🍥｜anime

[glacial bane]

Mb chat wrong image

[desert island]

"Whale, whale, whale. what do we have here?"

[gleaming flame]

These ain't anime, so I'll post only one, but I tried my old prompt on the vxp_il_xl anime model... and it makes a damn good job at creating funky steampunk golems as well 😄

[modest fog]

Where miku

[green temple]

anything smol is anime dont worry

[modest fog]

my headset left speaker kinda broke

im big sad

[dusky linden]

same, last month i had to replace it

got a dope one now tho

[modest fog]

I cant get a new one ):

[green temple]

@modest fog help

i forgot to delete "big "

[modest fog]

https://tenor.com/view/cat-gif-20118453

[gleaming flame]

ok, this one counts as anime

[glacial bane]

is there a problem "sharp (galare:1), smug look0.2

[gleaming flame]

prompt talking? 😮

[glacial bane]

NO MIKU NO MORE MIKU

I'M GOING BONKERS

BECAUSE OF MIKU!!!

[modest fog]

||including the cleavage above, that is probably not too sfw for this server, I'd tone it down mods are scary||

https://tenor.com/view/cat-running-away-escape-getaway-bye-gif-16631286

[green temple]

+911

why eep 😢 she should

[modest fog]

That is literally the default face this stupid model makes

Illustrious

[green temple]

try pencil mix , the smol one i`m using

[glacial bane]

You're right mb

[modest fog]

look on the left of the bed

easter egg

[dusky linden]

i just had a completely original idea

[modest fog]

wait

its sona from league of legends

amazing

👍

[dusky linden]

messing with my base prompts. (basically partially stealing ideas from others) getting neat detail on these now

[dusky linden]

whats nicer, dark or bright backgrounds

[green temple]

dark

[supple raptor]

https://tenor.com/view/pat-cat-pat-pat-gif-25250101

[modest fog]

[lyric zealot]

https://media.discordapp.net/attachments/1084147551537746042/1205246324803510292/evillicka3.1.gif

[gloomy scroll]

https://cdn.discordapp.com/attachments/1019446913268973689/1326270709542162452/ezgif.com-animated-gif-maker.gif

[desert island]

The shadowing is so good in that gif!
<- jealous

[desert island]

TOO CUTE! I can't stand it! * SQUEE !!!*

[desert island]

Ugghhh gotta make some airflow adjustments to my case. 😠 temps are hitting 72° C

[modest fog]

Would

[lyric zealot]

https://cdn.discordapp.com/attachments/1327242612205162526/1327242612515672097/eliv_snowball.gif

[gloomy scroll]

https://cdn.discordapp.com/attachments/1019446913268973689/1322239532547969075/YueWave-ezgif.com-crop.gif

[modest fog]

@gloomy scroll help

[modest fog]

uses her as a shield

[gloomy scroll]

I let you borrow her

[lyric zealot]

https://media.discordapp.net/attachments/1072697081443131476/1173827110335434792/ezgif.com-gif-maker-1.gif

[gloomy scroll]

But wether she's helpful of not is not something I can say

[modest fog]

[trim current]

which model you using these days smol

[lyric zealot]

been off ai art for like a few weeks now. After getting back again these few days, i was just using what i got previously... which is....

noobaiXLNAIXL_earlyAccessVersion

[trim current]

ooooo pog

[lyric zealot]

Will need to get back into it, but might first have to fix/reinstall my swarmUI, since my super outdated forge can’t support most new stuff

[trim current]

i fully swap to comfy tbh

[steady grail]

comfy is the only true answer

[modest fog]

ugly

Generating images through discord bot is the only true answer

[vestal ferry]

https://cdn.discordapp.com/attachments/1327242612205162526/1327242612515672097/eliv_snowball.gif

[fierce temple]

Which one you use for image generation?

[modest fog]

It was a meme

[fierce temple]

Stay safe 😂

[vestal ferry]

https://cdn.discordapp.com/attachments/1323068703717458041/1323068704467976362/colored_C014.gif

[glacial bane]

A villainous daughter

[gleaming flame]

voluminous daughter you say?

[desert island]

voluptuous?

[wispy canopy]

mermaid? XD

yeah, totally a mermaid.

[mystic flame]

beautiful cartoon character girl with princess dress, in the beautiful and big green plain,

[mystic flame]

How can I create picture?

[gleaming flame]

this isn't midjourney, you don't create here. you just share what you created

[mystic flame]

Aha. Ok thanx

[fierce temple]

Future

[modest fog]

Rouge

[wispy canopy]

you gotta admit, he does look ffing cool

[gilded quiver]

Bro ngl ur art is magnificent masterpiece

[gilded quiver]

https://tenor.com/view/sus-ayo-sussy-imposter-hello-gif-23089496

Umm

Its quite detailed

[wispy canopy]

it's a good model

[gilded quiver]

Ye

If its were little close it would be more great

[glacial bane]

I personally just gate keep my good gens

Daddy's home~~~~

[gilded quiver]

Cool

If it has purple aura around him it will be quite fascinating

[keen hollow]

This is so good. We need more of this character. I need a widescreen one of her sailing her pirate ship!

that was raw flux dev, no loras.. not bad, I didn't think flux did the anime style that well

trying the pirate lady....

[gleaming flame]

yea, I tried flux S... because that's max I can do with 12GB... and gave up... too slow, to generic... written words I can add in post if I ever need any

[keen hollow]

I didn't really like the flux s results either.

I can do dev on 12GB though - you can't? 🤔

[gleaming flame]

I tried once, and after long while and a long black screen my pc switched to 1024x768 and gave me a bunch of lovely errors

[keen hollow]

weird

I just did one at 3840x1600

granted it isn't my favorite image every, but yeah, no issues

[gleaming flame]

I'll try again, after I have more then minimal required vram...

sisters...

