🍥｜anime

[bitter latch]

derp

random brain is random

tries to sleep... thud

what ya say? say it again.. dare ya

thud

[bitter latch]

goes back to bed

[modest fog]

big

[steel zephyr]

I was wondering about something...
Is this technically possible in comfy to turn prompt into some kind of temp lora or embedding, make a few of them and link into one general prompt to get consistent custom characters?
(I know nothing about comfy btw)

[modest fog]

[modest fog]

What do you mean by that

Like

Are you just saying "ip adapter from controlnet"

Cause that's what it does

Tries to generate the same character for you

If you don't want / can't make a lora I would use that with the same prompts to help you get the same looking character

[steel zephyr]

Wow, yeah. Didn't knew it's a thing.
Yet there's still no way to say 'characterA does X, characterB does Y', isn't? Or I'm wrong?
Or ip adapter able to merge poses and stuff?

[vital raptor]

regional prompting for multiple characters

[modest fog]

If you want multiple characters in one image you use regional prompting

It's easier

Didn't understand you wanted that

[steel zephyr]

I tried regional once, it was terrible, like two different images poorly stitched

[modest fog]

You need professional assistance

Lmao

I couldn't get it to work either

Too much for my brain

But it should look perfectly fine

[vital raptor]

regional prompting can be tricky, but it works

[steel zephyr]

Good to know that pony loras work on illustrous just fine

[gusty palm]

ever had luck with getting hello kitty characters to work

[gusty palm]

ive had some results doing interior bedroom hello kitty theme but thats the close as i got

[supple raptor]

huh, never tried before but yeah, getting just hello kitty to work would probably be really tough. the correct tag is hello kitty \(character\) but most of them are mistagged and are actually other characters that just have something hello kitty themed like the bow, hello kitty print clothes, or maybe holding a hello kitty doll

[bitter latch]

challenge accepted

"Hello KItty" \character\

hmm so it wasnt the prompt it was a sytle... but which one?

[rough oxide]

same problem with hachune miku :'(. had to formulate a whole prompt and it still doesn't get spiral blush but close enough.... as for hello kitty...

probly would work better with "no humans" too

hello kitty works totally fine i think. just use sanrio as copyright after hello kitty \(character\)

and no humans makes it more stable

[rough oxide]

specific characters or just hello kitty? i think hello kitty works fine in noobai at least.

cinamoroll & my melody & kuromi should work fine too

[gusty palm]

oh i never trued the \character\ thing

thats wayyyy closer than anything i ever got ill have to try that although the hello kitty bedroom i made was awesome

[modest fog]

Big blue

[green temple]

she`s smol but close to the screen

[modest fog]

@green temple smol ready to attack you

Smol was defeated by sooma @green temple

The great battle of knee biting

[green temple]

knee biting 😱

[modest fog]

runs away

[green temple]

comeback , get my knees back

[modest fog]

nothing is allowed

[green temple]

[modest fog]

slug

[steel zephyr]

where hands...

[gloomy scroll]

Keeping them warm inside onesie

[modest fog]

Here big hands for you to grab

[zinc flax]

Huge

[modest fog]

@dry hatch

Bam!

[green temple]

[past junco]

@modest fog what checkpoint are you using?

[lyric zealot]

that was nai

[bitter latch]

down the hole! nooooo

[past junco]

Poor Toilet-chan

[past junco]

I don't believe I know that checkpoint

[past junco]

I love the goofy onesie

[grizzled moat]

same ... same

[bitter latch]

combines the two , kinda

[grizzled moat]

[grizzled moat]

cozy @gloomy scroll

[bitter latch]

ok.. sorry been sad lately.. sigh. will try to make it more happy

[modest fog]

answered for me, banned

[modest fog]

Novel ai is a paid model

Closed source

But basically looks like illustrious (sdxl)

[bitter latch]

reverses

[modest fog]

Food you're back

[modest fog]

Would help

[bitter latch]

odd hands

[bitter latch]

goes back to bed... now if my brain will let me...

