🍥｜anime

[wispy canopy]

that is a weird cow xD

[frozen olive]

photoshop generative fill is crazy 😄

[wispy canopy]

yeah, that's actually pretty damn good 😮 if you didn't know what to look for, you'd just miss it

[frozen olive]

jeah alot better as inpaint.. but there alot of work still

[zinc flax]

I forgot that I genned this yesterday out of curiosity

[nova remnant]

robot chicken RAHHHHHHH

😩

[brittle jay]

I am thinking of perhaps making a custom logo for Firefox. Do you think this can be done with controlnet and if so how would you go about it?

After I "render" it with SD I can then just transform into vectors

What if I make a lora on custom Firefox logos?

[lament trench]

I would do lora on application icons and use that

there are so many high resolution high quality icons for download, should be simple

[brittle jay]

But I only want firefox

I wouldn't really gain anything from other logos

Might actually make it worse

And with like 20 images you can make a pretty good lora so I don't need that much

In any case I will probably use it together with controlnet

[near elbow]

frog!

[nova remnant]

Lmao

[wheat falcon]

Controlnet normalmap should work

[brittle jay]

How to prompt? On what kind of model?

Controlnet I kinda have an idea

[brittle jay]

Nice

[wheat falcon]

Maybe Icon, Logo of ... , Material you style,

[compact warren]

https://tenor.com/view/why-huh-but-why-gif-13199396

[lyric zealot]

https://tenor.com/view/why-not-spongebob-why-not-magic-gif-18840559

[compact warren]

[brittle jay]

So it understands what is happening in the image more and therefore the output of the controlnet will be better theoretically speaking

[compact warren]

Uh…gotcha

Yes

[brittle jay]

Bruh text2vid actually good

Baby being flushed down the toilet

(not my creation, stolen from reddit)

[zinc flax]

Hella cursed

[nova remnant]

?

[compact warren]

This is wheeze meme certified

[nova remnant]

Any tutorial to train a lora of a character?

[wispy canopy]

where can i get one of these?

inb4 someone says this is a vicious predator

[nova remnant]

YouTube is probably your best bet 🌟

hopefully you have a decent computer

or if lizzie is online you could ask her, she trains loras

[nova remnant]

The tutorials are from 3 months ago and maybe now there are better alternatives.

[zinc flax]

https://civitai.com/models/22530/guide-make-your-own-loras-easy-and-free

[nova remnant]

thats so good

such a nice style

[jovial bane]

thank you 🙂

[nova remnant]

that background came out amazing

[nova remnant]

Thanks, i will try

I am studying illustration, do you think I will have a future?

yup

you will

AIs can practically be controlled by anyone with time and they will do better than me at drawing.

[zinc flax]

Eh, AI still can't beat illustrations imo

[nova remnant]

well do you like illustration?

In fact, I myself have orders from ai for $80-300, which is more expensive than what an illustrator charges.

[nova remnant]

Ai orders

[nova remnant]

I guess so

[wispy canopy]

👁️

[nova remnant]

well if you like it then, you will have a future as long as your happy doing it

eventually youll find out what you wanna do with your art

I hope so, now with the integration of artificial intelligence into photoshop, designers may become prompts engineers.

[gloomy scroll]

[wispy canopy]

that prompt gave me some weird creatures 😄

[gloomy scroll]

I feel like every update webui becomes more and more buggy

Just few more seconds until pnginfo has read exif

[wispy canopy]

or stuff plain breaking down

[wispy canopy]

ah, you can reload your gui -> that's a crashed module

[nova remnant]

maybe you have a point , but if it ever gets to that point you could combine your illustrations and ai to make something no one else can 🤞

[gloomy scroll]

I don't want to reload ui like every hour

[wispy canopy]

open 2nd gui, just to read the png info🙂

[gloomy scroll]

I'll just do this

[wispy canopy]

that's... a scary creature 😮

it's absorbing the treee?

[pastel pollen]

https://tenor.com/view/matrix-elmo-elmo-fire-meme-matrix-code-matrix-elmo-gif-24606391

[gloomy scroll]

Tree absorbed it at the end

[wispy canopy]

maneater tree 😮

[nova remnant]

How can i do this

[near elbow]

call it "hairy" for a start 🙂 and then choose inpaint masked so it's just the region, and generate until you get lucky

[lyric zealot]

[nova remnant]

Not working

[scarlet kelp]

omg thats so cute

[lyric zealot]

Appreciate it

[nova remnant]

Help pls

[pastel pollen]

https://tenor.com/view/i-personally-never-use-this-adam-engvid-i-dont-use-this-i-refrain-using-this-gif-24162233

[scarlet kelp]

i would say its on his maximum xD does anyone have any hints for me creating loras?

like settings?

[obsidian willow]

Just reading some books.

[merry tendon]

so.. sorry if im too dumb, but how the af do you run this model with oobabooga? https://huggingface.co/TheBloke/manticore-13b-chat-pyg-GGML
ik with llama.cpp or whatever, but i get like 0.01 tokens per sec. i mean what settings
i have 16gb ram and 8gb vram + intel core i9-10850K btw if it helps

[sharp bluff]

yall have any idea why im getting this all the sudden?

just started happening constantly

oh and this?

every time it generates an image i get that

[gloomy scroll]

I heard last night that latest update broke loras but I haven't tested any myself yet

[sharp bluff]

ah ok

thats probably why haha

was it the latest webui update or some extension?

[merry tendon]

so.. sorry if im too dumb, but how the af do you run this model with oobabooga? https://huggingface.co/TheBloke/manticore-13b-chat-pyg-GGML
ik with llama.cpp or whatever, but i get like 0.01 tokens per sec. i mean what settings
i have 16gb ram and 8gb vram + intel core i9-10850K btw if it helps

[trim current]

Probs better luck in their server's help section rather than the anime section of SD lol

[pastel pollen]

new oc huh LOL

[gloomy scroll]

New?

