🍥｜anime

[gentle musk]

Just a single KSampler at the high resolution with the model going through the HighRes Fix node.

[stable crane]

no, highres fix still works

just adding more details

[gentle musk]

Yeah you can use it after, but you don't need it

[stable crane]

this works well

[steady grail]

i spent the last 3 months studying highres fix on comfy

[stable crane]

only 105s

[steady grail]

let's see how this node influence the work

[brittle jay]

Guys do you even know what the purpose of hires fix is/was?

[steady grail]

arron, do you knnow the difference between gradient and not gradient node?

[steady grail]

adding details to upscaled images

[gentle musk]

The gradient is something new he did last night testing stuff, he said it only really works on SD 1.5

[nova remnant]

would work better in my hands

[brittle jay]

No

[nova remnant]

[gentle musk]

It's to work around the fact that SD goes a bit mad and starts making duplicates of things at higher resolutions than it's designed for.

[brittle jay]

Yes

[steady grail]

oh

[brittle jay]

It fixes high resolution images

[gentle musk]

I posted an example a bit ago of what this "fix" does.

[brittle jay]

Images are mostly trained on 512x512

If you make it create a 1024x1024 image for example, it duplicates stuff and does bad stuff

[steady grail]

yep

[brittle jay]

So you make 512x512 image

And upscale it to 1024x1024

That's basically it

[gentle musk]

And then do a 2nd pass over it as img2img

[stable crane]

so kohya must be scaling the model or something (judging by parameter names)

[gentle musk]

He's scaling the unet I believe

[steady grail]

what's the unet?

[gentle musk]

He's explained it on his twitter, but it's in Japanese and I'm not sure how well Twitters Translation works.

[brittle jay]

In comfyui terms. Create 512x512 image -> upscale esrgan -> downscale to what you want (1024x1024) -> img2img -> output

[nova remnant]

What does exponential mean in the samplers

[brittle jay]

This is hires fix

[brittle jay]

Unet is a hella complicated thing

Just give up

[gentle musk]

It's all maths and latents

[stable crane]

the default scale was 2, so I figured if you want 2x larger than 1024x1920, you have to use 4x

[brittle jay]

I've been trying to understand unets for half a year

[stable crane]

I think tiled vae screws it though

[brittle jay]

Only got like halfway there

[steady grail]

i know that feeling -.-'

[brittle jay]

Nubby please bless us

With your nubby skills and knowldege

Ok I go slep

GN everyone

[steady grail]

nighty night

[supple raptor]

it's the noise scheduler. haven't looked into the exact math, but i suppose it probably uses exponential scaling on the added noise.

[nova remnant]

@stable crane

[nova remnant]

Nubby is typing

cleans keyboard, grabs tea, stretches meanwhile

[native halo]

https://tenor.com/view/discord-who-asked-me-looking-for-who-asked-travolta-meme-travolta-confused-gif-27133622

[supple raptor]

the model is made up of 3 basic components.

• Text Encoder (sometimes referred to as CLIP) turns your prompt into vector embeddings. Basically numbers that can be interpreted by the UNET
• UNET takes the vector embeddings created by CLIP and does lots of funny math to find the output that best matches them (this is the image generation process). what it spits out are still just numbers though, these numbers are called latents and they each represent an 8x8 block of pixels
• VAE translates the latents into pixels so that you can have an image. in img2img it is also used in reverse to translate the pixels into latents that the UNET will be able to work with.

[nova remnant]

@native halo what is the theme for today

[supple raptor]

don't worry, nobody really understands them. it's largely a magic box that works, but we don't know why. some people understand them much better than you or I do though.

[nova remnant]

magic box

[steady grail]

ohhh

[native halo]

girls with crow head hat

[stable crane]

I've gott try this on realistic now too

[steady grail]

the TE is the tag of the image's data and the unet is the library of data images

[nova remnant]

I just got another load of thick witches to gen, will try some crow

hopefully not alive on their heads

[steady grail]

unet uses the TE to call the right image's data

[stable crane]

he crossed out thick 😢 😢

[steady grail]

or something close to that

that's why is called net

[supple raptor]

not really no. there is no library of images. and unet is what actually learns the text/image pairing. TE is kind of like VAE but for text.

[nova remnant]

(Its so admins don't bonk me)

surely works

[steady grail]

a model, is not a library of images converted in code, linked by tags?

an image is tagget with tags like green eyes

ecc

[stable crane]

so good

[stable crane]

it's not storing the images like a database

[supple raptor]

the model learns the relationship between the tags (more accurately the tokens that represent those tags) and the images. it should not be capable of ever recreating the images that it was trained on. this can happen sometimes, but it is not a desired behavior. that is what we call overfitting.