[modest fog]

https://tenor.com/view/bunny-bnuuy-steals-your-bunny-steals-your-bnuuy-gif-26209043

[gloomy scroll]

Well at least she knows how to come back home

[modest fog]

small

grabby

[humble current]

What prompts did you use for the eyes?

https://media.discordapp.net/attachments/1089014954927071362/1308826694462996511/tmp328wxbqi.png?ex=6740acd8&is=673f5b58&hm=28068b17e9b84048d347c3b41f42043e9a12dacb7451a89e4fd59c6d6184cf1c&

[thorny mango]

actually i didn't use any prompt for the eye. I was using the "NTR Mix" model which is really good a generating close view image

[humble current]

https://media.discordapp.net/attachments/1089014954927071362/1308833553504538714/tmpgtjbyt9w.png?ex=6740b33c&is=673f61bc&hm=d6ee9f13d79660740f6cf4ff18a6caf7f75fa94dd6bb5e9f49121d934c6d71eb&

[green temple]

big face

[vestal ferry]

artist

https://tenor.com/view/doro-assemble-stampede-raid-gif-3238891709128671482

[green temple]

nyalia i think

[modest fog]

Where smol artist

[green temple]

i want smol cg i can`t find one

something close to quan_(kurisu_tina)

[green temple]

but bigger eyes

[modest fog]

What is cg

[green temple]

PeM and asusnm style

[modest fog]

Many bloo

[modest fog]

Sad smol cat

[vital raptor]

i should make a good prompt and run a huge artist list through pony and noobai models so we have good smol examples instead of big bikini nsfw examples

[modest fog]

Smol farmer

[keen hollow]

evenin'

we need a mountain character

haven't seen that in a while

[modest fog]

Mountain?

[keen hollow]

Yeah.. like ... hang on

ill hit up dall3 😛 dont have sd running

good grief dall3 is still king....

[modest fog]

The outside is scary

[keen hollow]

black and white hair is neat

ok the horses are hysterical....

[zinc flax]

Super important big bikini mention

[vital raptor]

i have pirates

[modest fog]

pirte smol

[vital raptor]

hmmm, i don't have any steampunk oddly

[silver dawn]

cat

[keen hollow]

funny, no 'steampunk' in that prompt

[vital raptor]

i searched for 'steam' and this came up 😄

[keen hollow]

"A captain retropunk space lady on a large atompunk ship, retro space age"

spits out stuff like that....

pop stars?

[vital raptor]

. 😄

[vital raptor]

no popstars, but i found another one with mountains

that's one of the ones i did with song lyrics for the prompt on one of my old merges

this one too

[vital raptor]

oo, 2 days until arcane season 2 is finished... time to binge it starting from season 1

[modest fog]

Ugly

I was watching the best game renewal today

Path of exile 2

They started it with a few minutes cinematic

[modest fog]

I couldn't last any longer,I just wanted it to end and get to the good part, I'm not a fan of them can't imagine watching something like arcane

[vital raptor]

its really good

[modest fog]

@green temple magic smol

||she turns you into an ugly frog||

[green temple]

[modest fog]

[modest fog]

Last time I liked watching a movie was like 2014

[vital raptor]

frog not ugly

[modest fog]

still green worst color

[modest fog]

@green temple

Tasty

[modest fog]

@green temple

lick

runs

[zinc flax]

wow those are fried

[supple raptor]

Noob Vpred 06 is still struggling with white backgrounds a bit. It's a huge step up from how 05 handled them though which hopefully means that it will keep getting better. For the time being, at least AYS scheduler looks pretty good.

[bitter latch]

grinz evilly and change's it up

i'm trying to get every little idea thats keeping me awake out

trying to tire out my brain

hhhaha oh gawd.. wtF?

ZZzzzzzz thud

[steel zephyr]

super smol Miku

[terse ore]

If you want to train a model for styles and concepts, it's better to just train a Lora instead of an entire new model?

[zinc flax]

Yep, lora is better for quick training

[steady grail]

a very relaxing image

[terse ore]

And if I want to make a realistic lora, then Flux would be best?