[pastel pollen]

https://tenor.com/view/toddlers-and-tiaras-big-grin-excited-girl-giddy-smile-gif-25584914

https://tenor.com/view/konosuba-steal-kazuma-adventurer-anime-gif-17106503

[lilac totem]

Goofy ahhh

[unreal ridge]

Almost perfect hands

[compact warren]

No i claimed her first

Well secondhand claimed her

[stable crane]

is this nvidia driver update as baller as it claims?

[compact warren]

Have yet to run it

Didnt know there was one

[stable crane]

supposedly doubles inference in SD

[compact warren]

That would be much appreciated

[stable crane]

very cute

[rich tartan]

what the fk? I can't even reach 100% of my 8gb vram

ah wait, it's windows

oh

the model stays loaded in vram, that's why

[scarlet kelp]

uhhm i just made perfect settings with the batch size and so one. Needed some trys

[rich tartan]

ah, makes sense, your not using medvram. My gpu's cpu performance would die first, which never let me use till 100% vram (max 84%, afterwards it crashes my system bc of too heavy load), i ask myself, if the vram goes off first on high-end gpus nowdays

[scarlet kelp]

its a 4090 so yeah on my old it wasnt even possible

but 4090 create new generation

i´m saving now for NVIDIA L40 Ada 48GB

[rich tartan]

i would really want to feel the lightspeed too, but i think that i'll wait for the blackwell series. And i really have the need to switch to team green, my rx580 is already out of support(only about 2 years of support on rx5XX series, never gonna get amd again, but it's alr 5y old)

[rich tartan]

Ah, that would be interesting to watch, more vram but little less gpu performance

[scarlet kelp]

blackwell not before 2024

[rich tartan]

yes i have to be patient for now

i also use Blender, where rx580 is also out of support is the thing(on cycles)

[trim current]

what is that cursed imagery

[trim current]

minimalist?

[supple raptor]

<lora:EmoteMakerLora:0.8> <lora:minimalistAnimeStyle_v10:1>

[trim current]

have you tried negative add details?

[scarlet kelp]

just almost 5 hours later

[supple raptor]

at -1 it tries to add details back in. I'll try -2

[trim current]

yissss

still works okay

[supple raptor]

this one is also -1. it came out really good

[trim current]

oooo i like that

[lilac totem]

Wjat

No

Even 2070 does generate

[scarlet kelp]

your out of context xD

[lilac totem]

Like your existen~

[scarlet kelp]

wtf is wrong with u

[lilac totem]

https://tenor.com/view/avengers-endgame-marvel-thanos-disintegration-disintegrated-gif-17514639

Ur fake

[supple raptor]

-2 got it back to minimalism, but they all came out cursed.

[lilac totem]

She seems comming

[trim current]

yea....

that one is like

the

[lilac totem]

More like

The aughhh

[trim current]

pls dont include me in your cursed comments

lol

[scarlet kelp]

thats a really nice sticker, did you made it?

[trim current]

ye with emote maker lora

and my lora

[scarlet kelp]

nice, which emote maker

[trim current]

https://civitai.com/models/53294/twitch-emote-maker-lora

[supple raptor]

if that gives you sexy time vibes then you might want to reevaluate your techniques.

[trim current]

and my unpublished lora trained on my mj selfies to help retain consistency

[scarlet kelp]

ahh the one which dont work for mine

[trim current]

its a hit and miss definitely

as with the other emote makers

[scarlet kelp]

i´m using the other emote maker and another lora which gives me really nice outputs

[trim current]

aslong youre getting what you want all that matters

[scarlet kelp]

true

just a example from this second

[trim current]

cursed

was going into pow unsure, ended up really liking what im upscaling

[supple raptor]

sadge because they keep getting called cursed

[trim current]

not that one however

i like that one

random red eye rectangle but it rlly nice

[supple raptor]

yeh, if i cared more I would manually edit that pink rectangle thing over the left eye out. maybe the pink hairclip and dot in the middle of her bangs too.

[trim current]

She got pink eye is k

alright this time i wont crank x4 on a 1280,

cos it crashed my linux

[pastel pollen]

bro do you know how to fix it ?

[supple raptor]

I haven't used the additional networks extension in like 2 months. I'd recommend just using the built in LoRA implementation along with this extension (check the readme because alternative LoRA types need to go in their own folder): https://github.com/KohakuBlueleaf/a1111-sd-webui-lycoris.git

[pastel pollen]

https://tenor.com/view/thank-you-gracias-ty-thnx-takk-gif-26160678

[trim current]

very nice

still doing add details negatievly?

and just finished my pow, really digging it lol

[supple raptor]

no, it was making the style worse most of the time. I don't think it works very well with the minimalist Lora.

[trim current]

ahh make sense

[supple raptor]

nice submission

I might have to actually vote this week.

[trim current]

was done with darksushi this time around

And add details to give more humanoid-sculpting

Cos without it looked too stiff

[supple raptor]

I'll have give darksushi a try. I've seen it around but haven't used it yet. Did you use the regular or 2.5d version?

[trim current]

I use the darker version on civit

[sharp bluff]

anyone know how i could get a wolf cut?

[pastel pollen]

wolf cut hairstyle

maybe

[sharp bluff]

see i tried and it started adding wolves in the background 💀

[trim current]

Try mixing shag and mullet idk

[supple raptor]

Just had to look up what wolf cut is and it seems to be pretty varied. maybe try running an example that you like through wd tagger to see what it gives for suggested tags.

