🏞｜general-with-images

[hollow oak]

in love

[pastel mango]

He's someone who'll sadistically slay a few hundred people in an attacking army, playfully, while making them think they're winning at times to make their loss and death all the more horrible for them. Then he eats all the power of their souls, but not their souls (them) so they don't have the strength to pass into the Endless Stream (probably for the best though.)

Then after all the slaughter, joy and sadism. He'll go back to his friends and family and hug them and show them affection. He'll heal wounded and sick if he passes through a village in pain...But if that village were to collectively agree to a decision that hurt anyone he cares about. He'd wipe the village off the map quite literally. there would be nothing but dirt left.

If an army surrenders to him though and they haven't caused his loved ones or him grief, he'll let them go.

So quite some shifts in personality. I still have to nail down the specifics of how far he'll go in any direction cause I don't want to take him too far where he just becomes a parody of himself

[wispy ether]

Hi

[wispy ether]

https://tenor.com/view/hhhhomeboyyyy-gianluca-me-reading-wtf-gif-21273156

[pastel mango]

Yep

[frigid torrent]

What does it like. Mean?

What are u talking about

Oh oh the story

Ok sorry I forgot how to scroll. Anyways sounds cool

At first. I thought you were referring to the dragon

[pastel mango]

No just vaguely talking about the main character at a certain stage in the story...but I don't want to give away details on the story so it only makes it sound edgy rather than psychologically broken and disturbed

[frigid torrent]

The endless stream, is this an afterlife in your work?

[wispy nest]

[west coyote]

Yo

How to do img to img in this server

[proud dagger]

not available on the bot in here, i think you can on dreamstudio tho

[errant heath]

how can i avoid my non-hires pic turning horrible when i make it hires?

in non-hires it looks always proportionated etc

and after that it's trash

with lots of flaws

[wispy nest]

Never upscale 4x. Generate 512x512 upscaled 2x with 15 steps using ESRGAN. Then upscale to whatever you want in the extras tab. Use this negative prompts if you don't already:
(worst quality:1.4, low quality:1.4),

[livid cobalt]

Hi everyone! I'm new to stable diffusion and have just been playing around with it a little on my M1. I keep getting these types of results... does anyone have insights into what's happening here?

[livid cobalt]

it happens when I try to increase resolution past 512 or any time i try to increase batch size.

[wispy ether]

Average apple issue

@livid cobalt dont use apple

[astral stirrup]

What tool are you using I use a mac and hopefully I'm not a jerk like Nep.
I tend to use InvokeAi and simple python scripts for local generation. If you're looking for advanced usage I might not be the best person to ask.

[slow barn]

this is what i am playing with @tropic shell

@tropic shell

[tropic shell]

looks pretty!

[slow barn]

its very fun

you should try it sometime

[tropic shell]

I have 🙂

[oak osprey]

@smoky oak lmao the arc of selling ai art on etsy for a youtuber

[smoky oak]

Lmao

[oak osprey]

fuckin LOL

REMOVE AS MUCH THOUGHT

classic youtube

[errant heath]

thank you for your help, so it's like this?

[oak osprey]

why are those SO blurry?

[fading anchor]

vhs lora

[oak osprey]

oh

[smoky oak]

Hey, that sounds just like the assholes at WattleArtCreations and how they screwed me and continue to screw others by running an AI generated art sweatshop that pays even more horrifically than normal commission jobs

(side note to everybody here, steer clear of them, they are trashy, and shitty)

[oak osprey]

yessss they probably had the same arc

that's how movie villains are created lmfao

[quartz ravine]

2023 - Discord users stealing your prompts, colorized

[quartz ravine]

I just read that article earlier today. I love that last line

[smoky oak]

@oak osprey any thoughts on hyperdreambooth?

I would love to play around with it when I get my 3090

[oak osprey]

idk, it seems like something cool but their comparisons made no sense to me. not like they mean much anyway when reproducing basic results has always been difficult

[smoky oak]

Fair enough, though even if it's just as good as dreamboothing, but still faster and smaller, that could be very attractive

Could bring the floor to entry for fine-tuning down

Inb4 I start doing BS 100 1.5 LoRA's lol

[oak osprey]

L2 loss is an interesting concept tbh it looks like something SAI hasn't used enough of

i've always kind of felt that making image models in particular, easier to fine-tune, serves almost no productive purpose for society, but i don't care enough to try and force that into anyone else

the fewer romance scammers that have at their fingertips, 1000s of LoRAs/Hyperdreambooths to use to ensnare victims, the better

it sucks when they have 50-80 images to try and convince you they are real, with. just imagine, hyperdreambooth from one image lmao

i'm going to have to meet someone in person now? RIP

[smoky oak]

I'd love to give hyperdreamboothing a chance

[sharp sparrow]

Hello!
Can someone please say the name of this Stable Diffusion model? How to get such cartoons and shine?

[oak osprey]

looks like one of nitrosocke's pixar models with controlnet tile

[sharp sparrow]

Thanks!
But I don't think that this is Pixar
Another example:

Maybe this is some type of Lora?

[oak osprey]

i have no idea lol there's thousands of models floating around

you would need to ask whereever those image decks are from

[smoky oak]

Looks like a Sam does art style rip honestly

So many people rip him off, it's kinda gross

[oak osprey]

i kinda feel like that about most art style prompts

lmao

unless the person has said "please, use me"

[smoky oak]

I do agree there as well, but Sam does art is the most hostile example of people trying to screw an artist over

[oak osprey]

yeah like greg rutkowski being used in so many prompts and fine-tunes despite his feelings on it?

[smoky oak]

One of the furry artists I prompt art styles from has actually come out and said he's fine with it, and another I used to use said he's not, so I stopped using him

On top of that, I always try to use at least 4 artists to try and dilute styles as well

[oak osprey]

i'll never forget that Corridor Crew video where they used his name in the prompt

[smoky oak]

For him it's just people not really listening

For SamDoesArt, people literally go out of their way to say "fuck you, I'm doing it just cause you said not to"

[oak osprey]

same for Greg, homie

for 2.x finetunes in particular

[smoky oak]

I haven't seen anybody sharing models called "fuck Greg rutkowski"

Cause they do that for Sam, and it's so annoying

[oak osprey]

sounds like an even more toxic subset

[smoky oak]

I mean. Have you MET the AI community? Lmao

[smoky oak]

Me trying to think of things to try with SDXL

[oak osprey]

oh no 😛 you peaked

[smoky oak]

I've just been super discombobulated these last few days

Been feeling sick still, not sure what's up

I think people are commenting on my GitHub with issues, but I haven't been feeling up to checking. I'll be doing that soon

[oak osprey]

a stunning portrait of disney's uncanny valley

[lilac pilot]

Anyone living in Japan 🗾? Where is your favorite places to get Ramen?

This one doesnt scan. Fail

[pastel mango]

So I made this in Clip Studio Paint and I dunno what to run it through img2img with

[oak osprey]

@smoky oak am i a terrible person for assuming any torn-to-shreds image from the Cursed AI group is Stable Diffusion 1.5?

[smoky oak]

looks more like base 2.1 to me lmao

Very gritty and scratchy lol

[pastel mango]

cursed ai?

[oak osprey]

or 1.5 with a CFG around 12-15

[pastel mango]

the image is very lifelike though

looks just like him

[shadow grove]

Anyone an idea how to prompt this?

[calm canopy]

Use the interrogate button in a1111's img2img for a start

If it has metadata, could look for that too.