[modest fog]

what does that even mean

[dusky linden]

I post the good ones but i do remove the prompt at times

Especially if i post it to civitAI. Keep the cycle toxic 💯 💪

[modest fog]

Kinda cringe

[modest fog]

Her arm

[gleaming flame]

I guess she won

[glacial bane]

It means NO ONE AND ONLY ME CAN SEE MY GOOD GENS jk depends if i wanna show it if i do I'll just unsent it,

[modest fog]

You need more mikus

[glacial bane]

NOO PLEASE ANYTHING BUT THAT

[glacial bane]

Is what you thought i would say

[glacial bane]

DIVERGENT FIST

[modest fog]

Ugly

[dusky linden]

was working on lamia's but since the subject came up anyways

[glacial bane]

STOPP NOOOO I'M BURNING

[glacial bane]

@dusky linden@modest fog

[dusky linden]

would ||not||

[modest fog]

Very ugly

[modest fog]

[gleaming flame]

Mikolai, for friends (he has none): Miku

[modest fog]

I think I'd prefer miku snake

[gleaming flame]

good night ppl

[near elbow]

arrrr, prepare to be entered! (i think this is SFW enough, blurring for a bit of extra awareness 😉 )

[keen hollow]

I like that bird 😄

[dusky linden]

did you post this on civit by any chance earlier?

[keen hollow]

some of these pirate generations come out fairly... spicy. 😐

what is it about pirates that just does that?

[dusky linden]

female pirates didnt originionally exist

and then onepiece came with the bikini pirate

and this hololive

streamer

[near elbow]

no, i never post anything there - i'm playing around with zoinkscoobFurryNoobai 🙂
this is already a toned down version. just prompting for "flower bouquet" gives you a practically nude lady with a bunch of flowers. i had to add "wide sweater" to make her safe for here, and even then it took like 5 tries 😉

[dusky linden]

ah i saw a similar pirate girl on there earlier with the parrot and all

was wondering if it was your profile lol

[near elbow]

nope 🙂

hm, that's speech bubble text "not me" but i'm used to flux for text, sdxl doesn't quite cut it (and anime/manga is with tons of japanese in the dataset of course - can anyone of you tell if this has some "not me" in there?)

[dusky linden]

SDXL cant do text. it sometimes just freaks out

very rarely it spells a word right and im just shoked

[rough oxide]

sdxl can do like 1 word

sd3.5l/flux can do like a sentence, sometimes two

[modest fog]

I can do 3 sentences

Sometimes 4

[rough oxide]

🤔

it's possible but then accuracy gets way lower and have to try for a lot of gens i feel.

can sometimes pretty consistently get 1-2 sentences correct in a lot of the gens.

[modest fog]

I meant I can read 3 or 4

smol brain

[terse ore]

Do you think with the 50 series Nvidia cards coming out, the 3090 PCs will become cheaper?

[modest fog]

If you can actually buy them, yeah

From what I saw their prices are good

but that is not real if theres no stock

[terse ore]

I mean, I would want to buy a used 3090 pc, not a 50 series. Just that maybe with it coming out, it could mean the older stuff could depreciation further

[modest fog]

Thats what I'm saying too

if 50 series are actually available the older stuff will get cheaper

always been like that

[terse ore]

Oh I see

[modest fog]

It's bloon a while

[modest fog]

steals her coffee

https://tenor.com/view/cat-running-away-escape-getaway-bye-gif-16631286

[near elbow]

yeah, hit and miss, but much more hit than SDXL etc 🙂

[near elbow]

well unfortunately they let the 4090 supplies run dry... otherwise that would still be a very good GPU

[wispy canopy]

i feel the 5090 is gonna be a card with pure artificial demand -> don't produce any 40xx anymore, and don't release 24gb cards

if there was a 5080 with 24gb everyone would by that one instead of the 5090

[vital raptor]

tbh if there was a 3050 with 24gb+ vram no one would buy almost anything else

[wispy canopy]

yeah, nvidia knows what people want, and they know their premium card sales would tank if they did

[vital raptor]

very few people buy the top end cards anyway, i dunno why people freak out so much about them

[wispy canopy]

well, i have a 3090, and i'd LIKE to upgrade

[vital raptor]

right now the 4060TI 16gb is the best AI card for the money

[wispy canopy]

but they make sure there's not gonna be a 24gb card for an decent price

[near elbow]

hehe, and i have a 4090 (first one they released back in 2022), so i'm in the potential target group for a cutting edge upgrade 😉

[wispy canopy]

stop producting 4090's, so their price remains high, and overprice the 5090

[near elbow]

yeah, capitalism at its finest

[vital raptor]

for 24gb+, the 90 series has always been the only option besides workstation and/or datacenter cards

[wispy canopy]

i suppose they learned the lesson from the 1080ti -> that card was a long lasting beast

like, i think that card topped the charts for like forever

decently priced, and could pretty much run anything at the time

no artificial shortage either

[near elbow]

they clearly learned 😉

[vital raptor]

yeah, but the 1080TI only really hung that long because the 2080 was basically the same performance plus ray tracing

[wispy canopy]

which is the 4090 -> 5090 story all over again

[vital raptor]

we'll see

[wispy canopy]

the rasterisation perforance is 10% better ...

[vital raptor]

until there's actually benchmark data no one knows

[wispy canopy]

then again, i'm on a 3090, so a 5090 will be a upgrade for sure...

[near elbow]

well 8GB VRAM more means bigger models or several in parallel (LLM + T2I and maybe T2V)

[wispy canopy]

i already do that 🙂

32B llm model -> t2i sd3.5L

[vital raptor]

the vram is why i upgraded from a 3080TI to a 4090, i'd get a 5090 for the same reason

[wispy canopy]

just make sure models are kept unloaded while not in use

[vital raptor]

realistically, the only reason I'd get a 5090 at this point is if an actually usable AI video model comes out and runs meaningfully faster on a 5090 than a 4090

otherwise a 4090 is plenty fast enough for me

[near elbow]

hunyuan is quite nice and runs well with 24GB and less, but probably doesn't fall under you r"actually usable" (my results vary randomly between "wow this could be a marvel movie" and "ouch the horrors" and "this is just a silent zoom-in and totally boring")

[vital raptor]

I want something more along the lines of being able to follow a 'script' for 10-15 seconds at least, like "a girl walking away from viewer down a hallway, then turning left and looking back over her shoulder and smiling, then turning around and walking back"

I would also be ok with something that could take input/keyframes and animate the in-between frames cleanly.. there's a few systems out that try to do this already but they're not very good in my experience

[trim current]

i found using ollama to help prompt for hunyan i get better results than jsut myself

[wispy canopy]

after dealing with the adventurer's party.

