🍥｜anime

[steady grail]

i worked on confyui highres fix for 3 months before understanding how it works

i found so many funny tricks lol

[dawn jasper]

Who is this lovely lady?
Also @vestal crypt neat pic.

[wispy canopy]

just someone who exists because of song lyrics

[wispy canopy]

single pass render, had to give up lcm lora tho

[supple raptor]

bruh.... have you considered not living on the surface of the sun

[native halo]

coldest day in Texas

[wispy canopy]

"coldest day yet"

wait until international news picks up on new power failures there...

pixelarted another output of that prompt. amazing stuff

[wispy canopy]

10 steps, single pass 1824x1024. 20 seconds per image
(sdxl model + FreeU + LCM Lora + Kohya Highres Fix). Can highly recommend

[nova remnant]

my reaction when people think 36 is that bad

[supple raptor]

it is if you're waking up to it being that hot already

[native halo]

been chillin here at 40c 🛌

[nova remnant]

@native halo what about 40°c with her

[native halo]

too many clothes for this weather imo

[terse jacinth]

that's hot, makes me horny~

(the nerd talk, not the pics... 😛 )

[native halo]

balls

[wispy canopy]

"it can all be yours, for the special price of..."

[terse jacinth]

https://tenor.com/view/cat-fail-cat-fail-cheese-challenge-cheese-gif-17748665

you had my interest, now you have my attention ~

[wispy canopy]

https://github.com/ljleb/sd-webui-freeu (freeu)
https://civitai.com/models/195519?modelVersionId=219782 (sxdl)
https://civitai.com/models/195519?modelVersionId=225222 (sd1.5)
https://github.com/wcde/sd-webui-kohya-hiresfix (kohya highres fix)

[terse jacinth]

should give me this in DMs so i can pin it like it's hot

about to head to work 😦

[nova remnant]

[native halo]

those clothes are not good for this weather

[nova remnant]

when i put something thru im2im is there a way to upscale it?

[wispy canopy]

gib prompt!

[nova remnant]

Its not what it looks like

[wispy canopy]

i'll just interrogate it then

[supple raptor]

There are a few options. The simplest is to just increase the dimensions to the new dimension you want (will take some math to make sure you maintain the same aspect ratio). If you are just wanting to upscale without changing the image too much then I would suggest either tiled diffusion or sd upscale script.

[nova remnant]

i keep getting faces generated in the tiles tho lol

[native halo]

reduce denoising

[nova remnant]

no then it looks too low quality like the original

[supple raptor]

that means that your denoising is too high. you should also remove all mentions of the character from your prompt.

[nova remnant]

ohhhh

[native halo]

u have to play with denoise

[nova remnant]

it was prob the prompt

bc i had it at 0.25

[wispy canopy]

another optionp would be to first upscale in extras, and then feed that back into img2img

[native halo]

yea or leave empty prompt

[nova remnant]

||girl, window, sun, colorful, leaning in, table, sitting, sexy face, she looks at the viewer like she knows what they are thinking, (skimpy clothes:0.7), thighs, thick, witch, white hair, (high quality, best quality:1.2)||
I mean its just random words, I was testing different negatives and vaes

[steady grail]

upscale the dimension an than impaint details

[wispy canopy]

lets see what interrogator made of it 😄

[steady grail]

with lineart control net is even easier

[nova remnant]

@native halo idk what happened to the witch clothes, I guess she heard you

[native halo]

i will have to manually inspect the fabric to determine if its good for this hot weather

[nova remnant]

ah. so I have to upscale it 4 times? because when I did just x2 with i2i it did nothing, had to do 2 again with sd upscale

[wispy canopy]

2 glasses for thristy girl

[nova remnant]

godamn those are large cups

[wispy canopy]

ikr?!

[native halo]

(wet clothes:3.5)

[wispy canopy]

anime girl in a witch hat sitting at a table with a glass of orange juice, beautifully detailed render, perfect white haired girl, succubus in tight short dress, ultra-detailed. anime, female anime character, cat female with a whit and chest, “ anime, realistic cosplay, sitting in a lounge, with black is what interrogator made of it

[native halo]

i tried this once but without the chest part cat female with a whit and chest and she came out without torso so can confirm "chest" works

[wispy canopy]

😄

lol

[native halo]

anime girl with 2 eyes 2 legs 2 arms a torso and a head

[nova remnant]

"anime girl" sd: okay got
"with 2 eyes 2 legs 2 arms a torso and a head" sd: okay lets add more of those

[native halo]

anime girl with 2 eyes ,2 legs ,2 arms a torso ,a head ,skin , two lungs , a nose , hair , two ears , 2 hands and two feet,5 fingers in each hand ,5 toes in each foot,one heart,a skeleton made up of 206 bones made of calcium carbonate,a brain with 86 billion neurons and a nervous system this one better imo

[nova remnant]

@native halo you wouldn't do this to her right

[native halo]

would

[nova remnant]

[oak spoke]

heyo

[zinc flax]

Oo

Moth girl

[nova remnant]

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

[dawn jasper]

@nova remnant How have you been?

[nova remnant]

Would feed

[stable crane]

whoa cool

that's 1.5?

I would have had to guess sdxl

[supple raptor]

yep, 1.5 with LoRA and a prompt that I spent several days finetuning a couple of months ago.

[stable crane]

awesome

[granite shell]

lol damn ive been given a horni warning from this server

[dawn jasper]

That one is strong.