[stoic pivot]

Why do my waifu images get no votes... what is a waifu 🙂

[oak osprey]

i trust her as far as she can throw me, which looks, quite far

[grizzled sage]

https://tenor.com/view/muahaha-evil-laugh-seinfeld-wayne-knight-newman-gif-4185031

[fallen dock]

shurk

[fallen dock]

https://tenor.com/view/spiderman-funny-laugh-gif-11964564

[umbral lake]

hmm... am i doing something wrong?

[wild sorrel]

I don't speak that language, idk what is it saying

[fallen dock]

it's Spanish though.......

[wild sorrel]

"That language" - was safe way to say it, without looking dumb cause it's not what I said it is

[fallen dock]

dude Spanish is literally the easiest language to learn and recognize, bro

well your plan failed so, lol

[wild sorrel]

Why don't you answer him instead of trolling then?

[fallen dock]

Not my job to answer them, all I know is that it's in Spanish.

I've just never heard someone say they don't even know what Spanish is....

[steel apex]

Looks just like the real thing XD.
(Not trying to start political arguments, I am just joking around)

Anyway, here's a Midwest Thunderstorm

[oak osprey]

try a stunning cinematic portrait of a corn tornado cornado in a nevada field during a dark stormy afternoon

[umbral lake]

oh.. my bad for not translate the message xD
Cannot publish because it contains content blocked by this server. Server owners can also see it.

[calm canopy]

wrong channel. #1100170312106127410

[craggy schooner]

hahhaha I love it

[wild sorrel]

might be just weird auto filtering thing, try "home" instead of house or smth

idk who control what's filtered and whatnot on bot

[umbral lake]

nope

nothing works

[calm canopy]

ah i missed that. did you click inside the prompt box that pops up? looks like this

works fine when i do it

[broken grove]

color photo of: the sunrise over the peaks of Huangshan mountain range, golden rays piercing through a sea of clouds, a serene and peaceful atmosphere, captured by the lens of a Nikon D850, using Kodak Portra 400 film and a 24-70mm lens

[south quest]

was trying out a different checkpoint and wanted to share

[reef vortex]

is this look good

[left pond]

Plugging Midjourney outputs into ControlNet with Stable Diffusion

[umbral lake]

nope

[oak osprey]

people are so unoriginal lmao

seen so many of these "king of the hill characters IRL"

[left pond]

Just wanted to wrap my head around ControlNet with something familiar for a couple of hours, sorry it wasn't babes and Shrek memes.

[oak osprey]

well just do shrek memes and babes next time

[fallen dock]

you do realize you don't need controlNet for Shrek memes you know that right? especially with SD.

[oak osprey]

i don't need controlnet for anything

[neon crystal]

Some badly needed warm and fuzzy eye bleach after suffering through realism at work.

[polar star]

Rabbit

[frigid torrent]

https://cdn.discordapp.com/attachments/531959975514341386/1130685181305573446/output8LosslessYUV.mp4

time lapse of img2img loopbnack stuff

i love looopback images you get craqzy cool stuff as CLIP tries to describe more and more each time

and then at the end i have all these frames to assemble into an animation with ffmpeg

made it into a gif :3

[zinc lagoon]

/https://cdn.discordapp.com/attachments/1004159122335354970/1130688203880415263/phillydek_Create_a_surrealistic_landscape_that_blends_together__b8fc9104-ef24-4813-a131-829f0a222676.png

[warped pond]

GM

[frigid torrent]

Not sure where to send this so, this channel will do I suppose
here's a link to the training I've done off of scans of my own drawings, including a bunch of generations, source training images, and a sd1.5 hyper network:
https://archive.org/details/geometry_of_the_infinite_interior
https://archive.org/details/aibuildings_2023_06_30
https://archive.org/details/aibuildings_2023_06_28

While I lack any drawings of realism, the lineart concept of the drawings I do provides great structure when the weight of the hypernet is carefully tuned against another model. So, anyone trying to make some things that are labyrinthian, or sorts of concrete or complex structures, this might be of use

Here's a decently good example

[sterile temple]

That's cool! I had a similar idea to train a lora on thousands of randomly generated images of random shapes and colours, and the captions would be like 'red triangle against a blue background'.

[frigid torrent]

Like a geometrical structure Lora?

[sterile temple]

I thought it might help with attaching colors to different objects that were prompted

[frigid torrent]

I've been freehanding these drawings since I was a kid, so I took the scans I had at the time and cut out as many 512 pixel squares as I could to train a hypernet. I'll definitely scan all the rest of the drawings I have to train it up again, and probably with sdxl when that comes out

[frigid torrent]

Sounds like it would. Doubling down on teaching fundamental structural concepts to a model seems like a good idea, too, especially for landscape/architectural generation

[scenic umbra]

this is really cool

[frigid torrent]

ฅ⁠^⁠•⁠ﻌ⁠•⁠^⁠ฅ

This one I like a lot. Extended it out to a 16:10 wallpaper and upscaled

It'd also my phone wallpaper, zoomed in on that stack in the center there

For this twisted structure I used the latent mirroring extension on auto, 180⁰ rotation

That was before outpainting

More favs from me

That first one I printed out small on a piece of paper at work. And I'm drawing it out with pen

So it'll be interesting. I'll be printing out little segments generated by ai, then using them to start drawings. And eventually scanning them and using them for training!

Weird world we live in

Finding new ways to not only instruct the generation of images but actually collaborate as an artist with the machine is really quite fun and it's gotten me drawing more than I have before

Img2img is great with it

[fierce ridge]

GM all, first post here, Cyberpunk Barbie goes hard..

[daring kernel]

@frigid torrent I see a future for you

[frigid torrent]

A future where? Lol

Rn my future involves the current labor I shall work in perpetuity, but probably a shit ton of ai generation on my time off. Ive got an img2img loopback running at home rn, at work currently

I haven't tried using ai for it but I also write stories

Ai imagery I'll likely be using in the comics I am developing with a fellow author friend who draws way better than I

[daring kernel]

Still your life is better than mine

By like 1 million

[frigid torrent]

Wel,, thanks? But I hope you get to see a better life soon

[fallen dock]

and that you can for the lowest price of $9.99

[timid belfry]

i'm trying to get a spoiler to show up on this car but can't seem to get it to show up with inpainting?

i drew that to give it the basic shape of what i'mm looking for but nothing i get even comes close

[empty nymph]

Hello everyone! 👋
SD noob here looking for some advice
I've been trying SD for a couple of days and have been able to understand the basics, now I'm wondering what's the best way to generate images like the one attached using a specific subject?
I want all the images to be generated in that specific style.
From what I've seen so far, it seems I'd need to train a Lora for the subject and another one for the styles? I just wanted to know if that makes sense to you before getting on it.

[pastel mango]

There's something about Cyberpunk that's just really hot but I can't figure it out....This is Clara (Compliant. Liquidating. Analytical. Refurbished. Android) in various styles

[past radish]

Been working on isolated environments in crystals, here's the best of what I've made so far

Mountains and caves have been difficult to work with, completely failed to make a town

[pastel mango]

that's not bad

[slow cloak]

Holy thats great stuff, love thee idea, cool as fuck

[past radish]

Thanks!

Here's a couple others I made while experimenting with the prompt I've been using for them

[slow cloak]

That face

[past radish]

Yeah

[slow cloak]

I am trying to find out something I saw on civitia

[past radish]

It's the only successful attempt at getting a person inside one

What is it?

[slow cloak]

Trying to figure out how they managed to do this

[past radish]

bro what

[slow cloak]

(the picture is sfw, just no clue how they did this)

[past radish]

I'm not knowledgeable enough to understand this

[slow cloak]

Same

Only installed SD and tried it for the first time a couple of days ago

[past radish]

Yeah

Understandable

hmmm, weird

[pastel mango]

here's a few things that might help out anyone in here looking for references or whatnot

[past radish]

Ooo

thanks

[pastel mango]

and Im not sure how accurate this is, but it's for block weight merging

[past radish]

I think I have done it

It ain't the best, but I've gotten it

[slow cloak]

Ah ?

