🍥｜anime

[modest fog]

@desert island why are you thumbs upping my sleepy girls

let them sleep

[desert island]

Just saying I like the images. You're the one taking pictures of them and posting them on discord! YOU let them sleep! 💢
... 😁 j/k

[modest fog]

I will stop taking pictures and take action instead

👍

[desert island]

well it's got loads of problems but she blinks 😄

[modest fog]

It's alive

Kill it!

[zinc flax]

Cursed

[desert island]

https://tenor.com/view/stranger-things-fire-flame-gif-14920648

[keen hollow]

Thanks 🙂 I'll give it a whirl.

No specific scheduler or step count you recommend?

[desert island]

ughhhh jittery failure 💢 Now I remember why I stopped trying to do animations for animated characters

[agile meadow]

local catgirl gets seizure

[keen hollow]

yeah I dunno, I'm getting much better results with other samplers than EulerA CFG++... that one seems really blown out

[magic wave]

Just 30~ steps

[magic wave]

Are you using 1 to 2 cfg? :o

[formal shoal]

guys what do you think about a 12gb 3060 for generating images?

[vital raptor]

unless you're getting a really good deal, a 12gb 3080ti or a 16gb 4060ti would be better

[desert island]

you can gen with all the available models for the most part but you will not be future proof or have the speeds you'll probably want due to bus speeds. Can you use it to gen images? Absolutely but the second you're like "I want to try to do video or play with using an llm on local in conjunction with your generative models you'll wish you had more vram. All in all it's a great place to start and you'll defo be able to do stuff.

[formal shoal]

mmmmmm i dont want to spend a lot on money

[formal shoal]

it soo big the diference between 3060 and others?

[vital raptor]

realistically, you should wait a couple months for the 5090/5080 release and grab something good used

[formal shoal]

yeah that true

When the new graphics come out, they will lower the price of the old ones. Better wait.

[vital raptor]

in the end, your only real limitation is speed, if you don't have enough vram it'll overflow into system ram and go slower, but it'll still work, the difference between a 3060 and a 3080, both 12gb vram, is the speed

[formal shoal]

oh, i see

i'll better wait then

[vital raptor]

my roommate has a 3070ti (8gb vram) and it gens a sdxl image in something like 8-12 seconds

my 4090 gens in 2.5-3.5s, but not because of hte vram, it's just a much faster gpu

[formal shoal]

I think I'd rather wait a bit, spend $100 or $200 more and buy something that works better and lasts longer.

[vital raptor]

currently the best money for performance seems to be the 4060TI 16gb version, it's about $450 USD right now, getting one of those used when the 50 series gets released will be a steal

[desert island]

I'm on a 4080 so I can say that I envy your vram so it isn't like vram isn't important 😄

[formal shoal]

[vital raptor]

well, I went from a 3080TI 12gb to the 4090 24gb, and tbh 99% of the time the vram is nice, but not worth the money i spent

a typical gen only uses maybe 14gb vram at most, and usually much less

[desert island]

I went from a 4060 8gb to the 4080 super similar jump just one tier down. 😆

[desert island]

same except in video. Interpolation is easy but the initial gens are murder if I'm trying to do anything over a couple of seconds at a usable resolution.

[desert island]

out of curiosity what is lora training like on the 4090? My current pipleline takes several hours.

[vital raptor]

it's a matter of perspective too, I don't use flux models because I think it's too slow (40s for a gen with hires 2x), but some people can't even generate a 1024x1024 image in 40s

I haven't done any lora training recently, but when I first got my 4090 it was taking about 20-30 minutes to train sdxl

I also had a small dataset though, like ~20 images

[desert island]

DAMN! That's fast... similar dataset size but I might be doing to many steps. I'm working on cooking up a new method for lora creation since my last attempt didn't produce the results I wanted... anywho thanks for all the info 🙇

[vital raptor]

i think i was doing like 10 steps per image, but i don't remember, i was just training a specific face

in the end, i switched to sdxl because i was tired of all the prompts, loras, embedding, etc required to making anything decent with 1.5

but I make up characters that work well with the model I use instead of trying to get specific characters from the model, so it makes it easier for me

[desert island]

I'm a huge fan of SDXL I keep trying to get into flux but every time I play with it I end up just going back to sdxl. I keep trying to find a way to use it in a way that's beneficial but I find I can generate a comparable solution in sdxl most of the time. Every time I think I can do something in Flux better then I can in sdxl I find that there's something I can do if I just apply the right tactics when working in SDXL for less resources and faster gens. So many wasted hours testing just to find out that SDXL is a better use of time and resources. smh.

[vital raptor]

yeah, flux is great at context and comprehension, but it doesn't know very much ultimately, if there was a pony type finetune of flux that worked as well as both of them i'd probably switch

but on the other hand, pony/sdxl is still faster, so i'm still not sure i'd end up using flux anyway

[desert island]

came to the same realization a couple of days ago.