[zinc flax]

Depends on the model

What model do you wanna make a lora for

[terse ore]

I don't know yet. But I want the people to be the most realistic which is the priority

[zinc flax]

What model do you usually use

[terse ore]

Usually I use anime models but I want to try realism

[zinc flax]

Better be specific about the name since anime models nowadays have different bases lol

[terse ore]

Oh. I usually use crossmix autism pony SDXL

It's easy to get it to do what I want

[zinc flax]

You'll need to train on base pony then

[terse ore]

But for realism I only used a few models so far. Juggernaut XL, some 1.5 stuff. The skin is best on 1.5 stuff

[zinc flax]

ah

I just checked the realistic models, you'll need to train on base sdxl for that

[terse ore]

Maybe I should do more research? But yeah I was thinking either SDXL or flux

1.5 is getting old

[zinc flax]

Definitely not flux

[terse ore]

Oh?

[zinc flax]

Unless you're using a flux model

[terse ore]

Well I don't really mind what model I use

[zinc flax]

But I've never touched flux so no comment about that lol

[terse ore]

I see

So not 1.5. but either SDXL or flux Lora. I'll do some more research from there

Thanks

[zinc flax]

If you do use sd1.5 models then you'd also need to train using that lol

That's why ppl usually only stick to one base model

Less hassle I suppose

[steel zephyr]

pov: me generating new wallpapers

[modest fog]

Soooma

[green temple]

caaat

[modest fog]

Givesmol

I am alone cat

[green temple]

imma go to sleep

smol sleepy now

[modest fog]

[green temple]

[final vault]

was she forced to drink something nasty?

[green temple]

no

[bitter latch]

what you don't see she just got knocked down in a fight and now she's very angry and someone's about to have a bad day?

[final vault]

must be those slimes

[vital raptor]

it was the Hilichurls

[steel zephyr]

I've spent half an hour trying to understand why I'm getting white backgrounds until realized it was "pitch black backround"

[bitter latch]

yeah mispelling is common and gets some interesting results

[steel zephyr]

background is one token, but back+round is two, yet there was nothing related in my images
interesting

[bitter latch]

,maybe, my guess, the ai thought back around (behind) the black = white

[modest fog]

Back and round

👍

[green temple]

where is the g

[modest fog]

Oh

We were talking about backgrounds

[green temple]

[modest fog]

@green temple

[green temple]

[final vault]

or it was aether in hillichurls costume for Halloween

[modest fog]

steals her hat

[modest fog]

Grabby

[gloomy scroll]

It's hood and no breaking bunny onesie

Or Yue will get mad

[modest fog]

[gloomy scroll]

Here have a cookie instead

[modest fog]

[wet elm]

why are my generations just black and white artifacts? can someone help?

[modest fog]

Post one with png info

Most common issues are wrong vae, incompatible lora, models, resolution or samplers

[supple raptor]

way late to the party, but black and white artifacts sounds like possibly running a vpred model without the UI knowing that you're running a vpred model. will need more info to help

[zinc flax]

Use reforge and enable vpred

[keen hollow]

mornin'

[steady grail]

krita support vpred too

hello guys

[bitter latch]

.

[wet elm]

its not just black and white, sometimes its black and red, but idk what is a vpred model too. im using Noobai new version, the old versions work fine but its just the new one

[wet elm]

i deleted all of them but im launching webui and will send

[wet elm]

Oh.... it says vpred on the model page... im dumb, how can i fix this?

[zinc flax]

Enable advanced model sampling

Assuming you're using reforge

[wet elm]

im using auto1111.... idk how to use any other ui lol

[zinc flax]

reforge is basically a1111 but more optimized

iirc a111 can already use vpred too

[green temple]

better than forge ?