[pastel mango]

block weighted merging, bwm, or supermerger

where you choose each lane

[past radish]

Oooo the firey one looks really good

[slow cloak]

I grabbed supermerger

[pastel mango]

the last one is an original image + adobe firefly

[slow cloak]

That be looking mighty fine

[slow cloak]

I shall go to #🤝｜tech-support and coping for someone knowledgable (probably including yourself ?)

[pastel mango]

uh it's way too involved and convoluted for me to guide you in, you'd need to read up on it

[slow cloak]

Do you have a good source for it ?

also being dislexic doesn't help with understanding detailed written guides

I presume there isn't a program or extension that lets you enter an X amount for the merging ?

[pastel mango]

there is

Supermerger

or bwm

or auto bwm

(block weighted merger)

[sterile temple]

this naming style looks like simple model merges, then another model merge

[slow cloak]

I see!

Hmm, thinking about how it would have been merged then

[pastel mango]

well it was likely a simple merger, looks like

maybe add difference

it seems incomplete

because that's not a proper recipe

[slow cloak]

I also saw them using a combination like this

But not sure how helpful it really is

[pastel mango]

if I write it out it becomes 0.7(0.3(meinahentai_v4) + 0.7(0.35(meinahentai_v3)+ 0.65(tsubaki_v10))) + 0.3(animekawa_v10) and it doesn't make sense unless it's like doing several mergers in that extension I forgot the name of but you basically queue them up and then you can create a chain of it. It seems to involve tsubaki with meinahentai v4 and v3 with its ))

[slow cloak]

I see!

Hmm

[pastel mango]

would be (0.3(Meinahentai_v4 + (0.65(tsubaki_v10) , (0.35(Meinahentai_v3 + (0.65(tsubaki_v10) in that case I think

[slow cloak]

Oh ?

[pastel mango]

or something

I dont even know what extension it uses tho

so its very hard to tell

it doesnt look like anything I use

[slow cloak]

#🤝｜tech-support message

Saw it on this post

[pastel mango]

this is supermerger

the first goes on top of the second

I couldn't screenshot all the way

modelA=AAAAAmodel
B=BBBBB
position_id_fix=Keep Original
output_mode=Max Precision
0.67,0.53,0.4,0.41,0.37,0.48,0.45,0.52,0.58,0.62,0.69,0.66,0.5,0.04,0.13,0.18,0.48,0.19,0.19,0.45,0.19,0.49,0.22,0.2,0.56,0,0

thats what a blockweighted merger looks like when it only uses alpha channel

https://github.com/ashen-sensored/sd-webui-runtime-block-merge I prefer this one though

you basically load a model to go with your regular loaded one, then you fill in the blockweights and without merging them, you can generate and see what your merge will look like if you complete it

[slow cloak]

I can probably have both for extensions ?

[pastel mango]

sure they don't conflict just being installed

[slow cloak]

Where does it show up ?

[pastel mango]

which one? the runtime block merge?

[slow cloak]

Yea

[pastel mango]

it's on text2img

scroll down

[slow cloak]

Oh sorry I am blind

[pastel mango]

its very useful to try out different settings on a merge

[slow cloak]

Now I am just needing to understand how to exactly I am supposed to use it

[slow cloak]

Starting from there, I presume ?

[pastel mango]

uh I dont know

causeI dont know what extension they use

First like where did you even find those images showing that?

[slow cloak]

https://civitai.com/posts/392723 This one was for the one with the 3 models merged

https://civitai.com/posts/393192

This being the one with the 4 merged

[pastel mango]

yeah but where is the info on the merge?

[slow cloak]

On the bottom right, on the i

[pastel mango]

I dont know what that means because model is supposed to just show one model....the one that was loaded when you made the image. It looks more like they used some extension that can load several models and use them all in combination to generate

[slow cloak]

Now the question is just how in the world did they do that

[pastel mango]

I just don't know what they're trying to say. If they interrogated it and it gave the % chances of what model it is?

[slow cloak]

Maybe using controlnet multi-mode ?

[pastel mango]

I don't know

[slow cloak]

I guess I will try

[untold cobalt]

Hey!
Just wanted to share my new Pixel Art LoRa with u guys

https://civitai.com/models/111793/pixel-portrait

[slow cloak]

[pastel mango]

modelA=HyperVariant-AlphaV0.2.1.safetensors modelB=ExaltExarch0.3.safetensors position_id_fix=Keep Original output_mode=Max Precision 0.3,0.3,0.1,0.2,0.45,0.7,0.7,0.8,0.25,0.25,0.7,0.7,0.5,0.35,0.6,0.35,0.6,0.2,0.4,0.25,0.25,0.1,0.1,0.65,0.65,0,0
Resulting in my HyperExalt 0.1 alpha 😛

those two up there were already my own merges

so just joining up

[slow cloak]

nvm that is somehting completely different

Hmm, i see!

[slow cloak]

I just don't really know where to add these values in

Or is it just possible to just move a major slider to change the overall value of weight ?
Sorry, it is the first time I am merging models and I am still veryyy new to SD in general

[pastel mango]

well which extension are you using now?

where you add them will be different

[slow cloak]

I am installed both, super merge and the block weighted one

that you recommended

[pastel mango]

well on blockweight just paste the numbers in and hit enter. You have to load a model first though to enable it

on supermerger it's here

[slow cloak]

Imma try supermerger, since they seem very similar

[pastel mango]

supermerger is much more complicated though but yerrr sure

[slow cloak]

Ah-

Then we go with block weighted

[pastel mango]

load models up top for what you want to merge, check the box for use MBW if you want to use supermerger, then you can go down to that. If you use the sequential merge with elements etc...then it's really just a clusterfuq of convoluted stuff

[slow cloak]

I guess just a lot of merging in steps ?

[pastel mango]

no it's different lanes and each lane carries with it specific information from the model

such as anatomy, style, concept, background, color

[slow cloak]

I see I see!

My though process is that I would go through it like a math question

I guess it is in this order

Tsubaki + meinahentai_v3, followed by adding meinahentai_V4 and then animekawa_v10 ontop

[pastel mango]

yeah maybe they mean it like tsubaki and meinahentai v3 makes meinahentai v4, and then v4+animekawa makes something else

I dunno

[slow cloak]

Yea, like a math question xd

Thank you tho

[pastel mango]

but dude

that isnt a model

that's a lora

from the image you were looking at

[slow cloak]

I know that the character is a lora

Or which part ?

I might be smooth

[pastel mango]

yeah the character is a lora and the model thing makes no sense, especially if you look at all other ones. There's also no point in putting in that model info when you could just write the model. Unless you use a private model but then there are easier ways to write a recipe

[slow cloak]

Idk what would be best

[pastel mango]

ok now I think I might get it

0.7(0.5(meinahentai_v4) + 0.5(hassakuHentaiModel_v12)) + 0.3(animekawa_v10) would be meinahentai_v3+hassakuHentaiModel_v12 at merging 0.5 and 0.5. Then merging the result of that with Animekawa_V10

at 0.7, 0.3

[slow cloak]

I see!
Is there a big boy slider I can use to set the big values ?

[pastel mango]

0.7(0.3(meinahentai_v4) + 0.7(0.35(meinahentai_v3) + 0.65(tsubaki_v10))) + 0.3(animekawa_v10)

meinahentai_v4 0.3+meinahentai_v3 0.7, then meinahentai_v3 0.35 + 0.65 tsubaki_v10

Then results of both at 0.7+0.3

there should be on supermerger at the top, if you dont check the box for mbw

or if you have a checkpoint merger, do you have that on the menu @slow cloak ?

