn't work? 🤔

maybe you're coming at it the wrong way and don't got it figured out...

[stone fossil]

I have never.

[floral island]

you can use normal models for refining too. it can help some models which have trouble finishing up images in clean way

[crisp owl]

I suggest using a vae loader node to mitigate using the broken 1.0 vae a lot of checkpoints have baked in.
https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/vae/diffusion_pytorch_model.safetensors

[steady grove]

why are you using the BREAK keyword when under a dozen tokens? is it actually useful?

[floral island]

tokens get more weight the closer they are to the start, and less wwhen they are near the end. in song verses, i'd like to get each verse the same amount of power. that's why BREAK all song verses, so they get an equal amount of power, plus a single BREAK to get in the style i want

that way, i have better control over how much sets of tokens i send in

but you're right, for normal use cases, BREAK is not needed at all

[steady grove]

i wasn't arguing or trying to be right. legitimately curious. thanks

i'm not really sure how those words work but thats a great explaination

[floral island]

it's still generation, this is my current prompt:

BREAK I ride a silver angel, I've circled every star I've touched the edge of the universe But the closer I try to get to you The more far away I'll be I wanna scream!

BREAK [anime cell shading:0.2], surrealism, abstract inspired by Willem van de Velde the Younger, the little courageous angel chibi```

so the verser took 2 lines from my songmaster wildcard

each verse getting a single set of 75 tokens worth of weighting, and then the final style

each token set is then sent in average weighted

results are in, in about ~2 minutes

oh well, scuffed wings on 1 and 3

still, i like my new model merger 😄

[floral island]

u be stealing my prompt? 😮

looking wicked tho ❤️

[stone fossil]

How?

[wide crane]

this is what i got in 35 seconds

lol

[stone fossil]

This is what I do broski. 🙂

[wide crane]

break command definitely does

change the image

[floral island]

i'm just doing 50 steps restart with 30 steps highres 😛

that takes a REALLY long while

and lovely angel you got ❤️

[wide crane]

im doing 20 steps euler a 1024x1024

lol

[floral island]

switch to 10 steps restart

10 steps restart > 20 steps euler a

i'm not kidding

[wide crane]

im not sure

how i do that on comfyui

[floral island]

ah, comfy can't 😦

you have a restart sampler custom node

but i feel it's not quite as good as the one in a1111

i think best i got it was doing ddim/karras scheduler

[wide crane]

oh

yeah unfortunatley comfy has less

like

[floral island]

comfy has ait 😭

[wide crane]

extensions and other things as a1111

[wide crane]

whats ait

lol

[floral island]

aitemplates

[wide crane]

hmm

lemme look that up

[floral island]

basically, huge speed gain

[wide crane]

is it like

pre installed or sum

[uncut fiber]

it is for 3xxx and 4xxx now only

it is extension

[wide crane]

what does it do

[uncut fiber]

in my case 3070 it boost speed by 50%

[floral island]

huge speed up (100% in some cases)

[wide crane]

what the hell

how do i get that

[uncut fiber]

what is your card?

[wide crane]

3070

[uncut fiber]

same as my.

install custom nodes and search for ait.

[floral island]

dude, why you remove ur prompt, i had to check it first 😮

[wide crane]

this?

[vital ermine]

I am just saying a lot of people are giving up on the refiner.

[floral island]

clones 😦 -> this world is fcked tho

[uncut fiber]

But in my case i have to downgrade nvidia drivers to 531

[floral island]

yeah

[vital ermine]

Same, due to training

[floral island]

you have to set the model loader into ai templates

and perhaps checkout the right version of comfy

[uncut fiber]

@wide crane never forget switch keep in memory from enable to disable. It is must

[steady grove]

the newest nvidia drivers that bring all the dlss 3 stuff. Great for games, very crashy for AI

[wide crane]

where is this at

[uncut fiber]

in node

[floral island]

the aitemplate node

[vital ermine]

I got from 4s/it to 200+ if I leave 531.xx

[floral island]

200s+/it on new drivers?!

[steady grove]

wow . i dont have descripencies like that at all. just bluescreens when the power peaks

[floral island]

also, when doing sd on your own pc: fancontrol is your best friend

[steady grove]

descrepencies?

[vital ermine]

Yes, it takes longer and longer for the ADA cards so we all stay, and revert back, to 531.

[steady grove]

discripences

[wide crane]

just do this?

[steady grove]

doseces

[floral island]

yup. you can also do it with vae

[vital ermine]

Known issue Nvidia said will be fixed in some far distant, who knows when, driver update.

[floral island]

oh, and as bernix said, set it to disabled

[steady grove]

i got 4080. gen speed on the newest is fine for me. just stability issues

[wide crane]

so you said its way faster?