[stable crane]

you're out horni-ing me

[zinc flax]

:o

[stable crane]

maybe I was warned once, not sure

[steady grail]

hello

[stable crane]

hi

[grizzled moat]

hi

[wispy canopy]

hi

[wispy canopy]

thorny lady

[nova remnant]

generating in jpg saves me huge space on my ssd

[nova remnant]

for reference about sdxl https://youtu.be/vk_WhEiANgk

[zinc flax]

I saw NAI v3 gens and it definitely didn't fail lol
Assuming it's trained through sdxl

[wispy canopy]

new series when?

[nova remnant]

considering the resources demand and quality SDXL doesn't offer anything far too special that we couldn't achieve with SD 1.5

also the SDXL quality is debatable which may seem too artificial

[wispy canopy]

i'd like to see you try

this is actually 100% following the prompt

[nova remnant]

SDXL and SD1.5 has nothing to do with my personal prompt skill .. its a general idea where there are different views

[wispy canopy]

i vastly disagree, and i simply can't work without the increased knowledge sdxl has

[nova remnant]

you can vastly disagree like I do against sdxl

[wispy canopy]

i mean, this is just the stuff that's not coming out of sd1.5

perhaps on a specialized model

but for sure not on a generic one

[nova remnant]

ive seen better with sd1.5 using lora

and with much less resources

[wispy canopy]

lora is a resource too

which you evn have to search/download

and that's just a quick render, not a full-blown actual targeted image (like i almost never make xD)

[nova remnant]

i dont disagree on the context SDXL has it better but its not too far off for artistic variation

[wispy canopy]

how do you mean?

[nova remnant]

i can do w/o sdxl

[zinc flax]

Welp, I only saw the quality through nai so can't say anything further lol

[nova remnant]

in other words you can acheive artistic expression fairly well with SD1.5

[wispy canopy]

of course, that's something i won't, or even could deny

i feel the biggest hurdle for sdxl is just the massive baseline

[nova remnant]

i would embrace sdxl had it not been for resource hog .. trust me i've tried the comparison

[wispy canopy]

and the literal impossibity to train it properly on consumer grade hardware

[wispy canopy]

if i didn't hadn't have my 3090, i think i'd have skipped it too

[nova remnant]

our gpu is close ..i use 3060 i do farily ok with it for demanding tasks, but about the SDXL quality is another aspect i don't feel to fond of

[wispy canopy]

as in anime lineart?

[nova remnant]

that and even with realism

[wispy canopy]

nah, the realism is pretty good

[zinc flax]

But yeah, sdxl needs a beefy pc

Slow nai gens

[nova remnant]

too good gone overboard with metalic or plastic texture

[wispy canopy]

i mean, you complain about this one's realism... i'm really asking what you are nitpicking

[zinc flax]

Looks hella clean

[nova remnant]

that's cherry picking kind of but yeah most models you see on civita don't accomplish the same thing due to accessibility reasons, to train them

[wispy canopy]

uh.. cherry picking? dude?

are you for real right now?

[nova remnant]

haven you explored the sdxl section on civitai?

[wispy canopy]

i know

why do you think my merger took so long

i actually did get it right

and it does realism REALLY well

[nova remnant]

so ... does ... sd1.5

[wispy canopy]

on a specialized model

not a generic model

[nova remnant]

doesnt matter

[wispy canopy]

i know it doesn't. Everyone works as they like. i like to have 1 model, 0 loras

[nova remnant]

well im not looking at it as a bad thing, its part of the dev process and im hoping to see some good development in the next version that addresses the vital criteria of sd 1.5 and sdxl

but im not gonna use sdxl like i would sd1.5

[wispy canopy]

the training is all the problem, having that massive vram requirement for even training, just lowers the maximum quality achieved through training in the same amount of time

[nova remnant]

yeah that is a crucial thing the whole community feels about

[wispy canopy]

you'd either need one of the pro cards (quadro's and such) or access to other high end stuff

24gb 3090/4090 can't train properly, that's a problem

[nova remnant]

that would go beyond consumer reach, not a very good strategy for accessibility

[wispy canopy]

and that's the thing what's keeping the training of models so much slower on sdxl

anyway, i feel my model does realism quite well, and certainly doesn't ahve the plasticy feel (spoiler: non-anime)

[nova remnant]

dont get mw wrong, i appreciate your effort in creating these models, its a big contribution towards developing what could be more accessible and usable

[wispy canopy]

oh yeah, i was planning on trying to get some anime action going on my model

see if i can transfer at least styles

dunno what to do about the comprehension though

i mean, anime models are usually very booru-oriented, i suppose

[nova remnant]

cool but in all fairness i think you should keep a balance with how much you can push sdxl and how much it is accessibile

[wispy canopy]

i mean, for me it would be a matter of finding a model which has "compatible weights" and then trying to transfer the style, so it simply becomes an anime model, with the same amount of general knowledge

however, getting it to respond to anime tokens better like yandere, and all the booru tags and such, might be abit harder, as i'd have to transfer IN blocks, and that might influence model outcome too

[cyan burrow]

Good afternoon

[raven oasis]

just a bunch of my best renders

[gloomy scroll]

Here's Catcat for you

[sage mountain]

https://civitai.com/models/171778/add-texts-xl

anyone tried it with anime model?

only seen people testing with other kind of models, where its surprisingly decent. obviously its no dalle 3 but still

[bitter latch]

haha

[bitter latch]

the hair took on a mind of it's own

[supple raptor]

tripod for extra stability

[gusty palm]

chairwoman tripod

im trying the 'one button prompt' extension

takes roughly 4 minutes though for an image like that

[supple raptor]

this one was roughly 4 minutes including both the txt2img+hiresfix and img2img+tiled diffusion upscaling

[keen hollow]

we doin' widescreens?

[nova remnant]

what about large instead

[supple raptor]

[brittle jay]

Nice

[brittle jay]

Some tips from the time I helped a manhwa scanlation group. You perhaps wanna improve typesetting with these easy tips, 1. Center text in the middle 2. Change font.