[wet elm]

oh i might need to start using reforge then

[zinc flax]

both are the same thing lol

[green temple]

i thought it`s something else

[steady grail]

sorry, still not fix for dumbness <.<

[wet elm]

😭

[steady grail]

[wet elm]

sorry but how do i do that? i searched on google but i got no answer

[modest fog]

scary

[green temple]

[bitter latch]

gets up looks around.. goes back to bed.. "NOPE"

[zinc flax]

Scroll down txt2img tab

[modest fog]

@green temple

Many more smol

quan_(kurisu_tina) @green temple is this the artist you were trying to use

maybe it was someone else

I don't really know if it works in NAI

[green temple]

yes but with a bit of 3d/cg

[green temple]

it is yup

[modest fog]

@green temple tried some 3d stuff

[modest fog]

looks like some game

[green temple]

what game ?

[modest fog]

I dunno, just said it looked like a game cause of the 3d effect

[green temple]

ok

[keen hollow]

evenin'

[modest fog]

@green temple curious smol

[green temple]

the one on the right better 👍

[modest fog]

[modest fog]

@green temple angy smol

[green temple]

why angy

[modest fog]

You didn't give her the new most powerful iphone

@green temple you in the background it's over

[modest fog]

@green temple smol is calm again

[modest fog]

yellowsmol

[modest fog]

no cry allowed

[green temple]

sad smol

[rough oxide]

new model is gud. feels pretty significant for a .05 update

[modest fog]

veigar wtf

[rough oxide]

idk what a veigar is

i totally still just use my randomizer except in cases like bottle miku and iono

[modest fog]

Ugly

@green temple cake

[rough oxide]

i was also trying sumthin sweets related

but i think i need to work on prompt a lot moar

lol

[modest fog]

[unkempt cape]

Eh

[unkempt cape]

Lmao her head is going out of frame to poke the text

[modest fog]

who dis

[bitter latch]

come in come in

yeah i'm bored, brain is struging with theme.. sigh

[terse ore]

Anyone else use Segment Anything? I'm getting OOM but I don't understand why

[gloomy scroll]

We all sip together

[vital raptor]

yummy sippys

[green temple]

smol

[modest fog]

I call upon my protectors @green heart @bold ether bam him

[modest fog]

many smol

[modest fog]

Big cat

There's a streamer I like watching

[green temple]

[modest fog]

She's fairly smol

I call her cats giant cause of the size lmao

[green temple]

😂

[modest fog]

Looks like that

@green temple tired bloo bnuy

[green temple]

school

[iron ferry]

"Generate an anime-style image of a young woman with long, braided hair that is half green and half pink. She has bright green eyes and is wearing a dark school uniform with a white collar and gold buttons. The background is a blurry, green outdoor setting. The overall aesthetic should be inspired by 'Demon Slayer' with a focus on vibrant colors and detailed character features."

[green temple]

mitsuri ugly

[modest fog]

school means math

@green temple grabby blue

[green temple]

no grabbing

[modest fog]

Mine

[bitter latch]

like th is?

[supple raptor]

i made bnuuy too

[bitter latch]

haha

what you looking at?

[bitter latch]

ooh she mad

[bitter latch]

uhh oh shes brought out the weapons

😄

[mighty wharf]

Fleg

[bitter latch]

random

she did some art

foundem

miku bear

[bitter latch]

show off

miku mouse

[keen hollow]

kinda neat

[keen hollow]

this one too

[green temple]

[humble current]

made her green and icy

[humble current]

a lot is weird here...

[keen hollow]

make her into a candle made of ice

[humble current]

this sword looks a bit like a candle lol

[humble current]

well here it is

[green temple]

[modest fog]

smol

[modest fog]

@green temple evil

[green temple]

[dusky jolt]

Gokuuu

[keen hollow]

video gamey

[modest fog]

[green temple]

i`m just trying to learn how the artists affect on the prompt

[green temple]

[modest fog]

I'm big sad cat

[modest fog]

Just do 1smol and everything is fine

[green temple]

no sadness allowed unless it`s a math test

[modest fog]

All is sad

https://tenor.com/view/sad-gif-3743010737282172028

[green temple]

[modest fog]

[green temple]

@modest fog

[bitter latch]

hehe

yep i'm being random

[bitter latch]

makes it more anime

[modest fog]

Sued by Nintendo

[gloomy scroll]

But it's Ponto

[keen hollow]

I got the new OLED Ponto

[bitter latch]

huh chatgpt does images now

[modest fog]

same as bing ye

just more free now

[bitter latch]

It wouldnt' do miku. I tricked it 😄

[modest fog]

it used to be on the paid gpt option only I think

[sudden hemlock]

Hey! What model(s) would people recommend for anime-styled backgrounds, landscapes and interiors without characters in them?