[nova remnant]

@stable crane

[stable crane]

nice, getting thicker

[native halo]

where the crows at

[steady grail]

so, analyze the images, looks for relations between tags and forms and store them?

[nova remnant]

[stable crane]

https://github.com/yl4579/StyleTTS2 this looks fun

[supple raptor]

Think of it sort of like teaching a child their colors. You show them many different objects in different colors and eventually they will figure out that red means the color of the apple, not the shape and will be able to tell you that the firetruck is red even if you hadn't previously showed them a picture of a firetruck.

[steady grail]

and this bing net of relations and tokes is the unet, i suppose

[supple raptor]

yep

[steady grail]

mind numbing, but very cool

[stable crane]

and it's all powered by sand

🤯

[steady grail]

lol

[stable crane]

sometimes I really feel like I'm using magic

this experimental node seems to crash comfy a few bit

[supple raptor]

here's the a1111 extension for it if you want to try it there https://github.com/wcde/sd-webui-kohya-hiresfix.git

[stable crane]

ty, I might

well, probably don't dare

due to VRAM

I wish it were easier to re-use seed in comfy

or mark a node as dirty

[supple raptor]

what do you mean?

[stable crane]

I think if you generate once with "fixed" seed, interrupt it, then change that to "randomize", it'll try to start your gen from the saved state

there's a layer of memoization built in it seems

very cool but can be annoying

in this case, it's trying to do the highres fix step again (when it should really be doing the base gen w/ new seed)

[supple raptor]

oh, I get what you mean

[stable crane]

your crows @native halo

[native halo]

we need a b.a.c ( big ass crow)

[steady grail]

there is a seed node

i have every node set to seed input

[stable crane]

😮

[stable crane]

what is it called?

[light moat]

Whoa, that's awesome.

[stable crane]

I don't see how it would fit into KSampler node

[steady grail]

rgthree's ComfyUi Nodes

[stable crane]

great work

[supple raptor]

ty

[steady grail]

spectral

very touhou

[gentle musk]

Crow lmao

[stable crane]

lol

I like how it morphed into her hat

[steady grail]

pure meme material. but the question is, internet needs more meme?

[steady grail]

this node give sharp images, but weird mutations

[steady grail]

using that node, the result is so different from the original, that looks like it ignoring some tokes, or loras

but the end result is much more polished

this gives me an idea

[nova remnant]

Where crows

[nova remnant]

crow menacingly stands near her

[gentle musk]

More crows

[steady grail]

give her frizzy hair and she'll look like hermione

[nova remnant]

@native halo more crows

[steady grail]

not enough

[nova remnant]

[steady grail]

i said MORE!

🤣

that meme never dies XD

[gentle musk]

lol what is this

[steady grail]

the clouds are trying to steal her bike

i can relate

[dawn jasper]

This is amazing.

[steady grail]

creepy

[dawn jasper]

Two ravens?

[gentle musk]

I may have sort of mixed a crow with a cow

[dawn jasper]

@gentle musk A Crorow?

[gentle musk]

lot of crows sitting on heads

[brazen rapids]

Am i not allowed here?

[bitter latch]

hey howdy

the crow ma haw

[steady grail]

looks like lcm don't works well, splitting the work on 2 samplers

and even when it works, the ending result laks in shaptness

[stable crane]

ty!

@supple raptor your prompt ❤️

[steady grail]

cute

reminds me of elden ring

[keen hollow]

evenin'

[stable crane]

https://github.com/light-and-ray/sd-webui-lcm-sampler for A1111, do you use this with the LCM lora?

[steady grail]

ohy

yep

with max 6 steps and max 2 cfg

[nova remnant]

squish

[bitter latch]

hmm think I used wrong model

sighs. soooo bored

[dawn jasper]

This is adorable.

[zinc flax]

Super cute

[native halo]

it works even with high-res fix but several details such as the crows and the pyramids dont appear on the img i guess thats the tradeoff less details but faster imgs

no pyramids or crows 😔

[bitter latch]

bored....

[steady grail]

kohya deep shrink, totally changes the end result..... that node needs deep studies

[supple raptor]

Seed N at resolution A is not equivalent to seed N at resolution B. It might effect the output for more reasons than just that, but that alone is going to have a huge impact on the output.