[slow cloak]

I do

[pastel mango]

go into that

set up the first models in a and b (meinahentai v4 and v3

set the slider to 0.3

use weighted sum

safetensors

don't add any vae where it says bake

copy config should say don't

then name the result

custom name

[slow cloak]

Wanted to start with the shorter one first and then the second one

[pastel mango]

then it's meinahentai_v3+hassakuHentaiModel_v12 at 0.5 weighted sum

[slow cloak]

So like this ?

[pastel mango]

no now it's 0.5

[slow cloak]

With the slider at 0.5

[pastel mango]

since you switched

no copy config

[slow cloak]

Yea, corrected that

[pastel mango]

don't

[slow cloak]

Aye

[pastel mango]

name the model something simpler

like meinahashentai

[slow cloak]

This was I will know what it is when merged, since I am gonna merge and know the values it is for the future™️

Naming convention is something for later to worry about

[pastel mango]

yeah but long names might be something windows doesn't like

if they get too long

[slow cloak]

Gotcha, I will keep that in mind

For the second merger then put the slider to 0.3 ?

[pastel mango]

umh

[slow cloak]

I presume the slider is inverse aka 0.3 = 70% model A ?

[slow cloak]

Doing this one

[pastel mango]

no

0.7

the model you made is 0.7

and animekawa is 0.3

[slow cloak]

So like this ?

[pastel mango]

yep

[slow cloak]

Merged, lemme try and then go for the other one

[pastel mango]

yeh, using checkpoint merger to begin with might be a good idea

when you grasp how it does things (merging) you can move over to the rest

[slow cloak]

Yea, still very new to this

Seems to liiking nice/similar, thank you ^^
Just gotta find out the exact quirks and mess around some more

What does merging actually do ?

[pastel mango]

it takes both models and blends them together after the percentages set

if you use block weighted however you can go deeper

and if you like the style of a model but not the content, you can figure out where style is on it and only merge that

then if you hate the anatomy and style of another, but you want the concept. You can set that

that's why weighted blocks are much more powerful

I managed to make a natural language model with a really broad variety of styles and concepts and one that can actually hold swords

https://civitai.com/models/6209/dark-gemini-v3 this one (file menu on the right side has fp16 version that's smaller)

[slow cloak]

I see!
Something to dive in for the future for me then

[pastel mango]

that's what I made earlier this year

[slow cloak]

Damn thats cool as fuck

Just lacking saturation seemingly to the source material but something with different VAEs can adjust ?

I just got the answer by trying out

[pastel mango]

if you use the one suggested on my model page

you get much better results

https://huggingface.co/nubby/kl-f8-anime2-blessed or potentially this one

[slow cloak]

Should investigate it but yep probably it has to do with it

.ckpt also correct for vae ?

[pastel mango]

yeah

[slow cloak]

Aye, thank you very much tho!!

Also send my friend who is looking for a similar model to yours, your model ^^

Would you like to see the result from the merging compared to the one the one I wanted to copy ?

[pastel mango]

cool

sure

[slow cloak]

The first one, with the 3 models merged, on the left "ours"

Was also using a different vae on accident

And with the second one, that was merging the 4 together, the one from the merge on the left again

[pastel mango]

heh, ecchi and hentai models always the most popular

[slow cloak]

[pastel mango]

why arent you hanging out in #🍥｜anime ?

[slow cloak]

I just joined because I was lost xd

[pastel mango]

aha

well great people there

[slow cloak]

I will probably annoy them in the future

But I know the concept now thanks to you ^^

[pastel mango]

aye no problem, there's a lot to learn with stable diffusion but once you learn to combine control net with regional prompting, adetailer etc and then multidiffusion on img2img you can create some really amazing things

I use Clip Studio Paint or Blender to set up a 3d scene, sometimes in layers depending on what I need where. (static objects vs character) and I set up control net for the character pose, then I set up the objects for lineart to see if they can be made into that, or canny. Then if there's an art style I like I can set that as reference and then I prompt something similar to the characters animation and the objects

[slow cloak]

I have yet to figure out how to actually make two character not look the same if using an Lora for the character

[pastel mango]

etc

well if you use a single character Lora that becomes difficult

unless you use AND to separate them into secondary prompts

and use the lora on that

[slow cloak]

how

[pastel mango]

I just told you

AND

[slow cloak]

Teach me the dark arts

[pastel mango]

literally

type AND in the prompt

[slow cloak]

Oh

lmao, gotcha

[pastel mango]

you can sort it by doing like
Prompt
AND
Lora Prompt

the main prompt should have style, concept etc in it

then lora prompt can just describe the character and what she's doing or whatever

[slow cloak]

And then a second AND ?

[past pelican]

I thought you needed to use BREAK to separate.

[slow cloak]

Explain, would love to know more

[pastel mango]

AND is different from BREAK

AND creates kind of like a separate prompt but still not. It makes sure it doesn't affect any others. I think

BREAK triggers token rotation

afaik

[slow cloak]

My mind trying to understand it:

[lyric storm]

my zero snr finetune is progressing pretty well it seems

[fallen dock]

it's not ai generated I know that much, lol I'm jk with ya

that's what I just said......

[weak sage]

why is img2img sometimes making my image complete black

I'm old enough to remember a time when if something malfunctioned in a program it would give an error code

[young crescent]

Any idea for what word(s) are being blocked in my prompt?

Sometimes the same prompt goes through, but I’m not able to generate anything from it because generation keep failing for mostly everyone right now.

[scenic umbra]

The bot is experiencing some issues with completing dreams. Please be patient as we look into it! thank you 🤖

[wispy fog]

well i mean it can always error 418

[slow cloak]

@quiet current What dark arts do you know off ?

Literally 6x faster

I am eternally grateful

[meager elbow]

what are we looking at?

[slow cloak]

Generation speed

Same.promt and setting as before for the what is being generated

After Aryetis told me the extra parts for what to add

[dry crow]

Yea amd is fast with the right argumrnts

Sry was AFK but aryetis told you the magic

[slow cloak]

No worries, still trying to figure out all the details

[dry crow]

Whats your GPU btw?

[slow cloak]

6900xt

[dry crow]

Ah nice. Then you can also use Highres fix with upscale by 2 without problems

[slow cloak]

Yea, sometimes SD just shits itself and you can't even generate a 512x512 image fsr

Gotta love the lack of rubbish collection and questionable memory management, iirc

[dry crow]

Normaly it shouldnt do this

But if your vram is filled to max then it can occur sometimes

[slow cloak]

Yea, it doesn't clear itself after generation

I presume there is some caching going on

Also also still suffering from this issue

[dry crow]

Make sure the model you try to use is larger then 1.99gb

And yes the vram gets cleared if you restart SD.

[slow cloak]

I am, it is ~2.05gb

[dry crow]

Okay good

Do you have --no-half-vae in your Webui-user.bat ?

That normaly should help. Also what vae are you using? Maybe load an other one

[slow cloak]

I need to mess around with the .vaes aswell, currently using one that is just called animevae.pt

[slow cloak]

Yes

[dry crow]

Try the kl-f8-anime2 vae

That never gets issues

[slow cloak]

The one at the top if I got the correct ones ?