[near elbow]

yeah - i have limited success with describing a sequence (X does A then B etc) and camera control (camera circling, camera following, camera flyby) but the results still seem a bit random. since you can just generate a few seconds at a time anyway, you could use the last frame of the video to start a new one... but i think there is no img2video for hunyuan yet, just generation from scratch. LTX is much faster and does img2video but quality and concept understanding is much lower...

[dusky linden]

me generating cute anime girls

[wispy canopy]

"and here we have a house - i have no idea what i'm doing"

[summer night]

which inpaint masked content option is the best for fixing hands?

and also which one is best for adding something that wasn't there before?

[wispy canopy]

firefox

lol, alternative firefox xD

[green temple]

smol firefox

[fierce temple]

Circle Firefox 😆

[wispy canopy]

yeah, that tail might give you the idea she had some REALLY spicy food xD

[gilded quiver]

Ngl it's look fire

Make more with that prompt

[desert island]

This is what I got from mistral-nemo...

My Prompt: anime firefox logo on white background. very high resolution cool design. kawaii chibi anime style.

AI's Prompt: "Anime Firefox Logo, White Background, High Resolution, Cool Design, Kawaii Chibi Anime Style, Vibrant Colors, Rounded Edges, Glowing Aura, Detailed Eyes, Cute Expression, Simple Line Art, Clean Outline, Minimalistic Shadow, Bright Lighting, Straight On Camera Angle, Smooth Anti-Aliasing"

[fierce temple]

Nice

[desert island]

This is my super simple griptape workflow that uses ollama in comfy. I communicate with my ollama model through comfy, have it generate the prompt and send it over to the image generator... if you're interested.

[green temple]

big brain stuff

[desert island]

Griptape is so freak'n cool and I love how they made it so that you could use it in comfy to do a boat load of cool stuff.

[wispy canopy]

why not generate directly in comfy?

[desert island]

I am. I'm just asking my llm to make the prompt.

[wispy canopy]

oh like that. well, i just used direct ollama generate, took me a while to end up on a good system prompt and model tho

[desert island]

you can add rules through the griptape nodes to steer the model to do what you want. Or add instructions. Hell you can even have it write it's own rules and evaluate it with a vision model... <-- if vision models were better then they currently are.

It's meant to be an agent system but the fact that it has custom nodes that you can drop into comfy makes it so that you could do tons of neat things. Like scrape the internet for prompts etc.

[wispy canopy]

i'm lazy, so just used the ollama generate nodes xD

and all of a sudden sd3.5 isn't that bad, i should look into upscaling that shit tho...

SD3.5L 😄

[desert island]

I know right. I played with that model the other day and was shocked by how much better it got.

[wispy canopy]

3.5L is just too late to the party, with majority going with flux

and being a bit worse than flux at people

[wispy canopy]

anyway, gotta go do some groceries xD

[desert island]

I only like flux for very specific things

[green temple]

nice style on the left

[wispy canopy]

sd3.5 is really good with watercolor stuff

[green temple]

nice i really like watercolor

[green temple]

online or local ?

[wispy canopy]

local

used LLM to make and split the prompt (which sd3.5 kinda likes)

can tell u when i get back, if u want

[green temple]

i don`t think i can run local my pc is smol

[desert island]

I'm holding out till May to do an upgrade... I want the Digits box that Jensen announced in his Nvidia Keynote at CES.

[green temple]

no idea what ur talking about but 3060 is enough for me

[desert island]

yeah it has 12G of vram. That's good enough for most things.

[green temple]

yeah

[desert island]

I love the styling of the images your pushing out 😄

[green temple]

thx

[modest fog]

Smol PC

[green temple]

[modest fog]

eats it

yum

[green temple]

[modest fog]

What are those things anyway

[green temple]

dango

something tastes like marshmallows i think

[modest fog]

Never had one of those

[green temple]

me too

[modest fog]

It's sweet then

[vital raptor]

it's a sweet desert type of food made with rice flour and sweet syrups

[modest fog]

@green temple captain very smart

[green temple]

indeed

[vital raptor]

I'm just a fan of japanese culture and food 😄

I've made dango and mochi before

[fierce temple]

Heavenly

[stable crane]

does it unload the LLM in between gens or what

not sure how you can have vram for both llm and SD

my image gens are taking about 17gb vram now

[fierce temple]

which LLM?

[stable crane]

idk, every LLM I've used uses up most of my VRAM

maybe if I had 2 GPUs I could do local LLM + local SD simultaneously while maintaing good speed

[vital raptor]

that's what i do, i have a 24gb tesla M40 in my server running a LLM, it's a little slow, but generates faster than I can read so it's not bad

haven't used it for any image generation stuff yet though

[fierce temple]

Buns 🤤. Need upgrade I guess.

[stable crane]

I have 24gb vram I can't upgrade 😂

[vital raptor]

that reminds me, i still need to see if I can cram my 3080TI into this computer..

[fierce temple]

Maybe bun upgrade 😆

[stable crane]

can't post those ones here 😉

[modest fog]

@green temple

had to prompt for the whole face so the character actually faces the viewer but bloo maid

[fierce temple]

try for sweets and cakes. But a bit older.

[modest fog]

no older, smol

[fierce temple]

Try vegetables

[modest fog]

evilgtables

@green temple go buy her vegetables

[green temple]

[gleaming flame]

Tried one of those new fancy animating services... and got exactly what I wanted xD

[dusky linden]

witch one did you use lmao

even with LTX i didnt get such a result

[gleaming flame]

hailuo ai

[dusky linden]

ah yeah that one

its okay

sometimes

[gleaming flame]

hard to tell from the result xD

[finite herald]

!!!

[acoustic agate]

I didn't think you frequented this establishment anymore

[modest fog]

wtf did I just watch

[green temple]

trying to get haizakura

[green temple]

scary

[modest fog]

She's friendly trust

[gleaming flame]

more swords...

[green temple]

how

[wispy canopy]

less frend shaped

[gloomy scroll]

More frend shaped

[wispy canopy]

love yue 😄

"don't touch our yue's"

[green temple]

best friends

[modest fog]

[vital raptor]

is that a pose?

[near elbow]

presenting the cherryberry on top

[desert island]

I'm using mistral-nemo quant 4 and animagine. Each are about 8GB of VRam so its just enough to run inside my 16GB's.... The model will unload automagically after 5 minutes with Ollama but I think there is something I could do to force it to unload after each gen but that would take way to much time loading and unloading models every generation.