[stable crane]

yo these new drivers are popping

14s for a gen now

[supple raptor]

You really gonna make me switch off of studio drivers

[stable crane]

it does feel 2x as fast

I thought I'd have to pull latest pytorch or smth

maybe it'd be even faster if I weren't so lazy

[supple raptor]

which gpu do you have?

[stable crane]

3080

certain situations seem to slow it down again

maybe my card is throttling

[supple raptor]

Guess I'll update then. Just ran a benchmark in system info so I should be able to see if it actually speeds things up for me.

[compact warren]

Hires or normal

[stable crane]

hires

[compact warren]

Hmm

[stable crane]

I think it slows down with number of loras

[trim current]

cant wait for my 3060 12gb

[compact warren]

Isnt that normal

[supple raptor]

Yeah, multiple LoRAs have always slowed down gens.

[trim current]

@supple raptor what will be limitations with that card?

[stable crane]

TIL

yeah it's so fast without any loras lmao

[compact warren]

I run on 3070, which i had heard was best for inference time

So im curious to see how much better it can get

[stable crane]

https://blogs.nvidia.com/wp-content/uploads/2023/05/Stable_Diffusion_Performance-532.03_Driver.png

"twice"

https://blogs.nvidia.com/blog/2023/05/23/microsoft-build-nvidia-ai-windows-rtx/

[supple raptor]

Only thing you won't really be able to do as far as training is actual finetuning. Technically you could do that, but from what I've been told it isn't worth it unless you can finetune at fp32. As far as generation resolutions, I've ran 512x2048 with 2x upscaling in Hires fix on 12GB without tiled VAE.

[compact warren]

Pog

[trim current]

finetuning is what exactly?

model training?

[supple raptor]

"real" model training. Most of what we do (including LoRA training) is some variation of Dreambooth training.

[trim current]

mm, thats ok i think lora training is are far im willing to get into

was tryna make a statue with comfyui

ended up making a god

[cyan burrow]

a 4 fingered** god

[supple raptor]

god of kindness needs no middle finger.

[trim current]

man im so hyped ps5 showcase

[trim current]

can kick your butt anyday

[trim current]

i like that

[supple raptor]

I got about 1it/s speed increase by switching from the current studio driver to the new game ready driver.

[rich tartan]

thats

[sly rune]

There's some new FuzzyMixes! In addition to FuzzyHazel and FuzzyAlmond, there is now FuzzyLime, FuzzyYuzu, and FuzzyEclair. https://huggingface.co/Lucetepolis/FuzzyHazel

[grave snow]

maybe I should stop spamming 💀

[sullen hollow]

lool i love this style what model did u use ?

[grave snow]

https://civitai.com/models/35893?modelVersionId=78228

the aesthetic there came from ((cozy upscale synthwave cassette retrofuturist potted plantpunk)) in prompt (or that's the rough syntax equivalent in A1111 anyway)

[unreal ridge]

This prompt is insane ❤️

[grave snow]

back to spamming until I'm dragged away to download outlast trials

[stuck aurora]

Nice. What model did you use?

[grave snow]

https://civitai.com/models/35893?modelVersionId=78228

for reference the prompt there (converted) was absurdres highres portrait of a ((sharknadoism)) 1girl and neg bad-artist (low quality, worst quality:1.3), negative_hand

[supple raptor]

Thanks! I really like that one too. Heres a couple more from it

^without loras

[unreal ridge]

I have no Idea of what loras you used in the previous one, but even without them I get some pretty good stuff

[supple raptor]

same mix I've used on most of my recent stuff <lora:NAI-bluethebone:0.3> <lora:WhiteMelancholy:0.45> <lora:loraslayernoiseoffset:0.45> <lyco:lucky768-2:0.45>

[unreal ridge]

Suddenly undertale

[supple raptor]

https://tenor.com/view/sans-glowing-eyes-sanes-sans-eye-sans-power-gif-24694251

when the sword is taller than you it needs a handle in the middle of the blade so you can lean on it

[untold glacier]

It's heeeeeeeeeeeere

Time for testingggggggggg

[supple raptor]

It's not nearly as exciting as it sounds. We don't use the "Olive optimized version of Stable Diffusion".

[untold glacier]

Hmm, did I misread? It said with Auto's build but I guess I missed an important step.

Ah, ok.

[untold glacier]

Yeah a look at the steps they give quickly shows me that it's beyond my knowledge to implement.

[supple raptor]

It did give me a small increase in speed for normal usage, but it's less than 10% and nowhere near the 2x claim.

[unreal ridge]

It's not giving me the black background

But I'm not complaining

[supple raptor]

I've also started getting some unexpected NaN errors since updating my driver that I've never gotten before.

[supple raptor]

The model I'm using has noise offset mixed into it. If yours doesn't you could use a LoRA for that.

[unreal ridge]

Not sure of what it does

Tried one time before but never noticed any difference

[supple raptor]

noise_offset is a setting that can be used during training. It gives the model a wider range of average brightness levels that it can produce. Without it, SD tends to produce images very close to 50% average brightness. Basically, it allows images to be brighter or darker. This one is my favorite. https://huggingface.co/closertodeath/ctdlora/blob/main/loraslayernoiseoffset.safetensors

[unreal ridge]

I'll give it a try :D

[unreal ridge]

[unreal ridge]

Tried changing models

[keen hollow]

evenin'

anyone around?

[unreal ridge]

This channel is pretty much dead around this time

😔

[keen hollow]

I know, and it's the time I'm around. 😛

[trim current]

How to best performance with directml.
Step 1: you dont you go to linux

[unreal ridge]

Guess using it at <:1> was not a good idea

https://tenor.com/view/aoi-tense-scared-priconne-princess-connect-gif-24633357

[supple raptor]

Yeah, that one is strong. I use it around 0.3-0.4 usually. Some models can take it a little bit higher.