[sturdy abyss]

ip adapter tested in ainodes

[vital ermine]

Gen as in training? For inference no issues with any of them

[floral island]

i'm on a 3090, i think about 90% if i took the same settings

[wide crane]

wuh oh

[steady grove]

i trained last week before i did the newest update for starfield, but i was on the newest before that

[steady grove]

for baulders gate

[floral island]

git reset f92074b --hard

[uncut fiber]

you have to downgraded nvidia drivers to 531 and everything properly restart.

[floral island]

on your comfy

that's the thing that fixed it for me

[sturdy abyss]

hm its trying to use aitemplate

[vital ermine]

Well, on the 4090 the tickets are everywhere and the brand new one is no better we simply can't train. Instantly 10x, and slower as it goes, s/it

[sturdy abyss]

didnt know it was included lately

[wide crane]

does it matter where i run it

[floral island]

custom node by fizzledorf

[noble shoal]

I think its a batchsize error. Use batchsize 1 or 4

[floral island]

in the comfyui main directory

[vital ermine]

AIT max BS is 4

[wide crane]

?

[sturdy abyss]

neat, should import without much hassle to ainodes too

[wide crane]

alr

i did it

time to see

[steady grove]

okay, so it's a 4090 issue and they're mentioned everywhere? https://github.com/kohya-ss/sd-scripts/issues?q=is%3Aissue+is%3Aopen+4090

i really think this is a configuration specific issue if you're seeing that huge of a speed problem. i mean, you got a fix for it so who cares really, but i don't think your "diagnosis" was correct

[wide crane]

let me just rebuild my workflow from scratch

[steady grove]

right for the wrong reasons kind of thing

it's not an ADA wide issue

[vital ermine]

In various traininers not just SD, yes. At least Nvidia DID acknowledge the issue and said it will be fixed in a future update. Case closed

[uncut fiber]

@wide crane it is not needed, just put this node between model loader and model in ksampler

[steady grove]

the new drivers aren't ideal as i've anecdoted as well. They give stability issues here

[wide crane]

it was giving me a big error

[steady grove]

i think you're reading into whatever nvidia confirmed a bit too much

[floral island]

try default sampler?

[wide crane]

yea thas what i was gonna do

[vital ermine]

And I think you just like to argue. Case in point is they will fix it and I can't wait.

[uncut fiber]

have you disabled that keep in memory? @wide crane

[wide crane]

yes

[steady grove]

various trainers aren't "everywhere" i can yet find any ticket about 4090 speeds. kohya is the basis of "various trainers" so i went to the source. didn't see any mentions there. working my way downstream. still don't know what ADA wide issues you're talking about.

i myself have ada so i like to stay on top of these things instead of just accepting what some guy online said

[steady grove]

don't really think this is an argument. i haven't taken an opposing position

i'm trying to understand your statements better

do i like accuracy an detail? you betcha

[wide crane]

maybe this work

@floral island

[floral island]

lemme fire up my comfy, lemme see how i set it up

bossfighting time!

eldritch mecha?!

[wide crane]

ughh

another error

[uncut fiber]

i had today several errors, but now working o.k. Did you downgraded your nvidia drivers?

[floral island]

yeah, that's the right setup

[wide crane]

i dont know

[wide crane]

are you also erroring?

[floral island]

very important: 531 drivers

[wide crane]

oh

[floral island]

AND have a gpu that supports it

[wide crane]

yeah im on 537

[wide crane]

how do i downgrade drivers

[uncut fiber]

check for older drivers and install them

[steady grove]

it's a testimate to Nvidia's driver architecture team that we cna just downgrade so easily. doing this with an AMD setup over the years has been hell. AMD changes architecture so often and their dreivers hate to uninstall well.

another bogus with AMD is everytime you install the driver, it defaults the system spyware monitoring tools to checked "i agree". BOGUS. everyone seems cool with this practice of theirs. opting out on every driver install sucks. especially when the process causes system performance issues

[uncut fiber]

Bayerishe motor werke 🙂 ❤️

[floral island]

in dutch: Breiwerk motor werke

[wide crane]

welp

im downloading drivers now

541

[floral island]

uhh... i think that's 531

or you're fom the future

[wide crane]

yea

i meant that

time to restart pc yay

[steady grove]

https://www.reddit.com/r/Amd/comments/9tpyh6/amds_user_experience_program_is_using_a/ this stupid problem that AMD keeps disguising as something new. it's always a user experience data collection thing. it always destroys system performance

[steady grove]

it's been nice coming back to nvidia. none of this opt out on every install bs

[uncut fiber]

seamless or almost seamless @vital ermine ?

[vital ermine]

I wish I could see that seamless

[half ivy]

etsy lg!!!

[floral island]

whatever came out of that thing.lets hope it's long gone already

[floral island]

add rally! they need speed!

[steady grove]

kachow

[stone fossil]

Ok sir Eface.