[green temple]

cool

[fierce temple]

A familiar anime character 🤔

[green temple]

frieren

[fierce temple]

Ah, I see.

[keen hollow]

ah yes, worth every cent

[modest fog]

ugly

[modest fog]

much ugly

[green temple]

[trim current]

why do you call everything ugly? you dont like it just yknow keep that to yourself lol

[wispy canopy]

"adjusting tie"

but you probably already guessed

[desert island]

I'm not infront of my computer right now. The house cleaners are cleaning my office, but can I merge SD 1.5 with 3.5?

[modest fog]

smol elf instead

[green temple]

there`s something called "joking"

[desert island]

Ugly. (I can say that because he blocked me a long time ago, for telling him my wife uses cbd to help her sleep. Context he was having nightmares so I was just trying to be helpful)

[modest fog]

@green temple face of a beast ready to devour the bread

[trim current]

bruh who cares i smoke pot to help with my chronic anxiety. if someone gets mad at you for a literal health reason then that says more on their character than yours.

[desert island]

@trim current 💯

[modest fog]

Yeah I don't like that

You blocked too

[desert island]

Lol

[modest fog]

[desert island]

Stupid cat. 🤣

[trim current]

They post allot of questionable content anyways who cares lol

[desert island]

@trim current cheers 🍻

[trim current]

https://tenor.com/view/cheer-cheers-alymew-cheer-alymew-kanpai-kanpai-gif-25305089

its like getting mad cancer patients smoke weed to help them eat

🙄

[desert island]

Well to answer my own question about merging earlier... it seems like a no. I got an exception when I just tried but I'm not sure if I messed something up. Just gonna try one more time when my machine stops screaming at me.

[desert island]

Omg sooo pretty!

[green temple]

thaanks !

[desert island]

The lighting the textures the composition. All of it, are really nice.

[green temple]

thanks

[desert island]

What model are you using

[green temple]

theres no specfic things

only art style

[green temple]

noobVPencilXL_v051

[iron star]

poor cat always attacked by redditors

[green temple]

prompt : 1girl, solo, in medieval village, brown theme, outdoors, peaceful, warm, scenic, outdoors, city,

so simple

[desert island]

Well it came out gorgeous. Kudos 👏

[modest fog]

I play path of exile, I'm always being attacked by their stupidity

[green temple]

thx ❤️

[desert island]

https://tenor.com/view/boop-cute-poke-cat-gif-16281865

[green temple]

angry

[lyric zealot]

[modest fog]

What

[lyric zealot]

the cat is like "don't approach this girl, she's mine"

[green temple]

yeah lol

[modest fog]

Flat

[green temple]

i`m innocent (0)

[green temple]

more glow

[green temple]

smol glow not sword glow

[modest fog]

dangerous smol

[green temple]

[modest fog]

Even more dangerous weapon, have to spoiler || 🥖 ||

[green temple]

kof character ?

[modest fog]

Yes

https://tenor.com/view/shermie-king-of-fighters-kof-kof98-shermiekiss-gif-26120952

[terse ore]

Any recommended pony models that are monochrome? I wanna try img2img+controlnet to see if it will work as an experiment

[green temple]

big

[desert island]

couldn't you just add a color node to remove the color from the produced image?

[modest fog]

Grabby 👍

[terse ore]

Hmm never thought of that. I'll look it up

[desert island]

don't know what you are trying to do but give me a sec I'll find the node I use for this...

good god it's like looking for a needle in a haystack too many custom nodes and I can't find it in the images I thought it would be in... nor is it in any of my saved workflows.... I know I used this before...

[desert island]

Found it. It's in the was node suite "Image filter Adjustments". Do you want a work flow with it in it so that you can just install the missing node or are you just finding your own solution?

[desert island]

This has the embedded workflow.

[terse ore]

Thanks, I'll try it. I wanted to use the blender renders I made to then run it through img2img+ controlnet to make manga.ill try tomorrow

[desert island]

Cool keep me posted. Hope your project goes well! 😄

[iron star]

now do mai shiranui

[vestal ferry]

https://media.discordapp.net/attachments/908912109142827082/1079409922397184100/ezgif-3-af1195d740.gif

[summer night]

does generating an image with broken text using a model like illustrous and then fixing the text with flux inpainting work good?

[near elbow]

should work, but why not just use flux instead? 🙂

[summer night]

i feel like it doesn't work the same as the models i'm used to, i like tag prompting

still waiting on someone to finetune flux with danbooru datasets

[vestal ferry]

https://tenor.com/view/pekora-curry-dance-pekora-curry-dance-pekora-dance-hololive-gif-26798516

[wispy canopy]

xD tv-pokemon 😄

[near elbow]

that could be hard... there's SDXL based danbooru stuff but it retains not much of the original SDXL
personally i prefer the regular way of prompting and anime is just another style, so i'm not missing the ability to generate a super specific character from a super specific show, or strange postures, or tentacles doing interesting things 🙂

[wispy canopy]

you are absolutely right tho, making stuff like this is a pain in friggin ass when you have to work with tag based systems -> they're generally a lot dumber by nature. i wish we had a good finetune of sd3.5L, it's a lot better in styles than flux is, imho

[wispy canopy]

how lewd.

[trim current]

[gloomy scroll]

I wonder who this is

[wispy canopy]

i know, but i suppose you don't care at this point XD

[gloomy scroll]

I mean I have only one person blocked here and it's that one mentally challenged person

[trim current]

thats what they say lol

[wispy canopy]

i'm not about to get into that drama ^^

[trim current]

https://tenor.com/view/crying-happy-but-sad-face-gif-19422780

[wispy canopy]

kek, seems there's a word on the banned list here. anyway, a certain not as safe model is great for making pokemon 😄

great for making weirdness, my kinda shit

🤣

[wispy canopy]

well, that's the limitation of the wai model, here's sd3.5 with the same prompt fed into llm: alternate species, monsterification,hybrid,, holy creature:, nature, anime

tbh, i think it's cool ^^

switched over to 32b LLM -> they are so much better in creative descriptions

[terse ore]

I mean, I tried it with controlnet and a manga model in the 1.5 days and it didn't turn out well. If it was so simple by now, I'm sure others would have tried it

[modest fog]

Mysterious message, we can only guess

Also why does eface have a circle thing

[modest fog]

[summer night]

reforge fixed the extension sorting so i can finally upgrade from the old version of forge

huh... hires cfg scale is a new parameter

idk what to do with it so i'm just going to set it to my generation cfg

i was hoping reforge could do flux but i guess not

the only real reason i decided to install reforge was for the "RescaleCFG for reForge" plugin so i could use noobAI properly

the 1.0 version of the v-pred model was producing oversaturated and messy results compared to the 0.75 version

noobAI is still more "messy" than wai-nsfw-illustrious but it recognizes more things due to being trained on both the danbooru and the e621 datasets

[modest fog]

All are messy and dumb

:<

[wispy canopy]

the black circle?