[dry crow]

Yes but idk why the are called bless08 at the end

[slow cloak]

I got it from the hugging website, so not sure either

[dry crow]

Here is the original file:
https://huggingface.co/hakurei/waifu-diffusion-v1-4/blob/main/vae/kl-f8-anime2.ckpt

[slow cloak]

Imma grab that then

[frigid torrent]

https://archive.org/details/aibuildings_2023_07_01
https://archive.org/details/aibuildings_2023_07_10
https://archive.org/details/aibuildings_2023_07_11
https://archive.org/details/aibuildings_2023_07_17

Uploaded some morre

[frigid torrent]

I think these kind of images are disallowed here, tread with care (and check rules channel)

[wispy fog]

Did, put spoilers and warning

Still noone answered me

Why is the ai horny

I think if anyone clicked they knew it was nsfw

[frigid torrent]

Well it's just one of the rules.. that's all I wanted to point out.

And to answer your question, it'd be the training data (so, more broadly the general internet) which is the "horny" one. Association of a cat as well as an exposed belly might have done it. Quite an impure place the web is

[meager elbow]

there's no way SDXL produced that picture without some explicit prompt

[weak sage]

if u want share nsfw this the wrong server, no offense

[weak sage]

idk i wouldnt be surprised

[meager elbow]

I just tried all I could, even using "p*ssy" instead of "cat"

and still all I've got is lovingly delicious kitties

SD1.5 is another story of course

(PS: sorry my bad I thought we were in the sdxl chat btw)

[weak sage]

right

[meager elbow]

then yes, all the way boobland

[weak sage]

i prefer just to have some idea what images my model was trained on, so i can counteract whatever the bias is

[dry hill]

Hello guys, i ca&n't find the endpoint to create img in my local stable diffusion through http call

[meager elbow]

"naked" in the negatives is pretty much standard 😄

blame the trainers

[pastel mango]

snr?

[copper cipher]

someone sent me this AI generated image. I asked me friend and he told me that this should be generated by using "standard/public" LoRA model . Anyone know what LoRA model is used and prompts that I can use to generate similar image?

[pastel mango]

looks like a futuristic synthetic human in a white exoskeleton with black joint areas. Or any variation thereof or modification and of course it wont turn out like that with what I wrote up there as literal. You have to use it in whatever manner your model understands.

if I search for standard on civitai all I get is porn....lol

so yeah careful

[frigid torrent]

I'm here for crow guy

[wispy fog]

Rev 1.2.2 is horny

[wispy fog]

I put naked nsfw and explicit in negatives

Always

But rev is horny as i said

[meager elbow]

yeah sorry I thought we were in sdxl, it's all I check these days 😄

[wispy fog]

Yup

Select squared with capy

[kindred willow]

This is what birds would look like if they were real.
Why he a sailor tho?

[hoary monolith]

Hey there, trying out deforum for the first time, followed a few tutorials on yt, I finally can get pretty consistant animation, yet this "doubling" happens, tried messing with the parameters a bit but it doesn't seem to really go away, I still don't understand how the tool works fully so I'm just shooting in the dark. Maybe someone has any ideas why this is haappening?

[scenic umbra]

are you generating on top of a real video? seems like some kind of interpolation between frames. you can try asking on the https://discord.com/channels/1002292111942635562/1025266140445933648 channel, lots of deforum users there

[hoary monolith]

Yes, I'm using a video to generate this, I'll go ahead and crosspost then, thanks!

[novel shell]

Hello, can someone help me to run stbale duffision 2.1o. My computer

[pastel mango]

what's this type of armor called?

is it a Turgeon?

[hollow oak]

is it lotr?

[past radish]

Originally no

Funnily enough when I saw this I immediately thought of the ring of power as well

[scenic umbra]

Just a friendly reminder, we need everyone to go and vote on the bot channels, please select what image you like best: "A" or "B", this is for helping improve and finetune what will be the gold ⭐ SDXL 1.0 model.

[pure flower]

Im curious if anyone has experimented with precise, image to image infographics

[kindred willow]

Where did he sail to tho?

[cyan gulch]

Домалюй клоуна

[rocky cypress]

how to fix this error?

[steep shadow]

https://cdn.discordapp.com/attachments/923833387905581108/1131417366383824997/caption.gif

i love paperspace

good enough

[past radish]

Forest 1 VS forest 2

Both are decent

Can't decide which one is better

[pure bane]

W or L anyone?

[daring fractal]

I'd eat it! 😍

[compact terrace]

Serj

[lime igloo]

I was kinda liking these.

[frigid torrent]

https://archive.org/details/aibuildings_2023_07_19

[echo matrix]

Cute kitty, first imagine I generated after hours of making this crap work on amd

[cyan snow]

I tried using SDXL in A1111 with the new branch that supports it, didn't work. back to 1.5! =[

Vlad's UI isn't as good as A1111, and comfyUI works well, but not as consistant

[echo matrix]

SO CUTE

[echo matrix]

What does that mean

[echo matrix]

Wow wtf, it's beautiful, Especially last one the best

[eternal geode]

https://cdn.discordapp.com/attachments/1127297005941694505/1131607580989460663/view.png
8k

https://cdn.discordapp.com/attachments/1127297005941694505/1131602073788620870/view.png

[indigo aurora]

Hi. Can I know which model did you use to create this?

[past radish]

Dreamshaper_7
I used the Magmatech lora

[subtle ridge]

Anyone got some constructive criticism?

[subtle ridge]

also that is amazing

[agile idol]

ComfyUI with control net 😉 Don't ask for the Nuuuedist Beeach version. 😉

Realism model in development:

[wild sorrel]

🖐️

left hand - 6 fingers, right hand - seems like 5

on everything else not sure what I'm looking at and what it should look like to critisize

[rough stag]

i know, i know, enough with the posters but this came out so well

[wild sorrel]

TASAPOP sounds like futuristic movie name, when all actual words are taken for other, older movies lol

[stoic pivot]

Prompts I regret.

[rough stag]

whatever it chose to mix and blend from the prompt to come out with that it really sounds like a title for a horror movie, and i was going for that so....not disappointed

[ripe cedar]

For img2img to make it more stylized, I am using add detail Lora anime line-art ControlNet, anime esrgan upscale, and Face Detailer. Should I then add another upscale, tile preprocessor, or something else?

[versed hill]

Generate image of root written in roots

[stuck ferry]

Did you use SDXL to get those images?

[jagged shoal]

Is this SDXL based?

[jagged shoal]

looks great. couldn't import into comfyUI

Which model is htis?

[hasty notch]

？？？？

[nova basalt]

Indian a girl looks like Barbie Looking in the mirror with bleach applied to her face

[valid belfry]

imagine a dog

[valid belfry]

imagin a panda

[plain finch]

icon, art, aviaiton, collage, patchwork, safety, color #3D1313, color #0EA58A, color #FBEADD

[subtle ridge]

wtf was the prompt

[stoic pivot]

"feminized Kermit in a sexy swimsuit, pinup pose on a tropical paradise beach, warm color palette, sipping tea", it may be a dangerous prompt, use it carefully.

[near bough]

a scary story youtube channel logo

[frigid torrent]

Another printed out two segments generated off the hypernet I trained on my drawings, another drawing out of them by hand. To be scanned later and turned back into more training data :3

[near bough]

#🏞｜general-with-images a youtube channel banner for a scary story channel, unreal engine 5, high octane render --ar 5:1

a youtube channel banner for a scary story channel, unreal engine 5, high octane render --ar 5:1

[pastel mango]

Hello humans, want to come play with us in the marssssshhhheeeesssss? We've been lonely forever forever forever. What we won't give for someone new to play with forever forever forever

[rose bear]

Had to download the vid from my insta because I dont have it on my phone, but this is what I mean. And the art is not completed until you add some audio and video effects.

This lora is not mine but the crystals look amazing

[pastel mango]

which one is best to use as base for my forest generate? One of the four variations or the fifth, the original?