Rally need for speed?

[floral island]

ATTACK!!!!! (no lora's were abused in this image)

[noble shoal]

I didn't use a lora either 🤷

[wide crane]

im back

yaaay

[floral island]

@wide crane hope it works!

[wide crane]

ittt

did not

[floral island]

that sucks so hard 😦

[wide crane]

what eve

this mena

[floral island]

the good news, it's YET another error message

[wide crane]

the bad news: i havent a single clue what it means

[noble shoal]

Can you show the end of this wall of text?

[floral island]

is this the corona crisis i've heard about?

[noble shoal]

2 Years ago this would have been a priceless mansion

[wide crane]

the full error code

[nimble heart]

more denoise gives more details but also makes the image more eldritchian

[floral island]

using a esrgan type upscaler: 0.3-0.4 ar acceptable noise ratios, when using latents only, depending on sampler 0.48 is the lowest you can go

[nimble heart]

its their first time with SD so I left out esr and all that

my own personal upscale is typically ultrasharp then 0.3 denoise with 0.08% refiner

[stone fossil]

Might be a bug.

[floral island]

optimus butterfly?

[wide crane]

oh well

no ait for me

[noble shoal]

Do you have the latest AIT?

[nimble heart]

I'm a mammal not a crustacean

[wide crane]

yeap i just installed it

[floral island]

the cutest housewife doll ❤️

[noble shoal]

And your comfyui is also the latest version i guess

[wide crane]

yup

[floral island]

@wide crane did you do the git checkout it mentioned earlier?

[wide crane]

yeah

i can do it again

[nimble heart]

"housewife"

looks 14

[noble shoal]

And you run it with batchsize 1?

[floral island]

that's what chibi does

[wide crane]

mhm

[noble shoal]

What happens if you make a git pull in the AIT folder?

[floral island]

ai template empty latent?

[wide crane]

[wide crane]

nothing changed

😢

[nimble heart]

ait is buggy I'd just leave it be until it's implemented officially

[floral island]

the bug plaguing phinlorian

[wide crane]

f

[nimble heart]

its also even buggier the moment you try to do thing like controlnet or ipa

so even if t2i works doesnt mean other things will lol

[floral island]

yeah, with most loaders/samplers/techs -> they don't work nice initially

[nimble heart]

also dont you have to recompile your AIT kernels for custom models?

it only ships base XL

[floral island]

they only work nice if you call their own implementation

[noble shoal]

Nope, works fine

[nimble heart]

Ah so if the structure is the same weights dont matter

[noble shoal]

I think you recompile for different batchsizes

[native knot]

Now make that bug gold and red and you have an Iron Man bug.

[nimble heart]

is there an easy way to compile AIT kernels? my gpu isnt part of the shipped pre-compiled ones

[floral island]

well, right now it's using optimus prime as inspiration

[nimble heart]

so never bothered to try it

[floral island]

lets change the character 🙂

[sturdy abyss]

finally

[floral island]

uhh... did you just create an apocalyptic creature? these are genuinely terifying

[lusty wolf]

Where can I get this?

[noble shoal]

You might ask @indigo carbon . He's a professional in this subject afaik

[native knot]

Excellent. My plan is working.

[floral island]

🥬 groente!

[stone fossil]

Hollandse.

[sturdy abyss]

https://github.com/XmYx/ainodes-engine

[floral island]

new kind of chip.

[native knot]

That silver one is giving Ultron vibes.

[indigo carbon]

Yeah, I'm the one that provided the shipped modules for windows

[floral island]

upgrading my bugs

[native knot]

Now do one in a b&w photo from the 60's.

[indigo carbon]

However I think the next target should be making OneFlow a standalone for ComfyUI, even possibly an official release from @visual glade

[visual glade]

if you can get it working on windows

[floral island]

sorry, they escaped lab and started wrecking shit

[indigo carbon]

I will look into it, it should be possible to build from source.

[nimble heart]

is oneflow cross-vendor too or is it nvidia only

[floral island]

i'm gonna take a uninformed guess->nvidia only

[nimble heart]

AIT works on amd too at least

[indigo carbon]

@visual glade does the compiled OneFlow modules require OneFlow pip modules? Or is it like AIT that you can use precompiled modules without even having the modules themselves

[nimble heart]

in theory. haven't tested it cause i'd have to compile all the ait kernels myself

[floral island]

design schematics

[visual glade]

I think you need the runtime

[indigo carbon]

So we would need to make a scripts that builds the pip module from source automatically to make it a standalone implementation. It's possible imo

When I'll be home I'll try to build from source on windows to see if it's compatible. I'll also test if the compiled modules require the runtime or able to be used without having the runtime like AIT can do

[steady grove]

my experience exactly. if it's even installed when i'm trying to use useful extensions, it bugs out.

it seems to also only do performance boosts in some situations and not at all in others.