[steady grail]

yep

i think, it can be an important add, if studied well

looks like it can be used to add details to an image, with a relatively low time expenditure

[steady grail]

result with lcm looks very dry

i wil look for another way to use them tomorrow

see ya

[native halo]

u can change the way it looks by either reducing the weight of the lora or increasing it

[nova remnant]

@native halo she angy

be careful

[native halo]

https://tenor.com/view/rolling-dogeza-apology-sorry-funny-gif-8906456

[nova remnant]

Lmao

[native halo]

i swear i was lookin at her eyes

[gusty palm]

watch this

[nova remnant]

@native halo

Forgot to make her happy now trying again

@dull jackal @wooden coral

[wispy canopy]

freeu+sag is quite good 😮

[nova remnant]

what is sag

[wispy canopy]

self-attention-guidance

https://github.com/theroyallab/sd_webui_SAG

8 steps lcm lora with FreeU+SAG vs without

[bitter latch]

I think it got confused

[bitter latch]

had to tell it to have her wearing pants. oi

[native halo]

needs a latex bodysuit imo

[bitter latch]

ok...

or

[bitter latch]

she doesn't look amused

hmm no background.. odd

love this model

realcartoon

[brittle jay]

Anyone knows what this effect is called? (Not ai art, probably)

[nova remnant]

after-washing-machine-photo

[bitter latch]

negative photo ?

[nova remnant]

[brittle jay]

just so you know

I googled that

[brittle jay]

No, I mean like the texture thingy

[bitter latch]

she got old

umm.... haha

Random colors, miku, random age

best one yet

[vestal crypt]

https://media.discordapp.net/attachments/1027429244478955530/1176075591611600916/flat2DAnimerge_v30.png?ex=656d8cbf&is=655b17bf&hm=920cf6d549d2f261208635f99466edf07033aad20b65df3d86c8dc56007940d0&=&width=441&height=662

[bitter latch]

oooh

[brittle jay]

Nice

[bitter latch]

😄

[bitter latch]

goes to bed.. thud

[wispy canopy]

interrogate turned it into this 😮

[supple raptor]

You should try adding au ra, avatar \(ff14\) I'm curious if it'll get any closer to the right horns or add the scales.

[wispy canopy]

@supple raptor is this somewhat close, i don't think it is

(only minimal impact)

[wispy canopy]

no cherry picking:
3x single shot 1024x2048 , 10 steps lcm, dpm++ 2sa, freeu, kohya highres, SAG: 1 minute 20 seconds render time

[wooden coral]

how do you get such good outputs with LCM? Mine looks kinda terrible mostly

[wispy canopy]

(damn, after a gazillion retries, this prompt/composition is waaay harder thani thought it'd be, not using controlnet)

[wispy canopy]

freeu

https://github.com/ljleb/sd-webui-freeu
https://github.com/theroyallab/sd_webui_SAG

and sag (self attention guidance)

and in this case, also kohya highres https://github.com/wcde/sd-webui-kohya-hiresfix

oh.

[wooden coral]

Guess breakdomain isnt compatible with all that

[dull jackal]

WHO BECKONED ME!?

[wispy canopy]

which extensions did you use?

if you use LCM + freeu + kohya + SAG, you need to tweak the values a bit

[wooden coral]

I use the LCM Lora kohya SAG and FreeU as well

Maybe a problem with a lora I use or an embedding

Oh SAG throws me an error

[wispy canopy]

yeah, sag throws errors (but works)

@wooden coral if you use sag+kohya, you need to set the SAG guidance scale to 0.2

and that nets me this (single pass 1535x2048 10 steps)

[wooden coral]

Ill try that, thanks

Also youre using SDXL?

I mean I get images now, so there is that

[stable crane]

beauty

[lyric zealot]

this looks kinda off, like just something is wrong, no idea why tho

the body too long?

or something

[vestal crypt]

not sure

maybe the head is too small?

[lyric zealot]

maybe, or something, really no idea

and also, becareful with 🐫

[vestal crypt]

oh

I will try to avoid that

[lyric zealot]

[nova remnant]

Belly

[nova remnant]

2 meter tall girls also matter

[supple raptor]

No, not really. The horns should be in place of the ears and she should have patterns of scales on her neck and around the outside of her face.

[wispy canopy]

thought so. model simply doesn't know

have a dude

[supple raptor]

Yeah it doesn't surprise me. I was using a LoRA to get it to work.

[steady grail]

the distance between shoulders and chest is too long

[tribal ore]

That sometimes happens. When you change the aspect ratio, sometimes proportions are also adjusted. My guess is that some of the training images were resized incorrectly and stretched/squished. Sort of like how this Samus is too wide.

[steady grail]

wow, samus hit the Gim!

[nova remnant]

your model probably doesnt know samus well

[steady grail]

almost perfect, if not for the shadowy face

[wispy canopy]

mind the gap

[native halo]

i like those big trees

[wispy canopy]

and the shadows, sublime!

and check out all those tiny gaps, you can even see the sky!