[modest fog]

you are marked

https://tenor.com/view/cat-scared-cuteness-animals-aww-gif-17687147

[dusky linden]

You want the aheago or the mita ?

[vital raptor]

could literally be any generic anime model

[dusky linden]

^

The style and what artists work (and if it's a mix of styles bigger good luck. Got the image source site?)

[dusky linden]

The base model will still influence the style slightly causing your journey to be a long one

Good luck buddy

[dusky linden]

Like i used the mita style on this one yet the base model decided nah im changing stuff

[modest fog]

Where??

No smol

[green temple]

i started doing big

[green temple]

look , this is smol cuz its from above

[gloomy scroll]

I'm back

[gloomy scroll]

@midnight nest

[near elbow]

miku twins

[green temple]

[green temple]

@modest fog help

[modest fog]

Help what

[green temple]

give smool

[modest fog]

No

runs away

I'm going shower

[green temple]

ok

[modest fog]

I'm clean now

[green temple]

yay ! got smol

[modest fog]

Green smol

[green temple]

[desert island]

I don't understand prompting for NoobXL 💢 is it artist dependent?

[green temple]

yup

[desert island]

how do you prompt for better backgrounds? so far everything seems bland or boring or nonsensical?

[green temple]

use artists, i keep metadata in my images

[desert island]

cool thanks.

[gloomy scroll]

[modest fog]

squish

[green temple]

[desert island]

😭 I'm still getting this level of quality and I've copied your prompt over.... 😢

[dusky linden]

16:9 ratio?

Try smaller?

8:5 or 2:3

[modest fog]

@green temple

[modest fog]

forgot
big

[iron star]

default outfit

[modest fog]

(I have no clue how to do that outfit)

[green temple]

nsfw outfit

[modest fog]

runs away

shes just an innocent fighting game character 👍 nothing to worry about

[iron star]

https://tenor.com/view/shiranuiii-pose-kof-king-of-fighters-gif-16964776

yes shes a great fighter

[modest fog]

[desert island]

I don't think I like danbooru tagging... it's a bit frustrating to try to create what's in my head this way or give the model the space and freedom to surprise me. I'm still trying to figure it out but it's not intuitive.

[iron star]

well illust,pony and nai 1.5 models work best with booru tags,only model that works best with natural language is flux or sd3

[modest fog]

bruh I wasn't typing her name "backwards"

thats why the model was like "No I don't want it, its not exactly right waah" lmao

[desert island]

I guess it's just weird coming from sdxl and sd models over to this side... I'm not used to communicating in this way to try to get what I want. I feel like I have to learn a whole new way of speaking. I like the achievable quality of the models just I find it really difficult to understand or find the artists that will put me in the right space to start crafting the look and feel I want.

[iron star]

yea i think u have to type shiranui_mai cuz thats the japanese way

[iron star]

well sdxl and 1.5 anime models also use booru tags so idk which model u were using

[modest fog]

Pony model does it fine

everything is just more finicky on illustrious

puts cleavage in negative

[modest fog]

@iron star

[iron star]

https://tenor.com/view/community-ken-jeong-ben-chang-cant-see-small-gif-5154350

[modest fog]

can't peek

[humble current]

blood

[terse ore]

As expected, it didn't work. Looks like I passed the thing through a black and white filter with some changes. Doesn't look manga at all. I'll try some other stuff later but not likely to work out

[vestal ferry]

https://cdn.discordapp.com/attachments/1327242612205162526/1327242612515672097/eliv_snowball.gif

[modest fog]

@lyric zealot

Send snowball back

[lyric zealot]

https://media.discordapp.net/attachments/1084147551537746042/1205246324803510292/evillicka3.1.gif

[desert island]

I'm assuming you played with all the settings?

This might help you now that I know you are doing a manga thingy it has another node that can help you set up the panels if that helps.

[modest fog]

Smolahri

[vestal ferry]

https://cdn.discordapp.com/attachments/1249719621255168001/1328994966441037875/894F2627A9153E461DCBCD09DF57559D.gif

[terse ore]

I don't think it is likely I'll be able to get even in the ballpark of the style I want. But we'll see...I won't need the comic node because I can make the blender renders just the way I want and crop from there

I tried different settings yes but there are some other models to try still

Most likely I'll make it in color

Just nice to have a manga workflow

[desert island]

got it got it. All the best. Let me know if I can help 😄

[desert island]

ooohh pretty

[green temple]

thx

[modest fog]

https://tenor.com/view/flashbang-flash-cat-betterttv-gif-10690574645280585754

Flashbang

[modest fog]

Big

[green temple]

[modest fog]

[dusky linden]

Try Camille

Haven't gotten a model that does her justice

[modest fog]

I never bought Camille

She's in the ugly champion pool

[modest fog]

Lmao

[dusky linden]

That's leblanc mix

Bruh

[modest fog]

Aka it doesn't know Camille

[dusky linden]

Yep, shame

[vestal ferry]

https://tenor.com/view/maomao-hearts-excited-wakuwaku-gif-994138680801832398

[modest fog]

Is there even art of her around

She's ugly

[vestal ferry]

[modest fog]

pops the bubbles

[modest fog]

I was talking about Camille

[vestal ferry]

https://tenor.com/view/doro-goddess-of-victory-nikke-nikke-gif-18292928913275144164

okay

[desert island]

I've decided I'd start trying my hand at it. so far it isn't great but I think I'm moving in the right direction.

[dusky linden]

You blind man, Camille is like fine wine.

Then again, a man blind to taste cant appreciate it

[desert island]

Ughhh the quality just degrades... hmmm 🤔 ... time to try something else.

[summer night]

is there something that generates the images one part at a time at higher resolutions and then scales the parts down and stitches them together so you get a more detailed end product?