[modest fog]

any illustrious models should be good enough for that

you just have to make sure you try hard to prompt for things without a character, like put a bunch of "1girl, boy, person, human" on negative and no humans/landscape on positive

[modest fog]

[green temple]

why sad

[modest fog]

no smol to share wait

[green temple]

[modest fog]

Grabby ears

[nova remnant]

Hi

[terse ore]

I made this blender render

I used Grounding Dino and Segment Anything to make this mask:

I put in "red horns" and also put white black blue and some stuff in the negative. But it did not make the horns red and seems to have kept the same colors. How can I make them red?

[modest fog]

Is it too low denoise?

[terse ore]

I changed something but then it does this

[terse ore]

Or it will do like this

[modest fog]

Hmmm

Try higher denoise but with controlnet canny ?

It should keep the shape

[terse ore]

I tried Denoise to 1, but let me try canny. I did it with pony so I'll change to controlnet

[modest fog]

Denoise 1 will completely change the image

Or try to

Try at most like 0.75 so the background doesn't get screwed

[terse ore]

Ok

[terse ore]

But the mask is pretty precise with segment anything so the background will be ok I think

[terse ore]

Hmm the foreground horn is still broken up

[modest fog]

But it can not use the full mask to create only the horns

[terse ore]

Yes that is the problem

Also as you can see in the original render, and the mask, that the bigger (close) horn is also not inpainted correctly. The inpainted result makes it much shorter

[modest fog]

I don't really know

I just use controlnet and it works

[terse ore]

I tried controlnet before, didn't fix it...I'll try again

[modest fog]

@terse ore

Like

[terse ore]

You ing2img the whole picture?

[modest fog]

If you combine with image to image I think it works but will definitely lose the 3d render style with too much denoise

[modest fog]

Yes

Just showing the horns

[terse ore]

Oh ok it is not bad. But you can see the horn closer to the viewer is supposed to be higher near the top of the image.

The original render has the horn right near the top of the image edge

[modest fog]

Probably because I didn't put 1.0 strength on the model

[terse ore]

Well I get OOM with SDXL controlnet with 3060 12gb. Let me try 1.5. I already tried and it wasn't the best but who knows maybe this time...

[modest fog]

Also I'm using vertical resolution lmao

Am stupid

[terse ore]

I did resize scale 1:1 so I dont have to play with resolution

Better result this time...I switched to a 1.5 inpainting model

[modest fog]

img2img, canny and depth

[terse ore]

Oh multi controlnet. I had forgotten about it. It is taking my PC a whole to just do 1 gen with 1 controlnet...

Hmm this one did not color it but the shape is surprisingly correct

Well the reason I am doing it 1 inpaint at a time is because I want to try making a consistent manga

[modest fog]

@terse ore

👍

[terse ore]

Wow scary

[modest fog]

She's been posing there for 5 days straight

[terse ore]

How did you do that? Multi controlnet?

[terse ore]

What is your GPU? Takes my PC a few minutes to inpaint

[modest fog]

You know I'm Rich bunch of A100s

(am using online service)

[terse ore]

Oh I see

[terse ore]

Another correct shape but not color 🤔

[modest fog]

Are you doing image to image

Also you can zoom in on the horns if that's the only thing you want changed

Do like a 512x512 crop on it then stitch back

Image to image will try to keep the same colors depending on the denoise, I was doing at 0.9

[terse ore]

It is inpaint

[terse ore]

I could change the color of the horns to red in the character maker and then inpaint I guess

The stitching horns back like via an overlay is a good idea and I did it with the face. But it would take longer especially when making many panels

[slate narwhal]

mutant ninja monkey in the background:

[near elbow]

prepare to be boarded! (SFW enough i guess)

added "ninja turtle" because... ehm... ehm...