[tulip yew]

DO I NEED TO DOWNLOAD THIS?

[vapid crest]

just get the pruned model

[tulip yew]

Is it necessary to have some models? Or better a TOP 3 models that I like and different LORAS

[vapid crest]

im trying to inpaint away the thumb on the right arm to be the brown skirt, but... it keeps giving me the thumb again. any tips?

[vapid crest]

really depends what style you're looking for. and how much disk space you have

[white pivot]

.

[glossy herald]

meow to you Aku 🙂

[white pivot]

Meow 🤣

[sleek dome]

draw it out in an image editor

and rerender at gigalow denoise

[vapid crest]

how low? less than 0.3?

[quiet current]

THOSE ARE THE DIFFERENT VARIANTS OF THE MODEL

just download the one which fits your need (most likely the pruned one)

[rocky oriole]

Hi all. Does anyone know the requirements for image 2 image images

Everything I try doesn't work

Although any images generated by SD work as inputs

[brazen rampart]

i made this on bing image creator

i typed "bismuth lego minifigure"

[brazen rampart]

looks cool to bad i cant download it, i broke my laptop and im using my mom's

[dull notch]

I'm one of the founders of that software. ✨

Oh I'm sorry!
I'm going to delete that

[glossy herald]

well, sharing your work in a non spam way is ok usually

[dull notch]

... well.. it was a free project for local generations

Like automatic1111 I thought you might like it..

Yes... Because it's a different version

[glossy herald]

well, but I don't see a paying version for now. I could be wrong, but this seemed like a future paying software pushing a communication on their usable demo, maybe for feedback ? Not sure. I haven't tested it, but like, anybody could make a new tool and deservers their chance to share what they do.
The "close source" "potentially paying" side is not what we love around here, but as long as we don't go in ad spam territory, communication is OK

[quiet current]

Didn't have time to take a proper look at the entire message. But just in case, if it was a genuine side project/not an add/etc, it might be a better idea to share it in #1092446741984444416

[glossy herald]

no idea what you are talking about mate

[quiet current]

And overall, doing advertisement for other's projects without the approval of the author(s) might be a bad idea. So there's that.

[glossy herald]

good advice, this @dull notch

no idea, I just came back from a 2 months break, I have no tool installed currently

[dull notch]

Yes, I'll follow that suggestion and move that post there!

[quiet current]

git checkout dev and I'm not sure tensorrt requires it anymore as many things have changed since then.

be warned that things will break when switching to the dev branch.

it is not meant to be daily used by anyone.

[dull notch]

I'm one of the founders of the Arthemy Startup and we made a free software for anyone to use - we don't have paid features and we were just curious to give this alpha version of our software to the community as an alternative to EasyDiffusion, A1111 to understand if anyone liked its flow.

[quiet current]

yes master/main is the usual branch

[split rover]

Welcome back Guiz 😄

[glossy herald]

TIS ME !

GUIZ!

I saw you got leaked while I was out mate 😢

well "you"

I mean the XL

also, I saw some promising measurements on VRAM use :p

less than the 24GB I was scared of :p

[split rover]

All good dude, 1.0 is coming soon anyways haha

[split rover]

Yea runs just fine on lower vram 🙂

[glossy herald]

yeah, Saytan gave me some XL news :p

haven't tried 0.9 yet, since I just came back, but I may just wait for 1.0 now

[split rover]

Hope you were alright wherever you were off too!

[glossy herald]

4 days to go or something :p

well, I won't put the channel down so I may just not answer on that one. Doing better now 🙂

I hope it was doing great here ! times seems to flow at an other speed in this AI world :p

[split rover]

It never stops 🕰️

[split rover]

Well glad you are back 🙂

Also I remember you mentioning I should try comfy way back, well we liked it so much hes on the team now if you didn't see haha

[glossy herald]

yeah, I sent a thanks in a public channel to them earlier, but they seem popular. I'm really happy the team integreated them too, they have so much great ideas on this, and it opens so many possibility in terms of pipeline inovation

[split rover]

Totally agreed 😄

[glossy herald]

like, I'm not entirely sure what was Saytan's pipeline idea that got integrated and everyone is gushing about, but it's something like refining the first picture through a second model or something like that ? Isn't it through such tools that the community can participate, even with the lowest knowledge on the inner workings ?

Plus I mean....
I never had a visual novel tutorial before

[split valve]

My card has 32 of RAM. is that good enough to run SDXL?

[proud dagger]

where can i download the Dustin torrent? I wanna run dustin at home

[glossy herald]

if we could DL Dustin, our HD would explode made

[mild tusk]

lelouch

and a gunner

and a sniper

[tropic shell]

niceeeeeeeeeee

[strong sonnet]

Is this officials chat

Too many SAI staff in here

[prisma iron]

Too many, you say? 🙃

[mild tusk]

not enough

[smoky oak]

Why did it take me so long to find out that RVC is insanely good at singing coveres and music in general-

Its legit so good, oh my god

and fast too

This ariana grande model is insane

(takes about 3-4 seconds to render on a 3090)

took 1 second to render on my 3090

[strong sonnet]

atleast more than the staff talking on #✨｜sdxl

[strong sonnet]

its really great, the only thing that irritates me is extracting vocals

[iron kernel]

Hello! There is an Ai based Design Projekt which needs some votes to get the design world to recognise ai’s potential, please vote! Thank you ist would mean the world to me 😊 the Design award is based in Barcelona so the website is in Spanish i hope it is not a problem! Today is the last day to vote. Feel free to share the link and ask me any questions:
https://pr.easypromosapp.com/voteme/963372/646546313?lc=spa

[grizzled sage]

it really is so scary good. The influx of these systems is part of the reason network and studio execs are trying to bleed actors and writers dry. they're counting on AI being better than good enough sooner than later.

Movies and TV are going to start to suck as it becomes more business driven and the passionate creators are kicked out by business school graduates. I mean, a lot of it already sucked, but once in a while you'd get projects with actual heart to them.

The tide is turning. The business school types think they got the upperhand now due to these tools. I wonder how the pushback will creatively manifest.

Last writers strike we got things like "Dr Horrible" so i'm really looking forward to a new pushback this time

[terse solstice]

What's the model?

[wild sorrel]

I'd take that over autotune any day

[rain gazelle]

Been traveling for like 2 weeks - what's RVC? What'd I miss? 😛

[split valve]

SDXL 0.9 just made correct hands and fingers both sides... is that supposed to be possible?

[faint furnace]

it seems like that has been improving a bit in the last stages but every once in a while it comes up perfect

[cyan snow]

it even does text correctly ~80% of the time

all caps is incredibly easy with it. heck, I even managed to get entire sentences using SDXL

[pseudo solstice]

I'm using SD i2i to enhance images for a comic from pregenerated images, with loras I have got somewhat consistent over the facial structure. Is there easier way to control the eye colors through it without going to have to use masks or adetailer/ddetailer or controlnets etc etc

[chrome kite]

Lizard people

[raven bramble]

https://tensor.art/images/618044918000243253?post_id=618046236551011167&source_id=nz-0rl_lnUC3ovAibnz38RQm

https://tensor.art/images/617928803559181389?post_id=618045996032842177&source_id=nz-0rl_lnUC3ovAibnz38RQm

[livid crystal]

Bruh how do i get higher quality images

I know i am not doing something right

[chrome kite]

It takes practice. It also helps to learn from other people’s prompts. I’m not sure if you would use Reddit, but there is a Reddit Stable Diffusion forum. People post lots of images there, and you can sort by “Workflow included,” which means you will find images along with the prompts people used to generate them and any further processing steps they took.