[vital raptor]

it's also hard to switch to a new model, every model prompts differently so you have to relearn the quirks and 'bad' keywords/tokens for the new model

for me, I found a model and figured out how to make it work for me, so I don't feel the need to try different models all the time

i play with the new stuff like SD3/3.5 and flux, because I can and it doesn't waste too much time on my gpu

but they're not very impressive yet, so i just go back to my normal stuff and wait 😄

every once in awhile I try poking around with video/animation stuff, but it's all so bad I end up giving up

I just want to be able to gen a few images as keyframes and then try to make the AI fill in the blanks

[desert island]

toon crafter is pretty good at that. the pipeline I'm in the middle of working on is aiming to do just that. basically using DWpose to extract poses from video like a poor mans mocap then through the use of a lora apply that to the control net open pose rig stitched together with frame interpolation from toon crafter.

so far I have pieces but nothing works well together as of yet.

[modest fog]

eep

I wish I could sleep as peacefully

[desert island]

Aren't you a cat? Isn't sleeping like your thing? Or is that a racist thing to say?

[modest fog]

weather has been hot so I only sleep for like 2-4 hours

And I have nightmares every day

Nightmares probably make me sleep worse but I don't know

[desert island]

catnip helps since it's basically impossible to dream when you take some... Not sure what part of the world you live in and if it's legalized there, but food for thought.

[modest fog]

catnip does that wtf

to people ?

[desert island]

T.H.Catnip hint hint wink wink nudge nudge

[modest fog]

I will block you

[vital raptor]

toon crafter looks interesting, but the way it gobbles vram and the low resolution seems to make it mostly useless

[desert island]

don't use it for the actual generating just use it for frame interpolation.

the 512 image resolution doesn't apply to frame interpolation and vram usage goes way down aswell if you just use it for that side of things. Use svd to gen your main vid so that you can test it... you'll see the flaws but you can also see why better structuring your main video source is the most important part of the process since if you just use svd to go img to vid it loses the plot on many key details. I think the key to this is creating a lora and using open pose as well as something like a vision model to instruct keyframes...

[supple raptor]

Only if you're interested in buying used

[modest fog]

https://tenor.com/view/bonk-cat-slap-cat-gif-22044106

im angry now

:<

I hate that shit

if you do that just block me

goes for anyone

👍

[vital raptor]

well, tooncrafter doesn't work, just segfaults immediately with no errors.. seems like the comfyui method might work better, but I don't use comfy so I'm over it 😄

[desert island]

oh sorry what do you use? are you on swarm or forge I'm a comfy user so I didn't know it would cause problems for you.

[vital raptor]

i use swarm, so i could use the comfy method, but it's supposed to run on its own

but it doesn't, and it seems like the devs are all using windows so that's probably why it doesn't work for me

[desert island]

I'm on ubuntu but if I give you a workflow you can just use that right? Regardless it seems like your over it so I won't waste more of your time since the solution isn't what you're looking for anyways right now... However, if and when I do get it working I'll let you know. 😄

[modest fog]

@supple raptor eheehe

[steady grail]

what do you have against frieren?

meanie

lol

[modest fog]

she ugly

and dirtied lots of my elf gens

cause people thought it was her

or looked similar

[vital raptor]

considering all the anime elves out there.. frieren is not the one i'd pick for my waifu, i wouldn't call her ugly but definitely not the best looking

the anime is really good story wise though

[supple raptor]

https://tenor.com/view/lol-gif-21667170

[modest fog]

Not ugly but not the best looking

Me in life

[vital raptor]

[steady grail]

wow, that's a good style

[keen hollow]

yeah... I can create a grid example

[keen hollow]

@magic wave

that's with the silvermoon mix

lemme try illustrious default

[magic wave]

Well, I've been using NoobAI :o

[keen hollow]

with euler_cfg++ and karras?

[magic wave]

I think this is a pic I've done without any LoRAs, Euler Ancestral CFG++, 2 CFG, and Normal sampler

[keen hollow]

I still haven't tried that noobai one.

[magic wave]

If you want to grab the PNG Info from that one

[keen hollow]

I dunno why people like base illustrious, it's hot garbage... 😄 😄

illustrious pencil...

yeah, euler_cfg++ isn't my jam

[zinc flax]

You need to use 1-2 for ++

7 fries it

[steady grail]

that looks a lot like pony

[brittle rampart]

Doppel herz 😁 😍

[keen hollow]

this is CFG 2.

[zinc flax]

Something must've gone wrong with the grid then

[keen hollow]

I get the same with individual images.

[zinc flax]

That's odd, euler++ shouldn't look that much different than euler

[magic wave]

Is that Euler Ancestral CFG ++ or Euler CFG ++?

