🍥｜anime

[modest fog]

The what

[green temple]

the "🌚 "

[vital raptor]

it's not a valid tag on danbooru or e621, so it's probably just nothing

[bitter birch]

bruh

ok, understandable

[modest fog]

I can't translate emoji to real words I'm stupid

[vital raptor]

just because people put things in their prompts doesn't mean it's actually functional or relevant

[modest fog]

https://tenor.com/view/sanic-sonic-fast-run-walk-gif-17514382

[bitter birch]

I do learn aout it tho, sadge

we must learn the model

[green temple]

nooo

[modest fog]

This one came out better

lmao

She's coming for you

@green temple

[bitter birch]

💀

[modest fog]

She wants a piece of Sooma

[bitter birch]

why, is Sooma a cake?

[modest fog]

Yes

[bitter birch]

💀

[bitter birch]

vore 'em

[modest fog]

dinosaurs are back

[past junco]

But yes, if you have Tabletop Simulator, I want you!

[past junco]

Literally Akuma No Riddle

[gusty palm]

is the reason the upscaled is so different because noise is set to o high

[modest fog]

I don't know Japanese

[bitter birch]

I usually set denoise strength 0,5

[modest fog]

The higher, the more changes

[gusty palm]

okk

[modest fog]

Think of it as % of the image that will change

[bitter birch]

Bro, the AI wants to burn my laptop 💀

[past junco]

It's an anime, really good

[modest fog]

I don't like it

[gusty palm]

ok ya .45 was good

[vital raptor]

if you only want upscaling and don't need the changes, 0.2-0.25 is generally better, minimal changes but good upscaling quality

[gusty palm]

im prompting for mech armor tuxedo but this is good too

[bitter birch]

chibi

bruh, Chibi but with depth effect bg

[past junco]

You've seen it?

[modest fog]

No

I meant as in I don't really like animes or watching stuff in general

[past junco]

Sorry, just had to double check I was in the right channel

[green temple]

there is smol in it !

[past junco]

Death by thighs, worse ways to go

[past junco]

Akuma no Riddle is high up on my long list of watched anime and manga. Could not recommend it enough.

But, if you want smol (and similar vibe) I would also recommend Magical Girl Raising Project.

Both shows are dark and bloody but Magical Raising Project is basically at the top of my highest rated anime list

[green temple]

i`ve watched enough of smol

best smols are miri from buddy daddies & fran from Tensei shitara ken deshita

[past junco]

I swear, not like Magical Raising Project

Not a single anime even compares

[green temple]

ohh

[past junco]

No joke, I want to watch it again right now

I've seen a lot of anime

Beware to anyone who watches it though, you won't be ready for it

[gusty palm]

thats awesome

[teal pivot]

thanks :3

[bitter birch]

[gusty palm]

i love color sxplosion type things

ive been putting (dmt art style) and swqirling pattern in my prompt but not really nailing it

[teal pivot]

which model do you using

[gusty palm]

i kinda jump around

these are about my favorites

[teal pivot]

https://civitai.com/models/926443/ntr-mix-or-illustrious-xl-or-noob-xl?modelVersionId=1051221

try this one

[gusty palm]

ok ill give them a go , im not sure if i can run xl

my tummy hurt and i eepy

[teal pivot]

oh okie

[desert island]

Scissors are not coming out! -UGH!!!- I think I need to make a lora specifically for scissors 💢

[modest fog]

Not blue

[green temple]

bloo shading

[modest fog]

Not enough

https://tenor.com/view/rrane-battal-rrane-battal-kızgın-battal-kızgın-gif-25355985

@green temple

Smol attack

[green temple]

[modest fog]

we dead

[desert island]

(fear, horror, sadness, anger, depression)

[bitter latch]

she's like... NOPE

[past junco]

Lot's of spoilers here

[past junco]

I like how there's the sun... And the spooky

[bitter latch]

it's called shadows

😄

[past junco]

Not with that ambient lighting

Hmm

I suppose if you really stretch it

Might be possible

But I like the idea of reality being bent and distorted

[bitter latch]

"wait for me!"

[bitter latch]

hmm face is indistinct

[modest fog]

||Jhon dies in the next Superman movie||

[bitter latch]

shocked disapproving look

.. umm..

yep gone random

ponders different topics.. hmmm

she got'em

[desert island]

FINALLY! I've been trying to catch that lil' bastard for awhile now!

[bitter latch]

haha

[green temple]

food

[modest fog]

Me when eating grass too

[modest fog]

Food 🍴

[modest fog]

|| 🍴 || || 🍴 ||

[supple raptor]

happy tuesday @finite herald

[modest fog]

@green temple open the door

[green temple]

big-smol

[modest fog]

It's the same prompt I'm always doing

[green temple]

how ??

[modest fog]

[green temple]

nooooooo , don`t leave

[modest fog]

🔪

[green temple]

keep your true form away from me

[desert island]

Ughhh! Damn animation artifacts! (-sigh-) ... Back to the drawing board... 😭

[desert island]

This is better but the resolution and quality still kind of sucks. On the positive side I can do longer and smoother animations now.
192 frames took 7 minutes and 30 seconds to generate on my 4080.