[brittle jay]

Also this pretty funny

[gloomy lodge]

something about this makes my money brain happy

[steady grail]

how did you do this? it's simmply beautiful O.O

[dawn jasper]

@vestal crypt So may we talk?

[keen hollow]

today's Tifa

one more for good measure

[gloomy lodge]

daughters of our martyred lady the living saint, (Antonio J. Manzanedo, Anne Bachelier:0.4)
Negative prompt: (((multiple girls))), lowres, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, jpeg artifacts, signature, watermark, out of frame, extra fingers, mutated hands, (poorly drawn hands), (poorly drawn face), (mutation), (deformed) breasts), (ugly), blurry, (bad anatomy), (bad proportions), (extra limbs), cloned face, flat color, sketch
Steps: 25, Sampler: DPM++ 2M Karras, CFG scale: 5, Seed: 1804315485, Size: 640x960, Model hash: bb32ad727a, Model: darkSushi25D25D_v40, Version: v1.3.2 --- All the stuff ya need but tyty gonna make me blush

[autumn violet]

Nice texture

[supple raptor]

ty it's a new model from one of the mods over in the touhou server. I'm really liking it so far.

[steady grail]

so sweet😆

i love this model, what is it?

models are becoming more and more impressive

[gusty palm]

my thing drops to like 8s/it when i do hirez fix

i tried --medvram but i didnt notice a huge difference

[keen hollow]

I'm tired.

[supple raptor]

Yeah, sounds like it's probably falling back to system memory in your hires fix. You could try enabling token merging if you haven't already under Settings > Optimizations to help speed up the high-res pass a bit. Other than that the only solution would be lowering the vram usage which probably means lowering the upscaling amount.

[gusty palm]

i have token merging ratio exposed in my main ui ill try adjusting that

[nova remnant]

👍 hewo lewd

[nova remnant]

this is morning time

Just why

@wispy canopy look at it

[zinc flax]

Huge

[cyan burrow]

Thanks, yeah there's too many fonts to choose from

[stable crane]

https://freefontsfamily.com/cc-wild-words-roman-font-free/

[cyan burrow]

thanks

[stable crane]

also leave at least 1 character width (maybe 2) of padding between edge of text box and teh characters

[cyan burrow]

aah, if I can manage to get consistent appearances on characters so I can get them from different perspectives and I guess that will be enough to make an embedding or lora

and once I have like a dozen or more, I can finally just make a short story VN and release it here. Or at least a 15 minute long VN, which is like 5 minutes per route

[wispy canopy]

(kohya highres + lcm + freeu) combo is amazing

[gusty palm]

i was gonna brag about the perfect hand but i took the screenshot before realizing it was not perfect

[cyan burrow]

there's an upside to it. I can't state it in here

but there is

[stable crane]

lol

[supple raptor]

she can count to six

[stable crane]

she's really good at time

and base 12 number systems

[cyan burrow]

Nah, she has another mutation. Her thumb is actually a tendril that extends with the nail as protection and strikes the skin of a target, injecting them with a venom that shuts down their nervous system within 10 seconds. It has a 4 meter reach and a speed of 100 meters per second

alright I wonder how much I can get for my hybrid camcorder/dslr. It's only 1080p but it's prosumer level, comes with 2 lenses, a for time of purchase (2013) really good and a amoled viewfinder, touchpad window, SD card, 4.0 (becomes 5.0 cause it can centralize too) microphone (physically 4, not virtual.) Barely used. It cost 2800 dollars at the time of purchase with one lens. Now if I sell it with an extra lens and a good solid aluminium tripod from manfrotto(055xSomething) with a fluid head (for higher stability)(MVH500AH)...I hope I can get like 800 dollars for all of it. New it was all like 4000 dollars, all of it together. What shoots me in the foot is how fast cameras have evolved

[supple raptor]

search for the model on ebay. though, I wouldn't expect a lot out of a 10 year old camera. It's not old enough to attract the vintage crowd and not new enough to interest the modern crowd while also likely being outclassed by a modern iPhone.

[cyan burrow]

well both outclassed and not...In terms of resolution yes, but in terms of versatility and how it can fit any lens on it including anamorphic and such. You can get better images with it, even if only high resolution raw in photography, like 16mpx or something and I think HDRI

iphones camera is too small to function with swapable lenses and such

but you're right that it's likely not worth a lot. I think the highest value comes from the lenses, the tripod, the fluid head

[supple raptor]

yeah, agreed. the camera body probably isn't worth much. the rest could hold some value. ebay will show you how much pretty easily.

[cyan burrow]

yeh, good or ok lenses don't usually lose value as long as they remain in top quality condition

I mean there are 30-40 year old lenses that are still used today because they gave a certain type of look

I think the lenses, the tripod and such will get me maybe 400 dollars

and that's fine

maybe 600 with the camcorder

I'm gonna try that

start at 500, see if I get any bids

well alright, my MC giving the ladies (and whoever else) some candy

damn it sucks this channel is kinda dead now...I know ai art is very much alive but I prefered it when it was in here. People were much more mature

[nova remnant]

censorship will kill creative juice, always

[cyan burrow]

ai art is also censored...same rules

[nova remnant]

well gluck then

[cyan burrow]

eh?

what part of me wishing this place could be more active was unclear? O.o

[nova remnant]

im not talking about examples, im saying how censorship affects creativity

[cyan burrow]

yes, but I'm talking about chatting

Chatting with more mature people, friendly, less trolling, less toxic elitist behavior...I could forget I was on the internet (,,,,,,,,,and discord) for awhile