[raven bramble]

https://tensor.art/images/617409395984367304?post_id=617458994262354388&source_id=nz-0rl_lnUC3ovAibnz38RQm

[urban ibex]

Well, it's better than the mutations of previous models, but it still leaves room for improvement. Her right hand (on our left side) the thumb "pad" is too pinched, so anatomically, where the thumb would be (you can't see it, you have to infer) would be in the wrong place. Most people probably won't notice this, but artists will. Both hands look like they belong on a 20-30yr old woman, but the face looks more like a pre-teen girl, so that doesn't match up. A vast improvement over previous models, for sure, though!

[urban ibex]

https://rvc-models.com/category/downloads/musicians/ - I don't see Arianna Grande in the list of musicians?

[smoky oak]

they have a server for it, they have 1000's of models there, with people uploading several per hour. its called AI HUB

I can link in DM's if you would like

[strong sonnet]

@smoky oak did you use UVR to extract the vocals?

[smoky oak]

I didn't extract them, just use pre-extracted vocals

[strong sonnet]

where did you find them?

[smoky oak]

youtube

[strong sonnet]

oh ok
thank you

[smoky oak]

just look up a song and look for "isolated vocals"

Messed with this one too lol

21 seconds to render on a 3090

the slight squeaks on some notes are cause I was using not ideal settings

this is from one of my friends who is actually a tester for this software. They are demoing a couple new algorithms trained much better, and then better rendering settings as well

its a new tech for this effect, so there are some weird pitch loss sections, but the vocal cracks and such are near perfect

the weird low notes that slip out on the Zombie part are from the back acapella

[steep sleet]

Papercut Anime

[steep sleet]

Awwwww 🙂

[strong sonnet]

oh, your 3090 is perfect?

[smoky oak]

it has issues, but we figured them out

its working right now

[wispy nest]

Greetings, how do I create prompt, so there will be something in the middle but with removable background...

[strong sonnet]

thats great that you have figured it out before 26th

[daring kernel]

Is sxdl coming tomorrow or 26th?

[smoky oak]

26th

[steep sleet]

62th

[cyan snow]

real

[strong sonnet]

for sure?

100%?

i hope they dont postpone it again

[smoky oak]

26th is what I was told by staff on delay day

If they manage to disappoint the community again, then they are even more reckless than I first thought

[strong sonnet]

yeah, i wouldn't trsut SAI anymore if they do that, they should atleast postpone it few days before the launch

but never on the launch day

[cyan snow]

idk, 0.9 is way above expectations already. I'm waiting for the specialized finetuning tools that will be released

[smoky oak]

Here's helping we will be able to use them, cuz I'm not sure about that yet

[lost patrol]

The idea was to have a girl dancing with a skeleton.. maybe it's a fail 😅
An Italian woman wearing black lingerie is (dancing a walzer with a skeleton:1.1),

[grizzled sage]

kohya scripts have sdxl specific files already. they work good on 0.9 and will likely work with the 1.0 release

[cyan snow]

still, I'm stoked to see the tools SAI made

[grizzled sage]

i dont expect anything other than mvp to release there

and that don't mean kobe, god rest his soul

https://github.com/Stability-AI/generative-models#training as far as i can tell, their tools already are released

[strong sonnet]

@smoky oak are you still working with pseudo to finetune a model?

[smoky oak]

I already talked on this multiple times. We are not starting any work until 1.0 is out and we can test its shortcomings

[strong sonnet]

yeah you said that but pseudo was quiet disappointed with SAI(not about the release)

so i was asking if he was still interested?

[smoky oak]

we will have to see if 1.0 fixes some of the huge issues with 0.9, which it seems like it does

[grizzled sage]

i'm hoping for another sdxl_report.pdf i got a lot of great ifnormation out of the first one

[cyan snow]

idk what issues are there, SDXL reaches a level of detail and precision no other model got to yet

[smoky oak]

it has a long way to go

it will be much cleaner once people fix some issues lowering its pixel level quality

[grizzled sage]

i've been slowly developing a dataset plan for my DF model. i've put more spreadsheet work into it than image collection work so far. good thing too because now that sdxl is out, i think i'm going to pivot to using it to create a dataset a lot richer detail. willing to rent a gpu cluster to train this once i've got datasets going so switching to sdxl won't cause any problems with my plan either.

i dont' want to get caught in a duke nukem forever cycle though so i'm going to commit to making my 10k image dataset plan in sdxl, even if sd69 comes out

[smoky oak]

which pseudo was able to do with 2.1, which was massively more fried

10k images is pretty small for a dreambooth, what are you trying to train exactly?

[grizzled sage]

i heard someone referencing a step skip technique you were talking about so i looked into it more. leaving one step out between the refiner and the base. it makes really neat details. just wanted to say thanks for putting it out there because it eventually boomeranged it's way around to me

/random

[smoky oak]

if you are using a full step offset, then it will introduce a lot more problems than good

my method uses a fractional misalignment, not sure how you are doing it

[grizzled sage]

i just threw a step offset in. it didn't make too many problems in my generations. added a lot of granularity that i've often looked for. definately not something i'd use everywhere

tell me about fractional offset

[smoky oak]

may I see your sampler chain to see how you did it? You might already be using fractional, depending

normal, vs fractional, vs full step offset

some samplers are more sensitive to full step, but fractional works on most of them

[smoky oak]

if I can see your sampler setup, i could give some insight

[strong sonnet]

He said that he was able to generate realistic faces using 100k images or so

[smoky oak]

yeah, we will have to see

[strong sonnet]

I hope he joins back

[smoky oak]

he has no plans to, he is done with this server

[grizzled sage]

100% the layout i saw you release, with a few post process things tagged on which are irrelevant to diffusion. its all you that i'm using since i don't understand how to wire comfy very well yet

i change things like samplers and steps. thats about it

[smoky oak]

can I just see your sampling nodes? Just the k samplers are needed

depending on the sampler, step count, split location, and schedueler, it can change fractional or full step offset drastically

[grizzled sage]

i dont have it loaded right now tbh. away from that desk while i try get some things around the house done. on the auxilery desk where i can procrastinate around the work i have to focus on

[smoky oak]

all good, no worries

When you can, please send me a screenshot of your k sampler nodes so I can take a look and see if the setup you have is full or fractional

they are similar in extecution, but they reproduce very different results depending on a wide range of factors

[grizzled sage]

the way i did it was each refiner and base had the same scheduler. 25 steps total. base was 1-19 and refiner was 20-25

i adjusted around that theory

[smoky oak]

ok yeah, thats full offset

so for fractional, you would do it this way

base does 1-20/26, refiner does 20-25/25

The fractional misalignment between 20/26 and 20/25 is enough to inject some noise, but not too much

[grizzled sage]

thats what i figured. it seems like a self explanatory name. nice an intuitive

[smoky oak]

its a very small misalignment, but it brings some additional detail, without as many of the downsides (full step loves to destroy very simple images)

[grizzled sage]

yeah the fractional name suggests just offsetting a fraction of step. i see how you achieve that now i think

[smoky oak]

yup, I try to be as intuitive with my names as possible

[grizzled sage]

the theory behind it makes a lot of sense and demonstrates very well too. it's such a nice trick

noise offset. it's the new black.

[smoky oak]

Mixed diffusion is the name of the process of using two models on one image for SDXL. Its what Diffusers has adopted for their pipeline, and pseudo and I submitted it under that name haha

[grizzled sage]

step offset

you name it

you're better

[smoky oak]

So I have Mixed diffusion, and fractional offset under my belt ATM

and I have an idea for better upscaling, just need to figure out how to properly implement it, as it will be code, and it will be way above what I can do haha

I have not been that active here as of late, cause as I get into RVC, I am already talking to their devs for potential fixes and ideas.