[wispy canopy]

ultimate sd upscaler?

[summer night]

is that what that does?

[wispy canopy]

yup

[summer night]

i hate introducing more variables to the generation process
like i recently updated from the old forge to reforge and now there's a slider for hires fix cfg

and i have to use this to get the noobai model to work properly

[summer night]

is this basically a better alternative to hires fix?

or does it upscale the image and then downscale it

[wispy canopy]

It actually splits in tiles and then upscale the tiles.

And stitches back

Dunno what it uses for the upscale, an esrgsn or simpler algorithm

[summer night]

so in terms i could understand, it would be like cutting the image into a grid and then inpaint "only masked area" on each area of the grid?

my workflow has only ever really consisted of generating a batch of images, running hires fix on the ones i like, and then running adetailer on them

[wispy canopy]

Correct

[summer night]

but i notice that with using noobai, some details are still messy when i zoom in, and i can't seem to correct them with manual inpainting

[wispy canopy]

Works well because you can keep vram usage low

I haven't been able to get the noobai models to play nice either

[summer night]

it would probably help if i had hires fix set to 2x instead of 1.5x but i don't want the filesize to be too large

[summer night]

i wish my gpu had more VRAM so i could load 2 models simultaneously and use something like wai-nsfw-illustrious as a refiner for noobai

[wispy canopy]

Why not use it by default?

[summer night]

noobai is more capable at some things due to being trained on an e621 dataset

[wispy canopy]

And I think you can set it up it will only have 1 model in vram at a time

Ah, like that

Well, I haven't been able to get really clear output from it 😦

[summer night]

i might go back to one of the earlier vpred models like 0.75 because i remember it being a little more clear than the 1.0 model is

[wispy canopy]

I should probably give it a go again when I get back home

[summer night]

this was generated on the 0.75S version with hires fix and adetailer

i wasn't even using the dynamic CFG thing because this was back when i was on old forge

[wispy canopy]

I work in comfy nowadays tho xD

[summer night]

i just now generated it again on the 1.0 version and it comes out like this, with hires fix but no adetailer

missing finger, weird proportions

yea i'm goin back to using 0.75S

[wispy canopy]

That can always happen tho, missing finger, it is diffusion after all

I can't say for proportions

[summer night]

yeah but the 1.0 version of noobai seems to have consistent issues with hands
i noticed less issues with hands on the older versions and also on other models like wai-nsfw-illustrious

[zinc flax]

cute Kayoko

[summer night]

i consistently get good hands from wai-nsfw-illustrious, and the results also look cleaner, but it just isn't as capable at as many concepts as noobai

[wispy canopy]

Agreed on Wai, that one makes amazing hands

[summer night]

i would try my hand at merging them but i'm not sure how it would turn out or if it would even work since the "prediction types" are different

@wispy canopy do you know if it's better to hires fix before ultimate sd upscale or the other way around?

actually i guess you can't do hires fix after you leave the txt2img tab

[vital raptor]

you'd need a lot of vram to do hires fix after upscaling

strictly speaking, it would probably be better results, but unless you have like 48gb+ vram

[wispy canopy]

I use controlnet upscale

But yeah, takes a lot of vram

[summer night]

probably not going to use ultimate sd upscale anyways since i tried it (following the instructions on the github) and it just makes a mess of the image
this could just be an issue with noobai though

i have no idea how to do controlnet upscale, i honestly hate interacting with controlnet because i never have any clue what i'm doing

[supple raptor]

Ultimate SD Upscale has always been kind of bad. It rarely outperforms the built in SD Upscale script when configured properly and in cases that it does do better it takes considerably longer due to much more overlapping. Multi-diffusion/tiled diffusion was always better for large upscaling than both and much faster than ultimate.

[summer night]

is there any reason to use tiled upscaling solutions if my gpu can handle running hires fix for my target resolution?

like can they result in better quality than hires fix?

[supple raptor]

For resolutions that are much larger than native resolution it typically results in more coherent upscaling. For anything larger than ~2.5x native resolution I prefer to use tiled methods.

My full workflow in those cases is something like:

1. Generate at native resolution
2. Hires fix at 1.5-2x upscaling
3. Inpaint if major problems exist
4. Tiled upscaling at >=2x upscaling
5. Final inpainting and face detailer if needed

[wispy canopy]

it's probably not using those other prediction algoritms, so yeah 😦

my current upscale is native gen -> 2x controlnet tile -> 1.5 controlnet tile

controlnet tile upscales are nice tho -> you can do a denoise of 1 during upscaling, fixing a lot of things

[vital raptor]

i'm simple, I just gen -> 2x hires, I don't really do anything with my images that requires more than that

I think I have 1 image that i upscaled a lot, it's like 16x or something

but I was just messing with upscaler stuff on a nice clean image, nothing super special

this was generated at 512x512, 2x hires, then outpainted to 1024x1536, then upscaled (and downscaled) a few times, this is the 'final' version at 4096x6144 which is 8x the original, but I tile upscaled to 16x before downscaling by half to compress the details better

i can't find the larger version, so i probably deleted it, this one is already plenty big

[summer night]

i just found out that adetailer does not like 2048x2048 images

[summer night]

hmm.. maybe it's something else
for some reason the past few times i've been trying to use adetailer, it makes the eyes smaller and further apart

[terse ore]

In manga, there are typically a lot of white areas. Imo the problem typically associated with img2img is that it will shade lot of stuff that shouldn't be. Also, it's often gradient shaded rather than not. have some old 1.5 experiments at home, I'll send you those. It'd be a game changer if we found something that saved a lot of time even if it doesn't do everything

[stable crane]

@wispy canopy just thinking of Ultimate SD Upscale, I don't know if anyone made a version that takes a controlnet in, have they?

but yeah I haven't got it working well with noob models

[wispy canopy]

From memory, I couldn't change schedulers in comfy implementation of SD ultimate, which was a port iirc: probably same issue in (re) forge

[green temple]

smol

[desert island]

This Lora might do the trick. It's intended application is NSFW but it seems easy enough to trick it to do safe content like this.