[nova remnant]

i find it very disappointing how some people are talking down at Ai in general

AI is possibly the last chance for turning things around in the right direction

[native halo]

chatgpt is what you need then

[nova remnant]

speaking of that the CEO got fired

but i started with chatgpt then left when bard came out

[native halo]

https://tenor.com/view/shame-jerry-seinfeld-too-bad-thats-a-shame-gif-12571290

[cyan burrow]

I think AI is going to be abused by many companies and countries to gain the upper hand, as a response other countries will work on progressing their AI as well. A new type of cold war, until there's a near sentience AI and if given autonamous decision making...We're all dead :p

that would be the doomsday scenario'

another would be hyper-corporate influence and becoming a corpotocracy

I don't see any path forward that won't be filled with some very complicated and potentially risky situations

job losses, which for companies is great, they get more money. BUt what do the people do? Welfare? That's gonna cost countries a lot'

as for nothing else, making a VN panel now, need maybe 20 of em to make a 2 hour VN

[nova remnant]

you touched on a lot of things on the surface but didn't point to the underlying cause for conflicting takes on Ai.

although i don't think it matters much cause i have doubts on changing society's mindset

[cyan burrow]

the initial underlying cause is and will always be humanity, If an AI becomes sentient and hostile towards us, it will be because of us

even if not explicit, even if through a screw-up

[nova remnant]

Ai will never be sentient, you keep raising that point w/o realizing the scientific studies on consciousness.

what do you mean by he initial underlying cause is and will always be humanity,
are you suggesting fixing humanity or the system?

[native halo]

thats a good anime

[nova remnant]

good point, we should talk about it in philosophy channel

[native halo]

https://tenor.com/view/my-honest-reaction-lain-lain-iwakura-goofy-ahh-anime-gif-15752134645187818885

[nova remnant]

actually im in the mood to play some video game

[native halo]

play higurashi daybreak,game of the century

[nova remnant]

im playing guildwars2

i have eye on this monitor too

[cyan burrow]

My VN needs a 3000 year old vampire that doesn't look like a child

sorry lolicons

[native halo]

u need a better bait or someone with your same level to get a proper response

[cyan burrow]

I will have a young girl in it but she's not a loli...she's just a young girl, with serious psychological issues causing random outbursts of violence

perfectly normal

[nova remnant]

my dual monitors keep messing up when playing games with mouse going all over them

[native halo]

ultra wide miku when

[nova remnant]

image is little washed out but the content looks great

[supple raptor]

vibrant colors really wouldn't fit the mood.

[nova remnant]

color tone could be wee bit pronounced tho

[supple raptor]

I tried and unfortunately this prompt really doesn't like wide AR. Maybe I'll try outpainting something when it isn't 5AM.

[native halo]

speakin of wide

[summer shuttle]

wide

[cyan burrow]

did I manage to capture a grittier art style or does it just look poorly rendered?

[nova remnant]

adorbs

[dawn jasper]

This is beautiful.

[dawn jasper]

This is beautiful.

[brittle jay]

Did Sam ultraman get fired from closedai?

[cyan burrow]

What's closedai?

[brittle jay]

Openai

And Sam altman

The CEO of openai got fired

Basically

[cyan burrow]

oh

openai...dall-e, gpt etc?

[brittle jay]

Yes

[cyan burrow]

why was he fired? He didn't sell enough products and charge enough? Too little corporatization of open source?

[brittle jay]

Can't really know exactly why

You can read their stated reasoning

But also openai is supposed to be a non profit

Or something

[cyan burrow]

charging subscriptions is a great way to achieve non profit

[nova remnant]

lol

[cyan burrow]

I get that non profit means every profit is immediately reinvested into the company for the companies goals

and nothing is done for bonuses etc

everything is for paying employees and for funding their goal

but they must be making a shitload

reinvesting that isn't an easy task

[nova remnant]

i think one of the reasons board members were unhappy with Sam is that he was leading the company away from the their mission to use AI for the betterment of humanity

[cyan burrow]

he was trying to make it "profit"?

[nova remnant]

elon talked about that too, when he was with openai that later become too commercialized

[cyan burrow]

this is one of the scenarios that will make AI dangerous. Corporations using ever evolving AIs to gain control and with the money they have, the influence and power they will gain can slowly transform the world into a Corporatocracy

Cyberpunk will be real!

[nova remnant]

where pretty girls

[brittle jay]

OpenAI is an American artificial intelligence (AI) organization consisting of the non-profit OpenAI, Inc. registered in Delaware and its for-profit subsidiary corporation OpenAI Global, LLC.

-Wikipedia

So basically OpenAI, Inc. is non profit while OpenAI Global, LLC is for profit

[nova remnant]

Too much AI talk too little AI

[nova remnant]

@lyric zealot smaoul

[nova remnant]

whichever direction it goes, as people will likely venture into, the real test of merit would be its sustainability

[wispy canopy]

use kohya highres

[wispy canopy]

getting reeady to be rendered, obviously

[nova remnant]

the mirroring looks neat

[wispy canopy]

mirrors still fail pretty much 100% of the time tho

[nova remnant]

its not a mirror

how did you manage?

its a screen with the view of the image's perpective

[nova remnant]

[wispy canopy]

just put mirror in tags

[nova remnant]

ahh ok, haven't had luck so far

on sd1.5

[wispy canopy]

yeah, don't think i ever got it working properly on 1.5 either... that just makes things plain weird

[nova remnant]

whats wrong with my images

this is the kind of moment i feel like loading up sdxl again .. only im playing games right now

why they so weird like that

[wispy canopy]

cfg too high?

[nova remnant]

and dont think i've experimented with mirrors on sdxl

[nova remnant]

I never change cfg

[wispy canopy]

lora strength too high?