[magic wave]

Using the last version of NoobAI, something is different about it, but it still works just fine

[vestal ferry]

the last version is pretty good

i make this one with the 1,0 verrsion

[magic wave]

Yeah, from what I heard and tested, 0.5 to 0.75 is more unstable cause they trained TE, 0.75 to 1.0 should be more stable cause they re-trained the UNet to stabilize it, but I have to get used to it and relearn it xD

[vestal ferry]

yea i talk to LAX and i he improve my combination style alot

https://tenor.com/view/butter-trickcal-gif-670913706337336218

[magic wave]

Fair fair :o

[keen hollow]

same issue with both on those models, I'm afraid

samplers, rather

[magic wave]

Thats weird

Clip skip is right?

[keen hollow]

running sets?

[rough oxide]

running sets?

testing a lot of random things with 1.0

[keen hollow]

yeah just random sets of whatever 😛

[rough oxide]

prettym uch

surprised how much better my alyas turn out from just 10 more epochs xD

well "just 10", actually kind of a lot for a model of this size

[jagged field]

can someone give me a prompt to try

[rough oxide]

1girl, shinshuu maru (kancolle), kantai collection, braid, brown hair, long hair, twin braids, breasts, brown eyes, large breasts, ribbon, red ribbon, hair ribbon, ahoge, akakura

...but that only describes char

[keen hollow]

whicih model?

[jagged field]

anything you want

[jagged field]

do you want a specific one or random ?

[rough oxide]

¯_(ツ)_/¯

[keen hollow]

Anything with Tifa. 😄

[rough oxide]

i was gonna test it myself just now in

new model

akakura is pretty gud style

[rough oxide]

that model dun understand artist ;_;

looks like kancolle char is popular enuff tho

lol

[jagged field]

I'm using sd3.5 large

[rough oxide]

o

i have sd3.5 large in another tab

[rough oxide]

best animu model is something else tho

and way easier to run xD

[jagged field]

Wait lemme try

I have a big list of models i will try to choose one

[rough oxide]

this is the char i described

playing around with artist styles to see what works well with this new model per char

[rough oxide]

pwny is not gud

uwu

have u tried the new anime model?

[jagged field]

wait

[rough oxide]

it is better than novelai diffusion v3

and local

sd3.5l is good. nothing pony-based is good

check this model out https://civitai.com/models/833294/noobai-xl-nai-xl

[jagged field]

i lost the prompt 1 min

[rough oxide]

it's the main anime model rn

it can even do chars up to july

and adheres to artists/etc very strongly

xD

i r berry strongly recommending nai-xl

you won't even need to use loras

u can transition other stuff into any style no lora, like making chars psg style for instance

huke's style comes out perfectly

etc

[jagged field]

can you come to diffuser lonuge ?

I'm stuck with an error in comfyui

[rough oxide]

idk wym by that

[jagged field]

I have an error wait trying to fix

I loaded a model into wrong direcotry

[rough oxide]

post issue here?

[jagged field]

I think the issue is that i loaded a unet model in checkpoints

i have to move it becuase it says can't detect model type

[rough oxide]

put it in unet folder

and then refresh

should work

[jagged field]

yes that's what i did but refresh didn't work i had to restart

[rough oxide]

my main point is u dun even need a lora to do this

anymore

if u want best anime model

nai-xl is best one right now

[jagged field]

I'm just trying to run flux

I think i broke things up

[rough oxide]

flux is unet w/ specialized vae u need to run alongside it

• dual clip loader

i don't sitll have my flux workflow cuz i moved entirely to sd3.5l for realism

lol

[jagged field]

wait it's loading

I think it's fixed

something is wrong there is blur idk why is that

[rough oxide]

u using dual clip loader?

[jagged field]

yeah

[rough oxide]

u using right samplers/etc

[rough oxide]

use emptysd3latent

for flux

but that isn't the issue

hmmm

are u using flux guidance node

[jagged field]

lemme try

[rough oxide]

flux doesn't use cfg

emptysd3latent + fluxguidance

[jagged field]

where should i connect it ?

[rough oxide]

ummm

i dun remember xD

[jagged field]

xD

[rough oxide]

but it's necessary

maybe into prompt box

cuz iirc flux had no neg prompt

but uh i really dun remember where it goes xD

but that's 100% ur problem

use comfy's workflow?

[jagged field]

Okay wait

Oh

[rough oxide]

https://comfyanonymous.github.io/ComfyUI_examples/flux/

[jagged field]

MUHAHAHA

[rough oxide]

first one

[jagged field]

it worked

I'm comfy ui master

[rough oxide]

did u replace ur latent with emptysd3latent too

cuz that improves quality for flux

they use similar vae

[jagged field]

I replaced it yes

but

[rough oxide]

but

you activated my trap card?