[steady grail]

https://www.youtube.com/watch?v=xSFygd71JJI&list=RDQMBi_LpDEGVoI&index=13

[modest fog]

https://tenor.com/view/cat-wiggle-crazy-cat-zoomies-gif-9877358298643393466

anime people in real life

whats that video about

[steady grail]

music i'm into today

i so damn love that instrument

https://www.youtube.com/watch?v=yr3f5FCIFMI&list=RDQMBi_LpDEGVoI&index=16

[modest fog]

theres only one thing im into lately

and its they are billions

[terse ore]

What lora is it?

[modest fog]

(its a game @steady grail )

[steady grail]

what game is that?

[modest fog]

google it
RTS base building vs zombie apocalypse

[steady grail]

cute

[desert island]

@terse ore I think I'm gonna build out a Griptape story teller that will basically create the image prompts for all the panels to tell a story and then use the tool I just created to put all the panels together.

See if I can make something to where you just tell it the kind of story you want and it will generate the prompts for the scenes.
You put the panels together for the pages.
Inpaint out stuff you don't like in the panels.
Clear out the gibberish in the speech bubbles and write your own text.

Then Voila, manga glory.

[green temple]

not smol enough

[desert island]

Just a test/ proof of concept 😄

[keen hollow]

I expect to see the proof of concept by this evening.

[green temple]

ok

[terse ore]

I have seen that done in a reddit post before. imo you will get less control and consistency, but it will be much faster

[desert island]

I guess if you wanted more control you'd just manually prompt the panel that you needed. But you could have it just genning scene images all day and then look through the output to grab the stuff you thought was usable then fill in the rest with specifically prompted stuff that you purposefully generate.... I'll try my hand at it and make the mistakes to understand the limitations. This is fun. Thanks for the idea.

[terse ore]

Yep, that strategy has been tried as well, I read an AI webcomic that did just that. been looking into this stuff for a while and have made notes and listed ideas. Do you want to take it in DM? Here is ok too, I just wouldn't want to clog the chat if it bothered other people

[desert island]

https://tenor.com/view/that-gets-a-big-oh-yes-please-carly-shay-icarly-s3-e4-i-go-public-gif-16976083536695720112

[modest fog]

@green temple attacks you

[keen hollow]

trying to plan out a new pc build and can't figure out how I want to do my storage.... thinking I need a way to separate all the random data I keep around that isn't super high priority.

old game installers, dvd backups, stuff like that.

[modest fog]

I have a 250 ssd for windows and nvme for games

[keen hollow]

Yeah I'll have a dedicated OS drive and dedicated game drive.

I have 7 drives always connected.

I think I'd like to consolidate that.

I have a 1TB drive fully dedicated to AI stuff that is basically full. 😐

I'd take some of this storage and run a little NAS type box, but, ugh.. another device to maintain.

[desert island]

Maybe get a NAS?

[keen hollow]

not to mention cost for power over time

[keen hollow]

yeah maybe I should just get a synology or something

though I'd sort of prefer to keep NTFS filesystem... hmm

[modest fog]

Idk how you people even have this much stuff stored anyway I'm not the best to say

[keen hollow]

well three drives are essentially mirrors

But yes, lots of data accumulated over many years. Lots of photos and videos from vacations and stuff.

[modest fog]

I'm always in vacation

[desert island]

This case works really well for all my storage needs if you want to keep it all local in one box.
https://www.thermaltake.com/the-tower-500-mid-tower-chassis.html
I just set mine up a few days ago. Best case I've ever bought and it has so many bays for drives.

It's huge by the way it's like the size of a damn mini-fridge.

[keen hollow]

interesting case

I do need a case for the new build. Probably something a little more traditionally configured, though.

Most cases nowadays are glass paneled, RGB-obsessed, and I frankly don't need that. But like... supply is limited for high-quality cases that are just black boxes.

Fractal North XL seems to be a popular option.

[desert island]

I resemble that remark. 😄 I got tired of my little black box and wanted something pretty and shiny on my desk to look at since I already threw so much money into it.

[keen hollow]

trying to decide if I need to replace my desk with one of those fancy standing desks

[desert island]

my wife has one in her office that she loves. I'm on an 8 foot long cosco folding table. 😆

[keen hollow]

I use a fairly chongus wooden desk

not sure if chongus is a word, but seems right

[desert island]

I didn't know that word, and I still don't but I looked it up, and got a good laugh trying to decipher your meaning based off of my internet search. 🤣

[humble current]

https://tenor.com/view/dream-moon-clouds-gif-5101225

[desert island]

Team works makes the...

[modest fog]

Hot!

[supple raptor]

https://tenor.com/view/big-chungus-gif-21904924

[modest fog]

No sad allowed

[gloomy scroll]

https://cdn.discordapp.com/attachments/1019446913268973689/1322239532547969075/YueWave-ezgif.com-crop.gif

[modest fog]

steals her floating circle

[lyric zealot]

"you dare steal my halo"

[modest fog]

@lyric zealot have furry

@green temple have furry too

[lyric zealot]

https://cdn.discordapp.com/attachments/1067638175478071307/1292911100060373002/harpoon_him.gif

[vestal ferry]

https://tenor.com/view/doro-dorothy-goddess-of-victory-nikke-nikke-gif-5455004454303322529

[modest fog]

[modest fog]

ugly

[modest fog]

Ctrl z

[lyric zealot]

"stoopid"

[modest fog]

Ugly eyes

[vestal ferry]

https://tenor.com/view/snow-white-nikke-goddess-of-victory-doro-gif-17327124434392458870

[modest fog]

[vestal ferry]

https://tenor.com/view/maomao-hearts-excited-wakuwaku-gif-994138680801832398

[modest fog]

@green temple evil

[vestal ferry]

https://cdn.discordapp.com/attachments/1327242612205162526/1327242612515672097/eliv_snowball.gif

[summer night]

for how much pixai pro costs for a year you could literally just buy an office PC and put a decent used gpu from ebay in it and get something that can do local ai and more 😭
there would probably even be money left over to cover the power consumption

even renting a cloud pc is probably cheaper

you can get used/refurbished RTX 3070s on ebay for around $300 and they're decent for SDXL, when i used one for a bit, it took 15 seconds to generate a 30 step 1024x1024 image on forge

who is pixai for

[modest fog]

paid ugly

[summer night]

expect to see the used market for GPUs flooding with 4070s and 4070 supers in february

i have a 4070 super and it generates a 30 step 1024x1024 image in 7-8 seconds on reforge

[vital raptor]