[vital raptor]

nice, I actually like the first one better except for the random cape flying through, the flashes and streaks make it look like lightning or a battle or something

[desert island]

I'm just trying to work out a consistent pipeline and the artifacts weren't in the keyframes. 😄

[finite herald]

hi guys, can we please relax on imgs that lean too far in the loli direction, esp when showing a lot of skin- tyvm !

[vital raptor]

sorry

[green temple]

i hope my images are safe

[keen hollow]

what video workflow are you trying?

I messed around with mochi the other day but it was pretty taxing on my system.

[desert island]

right now I'm using svd and tooncrafter it's the same workflow I was using awhile ago but I wanted to work on it again.

[desert island]

Here's the workflow I put together. If you want the links to the list of models I'm using let me know. I'm sure I can find them all again.

I'm still working on it so it isn't totally optimized but it works... (-shrug-)

I think I want to try the svd interpolation in place of the tooncrafter nonsense as my next test

[desert island]

Well I learned something

It took 20 minutes and the quality is there resolution wise but the face morphing is terrible 🤮 also it lost the interpretive motion that was introduced with tooncrafter 🤔 I'm gonna have to think about this later to see if I can come up with a solution.

[modest fog]

eats her hand

[vital raptor]

new server toys

[vital raptor]

of course the power cables wont be here for 2 more days

[stable crane]

😢 😢

[keen hollow]

I was thinking most moved on from svd. mochi and that other one I can't think of 😛

[bitter birch]

Cirno

[bitter birch]

what's the prompt for that hand action?

[vital raptor]

heart hands

[green temple]

@vital raptor purple

[bitter latch]

faints

AH HA!

well.. almost

those hands.. eek

[desert island]

I got mochi to work but I don't have the horsepower to do anything meaningful with it. So I decided to go back to something I can work with 😄

[bitter latch]

umm what?

[vital raptor]

lol

[bitter latch]

"Milkky"?

ok going back to bed

[modest fog]

Purpol

@green temple

This is like funya when she's angry at me

[green temple]

aqua

[green temple]

miku

[green temple]

[bitter latch]

buzz buzz

[green temple]

@zinc flax

[wooden sail]

Today's experiments with artist tags

Model: IllustriousXPonyMix

[modest fog]

@green temple (blue)

[bitter latch]

hmm for some reason this just makes it look like she's real and staring into my soul...

[marsh grotto]

how fucking long does it take to refresh credits jesus christ

[modest fog]

@finite herald

🔨

does the ⚠️ not work anymore

[keen hollow]

I'm at the point where, with some of these newer tools and models, I might just buy a service if I can find a reasonably priced one.
My 3060 has 12GB of VRAM which lets me do a lot locally, but I feel like I could rent a much faster environment through these various services. (Not to mention, not have to worry about maintaining the tools.)

[merry tendon]

char for my game

[keen hollow]

what game?

[merry tendon]

my game

lol

[keen hollow]

😬

[merry tendon]

its still in dev

playing around with graphics and sound rn

[keen hollow]

that char needs some animation 😂

[merry tendon]

yes its very early into dev

its also miku LOL

just a placeholder

mostly everything you see is placeholders lol

[keen hollow]

I think tools to generate simple animations and such for "visual novel" games will be around before too much longer.

consistent characters still requires too many hokey workarounds

[modest fog]

looks like some monstruosities I made for vrchat

(They worked in the end)

[desert island]

I'm not a fan of having to utilize outside resources as I think it takes the power out of the hands open source community. At home we don't have to adhere to regulations or standards and can advance the space without red tape or oversight. I think we need power in the hands of the people for this technology so that corporations can't form powerful monopolies on this tech that is destined to change the fundamental landscape of our planet and society... of course all I'm doing is making images and videos but, you get what I mean. 😆

[keen hollow]

I do get what you mean. I do think folks providing hardware services to utilize the open source projects is somewhat of a compromise.

I'd love to run Stable Diffusion on an RTX 5090 some day, but that card will be $2500 😉

[desert island]

2500 before people buy and upsell them for twice that.

[keen hollow]

oh yes I don't pay scalpers just out of principle

ever

i'll just go without

[desert island]

I'm getting to nearly usable results with my img to vid pipeline still needs a little more tweaking but I think I'm on the right track.

[keen hollow]

hey that reminds me of Tifa 😄

Do a Tifa animation... battling something 😄

[desert island]

battling is gonna be tough since it's going to require more stuff. I have a thought as to how I can pull it off but I need to get this section to work first before I start experimenting with including masking and tagging other characters and assigning loras to those masks. Right now it's just image to video with a single character.

[keen hollow]

well how about just sitting?

there's just a supreme lack of Tifa animations in here

I don't even know how to do SVD. 😐 Just the Mochi one.

and I don't have 30 minutes to render a 4 second low-res video right now

[desert island]

gimme a 16:9 image of Tifa sitting and I'll send it through to see what we get

[keen hollow]

You gotta generate one! 😄

I mean, I have hundreds...

[desert island]

oh fine 😄

[keen hollow]

not sure how many 16:9 images I have though....