[pastel pollen]

https://tenor.com/view/coffee-good-morning-morning-patrick-patrick-star-gif-23886631

[pastel pollen]

wow the old one

[keen hollow]

love it! love halloweeny characters

[severe lodge]

wow that's beautiful. If this was a movie poster I'd watch it for sure XD

[pastel pollen]

https://tenor.com/view/toddlers-and-tiaras-big-grin-excited-girl-giddy-smile-gif-25584914

[severe lodge]

LMAO WHAT DOES THAT MEAN

[pastel pollen]

the old prompts always give cool image

[unreal ridge]

🤝
I'm still recycling prompts that I used on Novel AI a year ago 💀

[pastel pollen]

yup always save the old one for future

[runic vault]

things are somewhat stagnant yes

[pastel pollen]

lets trade old prompts

1girl, bag, bow, cape, ((water paint)), (watercolor painting), (water acrylic smear effect), (smearing), dynamic composition, motion, ultra - detailed, incredibly detailed, a lot of details, (amazing fine details and brush strokes), ((water_painting_art)), (with gradients), surrealistic, dystopia, (depth of field), mist, ((rich cinematic atmosphere)), (water_brushstyle), frills, ((texture hair)), watercolor oil painting, relaxing, calm and mysterious vibes, green_hair, hat, hat_bow, holding, gradient, (Superdeformation art:1.35), (Chibi Art Infusion:1.25), (Digital Illustration:1.20), (soft_hair), (fluffy hair), fur, disheveled hair, furry hair, floofy hair, (poofy hair), long_hair, long_sleeves, gloves, looking_at_viewer, ribbon, ((photo realistic)), (1girl:1.2), cute, (full body:1.2), white_eyelashes, ((chibi)), (cute), simple_background, thighhighs, wand, witch, witch_hat, (masterpiece:1.2), (best quality:1.2), (high quality:1.2), detailed, 8k, 4k, (extremely detailed), (highres:1.2), (detailed face:1.2), (sharp detail:1.2), (volumetric lighting:1.2), ((soft lighting))

[unreal ridge]

Most of my old prompts were prompt mixing

it's basically a shit ton of prompts merged together

[pastel pollen]

same

[unreal ridge]

hold on

gotta find my folder

[unreal ridge]

Masterpiece, High Quality/High Definition, extremely detailed CG unity 8k wallpaper, intricate details, Illustration, detailed scenery, detailed eyes, beautiful detailed, perfect anatomy, finely detailed, highly detailed, ultra-detailed silver hair, perfect anatomy, light blue magical eyes, cool aura, long hair, dynamic effects, space, lunar flare, magical circle, spirit, dynamic angle, medium breasts, standing, wind, abyssal atmosphere, full body, high resolution illustration, beautiful detailed hair, bloom, Extremely detailed|silver hair, long hair, perfect anatomy, light blue magical eyes, dynamic angle, lovely and Aesthetic, detailed eyes, 1girl, detailed scenery, solo, wind, , medium breasts, standing, full body, space, lunar flare, Brilliant vivid Colors, cool aura, [dark intense shadows], focus character, Detailed Cute Face, Depth of field, Breathtaking, Character Enhance, Beautifully Detailed face

Probably have a lot of repeated prompts since I was just copying parts of different prompts and fusing everything together.

This was the result of it

[pastel pollen]

https://tenor.com/view/my-girl-seal-friendship-its-a-deal-seal-spit-gif-17005706

[unreal ridge]

probably will give some really generic shit

[unreal ridge]

like 80% of the prompts I tried

recycling

[pastel pollen]

no prob share the old prompts is good deal

[unreal ridge]

got this from yours

[pastel pollen]

https://tenor.com/view/smile-laugh-cute-gif-20071688

[unreal ridge]

shit

making marionettes is hard

my model have no clue of wtf that is

[pastel pollen]

go this

[nova remnant]

I'm having so much fun with the webui

[pastel pollen]

nice one

[unreal ridge]

I'll se if I can find something else

[pastel pollen]

https://tenor.com/view/you-learn-very-fast-benjamin-skinner-rdcworld1-youre-a-fast-learner-you-learn-quickly-gif-26337313

[nova remnant]

thank :)) its cuz of you after all

[nova remnant]

https://tenor.com/view/cat-cat-jumping-cat-excited-excited-dance-gif-19354605

[pastel pollen]

https://tenor.com/view/your-welcome-gif-23991412

[unreal ridge]

Not mine. Got from some dude that posted it, still have it saved for some reason. Just look at the metadata

[pastel pollen]

neon effect is always good

[nova remnant]

how do you get nice generated girls.. I generated too many boys and now sd hates generating women

[unreal ridge]

💀

[pastel pollen]

https://tenor.com/view/lmao-spit-take-cracking-up-haha-so-funny-omg-gif-17921247

[nova remnant]

LMAOOOO

[pastel pollen]

just try prompt from my server in meta room

[nova remnant]

this one looks okay........for once 💀

[pastel pollen]

give me tomboy vibe LOL

[nova remnant]

gothic tomboy

[unreal ridge]

oh god

[nova remnant]

this model looks amazing, which one is it

[unreal ridge]

merge of 2 of my merges

[nova remnant]

nvm idk how to merge

[pastel pollen]

https://tenor.com/view/lol-emote-gif-24165461

[nova remnant]

my laptop will blow up if I do

[supple raptor]

Merging is far less hardware intensive than generating.

[unreal ridge]

merge of this one

and this one

for some reason the result was decent

[nova remnant]

woahhh the second one

[nova remnant]

third one reminds me of female gojo

[pastel pollen]

https://tenor.com/view/you-right-right-correct-gotcha-kevin-hart-gif-18115020

is from gojo

[nova remnant]

oh lord now I'm gonna generate female gojo all day

[nova remnant]

give her the black round glasses and I’ll be complete

[supple raptor]

lol adetailer identified the feather to the right of her hands as a face.