[jagged field]

I connected this

and connected the negative directly

[rough oxide]

that might be right

[jagged field]

and put cfg to 1

[rough oxide]

um

don't use negative

with flux

[rough oxide]

doesn't support iirc

well

it's not supposed to support it.

but looks like u got better qual

lol

[jagged field]

yeah but this is not flux

this is flux fusion

[rough oxide]

whai u no wanna try dedicated animu model xD

[jagged field]

oof it broke again

[rough oxide]

don't use neg

i think comfy's workflow had a custom sampler

node

copy the top fox image into ur workflow

[jagged field]

okay nvm wait

i will try the workflow

okay now testing

Hmmmmmmm

[rough oxide]

i think my desc doesn't include enough of the important parts cuz anime models fill the gaps ;_;

[jagged field]

Okay wait i will try your model

[rough oxide]

not mine lol

[jagged field]

i mean the one you sent

[rough oxide]

it will replace pwny for u. i swear.

no more need to have 400gb of loras

xD

[jagged field]

I don't have any loras yet

[rough oxide]

o

then u were using pony wrong

pokerface.png

pony is only good for loras

[jagged field]

I'm just playing around with H100

I think it's sdxl so i will load it into the sdxl workflow of sd3.5

[rough oxide]

one sec i'll send u my noobai workflow but

u'll either have to change the name of noobaipred10 whatever to NoobAI or expand load checkpoint thing

cuz this one needs clip skip

specific sampler/scheduler/etc

[jagged field]

Oh okay

Thanks

[rough oxide]

works very well if you learn some artists and knows every char/concept on danbooru with i think

[jagged field]

Prompt: 1girl, shinshuu maru (kancolle), kantai collection, braid, brown hair, long hair, twin braids, breasts, brown eyes, large breasts, ribbon, red ribbon, hair ribbon, ahoge, akakura

[rough oxide]

it's roughly 80+ now

so even like histoire works

from neptunia perfectly

i also have a sheet of chars i've confirmed to work in some capacity

...but

there's like 5800+ here

xD

easy reference sheet tho

[jagged field]

Lemme try

[modest fog]

All ugly

[rough oxide]

that shuld be fine. that artist was a bit too strong

i think i had to reduce weight a bit

[jagged field]

this is correct right ?

[rough oxide]

yea

[jagged field]

I'm trying dpmpp sde karras

[rough oxide]

dpmpp 2 i think

oop

[jagged field]

2m ?

[rough oxide]

2m yea

[rough oxide]

there's like diff setitngs for everything

a lot of us found euler ancestral beta to be best typically but the TE is a bit weird in this

sometimes u have to lower weight of artists/chars

[jagged field]

weight ? cfg ?

[rough oxide]

e.g. highlight a word

[jagged field]

i don't have weighting in the prompt

[rough oxide]

and press ctrl+down arrow

[rough oxide]

so like highlight akakura and hit ctrl+down arrow 4 times

things like that

[jagged field]

okay i see

[rough oxide]

the problem is

this mdoel

knows like

anything with 80+ quality posts on danbooru

so some things are too strong

it's basically a fully lora-less model

for any char/concept u could want

[rough oxide]

maybe down to 0.6

lol

but

that still has its own style

[rough oxide]

i like that output xD

[jagged field]

This is to 6

[rough oxide]

good shading therr

it's stuff like that

[rough oxide]

as an example just changing artists on this becomes....

(waiting for gen to finish, i batch 4 at a time lol)

[jagged field]

This is flux 1.1 pro

[rough oxide]

simple artist change becomes this

on otherwise same prompt

artists are super strong words in this model

[jagged field]

I can give you something

wait

[rough oxide]

using 1-4 artists & adjusting weight makes this model stronk.

[jagged field]

can i post links here ?

[rough oxide]

depends on the kind

[jagged field]

it's a link for my comfyui

[rough oxide]

like civitai page or sumthin?

[jagged field]

no

[rough oxide]

iun trust links like that xD

[jagged field]

errrm

try to scan it with virus total if you're worried

[rough oxide]

i just dun like visiting unknown sites ;_;

[jagged field]

it's okay if you don't want to

but the annoying part it will keep showing that it's opened in another tab because i opened it too :3

[rough oxide]

hmmm i'm still perplexed as to why

you had pony without any loras xD

pony shined with loras but its time came to an end when illustrious became base model

[vital raptor]

because pony doesnt need loras?

[rough oxide]

wat

it totally does

it lost almost all basic concepts

try to get a leek in pony

xD

a completely important part of any kasane te- hatsune miku gen

[jagged field]

lol someone just playing with it

[vital raptor]

technically I'm using autismmix, not base pony, but i've never needed any loras

[rough oxide]

as you see, hatsune miku holding a leek and spinning it

with no lora.

ez.

[jagged field]

I think i have this one

[rough oxide]

....and by hatsune miku i mean red-haired miku. kasane teto.

[jagged field]

oh rip it's removed

[rough oxide]

btw u shouldn't host if u have manager

tons of ways to rce.