[desert island]

This is nice! But I'm in the process of trying to make some 😄

[keen hollow]

I do ultrawides because my monitor is 3840x1600

[desert island]

ohhhh I can just center and use this 👍

[keen hollow]

kew

[desert island]

I should generate some tifa images though just so I can work on my generation skills for specific characters.... (I'm not very good at it, even when using loras) <-- I don't know why.

first run just parallaxed the background gotta find one that has her move a little... 😦 🎲 Rolling again.

[keen hollow]

"Tifa Lockhart, white top, black skirt, suspenders" is usually all it takes.

[desert island]

well damn! Even when I increase the bucket id it just increases the speed of the parallax effect. 🤦‍♂️ I'm gonna gen a more realistic starting point and hopefully we can trick SVD into animating her like a real person! 💢

[grave snow]

https://tenor.com/view/taroaxis-monday-left-me-broken-waiting-for-love-speech-bubble-gif-1267844232160950050

[modest fog]

[desert island]

starting image and generated key frames... I'm gonna roll one more to see if I can get better keyframes. 🎲 Regardless I'll push it through to the next stage and see what we get for the low res full frame animation. 🤞

actually here's the second set of key frames too.. hmmm I think I will do the second one.

Here's the low resolution frame interpolation... 😏 kinda wish she would blink but I don't want to keep rolling....

This was the final output so much fidelity lost 😢

[desert island]

but It's 8 seconds long and went from image to output in < 20 minutes. 🤷‍♂️

[bitter latch]

huh.. that came out not what I was expecting...

[keen hollow]

kinda has a ps2 vibe

[desert island]

I don't think CG cut scenes were that good back in the ps2 days... I'm gonna try a different one and see if I can get better results. ps2 vibes sounds like a knock on the video. lol 🤣

[keen hollow]

yeah I guess you're right... ps3 maybe... meant she has the video-game character feel for sure

ps4? I dont know... some ps

[desert island]

somewhere near Final Fantasy: The Spirits Within is what I'm shooting for.

[bitter latch]

not sure what it says but there ya go

[desert island]

what are you using?

[desert island]

Ohhh some naruto love!

[keen hollow]

google image search for a gif from the game 😉 😉 😉 😄

[desert island]

Henge!!!

[desert island]

animation and character design looks more like boruto now. 😄

[bitter latch]

never watched that one

well.. ok I tried but.. shrugs

[desert island]

...pst... it's okay I didn't either. Although I did read some of the manga for a bit... I should pick it back up though 😁

[bitter latch]

😄

hmmm.. I would watch it

oooh

hmmm ponders a story for these two....

Neo Tokyo cyber ninja village run-a-way beast kids

[desert island]

perfect! 😁

[bitter latch]

Miku-Neku

cackles

better

[desert island]

Really jittery... okay I'm gonna stop making videos for a bit. Gotta make some food and spend time with the wife. We're re-watching Arcane since she needs a refresher. Freaking love that show.

[bitter latch]

lowered to 10 from 15, and cfg is 3.... same seed as last

/me sighs, that time again.

hmm Narbatro?

[desert island]

I love these, so fun!

[bitter latch]

that would b ea fun anime

anyhooo gotta go

[gusty palm]

the mobile sd-ux-ui doesnt work very good

[desert island]

for which front-end?

[gusty palm]

https://github.com/anapnoe/stable-diffusion-webui-ux

this one its not that bad but the whole left side options menu didnt like being interacted with

[desert island]

cool thanks. Opening in a tab to look through later. 👍

[gusty palm]

whenever i tried to change any type of drop down setting it would close the panel , only on mobile over wifi . but the desktop ui is good and the github does say the mobile ui is in development

[desert island]

When I get some time I'll pull it down and look and see if there's a quick fix.

[gusty palm]

i was using brave over wifi if that matters i didnt know anyone here worked on that stuff

[desert island]

I just try to help when and where I can. Not promising anything I first have to see if I can replicate the problem and then familiarize myself with the code. If it looks easy enough I'll work on it, no promises 😄

[modest fog]

@green temple bloo smol spotted

[supple raptor]

ffx had some very nice cutscenes and that was ps2 https://youtu.be/oTloRDyYYtA?t=109

[desert island]

I forgot that ffx was still on ps2 😁 It's been a long time

[bitter latch]

o.O

[steady grail]

hello guys

@dull jackal @bold ether

[vital raptor]

i guess the reporting bot doesn't work anymore

[bold ether]

https://tenor.com/view/sailor-moon-cleaning-anime-jupiter-sailor-jupiter-gif-15159333

[modest fog]

Do you community guides have the power to bonk

[modest fog]

https://tenor.com/view/kok-cat-gif-25438805

[bold ether]

Do you wanna try it out 😄 ?
More seriously I'm not gonna answer what power we do or don't have. It could be counterproductive to talk about when there is or isn't someone with ban power active on this server.
Regardless it's ok to ping community guides or fruit whenever there are problematic activities.

[green temple]

🌚

[steady grail]

oh right fruit can clean too

[gloomy scroll]

[modest fog]

👍 Just wanted to know cause I don't see the mods anymore in this server, chaos

[modest fog]

why

no happy smol

[green temple]

i know nothing other than smol

[green temple]

big-smol

[modest fog]

It's smol

[green temple]

[green temple]

trying 🥲

[dull jackal]

A ping?