[nova remnant]

Its just a prompt no loras, I'll try the exact prompts as something I did that looks normal

I think its a combination of whatever negatives I have

[wispy canopy]

looks like high cfg, high weight, or strong lora, overkooked image

[nova remnant]

looking more normal

[nova remnant]

Would pull those ears

[lyric zealot]

[nova remnant]

[wispy canopy]

isn't that elf discrimination?

[nova remnant]

||more like elf love||

Right @lyric zealot 👍

[nova remnant]

yes its working indeed @wispy canopy

[wispy canopy]

what did you do? ^^

[nova remnant]

I think I had negative embeddings at too low weight

or a bad combination lmao

[wispy canopy]

probably a bad combo which made a particular weight very overloaded

[nova remnant]

https://tenor.com/view/elf-anime-ears-gif-20906472

lmao

[nova remnant]

@lyric zealot

[gritty sundial]

Does anyone know if there is a model that creates images in this professional digital art style?

[lyric zealot]

@gritty spruce @wooden coral @dull jackal ^

[wooden coral]

[lyric zealot]

thank you

[gritty sundial]

😔 everyone ignores me

[gritty spruce]

that was quick! @wooden coral 🤟

[gritty spruce]

I'd go to civit ai and do a bit of research there, models and loras

[gritty sundial]

I research what

"Digital art Model" already tried

[gritty spruce]

what you asked!

try something different, line art would be a good guess

[nova remnant]

just get a drawing lora, low weight and emphasize on digital

its what it looks like

[nova remnant]

What you looking at

[lyric zealot]

looking at you

[gritty spruce]

https://civitai.com/models/196752?modelVersionId=229782

[nova remnant]

nice forehead

[lyric zealot]

kinda like the moredetails lora right?

is there any differences ?

[gritty spruce]

just saw they publish a new one, haven't tried, 0.8 was fine with hair and clothes

[nova remnant]

It's hyper now

[lyric zealot]

that's high when compared to moredetails, since for moredetails, about 0.3 is about just nice

[lyric zealot]

surely that's the only thing that changed

Welcome to Smol's Corner

[dawn jasper]

I don't see anything particularly wrong with this image.

[dull jackal]

a ping

WHO BECKONS ME!?

[nova remnant]

had a spammer already taken care of

[nova remnant]

me now with this ping

hewo sunny

Its too sunny where I live help

[dull jackal]

well then, here's an anime waifu I'm working on

[dull jackal]

GENERATE SOME CLOUDS

☁️

[nova remnant]

Will do

[lyric zealot]

https://tenor.com/view/cloud-anime-gif-22623842

[nova remnant]

the clouds.. what do they say

[dull jackal]

Love the baggy clothing vibe thing goin on here

[cyan burrow]

here, something happier

[terse jacinth]

someone clearly has a type

[near elbow]

hmmm

[late fox]

more on my mochai! 😏

[lyric zealot]

your first gen is dangerously borderline nsfw

[nova remnant]

I'll take care of them 👍

[autumn violet]

Miku after losing her coffee cannon and being reduced to a mere cup

[supple raptor]

I've been using this one for the last few that I've sent. https://huggingface.co/Aotsuyu/test/blob/main/9-AAAAAAA.fp16.safetensors

[nova remnant]

pikasnewgeneration
positive: high quality and best quality
negatives: (worst quality, low quality:1.1), EasyNegativeV2, easynegative_8955, AuroraNegative, KHFB, negative_hand-neg_43127

[native halo]

we need more wide Mikus

[keen hollow]

Do Nixeu or Wlop have art books? Anyone know?

[gritty spruce]

VRoid + SD + LCM https://x.com/yokohara_h/status/1725975747209437621?s=46&t=0WU4EHlrcScluYAy-xmVqQ

[keen hollow]

what's this?

[gritty spruce]

Composing in VRoid, rendering with SD and your favorite models = total control

If you haven’t seen them, @dull jackal has great tutorials on VRoid

[keen hollow]

never heard of "vroid"

[native halo]

prob cuz u never heard of vtubers,its one of the programs they use for vtuber avatars

[dull jackal]

You can check out my tutorials here: https://youtube.com/playlist?list=PLbJAuQIEql3YuacjyEedREPnwgljH7CST&si=KOLPdTq8vY6iYrau

There are several prompts that work very well with VRoid--I'll be updating my prompting guide, and putting them there. There are many already in the human section that might fit what you need. However, I'll go ahead and add a VRoid section and throw in a few base prompts for those who might be interested in getting started.

[native halo]

plus if u combine it with full body trackin u can create even better animations

[nova remnant]

[native halo]

which one u would

[nova remnant]

@native halo needs more hair

[native halo]

needs afro style

[nova remnant]

Kek

[nova remnant]

The one taking the picture

The mysterious one

[native halo]

needs more fang

[nova remnant]

@native halo afro in prompt kek

[native halo]

(afro:2.5)

[nova remnant]

[nova remnant]

[native halo]

too sexy for this server imo 🌚

[nova remnant]

Give me some parmesan cheese

[dull jackal]

Alright---I threw some prompts together for you guys, for VRoid/anime. Feel free to try them out/see what happens: https://docs.google.com/document/d/1BxdWqfBJ3QPggHnBCBx3QIkUpHnBWShd_s3zEt2dTLM/edit?usp=sharing Byeeee now.

[nova remnant]

Mouse girl!!!

[nova remnant]

@native halo here is your afro goddess

is that enough

[native halo]

no i can still see her face

[nova remnant]

that looks nice, i see its also a trained model 👍

[nova remnant]

What do you mean

not sure what trained model means for it

authors don't have complete control overr merges

you'd have to work on what the trained model has to offer for merges

[nova remnant]

@nova remnant Those instruments look dangerously pretty good for plain text to image tex

Was more involved

i've been meaning to create some animated stuff with images too, but the animatediff bugged out on me

[keen hollow]

my contributions to large hairdos

and now medusa....