[unreal ridge]

https://civitai.com/models/60832/futuristicbot4 I love this lora

[supple raptor]

@pale monolith would probably like this one

[pale monolith]

you underestimate my power

[unreal ridge]

lol

[pale monolith]

also i need to retag this so i can search it for mechanical

so ty for reminding me it exists

xD

[supple raptor]

Or... does that just mean I guessed correctly?

[unreal ridge]

Anyone have suggestions for marionette prompts? The only decent one I got after like... 20 attempts was this one

And I still got the random ass dolls

[pale monolith]

like... a small marionette doll?

[unreal ridge]

a big marionette doll

[pale monolith]

https://civitai.com/models/21365/marionette-lycoris

there's the strings

then doll joints in prompt, and expressionless, and porcelain skin

now i want to make marionette

[unreal ridge]

(marionette:1.3), marionette girl, (dark and eerie atmosphere), porcelain skin, (hollow eyes), (strings attached:1.2), strings, intricate lace dress, tattered ribbons, delicate stitches, marionette joints, expressive hands, wooden puppet joints,
Used this on this one

[unreal ridge]

can you give this one a try and see what you get?

[runic vault]

been there done that

[pale monolith]

👌

[unreal ridge]

The preview images made me laugh

[supple raptor]

the marionette lora may have been a little too much to throw on top of the other 4 lora's i'm using.

[unreal ridge]

Seems like the little dolls are unavoidable lol

pretty good result tho

I just want to get the joints consistently, I don't really care about the strings

[pale monolith]

(doll joints:1.3)

and... i think there is a doll joint lora somewhere

idr where

[supple raptor]

i think it might be on the gitgud link

[unreal ridge]

I'll give it a try

[unreal ridge]

still no

[unreal ridge]

nvm it works

[unreal ridge]

I'mma go to sleep

https://tenor.com/view/sleep-time-gif-25989017

[supple raptor]

https://cdn.discordapp.com/emojis/1107715334489788426.webp?size=64

[near elbow]

maybe you need to use sketch with inpaint and draw a skin-color knee there, can be ugly as hell, to guide the AI. that's an unusual position for a knee in such an image, so it's not the natural place the AI will try to put it in

[wispy canopy]

using lyrics?

[storm ivy]

lyrics?

[wispy canopy]

song lyris

but i did look at the png, no lyrics involved 😄

[wispy canopy]

current pow submission, but fits anime too ^^

[wispy canopy]

how many lora's is that XD

"yes"

[crimson stirrup]

one actually

[wispy canopy]

bocchi?

[crimson stirrup]

yea

[storm ivy]

i dont remember this episode in the show

[storm ivy]

scathach?

[runic vault]

yes

not too hard to do even without a char lora

if you want a easy to inspect the prompt the project ai guys (waifu diffsuion devs) do have a prompt inspector bot

[storm ivy]

will look into that

[lyric zealot]

omg, posted wrong place

[magic wave]

Soosi

[lyric zealot]

finally updated my webui, and it's so much better now

[stable nimbus]

thoughts on sdxl?

[magic wave]

Oooo, nice :O

[gloomy scroll]

Nice pow image?

[lyric zealot]

[gloomy scroll]

You mean more buggy?

[lyric zealot]

is it more buggy?

I'm liking it overall

but it's probably because i update it by making a new clean installation

[gloomy scroll]

Getting closer

[lyric zealot]

how does this adetailer works?

[desert arch]

Bribery

[gloomy scroll]

It's just auto inpaint for lazy people

[lyric zealot]

so if i select that hand stuff, it's gonna work like magic?

[gloomy scroll]

Well it will make hands somewhat better but it wont fix missing fingers and deformed hands

[vale temple]

No, faces generally get better but I haven’t seen much improvement with hands honesty

[lyric zealot]

[gloomy scroll]

Maybe if you but high denoise and some specific prompt for it then maybe

[lyric zealot]

before vs after

[vale temple]

The face on that was ok to begin with though

[lyric zealot]

yup yup

[vale temple]

It’s more for janky pupils & skewed features

[lyric zealot]

[vale temple]

For instance (& this isn’t that bad either) something like this

Adetailed would go in and refine the eyes

[lyric zealot]

before vs after

it's definitely good for cases like this

[vale temple]

Again, these faces are fine to begin with

[lyric zealot]

well, true, but it can be better

[vale temple]

This would have been improved I’m pretty sure

[lyric zealot]

can you show the after for this

[vale temple]

I can try to later but not atm

[lyric zealot]

noted noted

if you ever do, ping me

[desert arch]

Evening chat

[gloomy scroll]

Miku

[desert arch]

What the miku doing

[merry tendon]

hey

i finally got a 13 b model to run on 7gb while still being super fast

like 3.50 tokens per sec, not the best, but its good

[vale temple]

adetailer on the right

[lyric zealot]

woah

a lot better

[vale temple]

yeah, but as mentioned you results were already past this point so not need 😉

[lyric zealot]

yup yup

you aren't using hires fix right?

[vale temple]

nope

[stuck aurora]

Hi. When using Highres, how many Highres Steps is recommendable when the Sampling Steps is 30?

[merry tendon]

can someone rate my oc(you will need a silly tavern or tavern ai instance)?

[gloomy scroll]

I use 8 steps for highres

[stuck aurora]

oh, is it lower the better?

[magic wave]

I use 10 steps

[gloomy scroll]

It's more like you don't rly need any more

[magic wave]

Yeah

DPM 2M Karras and samplers like it I use 10

[stuck aurora]

I've been doing 20steps.