[modest fog]

free nitros were in chat earlier, you missed it

[rough oxide]

character wildcard for noobai if anyone needs one, every char down to 79 tags. credits to laxhar for the initial csv, then wrote some python scripts to format it as a wildcard xD

also gud to just ctrl+f in if you need a char + all their description tags

[green temple]

ty

[modest fog]

so many characters

none are smol

[green temple]

there is smol

[modest fog]

no, only bloo smol matters

[rough oxide]

smol ame works

last i checked

[green temple]

there`s bloo smol ( i only know kafuu chino)

[rough oxide]

and smol ame is on the list

[modest fog]

[rough oxide]

and there's plenty of smol chars

hachiroku is small too

[modest fog]

food

[modest fog]

https://tenor.com/view/rrane-battal-rrane-battal-kızgın-battal-kızgın-gif-25355985

chars are ugly

[rough oxide]

of speaking of smol ame

might have to fix the line where it says blue eyes

[modest fog]

except miku (please spare me fruity)

[rough oxide]

xD

there's a lot of small things to be fixed in this list but i think i solved all the major issues the original csv had

[zinc flax]

Rare maitetsu mention

[rough oxide]

i ownz themz

they also have like

the best scene creators of any vn

i wish more vns added scene creation tools to screenshot

xD

although something silly in maitetsu's is that

u can add audio/voiced lines to the scene

but u can't capture it as video

...so wut's the point

pokerface.png

[zinc flax]

They missed the most important part lol

[rough oxide]

yeh. still u can play around with any scene the game has + make ur own scene similar to summer pockets

there's not much to be desired

[zinc flax]

And the other game from the same dev too

[rough oxide]

but since they added a way to add literally any voiced line to the scene too

they should've let u capture it as mp4 or something

xD

the vns themselves are actually

presentation wise 10/10

i wish more vns were in that style

[zinc flax]

I haven't played the vn myself though

[rough oxide]

it's essentially the step before steins;gate elite

but after muv-luv

in presentation

[zinc flax]

I did start it way back then, never continued it lol

[rough oxide]

nearly every single scene in both games is a cg

o

then u kno about the cg probably

xD

like how the chars r actually part of a cg

and they are slightly animated within the cg

there's still typical VN scenes but most scenes r full cg

[zinc flax]

It's a high budget vn for sure

Lagged my pc a bit

[rough oxide]

ahaha

if u got like the ultimate editions

[zinc flax]

Especially the intro part with the rain

[rough oxide]

i was surprised that there's hololive covers in here

sakura miko & azki sung some of the songs in the cover thing

xD

[zinc flax]

I didn't expect maitetsu to be a big deal

[rough oxide]

i'll try anything within like the top 200 vns

if the pics/story sounds interesting

i wanna try rance X too but waiting for official tl

cuz it's ranked like #4

[zinc flax]

I'd like to try maitetsu again but I'm too small brained for the train lingo

They love their infodumps

[rough oxide]

oph

it hink

hm

idk if i recommend jast ver

but jast has a version now.

[zinc flax]

Try dohna dohna. It's fun lol

[rough oxide]

even given the context of what maitetsu is

i don't like any censorship in muh vns xD

so i recommend dlsite + en tl somewhere

[zinc flax]

I'll get to it when I have the time

That 40hr on vndb is scary

[rough oxide]

i think i almost picked up dohna

o maitetsu?

uhh

maitetsu is like 70+ hr isn't it

92 hr

[zinc flax]

Is that with the last run

[rough oxide]

yes

last run is the only way u want to play it

it's improved in every way

it's like reflection blue to summer pockets

it contains original game

• all additional scenarios

also do not buy last run/monobeno on steam

as a super wanring

jast usa is one step better but still slightly censored

[zinc flax]

Wait holy crap last run is basically a whole new vn

[rough oxide]

(i think you know why given the nature of it lol)

and then dlsite is pure uncensored original ver

[zinc flax]

The maitetsu rabbit hole goes deep

[rough oxide]

yeh. haciroku is cute. but also for the wrong reasons

xD

train lingo is pretty over my head a lot of the itme in maitetsu too

i liked monobeno as a story more

to follow along with

maitetsu had better visuals tho

by far

[zinc flax]

Oh for sure

[rough oxide]

muh jast collection

nukitashi

is surprisingly

really good

but

it's very lewd? but in a comedic way?

like every scene ever there's nudity?

xD

[zinc flax]

Yes lol

Oh I played that first game on the left

[rough oxide]

that's the only one

i haven't rly gotten into yet

i liked majikoi a lot tho

u need

a lot of patches for majikoi

one to reenable all voices

one to uncensor it

well

the jast is the uncensor one

[zinc flax]

Dang

[rough oxide]

i feel like it needed a 3rd patch too for sum reason

how did this char become so popular to place #848

i will nevr understand

xD

[keen hollow]

I don't know those games. Are they good?

are those the visual novel games? Never really got into those.

[rough oxide]

monobeno & maitetsu? yea they're vns. on the higher-end of vns in terms of presentation so ur not just looking at character sprites over a background every scene but not quite at the level of school days/steins;gate elite

[keen hollow]

evenin'

[supple raptor]

YuukaWave

[keen hollow]

howdy - how was your day?