3080TI is about the same, but 12GB vram

[summer night]

the 4070 super has 12GB vram too

[vital raptor]

ah, thought it was 16 for some reason

might be thinking about the 4060TI

[summer night]

the 4070 ti and ti super have 16GB

[vital raptor]

that makes sense

silly nvidia making a super, ti, and ti super

[summer night]

the 4070 super was a refresh for the regular 4070, and the 4070 ti super was a refresh for the 4070 ti

the super variants released at the same msrp as their predecessors

so there's no reason to get a regular 4070 or 4070 ti unless you're buying used and they're cheaper

[green temple]

is 3060 good ( soon i`ll upgrade to 3060)

[summer night]

i believe the 3060 is one of the worst gpus you can get tbh

[modest fog]

dont ask these rich people

[summer night]

what is your budget

[green temple]

[vital raptor]

havent used a 3060, i went from a 2060 super to a 3080TI and then a 4090

[summer night]

like i said earlier, you can get a refurbished 3070 on ebay for around $300 usd

[vital raptor]

my roomate has a 3070TI, which is pretty decent

[summer night]

the price might go down when the 5070 releases, but i'm not certain

[vital raptor]

the main thing is the vram for AI, realistically as long as it has 12gb+ it's fine imo, though I'd go at least 16gb

[modest fog]

4060 ti with 16gb is very good for ai stuffs

[summer night]

i upgraded from a 1070 to a 4070 super

the rest of my pc is still as old as the 1070 except for some of the fans

[green temple]

using 1660 super rn

[vital raptor]

besides the vram, it's all just reducing time by a few seconds

well, anything is better than a 1660 😄

[green temple]

🤣

[vital raptor]

the 1650/1660 cards are a joke

[modest fog]

[summer night]

right now for image gen, i think the only real benefit of having more vram is for better flux quality and being able to upscale to higher resolutions without needing to use some tiled upscale method

having a lot of vram is really beneficial for text gen though

image gen models might continue to increase in size though

[green temple]

yeah you also have more time to test stuff without getting bored

[summer night]

the generation speed is more dependent on the gpu's performance than its vram

like how many cores it has

more vram just allows you to fit larger models

[desert island]

and bus speeds

[summer night]

i downloaded this because apparently it's the quant that will work best on 12 GB of vram, but i have no idea how to use it
i probably have to use comfyui or swarmui instead of reforge

i think modern forge also has flux support but i don't want to lose my extensions

[modest fog]

flux is eh

[summer night]

i was mainly just going to use it for inpainting text, but that doesn't seem worth closing reforge, loading up swarmui, inpainting the text, and then closing swarm and re-opening reforge

[vital raptor]

i expect flux to become more popular once the 5090 is out and people have 32GB vram to train with

or someone pays for the GPU time at a datacenter for training..

flux needs a pony or noobai type of model to be trained

[modest fog]

pony v7 when

[desert island]

Am I the only one excited about the Nvidia Digits box launching in May?

[summer night]

tbh i don't have high hopes for pony v7 considering that i now know that pony v6's incompetence for character and style recognition was intentional

i don't like having to download 200 loras

[desert island]

I feel like everyone missed that little nugget Jensen dropped at CES.

[summer night]

i made this comparison a few weeks ago

[vital raptor]

they're too busy whining about a $2000 gpu they were never going to buy to bother talking about a $3000+ box that they don't know what it's even for

[vital raptor]

that's not a very fair comparison tbh, you're comparing highly trained merges with a base model

but, pony was meant to be a base model, not a finetune

[summer night]

i compared autismmix too and it wasn't any better

[vital raptor]

it wasn't intended to generate your favorite waifu

it's more about artist styles and nsfw content

[modest fog]

boring characters

[vital raptor]

on top of that, if the name of the character is the only thing in the prompt, pony won't do well anyway, you have to add some artist tags and the garbage that was accidentally trained

[desert island]

148G of "Unified" memory on a Blackwell... I mean that's special. Sure it's DDR5 but the ability to do crazy stuff that requires multiple models on one machine inside a thing thats slightly larger then a mac mini... I'm gonna get one... 3000 dollars be damned! Take my MONEY NVIDIA!

[summer night]

these are separate images because i gave the pony models a fair shot by including their score tags

pony is still good for original stuff and does hands and anatomy better than most other models, but it isn't good at recognizing characters or styles

[vital raptor]

imo, a good model should be more trained on anatomy, concepts, etc. Loras are easy to train and use so leave specific characters and styles to that

any model overly trained on specifics is missing out somewhere else, there's only X amount of data

[summer night]

yeah but sdxl loras take up way too much space

i'd rather only have to use loras as a last resort for concepts or characters the ai can't quite grasp

[modest fog]

how many do you have

[vital raptor]

space is cheap, and loras are tiny

[summer night]

11 of these are for illustrious and noobai

[trim current]

christ lol

[vital raptor]

geez,i rarely use loras and mine take over 36GB

[trim current]

i only ever rlly use self loras, and train loras for fun, thats insane haha

[vital raptor]

meanwhile my models folder is 435GB

[trim current]

oh yeah same lol

but mainly full of flux models

those fuckers are big

[vital raptor]

you can get 1TB ssd for like ~$50/usd these days

[summer night]

i only keep checkpoints that i use

[vital raptor]

i have a lot of space so i don't clean out my stuff as often as i should, i only really use 3-4 models and usually just the same 1

[trim current]

yah i im at 8tb now lol

4tb is nvme

[summer night]

my PC does not have anywhere else to install drives

[trim current]

get an enclosure

and nvme

plug in usb 3.0

boom

[summer night]

my motherboard does have another nvme slot but i don't have the screw for it

[vital raptor]

i ended up building a NAS, it's at 12TB right now, with some mirroring and such, i have 8TB in NVME drives

[trim current]

thats a good idea for me if i ever consider getting more space

[summer night]

my motherboard still has 2 sata data headers so i could just plug in 2 loose sata SSDs

[trim current]

ya

thatll work

[vital raptor]

I bought an old dell r730xd server and put it in a rack, it has 24 hotswap 2.5" bays in it 😄

[summer night]

i don't think a loose unmounted SSD is as problematic as a loose unmounted HDD

[trim current]

just make sure nothing dangles or swiings when stantionay

[vital raptor]

i also added a 4 slot NVME pcie card to it

[trim current]

aslong its not moving when stationary