[real mango]

Can you gen in any anime art style?

For instance the art style of a few visual novels?

[native halo]

yes

[real mango]

Can I make a few requests because my computer doesn't work?

Or is that a big no no?

[native halo]

something like this style is the one u talkin about right?

or this one

[real mango]

Artist name is Fue.

[steady grail]

wow

[native halo]

yea similar but not exactly the same cuz theres no lora for his style

[steady grail]

they are more chibi in face

reminds me so much harves moon

or made in abyss

[native halo]

yea this is his most recent vn

[steady grail]

warm

[real mango]

Got it.

[light moat]

@wraith raptor those are all wonderful

[native halo]

another pic s

[real mango]

What about Miura's artstyle?

[native halo]

the creator of Berserk?

[real mango]

Yes

[native halo]

yea that one is easier cause theres a lot of loras of his style

[stable crane]

there's a really good manga artist by that name

[real mango]

hm?

[native halo]

like this for example

[real mango]

very good. Okay. Here's a request.

white hair, long hair, white skin, red eyes, black kimono, black sandals, black gloves, full body, blank background, fue artstyle

[bitter latch]

ok 😄

hmm didn't work out all ok.. umm

[native halo]

didnt gave me gloves or long hair 😔

[real mango]

If that's not long hair I would hate to see what you would call long

[native halo]

heres another bad one

[bitter latch]

some call bad I call. ART

😄

[stable crane]

I'm still not really sold on LCM

[bitter latch]

LCM?

[stable crane]

sorry random statement, a tech

[bitter latch]

nods

sorta like spouting randomingl about AHT, TOC vs TOH

😄

[native halo]

this one came out better but still needs upscaling to get more details

[bitter latch]

cx srv phn stuffs

[real mango]

Wait

Add 'black pants' and 'black sleeves'

[bitter latch]

read that as block pants and block sleeves and my brain went wait.. minecraft?

doh!

my brain being random again

.. ok.. .STILL

[native halo]

black pants 😔

[bitter latch]

sorta unreal anime ?

[real mango]

Watmer

[bitter latch]

ok using wrong model let me try an anime one 😄

[native halo]

last ones 😔

[real mango]

Good!

Now try genning Minato Hikaru from Full Metal Daemon Muramasa

[nova remnant]

can agree, its better without

[bitter latch]

hmmm?

[bitter latch]

ok last one of those

[nova remnant]

Nice hand

👍

[dawn jasper]

The bird is protecting the human.

[steady grail]

using ai, we all are becoming hand fetishists lol

but truly good hands XD

[dawn jasper]

They are quite handy.

[wraith raptor]

Anime screencap model with some adult fine tuning.

Also, thank you

[steady grail]

wow, this lora is gorgeous

look at the coolors

[steady grail]

charming

[steady grail]

she is so sweet 🤣

or he?!?

[lyric zealot]

https://tenor.com/view/gokushufudou-gif-21120473

im trying to go for this vibe

[steady grail]

oh, that anime lol

[lyric zealot]

[steady grail]

something like, my hubby is a yakuza XD

[lyric zealot]

yup

something along that line

[steady grail]

yeah, that story was a killing, i love it🤣

in that scene where he was sharing his spices, the police was sure he was selling d*ugs 🤣

🤣

[lyric zealot]

i love the webtoon

[dawn jasper]

Adorable. 😮

[gusty palm]

if i have a 1024 pixel image and want to upscale it twice and end up at 2048

i could upscale once at 2x,

2 runs at 1.5x wont do it though would it

1.45x almost does it i think i just have to settle for some thing else

[gusty palm]

love that one

[wraith raptor]

It was a cursed prompt trying to make a rat woman XD

[wraith raptor]

If you're baseline is lower you can upscale higher :)

Depends on what aspect ratio you wanna go for ofc

[bitter latch]

love this model! nuelement

[wraith raptor]

Which? :)

[wraith raptor]

Merging time

[bitter latch]

https://civitai.com/models/110537/nu-element

[steady grail]

the colors are fantastic O.O

[gusty palm]

i likem

[gusty palm]

i have to leave the house later today :<

[gusty palm]

if that was a game or anime id watch it

id love a honkai star rail or phantasy star online anime

in that style

[bitter latch]

thanks

[steady grail]

damn, the song of demon slayer are seriously like a punch in the face

maybe i should gen something in that style

but first...

guys, do you know how to change the name of a reroute node in comfyui?