[native halo]

now replace her hair snakes with hamsters

[keen hollow]

oof...

that sounds toughg

also tough

LOL

[nova remnant]

[native halo]

rat queen 🙏 👑 💅

[keen hollow]

hamsters.... come on 😛

[nova remnant]

cosmic mouse

[native halo]

👅 👂

[nova remnant]

https://media.discordapp.net/attachments/931925914470711346/1061076535525785680/321666B2-CC36-467F-B254-56055DEB1C8C.gif

@nova remnant this model is pretty gud

hmm, but i like pikas better for simplicity

[native halo]

https://tenor.com/view/anime-ear-ear-bite-anime-nom-anime-couple-gif-16088514

[nova remnant]

It is pikas lmao

really?

yeah

the preview images are different tho

I mean just prompt stuff and it works

its cool then cause i can play with noise and simple colors both

[native halo]

pingas new generation rly best model out there tbh

[nova remnant]

pingas ?

Is that the steven merge

[native halo]

Pingas generation

[nova remnant]

Probably easier to do stuff with loras, I'm not using any, also I don't have access to adetailer and stuff like that

Trying the complete opposite of that last prompt, see how it goes

ok, i only use two loras betw, lora:epi_noiseoffset2:0.4, lora:more_details:0.4 not too much image manipulation

@nova remnant did you try that hyper detailer thing

Its basically more details but a new version or so I heard

i havent yet

[native halo]

why use details lora when cd tuner has more detail options 🌚

[nova remnant]

more detail is fairly good too

oops left some moon related stuff and tried other things on top

with spoiler just in case

perfect pic for kundalini chakra

Shes a cultivator!

[nova remnant]

so above so below

[nova remnant]

i like that pic

[native halo]

she has the blue light on the cultivator

[nova remnant]

@nova remnant this one came out a big buggy but still cool I think

[nova remnant]

this one's heart chakra

you mix heart chakra with kundalini chakra that force is immense

This is more like the prompt I envisioned

[native halo]

🙇‍♂️ im a believer now

[nova remnant]

We always end up in the thigh church

[nova remnant]

this one looks like shes a beginner :p

only pose for alignment ;p

[native halo]

i love egypt now

[nova remnant]

@native halo

[native halo]

(latex,bodysuit:2.5)

[nova remnant]

@native halo The good one, or the bad one ?

[native halo]

shes a good one

[nova remnant]

Maybe

[stable crane]

hiya

[native halo]

need more egypt

[stable crane]

"Lotta Lewds"

[nova remnant]

tooooo many fingers

[native halo]

most clean and healthy r34 artist

[granite shell]

yo yo

[granite shell]

nice

[nova remnant]

whats happening shaz?

[granite shell]

not much, ive just been hanging out, playing some wow lol

[nova remnant]

oh cool

im off to play guildwars2 in a bit

[granite shell]

now it is time to gen 🙂

nice! i played on release, it was really fun

[nova remnant]

its a good mmo 🙂

[granite shell]

think i played the first xpac as well, then havent played it since

[nova remnant]

how come?

[granite shell]

friend group wanted to back to wow lol...

[nova remnant]

oh

well pugs in gw2 arent bad at all lol

[granite shell]

we usually mess with other games but always come back to wow

[nova remnant]

how is wow these days?

[granite shell]

true i should check it out at some point, brush off my guardian

[nova remnant]

i see lot of ppl play both wow and gw2

[nova remnant]

hi smol sus 🙂

[granite shell]

the sketch effect is nice

[lyric zealot]

[nova remnant]

15 images i gened now, hands came out pretty good

and very little finger deformity

[nova remnant]

Horns to grab

Or ears to pull

Which one is better

[lyric zealot]

why not both?

[nova remnant]

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

[vestal crypt]

https://media.discordapp.net/attachments/1027429244478955530/1175725727707054171/00036-222921476.png?ex=656c46e9&is=6559d1e9&hm=e20dc65486c983b816d0a4b054c20b39ba093b07291cccc0cc63299be9a3fb65&=&width=530&height=662

[wispy canopy]

anime!

[kindred meteor]

Yooo thats crazy, whats your model ?

[lyric zealot]

hi, here's the pic with metadata

it's mostly the work of the lora

[kindred meteor]

Ooooh the sketch effect is so cooool

Nice !

[brittle jay]

Not sure if this is alright to send here

Basically quest 3 at school

[gritty spruce]

let's please stay in topic, feel free to share those over https://discord.com/channels/1002292111942635562/1004159122335354970

[brittle jay]

Yeah alright

I am just too used to only/mainly chatting here

[brittle jay]

Actually #🌶｜off-topic is probably better

[supple raptor]

[steady grail]

looks like code

was she named v?

[native halo]

c.c

https://tenor.com/view/cc-codegeass-kiss-snuggle-gif-4677771

[steady grail]

oh, right, c.c.

[supple raptor]

yesh

[native halo]

new hat

[gritty sundial]

model name, pls

[wraith spade]

/🦊

3

[brittle jay]

since when you here?

[bitter latch]

a merge of several

[slender barn]

Trying some new resolutions and stuff.

[stable crane]

love the vibe

[slender barn]

Thank you🙏🔥

[stable crane]

anyone trained a sdxl lora? how much vram is necessary?