[indigo carbon]

OneFlow isn't compatible with AMD, sadly. But AIT should be a piece of cake for AMD

[steady grove]

there doesn't seem to be a lot of rhyme to the reason when using ait so i just keep it uninstalled. worth keeping an eye on but i'll let others iron out all the slag

[nimble heart]

yea installing the ait python module is as simple as cloning AMD's branch for my gpu then setup.py install. It's the compiling AIT kernels for sdxl and 1.5 and all that I dont know about

[steady grove]

custom nodes rather than extensions i mean

[sturdy abyss]

has diffusers xl multilora been tackled yet?

[nimble heart]

I didn't look too deeply into it but the docs were like 50 pages long so I just gave up cause I didn't wanna spend a whole day on something that might not work

[indigo carbon]

But yeah, OneFlow is definitely the fastest inference possible atm. I think SAI used it to train SDXL? I heard somewhere they use it for Clipdrop or something

[visual glade]

oneflow is for speeding up inference not training

[livid cradle]

anyone from here interested in productphotography using SD??

[lusty wolf]

Boo...!

[indigo carbon]

It can also be used to train according to their papers. But they said somewhere they use it for the backend

[floral island]

wicked!

[lusty wolf]

Thanks. Appreciate being seen ...😜

[floral island]

the moment that thing spawns out of a wall, you're pretty sure it is seen!

[indigo carbon]

I gotta look into OneFlow. If I figure out how to build it from source on windows I can keep you posted if that will be useful @visual glade

[low hill]

Hi, what model? and method?

[floral island]

i'm gonna take a guess, dreamshaper, lora

[indigo carbon]

I think they meant workflow

[floral island]

fair, i suppose, but that's the basics, i know thaevil1 made a yarn lora 🙂

[stone fossil]

This is 2 fine tuned models stacked with a personal lora.

[native knot]

@floral island - Need to combine the yarn LoRA with the bug images.

[floral island]

i'm a purist -> no lora's!

[native knot]

Not even add-detail-xl.safetensors? Come on now...

[indigo carbon]

Same, I tried the offset LoRA and it made results look washed out

[native knot]

I don't typically use LoRAs...except for that one.

[floral island]

i've used it ... 3 times? it's nice. but -> it goes for techno influence. i know because i'm using incredibly weak tokens. if a lora has influence, it will take over

hello kitty kaiju

[indigo carbon]

Huh, didn't know this existed yet. When I used 1.5 I also used that LoRA ALL the damn time

[floral island]

especially on the -1 setting 🙂

[native knot]

It's been out for a bit now. 🙂

Comes in handy.

[indigo carbon]

Doesn't that remove detail if on -1?

[floral island]

that's the exact purpose!

i've used it more on -1 than +1

[native knot]

Yeah, it works in both directions.

[indigo carbon]

Kinda dumb.. you usually want more details in gens

[native knot]

Depends on style. You might be going for various animation styles, which some require less details.

[floral island]

not if you want to persue a specific style and you are using generic models

especially lineart

[indigo carbon]

I suppose it could be handy for anime stuff. But if you're going for digital art or realism you usually want more details

[native knot]

Yes, but there is art of all kinds. 🙂

[floral island]

realism -> yeah.
surrealism/abstract -> not quite

i'm more in the 2nd camp

so for me the -1 detail was a godsend

[livid cradle]

Some results i was able to achieve with the software im working on based on SDXL 🙂 in #productphotography niche.

[floral island]

very nice

did you use controlnet to get the text?

[livid cradle]

using Image-to-image masking api endpoint... but anyonje knows how we can pass the init_image as a base64 in prompt or some way to tell SDXL the main subject is already in the image so i don't get another one generated behind the mask?

[livid cradle]

no controlnet used... just what i mentioned above.

If anyone interested to play around with the software do let me know in DM.

[steady sluice]

Hey everyone, I'm using the Stability AI REST API call text-to-image using SDXL 1.0 with the anime style preset and noticing that some seemingly innocuous prompts are getting blurred - no matter how much I try changing the prompt, it seems to be flagging the content of the prompt.

The prompt i'm using is "17th century painter male". Strangely enough the "male" characters are getting blurred and flagged but the female characters are not. Does anyone have any insight on what might be happening or the guidelines for what gets flagged?

[supple knot]

Hi I started using Comfy with SDXL, If I wanna use SD1.5 now with Comfy, Do I have to configure the folders differently or just use SD1.5 models?

[crisp owl]

just use the models

[stone fossil]

Dance with me.

[wide crane]

markiplier as an old ass man

but not sdxl

because im a lazy bastard

[native knot]

Hate to be that guy, but this is #✨｜sdxl.