[supple raptor]

pretty good, and you?

[keen hollow]

not too bad

same ol'

[supple raptor]

good to hear

[terse ore]

Is A1111 still good or is Vlad or something better? I've been in comfyui for a long time but I want to try segment anything+ grounding Dino like i used to before

[supple raptor]

• SD.Next (the one by vladmandic) has always been a meme. It should be avoided.
• A1111 mostly works reliably, but is not memory efficient. It's fine for SD1.x, but not good for SDXL.
• ComfyUI and SwarmUI (which is a more userfriendly frontend for ComfyUI) are much better. Best all around options.
• Some people use Forge, but it has some obfuscated code which is pretty suspicious. Could be unsafe, or dev is just trying to hide stolen code.
• Reforge is pretty decent choice, it's a fork of Forge from before lllyasviel decided to obfuscate code and is maintained by a person who is somewhat active in the Touhou server as well as the NoobAI server. Seems like probably the second best choice, but likely has incompatibility issues with some extensions.
• InvokeAI is well polished and has a much better feature set than it used to, but the documentation is very poor. It's inpainting/outpainting UI is very good, but support for any new features are very slow to be implemented if they ever are and custom nodes/extensions are virtually non-existent for it.

[modest fog]

Sd next is fine :<

Just ugly

[supple raptor]

sd.next is garbage. somehow he made his a1111 fork even less reliable than a1111 was back when it was getting 250 commits pushed directly to main every day

the UI (minus the color scheme) is the only decent thing about sd.next

[bitter latch]

shes coming for them

[dull jackal]

oh, but, why wouldn't we want that? lol

[modest fog]

They were scam links

[upper flint]

How to create

An ai amages

[modest fog]

Do you have a PC good enough to run it?

Online you can use Civitai for example

[past junco]

Got a new checkpoint merge I'm testing out, super cute

[steel zephyr]

returning to SD after years. damn, this became so good

[modest fog]

hello

many new things to play with in the last year of ai indeed

[steel zephyr]

Is there a good guide on how to use alternating prompt and from-to thing?

[near elbow]

i'm sure there is 🙂
(i just don't know any on top of my head)

[vital raptor]

using what UI?

a1111 (forge should be similar) has a pretty good wiki https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#prompt-editing

so does swarmui for that matter https://github.com/mcmonkeyprojects/SwarmUI/blob/master/docs/Features/Prompt Syntax.md

[steel zephyr]

trying swarmui rn
I thought I'm glitching when I saw a pics in this file...
I played a bit with fromto and seems it's a bit harder than just make char1 cosplay char2

[supple raptor]

yeah, fromto isn't very useful for that task. if you're using Illustrious or NoobAI and the character you want your character to cosplay as is fairly popular you can probably add \(cosplay\) after the 2nd characters name to get what you want. here's an example where i had Y'shtola Rhul from FFXIV cosplay as Vivi Ornitier from FFIX by using (vivi ornitier \(cosplay\):1.2), final fantasy ix, y'shtola rhul

pony models might get that right with super popular characters, but i wouldn't expect it to work on very many

[vital raptor]

eww, my 'new' M40 gpu takes 2 minutes to gen a SDXL image

not that I expected much from it, but damn

[rough oxide]

m40 is part of 900 series

....so 900 series speed.

vram isn't everything, the gpu architecture/speed/etc matters a lot

that's why 4090 is 2x faster than 3090

3090 feels right on the edge of what i feel's acceptable, like 8-9 seconds per image on XL

feels like way too long on flux/sd3.5l (like 30-40s an image ;_;)

[modest fog]

wtf is that x3

that thing on the right reminds me of that girl-dof from alchemist anime

runs away

[rough oxide]

https://tenor.com/view/cosmog-pokemon-sun-and-moon-pokemon-legendary-pokemon-alola-gif-25689672

and then yeh

cosmog found something weird in pokemon universe, an actual "animal"

an abomination of mankind

[modest fog]

what is the one on the right called

Is that a pokemon too

[rough oxide]

probly nina.

[modest fog]

noo

I want the name of what you genned

lmao

[rough oxide]

"ruins"

oh the cosmog entry in this wildcard has "blonde hair" for sum reason and "long hair" probly attributed to a trainer it often appears with or sumthin xD

[modest fog]

oh

[vital raptor]

yeah, I'm aware

I bought it for other stuff anyway, just testing out image generation

[rough oxide]

i mean u could still use it in a server, queue up a bunch of stuff throughout the day and just collect the pics later or sumthin xD

m40 is def a super bad gpu for pretty much anything nowadays tho

[vital raptor]

depends on what you're willing to spend

I already have a 4090 and a 3080TI, I just needed something to go in my server so my roommate could have his 1050TI back

[rough oxide]

wish i could've waited a month to pick up 4090 but my last pc pretty much died right before 4090 came out xD

and wanted to do full upgrade

crazy how it's like 2x better at gens ;_;

[supple raptor]

yeah, no fp16 cores is going to hurt it's performance a ton for AI inference

[vital raptor]

yeah, it's doing fine with some LLM stuff

not as fast as my 3080TI that I had doing this task obviously, but more than passable

[supple raptor]

yeah, for that kinda stuff i'd imagine it's fine

[past junco]

Oh my gosh so cute!