[nova remnant]

bob

[stable crane]

playing with sdxl base gen -> 1.5 upscale today

based in sdxl, then sdxl upscale vs 1.5 upscale

need a better base

[polar anchor]

How do yall get better details? What I mean by that is I can make an image of a Japanese street but everything looks like blobs and isn’t defined so it all looks unclear. How can I make that better?

Here’s an example of what I mean?

[stable crane]

how are you upscaling?

it looks a bit like a regular upscale

[dawn jasper]

DOGGO!!!! 😮

[stable crane]

for landscapes I would just use SDXL

[polar anchor]

I generally use Easy Diffusions WebUI for actual image generation. And then I upscale with it or try to ESGRAN but I’m not really getting great results so I’m looking for better

[stable crane]

that's the issue, you have to upscale it with SD

[polar anchor]

How do you use SDXL? What ui do you use?

[stable crane]

another diffusion process using the original image as the base

for SDXL I use comfyui, but I recommend A1111

you can get good detailed landscapes from 1.5 models too

[polar anchor]

I know A1111 I can’t use on my potato laptop so o would need my PC For it. But does comfyui by chance not use nvida and I can use web?

[stable crane]

idk about non-nvidia

sdxl requires an even beefier card so maybe avoid it

[polar anchor]

Poop ok. I shall try it then. My card is a little older and the only one I got so I try not to risk it too much. Thank you!

[brittle jay]

What is your card

[steady grail]

comfyui is the best performing gui i used till now

if wasn't that to properly set it, you need to be an engeneer

[brittle jay]

Real chads use no gui

The only GUI I need is notepad

Microsoft word

[steady grail]

damn... a true geek

[brittle jay]

I write my prompts in binary

I don't have a VAE, I encode / decode to and from latent space by hand

Just kidding, I just look at the embeddings in their true form. Only normies need pixels

[steady grail]

you sow too much matrix

[polar anchor]

I got a 2070 Super

It old

[brittle jay]

bruh

not that old

and it's nvidia

[steady grail]

using comfyui?

auto1111 have a reader in one of his tabs

out of auto, sd prompt reader

[nova remnant]

open it in your browser, download it to your pc and throw it in the png info tab

[polar anchor]

It’s old to me lol. It’s been a few years and new stuff has come out so I just say it’s old. Is that enough to run better programs though like A1111?

[autumn violet]

I run mine on a base 2070

[nova remnant]

2070ti is not bad at all for ai

many people use a 1060 for example

[stable crane]

tiled vae will help as well

[polar anchor]

I’ve heard of this. Do you have any good recommendations for one?

[polar anchor]

Oh sweet. Then I’ll give it a shot soon. Thanks for making me feel like I don’t have a potato lol

[steady grail]

tiled vae si an extension

[supple raptor]

This extension https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111.git
It will add two accordions to both your txt2img and img2img tabs. The one that will help with vram usage is this one and these settings should be good for 8GB vram.

[polar anchor]

Thank you so much! I’ll try and and see what happens

[native halo]

high res fix is a must for 1.5 models otherwise the details look bad

[steady grail]

i'm looking for a good balance between highres and model upscaling

for now, i'm undecided for an upscale between 1.3 and 1.5 and than a 2x with a model upscale

or to just push for a model upscale

[native halo]

try with high res fix first then img2img upscale

[steady grail]

usually, using a controlnet and than highres fix, i get good enough results

but with a model upscale, i get so much more sharpness

[native halo]

if u want more details u could try with high res fix (latent) but u gotta play with denoising strength to not get deformities

[steady grail]

with highresfix i use controlnet to manage this problem

with tiled model upscale is different

a good balance in denoise is fundamental to get good details, without weird mutations

[native halo]

yea also some loras can cause deformities if u use latent

[polar anchor]

What is high res fix?

[steady grail]

in the last months, i'm focusing on generating ultra wide upscales

[steady grail]

auto1111?

https://www.youtube.com/watch?v=ERUgZWNBn2o&list=RDGMEMhCgTQvcskbGUxqI4Sn2QYw&index=2

[polar anchor]

Sorry I don’t use it. Lol. I don’t know all the lingo for it. Thank you

[steady grail]

what program do you use to generate? automatic1111? comfyui?

this song, tonight, slaps like hell

[polar anchor]

I’m going to download Automatic when I have time but right now on my potato laptop I have Easy Diffusion

[steady grail]

never used

looks like a different model than sd

anyway, highres fix is an automatic1111 setting to upscale images

[polar anchor]

Gotcha. Ok. When I download it I will try that as well. Thank you

[steady grail]

the process is simple, starts to generate an image, at mid work, add some noise on all the image and upscale it, restarting the image generation

this adds details to the image

with too much denoise, you don't get just an upscale image, but something totally different, so setting it in his sweet spot, is necessary

[gusty palm]

is there any reason to update requirements.txt or whatever if update.bat says ur up to date

and has the lcm lora been added to auto1 yet

[steady grail]

lcm is just a common lora you can download

and load like any other lora

to use is well, you need the a compatible model and the lcm sampler

[gusty palm]

thast what i meant

the lcm sampler

[steady grail]

look in the samplers if i's there

on in the extensions if you find anything

[native halo]

dev branch of auto has it but idk if its inplemented correctly

[steady grail]

i don't have auto1111 anymore so i can't verify

honestly, i'm still looking for a good anime model supported by lcm

[native halo]

u could also try with this one https://github.com/light-and-ray/sd-webui-lcm-sampler

[steady grail]

they added an extension, so

[sage mountain]

oooooh

was waiting for someone to test it after i saw kohya's tweet

[native halo]

someone made an extension already https://github.com/wcde/sd-webui-kohya-hiresfix

[steady grail]

i should test it asap

damn, the son of goblin slayer is killing me inside

found for comfyui

[gentle musk]

I've found it's a few seconds slower than native at higher resolutions. It's a bit odd though because it starts out faster and then gets slower. Maybe I'm just hitting VRAM problems with 10GB.