Would love to get some SD and RVC credits under my belt. Would likely look good for ML versatility

[grizzled sage]

yeah i can see how loading 1.5 models into the process could work in comfy.

nodes are awesome

[smoky oak]

quick side note, I forgot

[smoky oak]

SDXL 1.0 in the bots here, when it performs well with the settings, looks about as good on pixel level detail as Pseudo and I were confident we could achieve

[smoky oak]

||oversexualized and fetishized women, what a unique and original concept in this community! lol||

[strong sonnet]

And plans to experiment with SDXL too?

[smoky oak]

We are both messing with SDXL

and we will both continue

[strong sonnet]

Oh ok

That's better

But it would great to have him back

[smoky oak]

He won't be back any time soon. Likely not ever

[cyan snow]

something here is unsettling. kinda like the backrooms

I mean- it's a good gen and all, but this is liminal

idk how to describe this man

am I crazy for thinking that?

[smoky oak]

no, I see it

add "hazy, foggy, low contrast" to the negative

I had that issue when I first started with SDXL

[strong sonnet]

@smoky oak you said that you are gonna finetune, so if you want any images please ping me, I will try to scrape websites and gather them for you

[cyan snow]

I don't need to, the rest of the gens were something like this

[smoky oak]

thank you, its very appreciated

[cyan snow]

I even did a 16img batch, this was the only liminal one. cursed seed??

[jovial umbra]

anyobe know why the heck theres no image2image on Clipdrop?

[smoky oak]

I think they all look a little unsettling with how they are hazy and washed out

[cyan snow]

this is unsettling?

[jovial umbra]

yes the dentist bill...

[smoky oak]

maybe not that one, but I would prefer them being a little less low contrast and hazy

personally, at least

[cyan snow]

idk, when using that one specific seed it just makes creepy pastas, it's like a curse. all the other gens I did didn't give me chills

[cyan snow]

like heck, the rest of the batch looks happy and all, but that one image is like a backrooms level

[tawny forge]

punctual as always

[smoky oak]

lmao

I love the idea of a bot not being on time

something about that tickles me lol

[quick lichen]

been messing with tiling recently

can make some pretty crazy and abstract stuff

[wild sorrel]

or textures

I generated plenty of them, but that's the only one I actually tested if it can be tiled

[quick lichen]

and i'm assuming it worked?

[wild sorrel]

not perfectly, but yea

need a lot higher resolution gen to make repeating pattern less notable

[quick lichen]

iirc there are non-AI tools that exist to make tiling less obvious

[wild sorrel]

yea, there are few papers on that, don't remember how exactly it works

[quick lichen]

use this tiling for your needs bro

trust me bro

[wild sorrel]

why tho

[quick lichen]

dammit i was going for a specific bit

this shape uh

tiles the plane oddly

[wild sorrel]

I'm to lazy to go and actually apply it 😄

[quick lichen]

the shape repeats but you gotta keep rotating it

[wild sorrel]

oh I see

[quick lichen]

oh also you gotta flip some of em

[cyan snow]

hey guys, what;s up with the gigabot reacting to some gens with envelope emoji?

[rain gazelle]

oo this one turned out well

[pseudo solstice]

WARNING: No libcudart.so found! Install CUDA or the cudatoolkit package (anaconda)!
Does someone know why am I getting this error on kohya ss?

[dry crow]

No cuda toolkit installed

If your on a GTX card Install cuda 11.8 if RTX then the latest 12.

[pseudo solstice]

I didn't need it before, apparently it wasn't copying the bitsandbytes files properly. Oh it's installed.
Basically, the bitsandbytes files needed to be copied properly before installing bitsandbytes

[dry crow]

Okay good to know

[cyan snow]

11.8 is still the fastest even on 40xx cards

[dry crow]

Really? What did they messed up on 12 xD

[pseudo solstice]

But need help on kohya still. All my trained lora through it are coming damaged, all they output back is a black image when I try to use it

[cyan snow]

same thing they messed up when they thought 4060ti is a good idea

[cyan snow]

however, cudNN is the fastest with the latest version of itself.

so, best inference is CUDA 11.8 + latest cudNN

[pseudo solstice]

can ya guys lemme know why my trained loras are giving black images 😅

wasted an entire day over it and still can't fix

[dry crow]

Do you load them correctly ?

And dont use --disable-nan-check

[pseudo solstice]

yes. I mean my downloaded loras work just fine

no?

is it for trained loras only or what?

[dry crow]

Okay do you know if there are cudnn files for GTX cards?

[dry crow]

No then it should work

[cyan snow]

I have absolutely no idea, lemme check

[pseudo solstice]

well it's not working then, all trained loras just give me a black image.

[cyan snow]

okay, it's not likely to make any difference for GTX cards, it's just faster pytorch files, and idk how pytorch behaves with GTX cards.

[dry crow]

Yea thats what i thought. Thx for checking still

[pseudo solstice]

ah man, I guess it might have worked now. previous installation was probably borked. lemme try running with my full dataset

I have 8G vram should I use lowvram?

[cyan snow]

however, the latest cudNN files are meant for specific generation of CUDA cores(the ones on 40xx cards), this might mean that if you dig deep enough in Nvidia's archives you might find a cudNN version for a GTX, that is, if cudNN was a thing when GTX was still made

this is the same reason that 40xx cards work ~13% faster with the cudNN files, it's just dlls designed for that specific gen of CUDA cores.

[pseudo solstice]

can you guys recommend a good config to train lora btw

[dry crow]

Im not into training sry :/ with 8gb you shouldnt need to use the lowvram option i think

[pseudo solstice]

does it slow the generation down though?

[dry crow]

Good point. But the Speed is okay for my GTX.

[cyan snow]

oh, why are you not going RTX?

[dry crow]

If they wouldnt be so pricey :/

GTX1080 is still good in Games

[cyan snow]

dang, you are stuck with gtx since when?

[dry crow]

2018

I want a new GPU. But i dont like nvidia price tactics

[cyan snow]

I think your already overdue. you deserve an rtx 40 smthn

[cyan snow]

fair enough. make sure to not go AMD though, you know what they think about AI

[dry crow]

Yea 12-16gb vram would be great

[smoky oak]

the used market should have many good options depending on where you are located

12GB 3080 should be like $450 ish

my 3090 was $600

granted, my 3090 has a ton of issues, but thats a fluke issue with shipping

[daring kernel]

When the 1080 launched it was 599

[wispy nest]

im disappointed

this is SDXL via bot

and this is SD 1.5 using same prompt

[smoky oak]

its an untrained base model thats not even done...

[wispy nest]

well that gives me a bit of hope

[smoky oak]

do you even know what 1.5 looked like when it was released? lmao

[wispy nest]

i know ... i hope you are right

[smoky oak]

I am already making LORA's for SDXL, we will be just fine

[wispy nest]

ok .. sounds good

glad to have your input 🙂

[smoky oak]

base 1.5 vs outdated SDXL Base 0.9 vs 1.0 bot previews

[wispy nest]

oooh, that looks good, im guessing that bottom right is SDXL 1.0

[smoky oak]

same here. 1.5 base, 0.9, 1.0 preview

so yeah, I'd say you have nothing to worry about lmao

SDXL can do a lot of things better out of the box than some of the best 1.5 finetunes

and its really pleasant to train

[wispy nest]

you have me smiling

now waiting for the dev to release it already lol

[daring kernel]

Tomorrow

[wispy nest]

ahh sweet, no more delays please

[daring kernel]

I hope not

[smoky oak]

Release is for July 26th. I can share that info now that people have since figured it out lol

My documentation should be out by then

it all got stalled by my GPU issues