[past junco]

I declare a CUTE BATTLE!!!

[past junco]

Wish I had a fast setup

[vital raptor]

generating fast just means i go through more images before I find one I like 😄

[past junco]

Yeah true

Why the... heads

[modest fog]

Hatsune mewku

[supple raptor]

berry cute

[green temple]

reimu will join

[green temple]

[steel zephyr]

Can someone help me to make a seed walk animation in swarmui?
I suppose it has to be done with comfy wf, but I don't know anything about it

[gusty palm]

how in the world is anyone supposed to run a flux model they are like 10+gb

does it just load a portion of the model in whatever available memory you have

[steel zephyr]

24 GB GPU: exists

[gusty palm]

who do i look like bill gates

[green temple]

illustrious is better ( for anime )

[gusty palm]

i guess theres no getting around the fact that the entire model needs to fit in vram in order to use it?

i was never really super clear on if thats the local limitation

gonna try an xl model today

[steel zephyr]

I don't think it's possible to calculate all the stuff if you need to read from the slowest device in your pc aka hard drive (even ssd)

It probably would take hours and days if even possible

Maybe it just needs all the data in the same time to process a single step

[gusty palm]

anyone know how t o use this stuff

or should i just enable and try it

[past junco]

Cute

[past junco]

Messing with things is the best way to learn what anything does

Granted, I could never get good looking tile generation, always produced weird results

Like a patchwork knitted blanket

[gusty palm]

yeah i was having that uissue with tiled diffusion on forge

i was able to get some interesting Gorillaz album cover art type stuff tho lol

by setting tile size super low

hey girl do u have a third leg or are u just happy to see me

[modest fog]

eat

Strawberries are overrated

runs

[green temple]

+1

[steel zephyr]

I think I've found a bug in swarmui wildcardseed caused by <random>...
No issues about it, I'm testing what causes it to loose seed

[modest fog]

@green temple smol got pranked

strawberry dress

[gusty palm]

ever since i ran a generation with demofusion ticked by accident ive been randomly gettging cuda illegal address kiindof semi randomly even with it not ticked

have to close server and relaunch (reload ui doesnt help) to fix it but i restarted my computer ill se if that helps

[gusty palm]

which is best

i like the perspective of higher numbers but

[steel zephyr]

How does one train LORA nowadays?

[modest fog]

hmmm

nubby was talking about it few days ago

lemme see

this

[steel zephyr]

thanks

[modest fog]

maybe you can make it work if you just look at the page and try it

genius stuff, I can't help

[steel zephyr]

btw, when there's no compiled exe it scares me

it almost never compiles for me

[gusty palm]

characteristic guidance extension produces good results if u dont mind it t aking forever

it took like 6 minutes though and heres without

[gusty palm]

actually nvm idk if it was that extension or the incantations one but it broke my ability to generate i had to uninstall them lol

[steel zephyr]

Should I avoid BREAKing prompt at any cost or it's not that bad and could be useful?
It feels irrelevant to let two different prompts affect one image, but maybe it's a common knowledge?

[near elbow]

whatever she is trying to say

[supple raptor]

first you must understand a couple of things. stable diffusion only supports prompts of exactly 77 tokens. no more, no less. the first and last tokens are the start and end tokens, which leaves 75 tokens for the actual prompt. so what happens if you submit a prompt that's only 25 tokens long? it fills the remaining 50 tokens with a padding token. now for the slightly more complicated part, how do we handle prompts that are longer than 75 tokens? we split the prompt into blocks that are 75 tokens long and send them each to CLIP separately, then combine the results before sending it to the UNET and this is what BREAK actually interacts with. when you insert BREAK into your prompt it fills the remaining space in the 75 tokens that CLIP needs with padding tokens so that the next token will be the beginning of the next batch that's sent to CLIP. BREAK is simply allowing the user to define where the prompt is split into different batches rather than allowing the normal rules to do it for you. on SD1.x models this function could sometimes help to separate parts of the prompt that you didn't want to affect each other as strongly, for example sometimes prompting for red hair, blue eyes, green shirt, pleated skirt, shoes the red from hair or the blue from eyes would cause all of the clothing to be red or blue, but inserting a BREAK between them would sometimes prevent that from happening. this behavior is much less common on SDXL models, and i've never seen any benefit of using BREAK on them, and in fact have generally seen quality of images degrade on SDXL models when using too many BREAK functions.

[green temple]

she is laughing at this

[supple raptor]

that's cause she's a baka. only baka would laugh at marisa

[green temple]

indeed

[supple raptor]

https://tenor.com/view/marisa-master-spark-touhou-spell-card-marisa-spellcard-gif-26768695

[modest fog]

@green temple where smol

these people write too much text

brain hurt

[green temple]

inc

[bitter latch]

https://tenor.com/kYctkZ2DgHL.gif

[steel zephyr]

thanks, I understand this concept in general. I'm just curious, is there any benefits of intentionally breaking prompt? Wouldn't sampler just treat entire picture as red hair then entire picture as blue eyes and apply this to entire image? Would it affect things that shouldn't be affected?
Or I better keep my prompt <75 tokens?