[bitter latch]

tries a merge :D

envydream nuElement and Azoyarpgtools

waits

watches the anime "Record of Grancrest War" while he waits

oooh ok, not sure if this is good.. no.. no.. i'm sure, it sucks... oi

https://tenor.com/fCeDP94EQAV.gif

oh... doh!

denoising str...sigh i'm a dummy.. or just tired

sigh

auto vae.. huh? sampler restart.. this isn't behaving like ther other models

[wraith raptor]

Is this too nsfw or still fine?

[brittle jay]

“In other words, if I had friends, I’d have to start worrying about them, right? If my friends were hurt, I’d feel hurt too, and if they felt sad, I’d feel sad too. You end up with more weak points, so to speak. I think that’s the same as becoming weaker as a person.”
“…But you have fun when your friends have fun, and you’re happy when your friends are happy, so it’s not all about becoming weaker, is it? You might gain more weak points, but you’d gain advantages, too.”
“No,” I replied, shaking my head, “I’d feel envious when my friends were having fun, and jealous when they were happy.”
“…How petty of you,” she nailed me.
Leave me alone.

[bitter latch]

shrugs, dont' ask me

[brittle jay]

Can always just ping mods

Or rather DM them

[bitter latch]

tho if you have to ask it prob isn't

[brittle jay]

Before even posting it here

[brittle jay]

Nah I once asked and they said it was alright

[wraith raptor]

@dull jackal
This still ok? Been a while since I was active here :)

Can remove if not.

[bitter latch]

well it's just the tops

the other word got blocked. odd

[wraith raptor]

What got blocked?

[bitter latch]

can't say cl**V*ge

[wraith raptor]

Words get filtered?

Hahah

[bitter latch]

yeah first I've seen that

[wraith raptor]

Well I'll hear it when sunny has time ^^

[bitter latch]

yeah think this merge is nucking futs

[wraith raptor]

Merging the ele model now :>

https://tenor.com/view/la-science-science-familly-guy-peter-griffn-kolakdm-gif-27448114

[bitter latch]

tho interesting style

yawns I should be in bed but meh dont' have to be in till later

[bitter latch]

back to nuElement

[gusty palm]

ah the spatial inconsistencies ><

i thought it was just the chair

didnt notice the tail

[bitter latch]

tries a new merge

oooh way better.

merged nuElement, AOM3, and hello25vintage

oh boy.. then again

[bitter latch]

yep I would say the merge is good 😄

[wraith raptor]

witch mixing potion and stuff with his magical friend owl, in a room field with books and potions, lamp on his sidee while working, glowing runes are floating on the book, elaborate scene style, glitter, orange, realistic style, 8k, exposure blend, medium shot, bokeh, (hdr:1.4), high contrast, (cinematic, orange and white film), (muted colors, dim colors, soothing tones:1.3), low saturation, (hyperdetailed:1.2),

Might wanna tweak weights xD

[bitter latch]

replaces last part after boken with his own hi-res style

waits

witch mixing potion and stuff with his magical friend owl, in a room field with books and potions, lamp on his sidee while working, glowing runes are floating on the book, elaborate scene style, glitter, orange, realistic style, 8k, exposure blend, medium shot, bokeh, fine art, masterpiece, epic detail, elegant, cinematic lighting, sharp detailed, focused, vivid colors, multi-tonal shading, great shadows, hi-res, HDR, 3.5D, Unreal render

falls into bed

witch mixing potion and stuff with his magical friend owl, in a room field with books and potions, lamp on his sidee while working, glowing runes are floating on the book, elaborate scene style, glitter, orange, realistic style, 8k, exposure blend, medium shot, bokeh, sharp focus, elegant detail, HDR, detailed shadows, cinamatic lighting, unreal render, realistic

ok last one

hhe trapped in a jar

cast sleep on himself

[whole portal]

does anyone have any idea which model or locon this is?

[dull jackal]

A ping, a ping

a ping for me

[wraith raptor]

That was me ^^

[dull jackal]

!! GASP!

[wraith raptor]

[dull jackal]

Where we runnin'?

[wraith raptor]

https://tenor.com/view/take-me-home-country-roads-john-denver-singing-performance-gif-16067984

[dull jackal]

https://tenor.com/wpxZ.gif

[wraith raptor]

Could you sfw check the image I posed a lil above the ping?

Been a while, not sure if it's in line with the channel policy :p

[dull jackal]

N e v e r! (Jk, I already did!) 👍 If there's an issue with something, we'll let you know!

Those are some beautiful eyes.

[wraith raptor]

Was just for reference for future posts :B

[dull jackal]

I getcha!

Heterochromia, my fave

[wraith raptor]

too bad of the tiled glitch

eyes <3

[dull jackal]

The gold/orange eyes are really coming out well these days.