[magic wave]

Samplers like SDE, 5 work

2S a (not karras) is wonky on 5 and needs 10 tho

2S a is a weird boi

[gloomy scroll]

Ofc if you have GOOD gpu then it probably doesn't rly matter just depends how long you want to wait per image

[stuck aurora]

How about Denoising?

[magic wave]

I mean, yeah, but it is a waste of time

[magic wave]

I do 0.5

[gloomy scroll]

.5 for me

[stuck aurora]

Thank you!!

[gloomy scroll]

@stuck aurora what is that char is your name?

[stuck aurora]

exclamation mark and a bunch of spaces

[magic wave]

O

[gloomy scroll]

I'm confused since discord shows it differently in when I click your name

[magic wave]

I thought it would have been some invisible ascii character :o

[stuck aurora]

yes, it's weird. but good for hoisting

[merry tendon]

can someone rate my oc(you will need a silly tavern or tavern ai instance)?

[magic wave]

Hoist?

[stuck aurora]

yes, like it brings you at the top of the members list.

[magic wave]

Ohhh

Fair then

[gloomy scroll]

AI: You wanted white hair... Black it is then.

[lyric zealot]

pink hair

[stuck aurora]

Hi, how do you get more detailed result? May I see the info, please? cuz mine is like this.

[lyric zealot]

Helloo, can you check if there's pnginfo inside?

[lyric zealot]

good generation tho

[stuck aurora]

No pnginfo

[lyric zealot]

oh right, i forgot i updated my webui, and removed that extension

[magic wave]

Stealth? Isn't it not needed anymore tho? :o

[lyric zealot]

why the fuck is my pnginfo not working

[lyric zealot]

yup, but i kept it cos it allows copying pasting into webui

[lyric zealot]

1girl, pink hair, twin braids, bangs, yellow detailed eyes, japanese street wear, print shirt,
Negative prompt: (low quality, worst quality:1.4), (bad anatomy), extra digit, fewer digits, (extra arms:1.2), bad hands, by (bad-artist:0.6), bad-image-v2-39000, censored, monochrome, greyscale, nsfw, nude, loli,
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 22, Seed: 3757262444, Size: 512x768, Model hash: 4d651c7638, Denoising strength: 0.53, Clip skip: 2, Version: v1.2.1, ADetailer model: face_yolov8n.pt, ADetailer conf: 30, ADetailer dilate/erode: 32, ADetailer mask blur: 4, ADetailer denoising strength: 0.4, ADetailer inpaint full: True, ADetailer inpaint padding: 0, ADetailer version: 23.5.18, Hires upscale: 2, Hires steps: 10, Hires upscaler: 4x_fatal_Anime_500000_G, Dynamic thresholding enabled: True, Mimic scale: 7, Threshold percentile: 96, Mimic mode: Half Cosine Up, Mimic scale minimum: 4, CFG mode: Half Cosine Up, CFG scale minimum: 4```

not the same seed, but should be same prompt

[stuck aurora]

Thank you!!

[lyric zealot]

you are right, my pnginfo is broken

[merry tendon]

can someone rate my oc(you will need a silly tavern or tavern ai instance)?

[magic wave]

Ohhh fair then :o

Smol thigh gang

[lyric zealot]

[graceful aurora]

Kurokawa Akane

[merry tendon]

can someone rate my oc(you will need a silly tavern or tavern ai instance)?

come on guys

[lyric zealot]

sorry, i don't have any of those ai instances so yeah

[merry tendon]

[brittle jay]

Very nice

[glacial bane]

I need assistance

[merry tendon]

at your service my friend!

what do you need?

[frozen olive]

GWAAAAAAA

[supple raptor]

I just got SillyTavern installed, give me a little bit to figure out how it works and I'll try out your OC.

[merry tendon]

man you gotta more deserve love

thanks

lemme know when i can send you the link for the char

[frozen olive]

i dont understand

[supple raptor]

Chat bot character.

[frozen olive]

is that something new

[merry tendon]

no

[frozen olive]

ah just dont mind me im not in all the ai stuff is so much for my brain

xD

[unreal ridge]

If it takes you too long you can always pick chatgpt API and jailbreak it.

It's a lot easier

[supple raptor]

I've been using Oobabooga with local LLMs for a while now and have used TavernAI a bit. I just haven't used SillyTavern so far.

[magic wave]

I trained Nijijourney style (i think)

But I may re-train with noise offset

[unreal ridge]

Silly is basically a better version of tavern

[magic wave]

HOLY SHIT HE FAST

MeguBro is a certified Megumin sniper

[unreal ridge]

Cultured 👌

I'm about to create a bot on silly tavern that generates me prompts cuz I'm too stupid to use my brain for something.

https://tenor.com/view/arknights-surtr-gif-20942897

[magic wave]

Fair XD

[unreal ridge]

🎈

[magic wave]

Think my LoRA works well tho

[merry tendon]

looks like my model when i do (realistic:1.1)

[magic wave]

I trained it on pics like these, I think I need offset tho

[merry tendon]

can i send it?

[supple raptor]

Yeah I think I got it figured out now.

[merry tendon]

i better send it you as a DM tho

[supple raptor]

That's fine. My DMs are open.

[merry tendon]

not nsfw, just wanna be sure

[magic wave]

I like the face style and shading

[trim current]

Looks nice

[magic wave]

Yeee, gotta test more though, and leave offset ver training while I sleep

Its missing a bit of the bright edges

[unreal ridge]

[merry tendon]

gimme that char

[unreal ridge]

NO

I'm still creating it

it still gives those long ass descriptions that gpt gives you whenever you ask for something

I hate that shit

I just want a normal/simple looking prompt

[unreal ridge]

not bad

[magic wave]

My femboy has boobs

[molten shadow]

THIS IS NOT MINE. But do you know what model or vae can make the picture so shining like this

[unreal ridge]

If there's no metadata there's no way of knowing for sure

[lyric zealot]

not femboy anymore

[magic wave]

Seems like some sort of pastelmix mix

[molten shadow]

if only it has 😭

[magic wave]

Tru

[unreal ridge]

But yeah, It looks really fucking good

[merry tendon]

what model and (if you use one) what lora?