[sage mountain]

maybe an implementation issue for auto1111 vs comfy's

[gentle musk]

It's on Comfy, but I'm meaning if I do 1920x1080 native vs that fix it's slightly slower.

But it does tend to fix the weirdness you get

[stable crane]

is there a link to the comfyui node?

[gentle musk]

Its built in

it's in the For Testing nodes

[stable crane]

😮

oh so I have to pull

[gentle musk]

Yes you need to update Comfy

[stable crane]

trying that now

any tut on how to use it?

[gentle musk]

But yeah in SDXL I think my issues with the slowdown is just down to VRAM

[gentle musk]

Just run a model through it. I've not seen any explanation of the settings. Changing them does "things". Default tends to fix the usual issues.

[stable crane]

at the highres step or even the base gen step?

[steady grail]

weird, i just found a custom node

with the same name

[gentle musk]

It needs to be run from the start, otherwise it's still going to generate duplicate stuff.

Example, this is without it running at 1080x1920

And this is with

[stable crane]

that's a raw 1080x1920? wtf

interesting

[gentle musk]

Yeah

[stable crane]

so I don't even need highresfix

[native halo]

that would take longer if u have 6gb of vram

[gentle musk]

If this works for you no, but it uses a lot of VRAM

[gentle musk]

So if you were using highres fix or a tiled sampler before because of VRAM issues, this won't help.

But if you have the VRAM, it will

[stable crane]

damn, it works

insane

[gentle musk]

With 10GB and using IPAdapters I unfortunately start to run out of VRAM, so it takes almost twice as long as a normal generation.

[stable crane]

I have 10gb as well, but seem to see no problems

[gentle musk]

Are you using IPAdapters?

[stable crane]

no, idk what that is

CLIP replacement?

[gentle musk]

That's why then, they use other models that gets loaded into VRAM.

And you should go look them up they are cool. Lets you use images to influence the generations. There's even a face one that does facial features really well.

[steady grail]

i can't find the tab with the node

[stable crane]

absurd, 121 seconds to do 2048x3840

[gentle musk]

What GPU

[stable crane]

the gen sucks

3080

but the res is nuts

gonna have to keep playing

[gentle musk]

I've got a 3080 as well, so yeah it's just me using extra stuff that's slowing it down

[stable crane]

you having troubles in the upscale or the initial 1024x1920?

[gentle musk]

If I take off 1 of the IPAdpaters so I don't run out of VRAM it does 1920x1080 in 48 seconds

[steady grail]

wow, tbat's big

i need 25 min

for a 2k 2k

[gentle musk]

You could try do that without an upscale pass, but I think that might run out of VRAM

[stable crane]

the esrgan?

[gentle musk]

As in just do a raw 2048x3840 generation, don't do a second pass, with that high res node.

[stable crane]

oh wtf

[gentle musk]

It might run out of VRAM though

[stable crane]

will try it

usually I found sampling OK, the vae decode was the problem

[gentle musk]

Well it's trying to run on mine, it's going to take a long time though lol

[stable crane]

did you try combining with the LCM Lora?

Warning: Ran out of memory when regular VAE decoding, retrying with tiled VAE decoding.

[gentle musk]

No, but I could give it a go

[stable crane]

figured

[steady grail]

what's the tab of koya downscale?

[stable crane]

Killed

[gentle musk]

Yeah it's trying to use 16GB VRAM

[stable crane]

trying again w/ tiled vae decode

impressed it just died in tiled

[gentle musk]

It will do 2560x1440

Just

[stable crane]

it did not go well

[gentle musk]

lol

[stable crane]

or do I have to tweak downscale factor?

[gentle musk]

Honestly not sure what that even does

[stable crane]

trying again with factor of 4

[gentle musk]

mostly works at 1440p

84 seconds

[stable crane]

it's incoherent

stil

kohya is a baller

[gentle musk]

I wonder if that resolution is just too much for it

[stable crane]

even 1024x1920 is misbehaving now

oh it's the tiled vae decode probably

yeah it decoheres with tiled

[gentle musk]

I'm not sure what that downscale factor setting does, this is the same image as previously but with it set to 3.5 downscale factor

[stable crane]

seems better

[gentle musk]

Not sure it's better, but it's different that's for sure.

[steady grail]

oh, it's in for testing tab

usually, such upscaler are bad

[gentle musk]

It's not upscaling. It's doing some stuff where it resizes the unet

And then generating natively

[steady grail]

weird

i don't understand the teory behind it

let's add it to the highres fix and let's see what it does

[gentle musk]

Your supposed to run it on the initial generation

[stable crane]

damn my upscale failed 😢

[gentle musk]

Not on upscaling passes

[stable crane]

oh, use the original model on highres?

absurd

[gentle musk]

I don't know how you have your workflow setup, but pass the model you are sending into your first ksampler through that node.

And set the high resolution on the first ksampler as well

[stable crane]

sounds good

[gentle musk]

If you use it on a 2nd pass, then it's kind of pointless, because unless you use really really high denoise, it's not going to make the weird duplicate stuff it does at high res.

[steady grail]

i'm using 2 ksalmples as highres fix

[stable crane]

can comfy load a png as a workflow?

[gentle musk]

Yeah

[stable crane]

idk their "PNGInfo" equivalent

[gentle musk]

You just drag and drop a picture on that has the metadata in it

[stable crane]

beauty

[steady grail]

so, this stuff makes highres fix pointless

[gentle musk]

If it works properly then yes, you wouldn't need a highres fix pass

[steady grail]

i need to experiment... again lol