[jagged field]

i trust you guys

[rough oxide]

u shuldn't.

lmao

[jagged field]

[vital raptor]

there's spammers and scammers in here all the time, don't trust anyone

[rough oxide]

yeh

there's way smaller more tight-knight communities u'd probly have no issue doing that in

but uhhh

this is the main sd server

there's tons of malicious lurkers

[jagged field]

[rough oxide]

just what happens in big servers

not a slight vs this server specifically

[jagged field]

I mean I have a backup so it's fine

[rough oxide]

nonono

unless ur on a sandboxed environment

i'm saying

[jagged field]

yes

[rough oxide]

they can affect ur pc

lol

[jagged field]

no it's not my pc this is my h100 server

[rough oxide]

how long does it take u to make 1 35 step dpm 2m/euler a image for noobai?

curiously

when u were testing it

[jagged field]

Errm wait lemme try

[rough oxide]

for me it's about

37/4 since i batch it

9.25 secs

[jagged field]

you can try it yourself

[rough oxide]

better than novelai used to be but worse than they are now xD

[jagged field]

if you want

since i already posted the link lmao

[rough oxide]

just wondering speeds a h100 gets

lul

[jagged field]

it's very fast on euler but dpm++2m takes longer

[rough oxide]

euler a is similar type of model to dpm++2m

i think

[jagged field]

for cogvideox5b it's 6 seconds and 50 steps it's 160 - 200 seconds

[rough oxide]

regular euler is super fast.

[jagged field]

euler and eulera both are fast

so this settings

[rough oxide]

nuu not euler

euler a

[rough oxide]

hmmm

[vital raptor]

slow

[rough oxide]

pretty fast.

about 4090 speed

[jagged field]

but i have like 5 other models loaded in vram

[vital raptor]

4090 is faster than that

[rough oxide]

innerestin'

[jagged field]

but this is autism

i think model size makes a difference

[rough oxide]

autism and noob r same architecture

[vital raptor]

it's still sdxl

[jagged field]

idk maybe something i'm doing is wrong

[rough oxide]

4090 specifically had massive ai improvements (and lol now i wish i went for it over 3090 more than ever)

so much stuff coming out that could benefit from speed

[rough oxide]

pretty sure if u batch 2 it'd be even faster

[jagged field]

what are you using ?

[modest fog]

Only a 3090 smh my head

[vital raptor]

4090

[jagged field]

no i mean

[vital raptor]

swarmui

[jagged field]

the software

oh maybe i need to change and try

[rough oxide]

swarm is same backend

comfy

[vital raptor]

doesnt matter, it's still comfy backend

[rough oxide]

i haven't updated my nightly torch stuff in a while

i shuld update errything

wait comfy added new text_encoders folder

r we suppose to put clip stuff there now?

hmmm

[jagged field]

is this 35 steps?

[rough oxide]

seems like yea

transfer stuff to text_encoders now

[vital raptor]

25

35 steps is 4.8s

[jagged field]

oh that's pretty close

i will try to make 4 images

[rough oxide]

then it's same speed

like i said

pokerface.png

[rough oxide]

still 2xish faster than 3090

wat

WAAAT

that's insane

lol

[vital raptor]

h100 should be faster i'd think

[rough oxide]

yea

that's way faster

2.5s an img

less

like 2.4s an image

xD

that's also why i said it'd be faster to batch

longer to get 1 image but in this case for 2x the time u get 4x images ;_;

[rough oxide]

yeh it hits a limit at sum point

that's why i dun batch after 4 usually

but

that's still a speed increase

of a further like .1s

[vital raptor]

yeah, in a batch of 4 it's twice my speed

[rough oxide]

u only saved like .4s

i save a lot more i think

[jagged field]

But for 1 image is pretty close maybe it has something to do with model loading and stuff like this

[vital raptor]

could be

[jagged field]

I like ur pfp

[rough oxide]

whose

[jagged field]

@vital raptor

[vital raptor]

lol, thanks

[jagged field]

and you too

[rough oxide]

lol

mine is one of those sd3.5l ones with photoreal bg with anime girl forefront

but

once noobai hit 0.75 i stopped putting as much time in 3.5l

[jagged field]

Maybe i can try to animate things

[rough oxide]

i have mochi workflow

u can use

from comfy himself

[jagged field]

can someone send me any pic to put it into cog ?

[rough oxide]

o

ur h100

can run mochi

[jagged field]

i have cog i run it but someone told me u can turn on onediff

will make it faster

[rough oxide]

cuz u can mochi on ada6000

[jagged field]

does this auto download mochi ?

[rough oxide]

no

[jagged field]

oh

[rough oxide]

mochi 1 is the best video model right now but

it's txt2video

not img2video

https://huggingface.co/genmo/mochi-1-preview

[jagged field]

okay wait

[rough oxide]

i don't think i modified the mochi workflow at all from when i got it from comfy

u can probly do more frames

than comfy put

[jagged field]

the problem is huggingface download speed is limited to 40mb

[rough oxide]

yeh

and it's a 20gb model

takes a while

[jagged field]

I have an idea

[rough oxide]

oh

50 gb model"

40"

....i cannot read fast today xD

but yea that 40gb is why you need like ada6000/h100 to run it

in the first place

to load all that into vram

[jagged field]

it's this one right ?