[runic vault]

based66 and the LowRA lora for darker images

[molten shadow]

I tried my merge with pastelmix and it somehow more shining 😂

[unreal ridge]

Hime cut + red spider lily never gets old

[runic vault]

i guess that trope was started with jigoku shoujo

[unreal ridge]

Hime cut is probably my favorite hair style

[molten shadow]

ye, hime cut naver die. Some of coworker in my company have that and I swear that so attractive

[unreal ridge]

Hime cut or bob cut makes me weak af

Probably the most normal looking "sword"

[pastel pollen]

https://tenor.com/view/sml-jeffy-whats-on-the-menu-today-whats-on-the-menu-whats-cooking-gif-24926848

[pastel pollen]

https://tenor.com/view/hmm-dot-dot-dot-stick-figure-intresting-gif-23376515

i need something darker

[lyric zealot]

[pastel pollen]

https://tenor.com/view/star-trek-something-is-missing-kirk-generations-gif-22127413

[lyric zealot]

?

[pastel pollen]

https://tenor.com/view/dumb-and-dumber-dumbanddumber-gif-19488170

so hard to get blood

[pastel pollen]

is there is lora for blood effect?

[lyric zealot]

not sure

[merry tendon]

?

when are you gonna show me some results?

[supple raptor]

Sorry, I got distracted playing with the group chat feature. Let me make a new one with fewer characters. This one with 4 characters isn't going to work well in screenshots. She's been good so far.

[merry tendon]

yep

[magic wave]

Green lips do green skin

[merry tendon]

and?

[magic wave]

[unreal ridge]

bro...

[merry tendon]

wow, pretty good, what model do you use?

[supple raptor]

gpt4-x-vicuna-13b

[merry tendon]

can you make a screenshot of all your generation settings?

/screenshots

[pastel pollen]

i just watch saw and this the result

[supple raptor]

I just use the NovelAI (Pleasing Results) config with response length turned up to 400

[merry tendon]

can you send a sceenshot of the A button on the top?

[merry tendon]

thx

[supple raptor]

Only change was setting tokenizer to LLaMa

Oh, and presets to Vicuna 1.0

[merry tendon]

you didnt enable it tho

[pastel pollen]

ok last one

[supple raptor]

Yeah, I've never figured out how to get instruct mode to actually work well for chat characters.

[zealous arch]

something I think is worth sharing.

[merry tendon]

you got some more results?

[merry tendon]

thank you so much bro ❤️ what do you think of her?

[supple raptor]

She's fun. I might steal some parts of her prompt for my 2B prompt.

[merry tendon]

yep, you can do that ofc. Thanks for testing

[supple raptor]

My pleasure! Gave me a reason to finish installing SillyTavern.

I should have done that a long time ago. The group chat mode alone makes it worth it.

[unreal ridge]

Yeah, group chat surprised me a lot the first time I tried it

Way better than CharAI group chat

[supple raptor]

I was expecting a lot more bleed through between characters, or responding as the wrong character, but I've only seen it happen a couple of times so far and just regenerating the message has fixed it each time.

[unreal ridge]

I just whish there was a way to pick who's going to reply next without having to type the name or something.

[merry tendon]

yep, works pretty good

[supple raptor]

Yeah, that would be a cool feature. It does surprisingly well at deciding who should reply and which parts of the message they should reply to when you do use their names though, even if you mention multiple characters at once.

[unreal ridge]

I really wanna try the "Character Expressions" extension

haven't figured out how it works yet

seems pretty nice

[supple raptor]

It looks like it would only work in one on one mode. But I think you can just set the images you want to use for each expression in here.

[merry tendon]

wow, they both work pretty good(i made both characters)

[supple raptor]

Wait, no. I just looked at the README on Github and you can put the images in a folder so that different characters can have different images.

[unreal ridge]

which folder?

should I put the folder in

[supple raptor]

1. Create a folder in TavernAI called public/characters/<name>, where <name> is the name of your character.
2. For the base emotion classification model, put six PNG files there with the following names: joy.png, anger.png, fear.png, love.png, sadness.png, surprise.png. Other models may provide other options.
3. Images only display in desktop mode.

[unreal ridge]

I'll give it a try later, ty 👍

[supple raptor]

Here's a tool that can help remove backgrounds from your images which could be helpful for making the expressions images https://huggingface.co/spaces/skytnt/anime-remove-background

[lyric zealot]

are yall using sillytavern?

[merry tendon]

yea

[supple raptor]

yiss

[lyric zealot]

sounds interesting

[merry tendon]

ive been using it for a while now

[unreal ridge]

I'm pretty sure there's an extension on SD that does the same thing

forgot which one it was

[merry tendon]

way better than tavernai

[wispy canopy]

sillytavern way better

i hate having used any of the 30b models on sillytavern 😦

[lyric zealot]

will have to try that

[supple raptor]

Yep, there's two of them. One is exactly the same as that HF space, and the other uses a segmentation model. The segmentation one works better, but I like using the one on HF because it means I don't have to load a different model on my GPU.

[unreal ridge]

Please do

I'm using mostly to create QoL bots for myself

I'll probably start creating actual characters later

[supple raptor]

I had tried out TavernAI before and didn't really see the point. It looked a bit nicer than chatting in the Ooba WebUI but other than appearance I couldn't really see the benefit.