Gotta love the hair, too

[gusty palm]

shes a painter

orange eyes by coincodince

[native halo]

spoopy 👽

we need more rats in here

[gusty palm]

yes

rats gotta be my favorite white meat

[native halo]

crows too

[gusty palm]

crows are smarter than some humans

i saw a video of a crow

putting rocks into a bottle of water to raise the water level so it could drink

[native halo]

yea they know how to use tools

[gusty palm]

i dont even know how to use tools

i wouldnt be surprised if crows were on the internet

[native halo]

maybe we are all crows here and u the only human

my hand hurts

[gusty palm]

i wish i could get another one of those "animal body human head" gens with a crow

or knew how to do it consistently

[native halo]

yea i sometimes get them when i prompt for a girl hugging a dog

dalle is better at doing them

[gusty palm]

whatd u do to ur hand

[native halo]

she only has 3 fingers

[gusty palm]

i mean to hurt it

[native halo]

i was talkin about girl in my gen 😔

[gusty palm]

ohw'

lol it reminded me of those warnings on like, heavy mac hinery

🚫hand warnings that show the hand being pulled into the geartrain

https://tenor.com/view/spongebob-broken-hand-pain-broken-fingers-my-hands-hurt-gif-27197901

https://tenor.com/view/biscuit-oliva-oliva-shobun-ron-baki-finger-gif-25818182

https://tenor.com/view/arthritis-finger-broken-fingers-finger-pain-gif-13027392

ive been learning controlnet more

[native halo]

https://tenor.com/view/joseph-joestar-gif-18910941

[gusty palm]

i wanna do the whole set of the "<adjective> girlfriend" memes template

[native halo]

i tried with piper perry memes but it gets confused when theres too many ppl

[gusty palm]

it didn't get the kitty stockings unfortunately

[native halo]

try with ipadapter and softline

[gusty palm]

https://github.com/lllyasviel/ControlNet-v1-1-nightly/raw/main/github_docs/imgs/openpose_2.png

i wanna learn ipadapter

i downloaded a huge like 2gb model when i switched to it but i dont really know how to use it adequately

does this image suggest that you can extract pose data from an image?

[native halo]

idk i have never used that ui,im using the webui extension but yours looks different

[gusty palm]

oh thats a screenncap from the controlnet wiki

https://github.com/lllyasviel/ControlNet-v1-1-nightly
its in the openpose section

i was reading through it for tips and it gives examples and stuff for each CN model

[native halo]

then yes u can extract poses from images if u select openpose preprocessor

[gusty palm]

woah

[gusty palm]

they are best friends

[native halo]

i love kisses

[gusty palm]

this seems excessive, but ive been getting great results all day

[wraith raptor]

[vestal crypt]

https://media.discordapp.net/attachments/1027429244478955530/1176527805870247936/00169-3823086525.png?ex=656f31e7&is=655cbce7&hm=5111ae942a19d0efd4a53bd2b7efbfe8d4539557a945af281db495605e6c79ab&=&width=496&height=662

[dawn jasper]

That is one big bird. 😮

[real mango]

Can you gen Minato Hikaru from Full Metal Daemon Muramasa?

(I know I asked this before)

[wraith raptor]

Who are you talking to?

[real mango]

you. Although the last time I asked this question, I addressed it to anyone interested in this channel.

[wraith raptor]

I can try I guess, got a lora?

[real mango]

no

[wraith raptor]

Remind me later, I can't refresh loras so I'd have to reboot sd :(

[dawn jasper]

This is a beautiful image.

[gusty palm]

the clothes look real

[bitter latch]

close?

[bitter latch]

yeah thats what I used to try to make her without a lora 😄

I think this new merge I made did pretty good

[bitter latch]

think I found my new fav model.. a merge of course 😄

Mum-Ra the ever living!

[magic wave]

Looks pretty dope

I just post Mini Milk

[nova remnant]

MINI MILK

hewo smile

[magic wave]

Yis

Hewwo cat

[wispy canopy]

hmmm, lcm sampler might be listening a lot better to the prompt

anyone did some major testing on that one yet?

[bitter latch]

ok was going for a mummy but got one punch man? what?

ah ha, turned steps down from 40 to 10 and cfg upto 9

[wispy canopy]

lcm?

[bitter latch]

restart

[wispy canopy]

reastart crazy good at low steps

[magic wave]

Restart:

https://tenor.com/view/gigachad-chad-gif-20773266

[bitter latch]

eular

[wispy canopy]

@bitter latch lcm might be the new kid in town tho 😮

i mean, check out the 12 step seed 202 lcm image

that's a lot of good faces 😮

[bitter latch]

ddim 20 , cfg 12

[wispy canopy]

-20 steps 😮

you went back in time?!