[rough oxide]

yes

i am unsure if u need encoder/decoder

[jagged field]

this will take ages if it got downloaded with 40mb

[rough oxide]

but comfy's workflow had something for it

....if ur like

hobbyist tho

it's best video model rn

¯_(ツ)_/¯

[jagged field]

wait i will try

smth

[rough oxide]

and u have hardware that can run it xD

i can do it too but only for like

2 seconds

and it takes me 4 mins

to do 2 secs

also it'll output a png

it's actually an apng

but won't work here

u'd have to convert it with ffmpeg or something

to like webp/gif

or mp4

[jagged field]

it will take like 50 seconds now

[jagged field]

they should create a custom node for this one :3

[rough oxide]

o nice, updating to newest pytorch nightly got me a speed boost

about 10% faster

[jagged field]

what is it ?

moochi ?

[rough oxide]

mochi 1 preview

ye

[jagged field]

How to get the nightly ?

[rough oxide]

i mean pytorch nightly

i hadn't updated in a while

[jagged field]

oh okay

[rough oxide]

but uhhh

might be smaller than i thought

shaved about a second off based on what i was just getting

if h100 qualifies as a nvidia gpu for pytorch stuff

then the command is
pip3 install --pre -U torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu124

[jagged field]

I think i have already done this one

when installing comfyui

[rough oxide]

if u did it recently then ye

otherwise it's "nightly" cuz they release updates like every day

[jagged field]

yeah i see

[rough oxide]

doesn't really seem to be something u need to do too often

sometimes there's small optimizations

[jagged field]

do i need anything else other than the moochi dit thing ?

[rough oxide]

i don't think so

i think it includes everything else innit

love how it gets diabellstar's masks now

xD

i forgot to remove restaurant part of prompt i was writing just now ;_;

[jagged field]

Okay i will try now

everything is ready

[rough oxide]

u can do higher length

i think

but not higher width/height

dun mess with that

[jagged field]

wait i forgot about the vae

[rough oxide]

this is only a 480p mdoel they'll release hd model later

no

i think

hmm

i feel like it's all in mdoel but mochi1 page is loading so slow

[rough oxide]

ye that's like auto loaded

from genmo

i think

[jagged field]

yes i thought it's my internet

[rough oxide]

everything is in the dit file

[jagged field]

nope

[rough oxide]

no?

ok then u need encoder/decoder

but

i'm 99% sure it was all in genmo itself

when i tested it

[jagged field]

okay now i've got it

[desert island]

I'm still downloading all the files 😄

[jagged field]

what is this

i can't find a palce to add the encoder

[jagged field]

rip xD

[rough oxide]

no that's correct

u have to convert it

to some format readable by discord

that's an apng right now

discord doesn't do apngs

[jagged field]

yeah

it's moving in comfyui

but when i click download it gets the png only

wait

[rough oxide]

ffmpeg can convert it

the png is the correct file

but it needs to actually be

converted by you

lol

[jagged field]

this is 1 sec only ?

how can i increase the seconds ?

[desert island]

can I get your workflow so I can catch up?

[jagged field]

it's the same workflow that @rough oxide sent

[jagged field]

This one @desert island

But you have to download the model in unet then choose it and download the decoder in vae folder and choose it also

[desert island]

yup just finished downloading everything just needed a starting point thanks 😄

[rough oxide]

you need 3090/4090+

this is nonnegotiable for this one

lol

they recommend 4h100s even

but it works with 1 ada6000 well enough it seems

[rough oxide]

increase framess!!

i told you!!! lol

how long did that take btw?

[jagged field]

you're welcome and also thanks for @rough oxide for sharing the workflow 💗

[rough oxide]

on ada it was 2 mins for that length

[jagged field]

do you mean this ?

[rough oxide]

yeah

[jagged field]

what does 37 mean ?

is this in miliseconds or what

[rough oxide]

it might be 24 frames a sec

or something

but 5.1 sec is max

iirc

[jagged field]

so what should i put here ?

for 5 seconds

[rough oxide]

try 120?

[jagged field]

okay

do you want me to try a specific prompt ?

[rough oxide]

but how long did it take u to make just the 37 sec?

nah i'm just curious about speed & stuff right now

xD

[jagged field]

I added 120 but it puts 121 automatically

[jagged field]

it's not 37 seconds

[rough oxide]

probly fine

i mean 37 frame"

[jagged field]

it's like 2 seconds

or 3

idk

[rough oxide]

i'm importing a lot of gakuen idolmaster chars that weren't in my sheet lol

[jagged field]

it took 88 seconds

but i think it had to load the model

[rough oxide]

it's gonna take that long

every time for 37

for 120 gonna be even longer

[jagged field]

i will try again with 121

[rough oxide]

i'm not entirely sure

what the frame count to sec is

i couldn't experiment that much

was going out of RAM really fast

[rough oxide]

for 121?