[vital raptor]

generally, if you're using sdxl it doesn't matter

you can use long prompts and you don't need breaks

I've heard that a1111 (and possibly forge) do some normalizing to the entire prompt weights which leads to tokens at the end of the prompt not being as effective on very long prompts, but I don't know how valid it is.

Overall, it's generally better to keep your prompts as short as possible, just enough tokens to get the results you need and nothing extra

[steel zephyr]

got it, thanks

[modest fog]

@lyric zealot help

[vital raptor]

that said, breaks can have an effect so it could be something to play with, but imo that goes for everything, I test multiple tokens and combinations to get similar results and then use what works best for my needs

[lyric zealot]

Help with what

https://tenor.com/view/miku-hatsune-miku-meme-looking-at-camera-looking-gif-14156409521403913189

[vital raptor]

like, using smile or happy could have very similar results depending on the model, but combined with something like shy they have very different results

[modest fog]

im innocent, a gun was pointed at me

[lyric zealot]

https://tenor.com/view/neuro-sama-vtuber-shoot-gun-pov-gif-7305956933862925905

[supple raptor]

nyo... two guns were pointed at you

[steel zephyr]

praise the neko-chiken-king

[modest fog]

Cute pockets

[steel zephyr]

Well, turns out latent image is not just a noisy texture...

[green temple]

where`s fox version

[bitter birch]

has anyone made a couple art? and perhaps it's like specific actions :v

[near elbow]

and with better CLIP models like SD 3.x or flux it blends into each other much less...
"anime drawing, cute yellow dog riding on red horse with a blue cat next to it" flux fusion

[supple raptor]

Those actually use the same CLIP models as earlier SD models used, the extra text encoder you were probably thinking of is T5-XXL which is an LLM.
SD1.x = CLIP_L
SDXL = CLIP_L and CLIP_G
Flux = CLIP_L and T5-XXL
SD3 = CLIP_L, CLIP_G, and T5-XXL

[near elbow]

it does its magic well though 🙂

[steel zephyr]

apparently it works
yet, I'm not sure if it's possible to specify something more specific for each char without some regional shenanigans

[supple raptor]

works well with characters that can be made with just their name. gets much tougher without regional shenanigans if you need to add description tags for the characters. at that point you just kinda have to get lucky or use regional tricks.

[steel zephyr]

smol fox Miku

struggling with sampler, tested a few, but euler a at 40-60 steps looks best for me

probably not makes sense because it's ancestral tho

[supple raptor]

depending on model, i usually use either euler or euler_a with normal or beta scheduler at 28 steps for first pass and 20 steps for 2nd pass with 1.5x pixel upscaling between passes

[steel zephyr]

Sounds like comfy. Should I sampler->latent upscale->sampler->output?

[modest fog]

[modest fog]

Smaller

[steel zephyr]

that's kind of terrible

[steel zephyr]

challenge accepted

[modest fog]

Challenge what

Just a smaller bonk for you

[steel zephyr]

oh, I'm already generating smoller Mikus

bigger bonk?

[modest fog]

Don't post that here

No kiss

[steel zephyr]

no kiss, just noses touching

no cute allowed

[steel zephyr]

how to make glitchy image?

[steel zephyr]

the most glitches I've got without ruining image

[modest fog]

most models you can do "glitch" on prompt

[rough oxide]

1. you might wanna upgrade to 0.6 v-pred model (just came out today). 2: a combination of glitch+chromatic aberration then using artist styles that'd push it more towards what u want usually helps.

if you add disintegration too it can get more bizarre

but the artist(s) will be one of the main things

[bitter latch]

Ive been using glitch style for a while

glitchcore works too

[bitter latch]

ok back to regular anime 😄

haha

[modest fog]

https://tenor.com/view/mirror-monday-morning-shocked-dog-funny-dog-gif-22203011

[rough oxide]

silly smol creature not understanding human-built objects

[modest fog]

Here is your image (silly small creature not understanding human-built objects)

My first image on 3.5

[terse ore]

This was original image made in blender (render)

Then I made this. I tried to inpaint the horns red using Segment Anything and Grounding Dino. But for some reason it looks bad

I used with and without controlnet. But neither look good 🤔

[rough oxide]

probly better if u use like koikatsu since danbooru has tags for koikatsu \(medium\)

so it can fit the style a little better

[terse ore]

This is koikatsu character imported into blender. I want the same composition and inpaint each part of the image to make it 2d. So I can make consistent panels

[rough oxide]

hmmm. you can export with the correct shaders & all?

i remember last time i tried shaders & stuff broke

eventually got a working model in vrchat but it took ages xD

[terse ore]

Yeah, it just needs the right compatibility

[rough oxide]

will have to look into that again if modding their new games ever become popular again. i got HC but the modding community wasn't nearly as stoked over it as they were with koi... even tho shaders&etc were 3x better

and now there's a new one after hc which i forgor name of xD

[terse ore]

What is HC?