[wide crane]

hate to be that guy but im banging your father

[glad grove]

https://tenor.com/view/me-when-the-markiplier-is-that-the-bite-of87-gif-19770814

[native knot]

Imagine using that response and thinking it's funny and/or original.

[wide crane]

never said i thought it was funny

[crisp owl]

And back on topic, check this out!

[wide crane]

thats cool

[native knot]

You didn't do any thinking at all.

[wide crane]

definitely bro

[nimble heart]

i broke my toe yesterday and it still hurt less than reading this conversation

[wide crane]

wishing you a speedy recovery

not your toe though

[glad grove]

my arm hurts

[hardy cipher]

my image blending machine is growing stronger

brilliant

[glad grove]

speaking of blenders

[hardy cipher]

top notch

[glad grove]

oh no 🤣

[native knot]

The monster has won.

[hardy cipher]

yeah, that monster

[supple knot]

https://www.youtube.com/watch?v=dDvEY8LT4x0

[hardy cipher]

I should probably use more coherently themed sets of images

[glad grove]

most normal outfit lady gaga has ever worn

[hardy cipher]

his hand has evolved past other human hands

[glad grove]

bro has bad breath

[glad grove]

real

[steady grove]

oops i'm using animatediff. not sdxl.

[glad grove]

my honest reaction

[native knot]

I figured you might just be posting reactions to the Spock stuff.

[steady grove]

forgot what chan i was in. i'm generating these as i go to match the spockergy

[night kayak]

are there any tips/tricks to make sdxl work better with text?

[native knot]

I don't remember this crossover.

[steady grove]

i'm still brinign spockergy to the #🏞｜general-with-images channel

[native knot]

I like to put the text on its own line in quotes. I also try to find models that do well with text. JuggernautXL (at least v3...can't say about 4 or 4.5 yet) did fairly well. Some others do not.

[night kayak]

yeah I even tried controlnet with text pics but that did not work too well

[native knot]

Depending upon what you're aiming for, some things are easier than others.

[native knot]

New promotion there from Burger King.

[hardy cipher]

that turned out way better than I expected, lol

[hardy cipher]

the bed of cheese curds really pulls it together

[nimble heart]

i thought it was packing peanuts

[hardy cipher]

I added some things

it's whatever you want it to be

[nimble heart]

robots need drip

[glad grove]

speaking of tasty burgers

[nimble heart]

is he trying to take a picture of the cronenburger using a grilled cheese

[stone fossil]

Dang thats a burger I want it.

[glad grove]

this one lookin even tastier

[nimble heart]

kids these days are too soft. back in our day we had to fight our food

[glad grove]

if u cant eat this u are a softie

[glad grove]

cryin tears of happiness for this tasty burger

[wide crane]

what is the prompt for this

[stone fossil]

Hand wirtten with many love as always.

[glad grove]

dinner with my friend 🍔

[stone fossil]

I strip my meta deta because I think its late if everything becomes one big copy job im sorry. 🙂

[noble shoal]

0/10. You are a liar. I got this when i prompted for "Magic"

[glad grove]

my skin feels kinda itchy

[hardy cipher]

he's becoming one with the monster

[crisp owl]

😆

[glad grove]

tasty rooster soda

[nimble heart]

buddy needs to see a chiropractor

[hardy cipher]

after the come down of a 3 month bender

[glad grove]

bro looks constipated

[noble shoal]

Just one more Burger

There he is

Just kidding. There is room for more.

[glad grove]

thats just a tiny meal for him 🙏

id love to eat this one though

[steady grove]

this summer, under cheese 3

this time its burger

[hardy cipher]

whoa

[hardy cipher]

,

[glad grove]

idk what happened here

[glad grove]

all american wedding 🫡 🇺🇸

[crisp owl]

https://tenor.com/view/zoolander-ben-stiller-will-ferrel-mugatu-relax-gif-4534388

reminds me of this scene

[noble shoal]

Longer you said........

[upbeat summit]

preparing the wedding cake

[noble shoal]

Only a little bit disturbing 🤏

[glad grove]

this one a little big

[crisp owl]

Are these those under the bridge rave kids?

[hardy cipher]

they've taken on a life of their own. so maybe

started out as rainbow goblins

[crisp owl]

I guess all things eventually evolve

[nimble heart]

Anyone know why the refiner just makes images chunky at high res? 4k image, left is only base, right uses base + 8% refiner

I'm setting the refiner's clip resolutions to match the latent size, in this case 3840x2160, so maybe that's goofing it???

[hardy cipher]

I've messed with the workflow more. doubled up on clipvision with the second image. did the clip g and the ipadapter

[crisp owl]

I keep my clip resolutions at 2048 and never touch them beyond that.
So, perhaps?

[dim steeple]

Why exactly does a higher 1:1 resolution than 1024x1024 not work well?