[wispy canopy]

conversational agents to help you out with various tasks?

i should give some of my ooba characters a whirl in silly, perhaps they'll feel better that way

llm is my sd propmting bot though 😄

with my creativity of a goldfish, that's a godsend

[unreal ridge]

@wispy canopy Laziness be like

[merry tendon]

when do ya guys think gpt4 quality models will be available for consumer hardware(8gb vram)?

[unreal ridge]

iirc Novel AI released a 3B model that is pretty fucking good

[merry tendon]

where can you download it?

[wispy canopy]

different prompting style ^^ (natural language)

[untold glacier]

Novel AI stuff is proprietary through their paid service. You're not going to be able to download it.

[merry tendon]

bruh

[unreal ridge]

8K context

[merry tendon]

i wanna have it

[supple raptor]

I think for consumer hardware we'll see more specialized models that can reach gpt4 levels of quality in their specific tasks, rather than generalized models. Maybe the 1bit quantization talks will actually work out and prove me wrong though.

[merry tendon]

yea, but WHEN do you think it will be?

[wispy canopy]

wouldn't 1 bit quantization just leave too little accuracy?

[unreal ridge]

Judging by how fast everything AI related is advancing, probably not that far from now

[merry tendon]

but HOW far

[wispy canopy]

although, massive increase in parameter counts could help fix that

[unreal ridge]

no fucking clue

[supple raptor]

That'll depend on us getting trainers that are more reasonable to use on at least high-ish end consumer hardware. I haven't followed LLM training news so I have no idea if anything is on the horizon for that.

[merry tendon]

👍 i hope soon

[wispy canopy]

afaik, you still need HUGE vram for training

[supple raptor]

From what I've seen people saying, it's much more accurate than you'd expect but I haven't seen any actual evidence.

[wispy canopy]

well, i suppose a quantized 4bit 30b model outperforms a non-quantized 13b model quite hard, so i suppose a quantized 1bit 65b model should be able to slap a 4b 30b model

[supple raptor]

Yep, pretty much all of the training is being done on one or more A100s or H100s to my knowledge and takes a fair bit of training time which makes leasing hardware less practical than it is for SD training.

[wispy canopy]

that's just the difference between 900m parameters or 30000m parameters xD

[supple raptor]

I think LoRA training can be done on high end consumer hardware like the 3090/4090 though.

[wispy canopy]

never tried the llm loras, they take up more memory, right?

[supple raptor]

I'd assume so, the SD ones do.

I haven't tried them either.

[brittle jay]

wait fr

[merry tendon]

yes

[brittle jay]

only used tavern or just normal webui (ooba / kobold)

[unreal ridge]

much better

[wispy canopy]

silly just has some really good QoL features which seem kinda missing on regular tavern

[brittle jay]

oh

[brittle jay]

lit

[brittle jay]

yo I should use this for my game

though it will make it less like my original idea of having everything procedurally generated

[untold glacier]

No idea why people are liking my image so much lol. I've posted it twice before and I find it useful as a meme response, but I guess I should've made the intent clear.

[unreal ridge]

It looks good tho

[untold glacier]

Heck, it's back from the days of when NovelAI just released their implement.

[brittle jay]

this is literally akane lol

[unreal ridge]

First time I saw it

[brittle jay]

or us using internet

or idk

[supple raptor]

If you can figure out a way to implement it into whichever engine you're using then it would probably be a more reliable experience than having it generate the images, and would also significantly reduce the barrier of entry since the user wouldn't need to be able to run an LLM and SD at the same time.

[brittle jay]

should be very easy

if you can implement in python

I can implement in engine

renpy too good

[supple raptor]

👍

[brittle jay]

problem is that it won't be completely procedurally generated

like the characters won't be to do whatever they want (image-wise)

[supple raptor]

Yeah, you might want to look more into it though. Inside SillyTavern it lists 28 different expressions that you can assign so I'm guessing it's probably more flexible than the SillyTavern README implies. It still wouldn't be the same having them be generated on the fly, but it would eliminate the possibility of it generating poor quality images that don't fit the scene anyways.

[wet elm]

am i the only one whos getting CUDA memory error codes but wasn't 2 days ago? using the same parameters

[unreal ridge]

Was trying to decide on a profile pic for the bot myself, but I thought this would be more interesting.

DoEs ThAt mEaN ChAtGpT iS aN aNimE gIrL?

[lyric zealot]

might be a variety of reasons. Could be 2 days ago, you weren't running other stuff, and now you are

could be memory leak

[supple raptor]

Check to see if something else is using your VRAM. Browsers and Discord can each eat up 1GB or more easily if you have hardware acceleration turned on (which it is by default normally).

[lyric zealot]

try going into your task manager, do under Details and look at Dedicated GPU Memory

[supple raptor]

Nvidia Broadcast will also use a bunch of vram if you have any of the effects turned on. You can easily turn those on and off by right clicking the icon in your task bar and clicking the device you want to turn them off for.

[brittle jay]

guys what was that website to resize images

[supple raptor]

https://www.birme.net/
I don't like it though. It makes images blurry.

[brittle jay]

oh really

[wet elm]

gonna try closing all other apps

[brittle jay]

thanks anyways

[slate night]

So I updated a bunch of stuff and controlnet and adetailer are gone? nvm restarting everything fixed it

[wet elm]

didnt work lol

[lyric zealot]

look at your task manager > Details > Dedicated GPU memory

[brittle jay]

wallpaper engine taking up half a gig

[supple raptor]

Yep, that's why I quit using it.

[brittle jay]

yes

ask veggie

he professional wallpaper engineer

[supple raptor]

It's a very kid friendly community. Only like 3/4 of the wallpapers on there range from lewd to straight up porn XD

[brittle jay]

only

[supple raptor]

[wet elm]

what do i do there?

[lyric zealot]

take a look and see how much vram you are using