[rough oxide]

their 2nd to last game now

they took ages to add unity studio thing to it too but now it has that too

[terse ore]

I see

[modest fog]

Smol but with headache

[green temple]

me the last 3 days

[modest fog]

Get better

Now!

it should be gone now

[green temple]

[modest fog]

@green temple time for treatment

[green temple]

strawberry

[modest fog]

You'll become a strawberry

food for smol

[green temple]

[modest fog]

would help her with homework

(just not math, im too dumb )

[green temple]

me too

[keen hollow]

evenin'

[keen hollow]

nice lights

[keen hollow]

these are awesome, need more 😄

[terse ore]

Can invoke AI use grounding Dino? I tried in a1111 but I get OOM if I try to inpaint using SDXL which I heard it is bad at anyway

[green temple]

[slender barn]

Thank you 🙏

[supple raptor]

Ty. The newest vpred noobai model is looking pretty nice. I think the vpred could still be stabilized a bit more, but it's already really nice for dim images.

[bold ether]

My name is pain. Back pain

[vital raptor]

Back pain and claustrophobia

[green temple]

the better version

[vital raptor]

omg so cute

[vital raptor]

I was sitting here thinking about making a new character and saw your pic, so this is all your fault @green temple

[green temple]

this is me crying from cuteness

[green temple]

🥺

[green temple]

dark

[zinc flax]

vpred is nice

[vital raptor]

that's the one thing i don't like about autismmix, the lighting control is garbage

although, it's shadows and general lighting is really good sometimes

always a key light, no matter what

[zinc flax]

Maybe you can try cattower vpred, that model almost has the same feel to autismmix

[green temple]

still smol

[vital raptor]

i haven't been able to get the right style from a noobai model yet

and I don't really do dark images enough to spend the time tbh

[zinc flax]

looking for the artist combination is tough

[gusty palm]

got an xl model to run

[vital raptor]

yeah, i ran a few thousand artists and combos and got pretty close, but it's taken me months to end up with what I have in autismmix

[zinc flax]

morning

[zinc flax]

what happened with the gen lol

[zinc flax]

probably the toughest part of a new model for sure

[vital raptor]

and I don't really have any reason to change, so I haven't spent the time.. probably won't until something like flux (or better) gets a good finetune

[gusty palm]

idk this one came out alil better

[zinc flax]

I don't think it's worth to use xl if you're gonna gen with a low resolution

[gusty palm]

ohh

[vital raptor]

yeah, the whole point of using sdxl is the higher quality

on the other hand, if you have a good upscale workflow, sd1.5 can be just fine

[heavy cobalt]

Great

[gusty palm]

i might stick to that this took almost 4 minutes

[zinc flax]

sd1.5 is still fine

Unless you need some of that newer character concept

[vital raptor]

tbh, i only switched to sdxl because at some point a1111 broke the way embeddings worked and it destroyed my sd1.5 output, so I figured if i was going to retrain a lora anyway, i might as well use sdxl since I have the hardware

if they never did that update i'd probably still be using my 1.5 model

[green temple]

it`s fine with lowres

[vital raptor]

i mean, yeah, but if you're generating at low res anyway, why not do it faster on a 1.5 model

[gusty palm]

howd they break embeddings

[zinc flax]

it's dogwater lol

no need to sugarcoat

[zinc flax]

probably the newer version

[vital raptor]

don't know, they never acknowledged the problem and ended up ignoring my bug report

it's possible the specific embedding I had was the problem, but it basically made it so i couldn't reproduce anything the same way any more

[green temple]

i find it easier than 1.5 🌚 and because of character knowledge too

[modest fog]

Grabby

[gloomy scroll]

Get grabbed

[vital raptor]

angy gremlin

[fossil fox]

cat

[gloomy scroll]

Yue has fish also

[gusty palm]

this came out like crap but im getting better

[twilit portal]

Her first time at the aquarium 💧

[green temple]

old gen

[modest fog]

Old smol

Blue

[green temple]

[green temple]

who is this

[modest fog]

1girl

[green temple]

no she is funnya ( missing strawberry )

[modest fog]

I'm 1sadcat

[gusty palm]

something kinda....

suggestive about this one

[supple raptor]

@green heart the spammers are getting more creative in hiding their scams now

[near elbow]

yeah, also reported it as spam the old discord way

[keen hollow]

Yeah I saw something got posted. I'll have to try it out. Same setup/structure as the previous versions?

[near elbow]

it's on many channels

[keen hollow]

I've noticed quality and artist tags are basically mandatory.

[supple raptor]

yeah, with the exception of it being a vpred model. SwarmUI, Forge, and Reforge should all load it correctly with no extra steps if up to date. A1111 will as well if on the dev branch. Comfy just uses a node to specify that the model is vpred+ztsnr. As for generation parameters, it works best with Euler and Euler_A currently, other samplers may behave poorly (later versions of the model are supposed to fix that) and it likes lower cfg between 4-5.

also, yes all of the illustrious and noob models are best with some artist tags and quality tags in the prompt

@bold ether we got a spammer in many channels

[modest fog]

@green temple

[keen hollow]

ok, I'll give it a whirl, if you INSIST... 😛

[south trail]

a cute girl

[bitter latch]

cant sleep

😄

haha