[bitter latch]

I dn't see lcm in auto1111

I see lsm

my bad, lms

[wispy canopy]

need to install the animediff extension

they bundled in the lcm sampler

[nova remnant]

@magic wave

[bitter latch]

ahh

[bitter latch]

yep getting that now with the lcm lora

hmm says Set CFG to ~1.5 and Steps to 3

really?

this i gotta see

[wispy canopy]

not sure if that's true

[bitter latch]

hmm wonder how it'll work with my model

[wispy canopy]

for my model, 2/12 seems to be overal best quality tho

3 steps won't do it XD

[bitter latch]

steps 5 , lfg 2 . hmm

blah

steps 10, cfg 5

doubles them

hmm?

[magic wave]

Mini Milk sis?

[bitter latch]

oh I didn't addt he lora. DOH

let me try that again

[wispy canopy]

hmm, that doesn't seem very convincing

unless you specifically prompted that

[bitter latch]

actually thats prompting

seems black aura is to dark 😄

[wispy canopy]

just testing that lora sampler. needed high quality science.

[bitter latch]

without lora

hmm then again it could be the polocromatic color

polycromatic.. oi can't spell

[wispy canopy]

@bitter latch i feel the LCM sampler is a lot better at following your prompt, like ddim used to be (until other sampler just got better quality output)

even without the lora, that's pretty good 😮

[bitter latch]

steps 10 cfg 5

hmm need to try without a merge let's see hmmm

[wispy canopy]

21 seconds 😮

and i mean, check those hands 😮

even proper holding stuff?!

what is this black magic?!

[bitter latch]

NuElement model

19 secs.. wow

tries 1024x1024

[wispy canopy]

kohya highres

[bitter latch]

on second thought better not

[wispy canopy]

@bitter latch throw me the prompt 😄

i wanna see what it does

[bitter latch]

sent

[wispy canopy]

i'm sold. lcm lora with lcm sampler -> the new king in town

[wispy canopy]

kohya deepshrink + freeu + lcm combo -> 1 pass highres cheatcode 😮

[steady grail]

wow, it's very polished

i don't get such polished resuults

[native halo]

now do it in Jojo DIU style

[sage mountain]

now draw her paying taxes

[bitter latch]

and back

[vestal crypt]

https://media.discordapp.net/attachments/1027429244478955530/1176727013940273212/00035-2299420437.png?ex=656feb6e&is=655d766e&hm=96a6db0630c991ce5fd45432a93be22e6985a16d10af16f1fa92fa9435738459&=&format=webp&width=530&height=662

[keen hollow]

evenin'

Today's Tifa

and one bonus Tifa

[bitter latch]

hmmm lcm gives it a soft effect with this model i made

cackles.. steps 100 cfg 30

gah

[keen hollow]

those are quite soft

[bitter latch]

umm

hmmm?

oooh

hmm LCM is just... shrugs

tho restart is good

withthe lora for weights.. hmm

lora loaded first

interesting

[wispy canopy]

gm everyone

[keen hollow]

more like gn 😛

[nova remnant]

moo

[vestal crypt]

https://media.discordapp.net/attachments/1041682785418629125/1176761612883996682/00041-3133681236.png?ex=65700ba7&is=655d96a7&hm=550a2c9bafc4e4324d67970d95cac85256bc6f8f642da5af447c1f3016161b78&=&format=webp&width=530&height=662

[wispy canopy]

seriously man, what the fuck did they do with LCM sampler, i only get good hands

😮 magic... where is all the effort i have to put in to get even a decent image

[nova remnant]

@wispy canopy

drink it

[wispy canopy]

she looks awfully happy with her science, what did she make?!

[wispy canopy]

sad about right ct, i love left cat, i want a cat that looks at me like that ❤️

dunno if anime or not, but this feels like some anime shows i've seen

[slate narwhal]

This reminds me of that Bruce Lee quote about becoming water, she kinda became the cat in that one

[nova remnant]

a way to skip lab trials

I just did merc and it had a golden door

[wispy canopy]

ah, the most lovely activy every league start

[nova remnant]

@wispy canopy It won't hurt I swear

[nova remnant]

thighs

[wispy canopy]

she just has to show of her goods

[wispy canopy]

u sure it won't hurt?

[bitter latch]

the hands. omg

[nova remnant]

THE hand

and THE tail

[bitter latch]

well the tails accurate

[nova remnant]

@wispy canopy

[bitter latch]

switched to lcm

[nova remnant]

the experiment failed

[bitter latch]

Miku leopard ?

[wispy canopy]

mecha musume?