[jagged field]

no

it produces like 3 files

someone told me moochi can do up to like 1 minute

i'm not sure

for 121 it's taking long

at 48% now

[rough oxide]

i'm not 100% sure about all of this. my limit ould probly be like 40 something

cuz my ram literally almost maxed out

and i have 64gb ;_;

[jagged field]

https://www.reddit.com/r/StableDiffusion/comments/1gb07vj/how_to_run_mochi_1_on_a_single_24gb_vram_card/

this might be helpful for you

[rough oxide]

i also forgor to add saya to my list lol

o 30 frames per sec

my bad

[jagged field]

OOF

it says ran out of memory

[rough oxide]

yea

see

lol

this is why they say like

4 h100s

but

u might be able to do hmmm

90 for 3 secs?

[rough oxide]

idk how far u got

[jagged field]

it took like 257 seconds

[rough oxide]

how'd u convert to mp4 or whatever u did btw lol when i tried that my ffmpeg kept breaking ;_;

but this isn't something i'd use myself anyways at this point

[jagged field]

wait

[wooden sail]

NoobAI-XL is hard to handle so far

[jagged field]

ffmpeg -i [inputFileName].png -vf "format=yuv420p" -vcodec libx264 -crf 23 -pix_fmt yuv420p [OutputFileName].mp4

try this

[rough oxide]

i'll save it for if i ever try it again

[rough oxide]

hard to handle? in what way?

ur trying to do something that's better in a vpred model i think

[rough oxide]

they're going to do vpred conversion soon

damn

it's pretty good

for that retro look

but mochi1 is uhh

[jagged field]

try to give me a prompt

[rough oxide]

general model

it can do like anything

idk how to prompt mochi :<

[jagged field]

try anything

[vital raptor]

noobai seems to be heavily artist tag dependent, and if you don't know what arist you want it's mostly just trash output

[rough oxide]

nyooo. i'm fixing up my character database right now xD

[jagged field]

okay i will try something maybe

but daymn that took all the 80gb vram

[vital raptor]

I'm currently running a few hundred gens through an artist list I have to find something good since it doesn't give the same results as pony models with the same artists (some better, some worse)

[jagged field]

but the quality is good

[rough oxide]

i have an artist list too xD i keep adding to it as i add more chars to my char list

these r way too big now

well. mainly char one is too big at 5907ish now

[jagged field]

@rough oxide maybe i can try to to animate one of ur characters

[rough oxide]

mochi isn't img2vid ;_; and i dun care much for curr img2vid stuffs

[vital raptor]

I'm not particular about characters personally, I don't gen known characters

[rough oxide]

i keep filling wildcard, good for inspiration or for a quick prompt

if u wanna animate make the feathers move here? xD

idk

[jagged field]

okay wait i will try something

[rough oxide]

now ic probly why you like pony-based models more xD

cuz this model excels at known chars/shows/styles like that

...which is why this list keeps growinnng

for my wildcard mostly

[zinc flax]

1girl 1other

[rough oxide]

xD

i do have 1other

i have a single 1other

1other, testament (guilty gear), guilty gear strive, black hair, red eyes, androgynous, long hair, choker

saya isn't a 1other right

pokerface.png

lol

i never actually read vn so idk wut saya is

but all booru tags for her r 1girl

[zinc flax]

Teeeechnically lol

[rough oxide]

hmmm

finally dun with gakuen idolmaster stuff. idk why i didn't have it before

with that many tags

xD

[vital raptor]

the other thing i don't like about noobai/illustrious is the heavy token counts for quality tags in both prompts, but it's also civitai so no telling what's actually functional or not

[rough oxide]

quality tag isn't necessary tho?

i haven't used them at all

but i have used long-ish neg prompt

and i do use worst quality in neg prompt ig

xD

i very very rarely use quality tags in positive prompt

it really isn't necessary

[vital raptor]

just adding 'worst quality' to neg is a huge improvement

[rough oxide]

the examples didn't use them either

yes just that for neg prompt

but

even noobai creator dun use them in the example prompts

xD

[jagged field]

a dynamic and vibrant animated scene featuring a red-haired heroine with cat-like features. The heroine should be in the midst of an action-packed moment, leaping forward with determination and focus

I tried with this prompt

[rough oxide]

hmmm

[vital raptor]

well, you don't really need anything in the positive prompt when you have paragaphs of word salad in the negative

[zinc flax]

also furry, anthro in noobai case lol

[vital raptor]

one of the greatest benefits of sdxl is that you aren't supposed to need quality tags or negatives at all

but one or two things in the negs if you really need to remove something

this model seems to take a step backwards as if everyone liked the bloated prompt mess from 1.5

[rough oxide]

i like results. noobai shows results

¯_(ツ)_/¯

[vital raptor]

i like results with less effort 😄

simple prompt, no negatives, it just works

pony does have a few required tokens, but not much

[zinc flax]

Fair enough

If it looks good with minimal effort then it's already a win lol

[rough oxide]

something like novelai diffusion went through 3000+ epochs(?)