I am trying 1536x1536

and it distorts

[nimble heart]

1536 works fine with a 1024 first pass

[glad grove]

wasnt trained at that resolution

[nimble heart]

1536

[dim steeple]

so if I do a high res fix?

[nimble heart]

yes

[dim steeple]

should I be using Latent?

[nimble heart]

those images I sent were latent, but depending on the scene latent may not always work well

latent needs at least 0.5+ denoise

non-latent can go down to like 0.3 or so

[dim steeple]

I am using a grid with controlnets

Still testing the tokyojab method

for video

[nimble heart]

I dont know what any of those words mean you're on your own

[dim steeple]

https://www.reddit.com/r/StableDiffusion/comments/11zeb17/tips_for_temporal_stability_while_changing_the/

I've been testing this in XL for the last few days

its been tough

[hardy cipher]

winning formula

[upbeat summit]

you can make images in higher resolutions - also without a high-res fix pass. but it heavily depends on the model, settings and prompt build you're using.

you will definitely have more disproportional content when you use an out-of-spec resolution, but it can be totally worth it because the details can be pretty great.

I've been working in and testing a lot of different resolutions with SDXL. 1536x1024 (3:2), 1920x768 (~21:9) and 1920x1024 (~16:9) are currently the high out-of-spec resolutions I'm getting the best results with.

native 1920x768 (1 pass)

[hardy cipher]

perfection

[glad grove]

funky

[glad grove]

this burger bit me

[noble shoal]

@hardy cipher I am onto something

[glad grove]

here we see steven seagal ancestor hunting for its prey in the wilderness

[glad grove]

looking sexy 🥵

[upbeat summit]

maybe in a couple of years this will be a normal sight in NYC

[hardy cipher]

lets hope

[glad grove]

my mouth hurts

[noble shoal]

🐴

[hardy cipher]

whoa

winner

[noble shoal]

Congratulations. This is utter garbage! I love it

[hardy cipher]

it really worked out well. I agree

whoa. it went into the NSFW zone for some reason. if you're sensitive to things please don't look

[crisp owl]

The scale of the objects really trips my brain

[noble shoal]

I have no idea whats happening in ths picture

[hardy cipher]

oh my

things we were never supposed to see

[crisp owl]

I see the fit men in that image for sure

[glad grove]

guys is there something on my teeth?

[hardy cipher]

what's going to happen?

[glad grove]

🤣 i dont wanna know

[glad grove]

🤣

[hardy cipher]

that middle tooth is not looking good

[shy kelp]

Yo chat wanna one discord alt account drop?

email and phone verified in format mail:password:token

Info for mods and admins and other staff:

• this is not an giveway advertise,
• this is not a scam (account info will be dropped here),
• this is randomly generated acc so not stolen or somethink like this,

so no need to ban or mute me for this i just don't have anythink better to do

(If the administration, moderation or staff have any reservations, please tag me and let me know)

[hardy cipher]

"this is not a scam (account info will be dropped here),"

well there you have it folks, not a scam

[shy kelp]

bro wth 😭

[shy kelp]

i said that bc i was banned 5 times in a row for "scam"

[shy kelp]

when i just said this

[shy kelp]

im adding more and more info t-

[shy kelp]

BRO WTH IS THIS

[hardy cipher]

joy

[shy kelp]

devs mods admins any reservations?

okay drop will be in like a 1 minute

[hardy cipher]

I don't know of any mods in here

[shy kelp]

idk too

[hardy cipher]

what's an alt? my iq is pretty low

[shy kelp]

second account

[glad grove]

my honest reaction

[shy kelp]

there are many ways to login with token

it will become visible for 15 seconds so just copy it

after 15 seconds it will be deleted

[hardy cipher]

oh, that sounds legit

[shy kelp]

i can recomend browser extension to login with token

[hardy cipher]

oh, and I'll win money?

[shy kelp]

i mean i dont have anythink better ti do do 😭

[shy kelp]

no you'll win discord account

okay it was visible

you copied it right 💀

[hardy cipher]

of course. I trust everyone on the internet

[shy kelp]

that sound's very ironic

[glad grove]

i won a seagal burger

[shy kelp]

ill just do 5 minute timer

[hardy cipher]

why are you giving away accounts, bud?

[shy kelp]

idk what to do with my life i don't have anythink better to do

i have around 200 of them and i just don't need that many

ANYWAYS

[hardy cipher]

200 accounts? is that legal?