[wispy canopy]

scavenging for good prompts

damn, lora even makes darkalfa shine again

[steady grail]

charming and warm

what lora is this? the results are very poolished

[wispy canopy]

no lora

my default model, bright protonuke

[steady grail]

out of my reach lol

[wispy canopy]

and the other one is darkalfa

[steady grail]

i can see from the style

[wispy canopy]

@steady grail using the lcm sampler and lcm lora is a MASSIVE increase in quality and speed tho

[steady grail]

the gens are faster, but the results looks rough, to me

[wispy canopy]

combine with kohya deepshrink and freeu

[steady grail]

maybe in sdxl is different?

[wispy canopy]

even on sd1.5

[steady grail]

yeah, i'm testyng a workflow with freeu now

[wispy canopy]

lol, trying on base sd1.4 XD

[steady grail]

later maybe i try to start just with a ksampler, pass the result on a fu and deepshrink and than end the last steps with just another simple ksampler. i want to see if i can improve the quality, reducing the time gen

[wispy canopy]

😮 default base sd1.4 😮

[steady grail]

for now, i just test a basic working workflow

[wispy canopy]

holy shit, the comparison with euler a, when sd1.4 was just out

[steady grail]

eheheh

[wispy canopy]

oh right, cfg is too low

needs to be 7

oof... even at cfg 7, that's rough

1.4 anime, in it's full glory!

[steady grail]

sd improved so much

[wispy canopy]

switching over to lcm, already huge gains

switching to darkalfa 😮

adding in kohya highres

[steady grail]

warm 😌

[wispy canopy]

@steady grail kohya deepshrink. single pass sd1.5 at 2048x2048

[steady grail]

i can see the improvement

[wispy canopy]

Negative prompt: bad, wrong, boring, badly cropped
Steps: 20, Sampler: LCM, CFG scale: 2, Seed: 3765791862, Size: 2048x2048, Model hash: d9538e0883, Model: darkalfav11 fp16 pruned-correctedvae, VAE hash: d31a5bb481, VAE: kl-f8-anime2.ckpt, FreeU Stages: "[{\"backbone_factor\": 1.2, \"skip_factor\": 0.9}, {\"backbone_factor\": 1.4, \"skip_factor\": 0.2}]", FreeU Schedule: "0.0, 1.0, 0.0", FreeU Version: 2, Lora hashes: "LCM_LoRA_Weights_SD15: aaebf6360f7d", SGM noise multiplier: True, Version: v1.6.0-356-gfc83af44```

[wooden coral]

ngl krita is crazy

[wispy canopy]

boto, setup LCM asap

[wooden coral]

This is with LCM

[wispy canopy]

ah

gigachad

lcm lora AND sampler?

[wooden coral]

Yep you can live generate or normal generate with the plugin

Its sick

Also all controlnets

[wispy canopy]

wow

so near instant image updating?

[wooden coral]

Yea while drawing

https://github.com/Acly/krita-ai-diffusion

[steady grail]

wow

that's an improvement

[wooden coral]

It connects with your comfy install

[wispy canopy]

damn 😮

it's like that very early krita plugin, just

better

by.. a lot

[wooden coral]

Getting an iPad pro later today, then I can draw and dont have to use my mouse

[wispy canopy]

just doodle with the finger and SD will sort it out? ^^

[wooden coral]

I mean thats what im doing rn

Also you can make presets and with loras, models, vaes etc

[wispy canopy]

same prompt with sdxl -> 3x faster lol

sdxl faster than sd1.5, you heard it here first

[wooden coral]

Made a simple anime preset for example

[wispy canopy]

boto, lcm sampler better, full stop

not just better at low steps, also at higher steps

[wooden coral]

lcm isnt good for details imo

Also SDXL doesnt seem to work properly yet in krita idk what im doing wrong

[wispy canopy]

my fairy disagrees with you

and my critter does too

[wooden coral]

not anime

[wispy canopy]

well, if you want science

[wooden coral]

Whats your LCM settings?

And which model

[wispy canopy]

this is all sdxl, brightprotonuke 1.2

[wooden coral]

I also need CFG steps etc if you can

[wispy canopy]

lcm weigth 0.8,
lcm sampler,
kohya deepshrink,
freeu
cfg 2
12-20 steps

[wooden coral]

Problem is Krita doesnt let me set the LCM weight I think

[wispy canopy]

science

[steady grail]

the setting for lcm is not with steps between 3 and 6?

[wispy canopy]

yes it is. but it doesn't mean you can also do more

[wooden coral]

Depends really I think

[wispy canopy]

6 gets you good images

12 gets you near perfect images

and between 12-20 it doesn't get better

[wooden coral]

got XL somewhat working but the quality tanks unfortunately

Hilarious pic tho