can't rly afford that benefit for smoller models

[rough oxide]

and rly

i rarely ever change my neg prompt

i've just been using the same one i setup from beginning and it works for me

"worst quality, bad anatomy, bad hands, lowres, text, artist name, signature, jpeg artifacts, simple background, white background, abstract"

i tried removing abstract a bunch of times but it seems to help a lot

xD

[vital raptor]

steals

[zinc flax]

I use novelai negs

It just works™️

[rough oxide]

i forgot wut they even use

and i can't inspect ur image

xD

but idk, using furry, anthro seems to lower qual. + i like using stuff like renamon a lot

[rough oxide]

damn reminds me of like

80s anime maybe xD

[vital raptor]

slayers

[rough oxide]

but uhhh

i think mochi will be better on like realism stuff

[zinc flax]

what i use
lowres, bad, error, fewer, extra, missing, worst quality, jpeg artifacts, bad quality, watermark, unfinished, displeasing, chromatic aberration, signature, extra digits, artistic error, username, scan, abstract, gradient background

[rough oxide]

o ok

that's too much for me

i like chromatic aberration

i use it in prompt a lot too

[zinc flax]

the shorter version i also used
lowres, text, error, worst quality, low quality, jpeg artifacts, chromatic abberation

[rough oxide]

still.

xD

ok like

if i removed chromatic aberration

that'd probly be good

also

u misspelled it

in urshort one

remove extra b, add extra r

[zinc flax]

oh true

[rough oxide]

hmmm that one doesn't work for me; too many white bgs but i can probly remove some stuff from mine stuff.

...starting to think if it works i shouldn't change it, mine works for me xD

and by not messing with furry

i can still get gud renamon

xD

[brittle rampart]

wow! a list of several hundred artists! I still have so much work to do!

[rough oxide]

o

i has a list with 300 artists

also 5900+chars

a bit less than half have full description

but my artist list is kinda more towards what i like

no 2.5d stuff

like wlop or ones like that

[brittle rampart]

Understand. to each his own. but still this is a pretty impressive list

[rough oxide]

credit to both uhhh

sanaeXL and animagine for parts of the list

then lots of my own stuff in both

[brittle rampart]

I have several Wildcard files with a list of artists. but so far it hasn’t been their turn.

[rough oxide]

i've been using a homebrew python script

so i ahve more control over my promtp

it randomizes char + 1-4 random artists

...that way i can alter prompt if it's a male char or something

vs just wildcarding

..."coding" with chatgpt is fun

autocopies to clipboard cuz of pyperclip too xD

[brittle rampart]

to me it all looks like magic. I'm not good at programming 🤪

[rough oxide]

same. that's why i tell chatgpt when it fucked up and i'm like fix it >:|

lol

or maybe i wasn';t descriptive enough

but i think this was just 2 iterations in

pretty simple to make

i also used it to make my merger script

i merged sanaexl list with animagine char list

but had chatgpt program it so it could detect duplicates even tho similar style

worked almost flawlessly, somehow vergil from DMC got duplicated

but that was all

just tell it the problem and it makes solution

lol

and yeh the wildcard script was literlaly just 2 text prompts

cuz i realized i wanted it to auto copy to clipboard after

can kinda learn what it's doing by analyzing it too

[brittle rampart]

This is Magic!

[rough oxide]

it's like stripping 1girl/1boy, male focus and then looking at next part

to see if it matches anything else

between one split, determined by comma

so idk why vergil (devil may cry) got duplicated

but that was the only one xD

[jagged field]

@rough oxide I'm trying the new setup with the moochi wrapper and i will tell you if it's faster or not but i got the fp8 version now

[rough oxide]

will be lower qual if u do that

[jagged field]

there is fp8 and bf16 and fp32

[rough oxide]

but u'll be able to gen longer

[jagged field]

i will try all

[rough oxide]

yeah

fp32 best, bf16 is what i had to use

bf16 seemed pretty bad sadly

already xD

[jagged field]

okay maybe i can try with fp32

[desert island]

Cool it's ugly and terrible but I finally got it to work on my 4080 😄

[jagged field]

👏 nice good job 🔥

[rough oxide]

pixie and alice kinda work now

noice

[rough oxide]

i forgot to wish y'all a merry christmas, i'm like 3 days late

...christmas is the day after halloween leading all the way up to dec 25th right?

[vital raptor]

according to the local stores, christmas started 2 weeks ago

[modest fog]

fu

long time

[steady grail]

ollama stopped working on comfyui...

[steady grail]

i surrender...

[modest fog]

https://tenor.com/view/get-get-pet-cat-headpat-pet-gif-7927420809381407547

Are you using noobai now

[brittle rampart]

yes

[modest fog]

steals her plushy

[gloomy scroll]

Activates magic spells

[gloomy scroll]

Mofu came to say hello

[brittle rampart]

we're finished! Mofu has arrived!