[shy kelp]

yes

i forgot to change time to 300 seconds

[hardy cipher]

whoa

[shy kelp]

i think 5 sec update time is not that bad

okay done

[shy kelp]

i think about giving away discord nitro but not on this server

[hardy cipher]

whoa

[shy kelp]

anyways account will drop in 50 seconds

[hardy cipher]

I don't know, I can barely handle 1 account

[shy kelp]

that's the same i thought you didn't copy it

[glad grove]

https://tenor.com/view/waffling-gif-26189788

[shy kelp]

https://media.discordapp.net/attachments/379732843041128452/1080915321801678898/0C0362C0-521C-405C-B5EB-640559E900F0.gif?width=646&height=160

[nimble heart]

as of 7 minutes ago they just deprecated the mi50 too lol

rocm 5.7 out and its the last release the mi50 works on ig

also no more radeon VII

[nimble heart]

:(

[upbeat summit]

well if it's a 8k tiled latent upscale - that would be fine 😉

[eternal fog]

ooo, if anyones using ControlNet LLLite, there's a new replicate model, that tries to replicate the input image along with prompts you've provided

Input

Result

[upbeat summit]

cool!

[eternal fog]

That's with txt2img, you can also use img2img if you want

[nimble heart]

nah just 4k non-tiled

i changed my vae decode to 1536 tiles instead of 1024 and it had to compile new kernels which is where the 8 mins came from

[native knot]

I'm really getting sick of all this child labor just being swept under the rug. Clearly dangerous work like soldering volatile electronics to high-voltage batteries. What a real shame...

[uncut fiber]

😍

[vital ermine]

Shit :/

[vital ermine]

Radeon 7?

MI50 loss of support is like Nvidia dropping V100 support age wise. That is insane. MI25 would be the Pascal age for Nvidia.

Such as the T4

[vital ermine]

Shit, he evolved.

Or devolved, ngl I am not sure

[nimble heart]

yea that hbm consumer card

i think all vega stuff is deprecated not sure though

[uncut fiber]

last one is result

[vital ermine]

Vega deprecated before it was released.

[nimble heart]

i guess they'll still get bugfixes but no new features

[vital ermine]

You know my 4090 hit 73c genning all these reg images and never that in any training. My presumption was correct as gen my card is pegged 99-100% (system is a bit sluggish too but cpu is idle). What I do not get is why when I train I hoover more around 60-80% and spike to 100%. If I could train at 100%/99% all the time I would be a speed demon.

[nimble heart]

is there a bottleneck on the cpu?

or maybe a specific gpu bus?

[vital ermine]

nope

for training I was told almost all of that is on the card whereas gen is more push back and forth

I know it is on the card because 23.5GB of 24GB is used

I swear something feels wrong as if a LHR like situation is happening where only training is affected.

[vital ermine]

I ordered my new case fans yesterday to help the heat with this blow torch and from amazon, delivered by Amazon.

[west breach]

maybe it's coming from international?

[vital ermine]

Even with prime only 2 days sooner

[vital ermine]

It comes into stock a lot and leaves fast too and stays out of stock for a while so maybe. Fans are from Arctic.

not the best, nor the worst but best for the price

but sheesh at 6 weeks

[west breach]

Sometimes I got long delivery estimate, but it still only took a week

[vital ermine]

yep

happened to me twice over the last decade

reason I went ahead and bought them

the fans that came with my case are not made for this and frankly sucky a bit. 120mm an 1200 rpm max but barely can blow. these are 1900rpm and 64cfm each at 140mm made for pressure. That extra 20mm means more air will blow by the exhaust of the gpu.

[uncut fiber]

👁‍🗨

[vital ermine]

the glass side panel gets really toasty even with the front fans full force so air is just not flowing past

[uncut fiber]

i am thinking about whole new PC. Not sure it is right time. Maybe some 4k € from which gpu will be about half of price?

[vital ermine]

Do realize PCI-E Gen 6 arrives next year

127ish GB/s

Don't care about that just the hardware needed to support that will make the entire system snappier

I am generating a lot of reg images and for XL of a man for class. wth is up with all the wonky alien eyes?

I was curating what I already had and dumped a lot due to that

alien/snake eyes

[uncut fiber]

yes just when it last so long, just to amuse you 🙂

those eyes are as well bad, just platini robot, so can be some sort of cameras.
@vital ermine it is result of blending, which took about 2 mins.

[vital ermine]

LOL

My new nvme 2tb to replace this HDD went up 20% so it waits until it comes down again. After that is a new 4k 27in monitor (or at least 1440p) high refresh. After that I begin to save for a new computer system next year unless the world falls apart.

[uncut fiber]

yes i dont need 4k, having VR, not using it for no reason now. But loved to driving in it. AC, AMS2 and much more.
Advantage to have this whole new computer would be i can move this on cottage <- not sure right world.
4090 with some ryzen How about multicore 64 and so i mean, any benefit for home work?

[vital ermine]

My next Ryzen will be at least 12c/24t

6c/12t is just not working out these days with the 4090 and stuff

[uncut fiber]

i ment something like this
https://www.amd.com/en/products/ryzen-threadripper
just probably very bad for gaming. Or simulation which is mostly single core @vital ermine

[vital ermine]

They have a TR coming out 1q 24 that is 128core and a 96core version

[uncut fiber]

@native knot is it cutted off? i mean if background and foreground is same generation?

[native knot]

All one generation of txt2img

[upbeat rivet]

what do you prompt

[vital ermine]

Modern Disney in a nutshell just poppin' caps in asses.

[uncut fiber]

@upbeat rivet it is image blending. So first two images results in 3rd image.

[uncut fiber]

@upbeat rivet here is specific combination of authors for wedding and mad scientist. Just drag and drop it in Comfui.

[nimble heart]

7900X very spicy. Mine boosts to 5.9GHz without any manual clock or voltage adjustments

[nimble heart]

that a denoise of mine or is it IPA/something?
#✨｜sdxl message

[vital ermine]

yep, 7900X3D is the one I was looking at.

[nimble heart]

why x3d

[vital ermine]

I like them and they do help with with the 4090

over plain X

[nimble heart]

pretty sure the higher clock of the non-3D would help SD more than the cache of the 3D

[vital ermine]

nope

that is old school thinking

doesn't matter as the 8k is due shortly

shortly being Q1 24

now in Blender or Cinema4D, etc... I think clock may be better but not sure. I know is gaming and in SD the buffer matters more.

7950 is too expensive for what it is

[hardy cipher]

yes

did some remixing

[nimble heart]

im assuming with ipa cause I opened it and 90% of the nodes were red lol

did you render straight to 4k with the reference or was it upscaled?

[nimble heart]

pretty sure the non-3d benches higher in code compiles and things

[hardy cipher]

upscaled

it's a pretty involved setup

[nimble heart]

if it's got enough information to keep the same composition you might be able to directly render to 4k

i did that in 1.5 with 1080p a few times

[hardy cipher]

hmm. I might actually try that

[nimble heart]

using controlnet to keep composition cause ipa wasn't a thing

[hardy cipher]

well that uses lots of things. ipa, clipvision, controlnet canny

also blip interrogator

[nimble heart]

damn

you probably have enough information lol

try 4k native

worst case scenario you spend 20 seconds making an eldritch abomination

[hardy cipher]

also split the conditioning, do some tweaks, then concatenate

really have no idea what impact it's having. but seems to avoid the black boxes I was getting before

[nimble heart]

yea I dont even separate my G and L prompts lol

[nimble heart]

literally my whole workflow for my 4k wallpapers

[hardy cipher]

you should see this conditioning nonsense i'm doing. I'll spread it out real quick

[crisp owl]

Getting closer to finishing my workflow where nearly everything is controlled off a single switch

[nimble heart]

at that rate it'd probably be easier to just make a custom node...

[crisp owl]

working with the Impact Switches that don't load in nodes that aren't necessary.

[nimble heart]

so you just apply the same unclip twice, once @ 3x strenght once @ 2x strength?

or something to that effect

[hardy cipher]

nah, one clip g unclip, one ipa unclip

but from the same input image

[nimble heart]

I mean the split

before concat

its the same unclip, multiplied by 2 then 3 then concated

[hardy cipher]

well I have it as the first thing after the conditioning output from the tet prompt

so yeah. that's what's going on essentially

at first I was just experimenting to see what would even render

but seemed to actually work out so ran with it

[nimble heart]

so with all the multiplying normalizing concating...

do you even objectively know the impact its having?

cause I'd have no clue without rendering a 100-image grid or something

even that'd be sketch

[hardy cipher]

well I could test it I guess. but I can say that each image has a distinct impact on the output

so it's not just randomness

[nimble heart]

right but how does the split/multiply/normalize/concat compare to just increasing the strength slider on the initial node

[hardy cipher]

you know, I just put that clipvision there earlier today. so actually don't know

originally put it there because I wanted to increase the impact of the text prompt, but didn't want it to get too out of wack, so shought by normalizing the one side it might increase it's impact without making causing the data to lose coherency

I just made that up in my head, but seemed plausible

[nimble heart]

with text

isn't what you're effectively doing with the concat

[hardy cipher]

well the first two images sort of have a double impact

first image is ipadapter and also clip g interrogator input

second image 2 clip visions and clip l

3rd image base for denoising and controlnet

[nimble heart]

the equivalent in a1111

(my fun prompt:2.0)BREAK(my fun prompt:3.0)

[hardy cipher]

you might be right. I'm kind of figuring out what's happening as I go along

[nimble heart]

if the prompt aint strong enough I'd just CTRL+up/down on some words to tweak their weight by hand

without blowing up the CFG

[hardy cipher]

well it's more experimentation to better understand how conditioning data works

[nimble heart]

I get that. I'm just trying to figure out what data is being taken away from it

[hardy cipher]

ahh. well I could run something with and without all that extra stuff. see what the difference is.