[steel zephyr]

I did something to my prompt and now it's near perfect

[green temple]

big but still nice 😃

[steel zephyr]

can't make nai make her smaller, only a lot smaller

shame there's no gramophone Miku lora. I'll try to make one myself one day

[green temple]

go for ( child : 1.5) ( loli : 1.5)

[steel zephyr]

I've made some tests, loli, child and young girl is about the same. Only toddler works, but that's not my type

[green temple]

[steel zephyr]

Seems fine if you mix all of them

[modest fog]

Try aged down

[modest fog]

[steady grail]

@wispy canopy @bold ether @wheat falcon

[modest fog]

What going on

[steady grail]

spam

the gentleman already cleaned it

[gloomy scroll]

Migus

[keen hollow]

no Thanksgiving anime images?

com'on crew

you let me down

[modest fog]

these don't happen here

[wispy canopy]

@modest fog poe2 hype. do you have your key?!

[modest fog]

dont have any moneys but friend got me one

[modest fog]

smolsus wants one

[wispy canopy]

that would mean i'd have to upgrade. again.

i already gave out a key 😮

[vital raptor]

I tried, but aside from 'turkey costume' (which doesn't work well) I couldn't think of anything that signifies 'thanksgiving' isn't just a generic table of food

this is turkey outfit 😄

it just does a lot of bread and sometimes a mushroom hat

i guess now that it's over I can start doing christmas ones though

[keen hollow]

I appreciate the effort. Well done.

[modest fog]

@green temple

[bitter latch]

umm ...

haha

[bitter latch]

For Freedom!

hmm kinda spider like... shivers

[near elbow]

just combine it nicely 🙂

[green temple]

reminds me of shary from seiken gakuin no makentsukai

[supple raptor]

"stuffing the turkeys" am i doing it right?

[vital raptor]

works for me 😄

[keen hollow]

is this the one you did before?

[supple raptor]

it's a prompt i've done before, but that was the first try with it on noobai eps 1.1 and with a style lora instead of the artist tags i've used with it before

[keen hollow]

looks kinda like a Guweiz style

[bitter latch]

loads up koboladAI

watches gpu run out of memmory. .doh!

[visual sapphire]

eva in the dragon ball

[bitter latch]

ok umm.. nm

Miku DeadPool Crossover

CHAAARRRRGGGGEE!

DOOOOOD

haha breaking the 4th wall

ok thats cool

[terse ore]

What controlnet should I use if I want to inpaint a mask but it's outline, not the stuff within it?

[modest fog]

[terse ore]

I changed resolution to 1024 so I could use anime lineart preprocessor and controlnet. Played with settings and this is one of the better results of changing the blender render's horn color to red.

Next step is I think to refine it. You can see how the color doesn't fill in the horns. I think it would need iterative upscale to fix it?

[modest fog]

[modest fog]

looking decent

could do a second pass of the same thing again on the new image ? maybe that works

[terse ore]

I think it changes the image too much? Idk

I'll try

[bitter latch]

hmm looks like Second Life

oooh I tried to do img2img with no prompt and it made'em topless , can't post tht

[terse ore]

I think it's better if I segment the entire character and then put a prompt for the whole, ie red horns, white buttonup, etc.

Easier to segment and prompt blocs at a time. Plus it has more context so it works better

[terse ore]

No controlnet? But for single images it's fine. Just if you're making multiple panels it's different

[bitter latch]

I'm using forge. never messed with control nets

shrugs

[terse ore]

Oh I see. Yeah that image looks much better than render

[terse ore]

Not perfect

If I can generate something that will be good enough to clean up by hand then it will still save time

The face can be done separately via an img2img close-up and pasted on top

Also, can't use detailed patterns like those striped pants. Single color is easier

[green temple]

big-smol 🥲

[modest fog]

noo

@green temple now more smol

@green temple why that face

What did I do wrong

[green temple]

still big-smol

[modest fog]

How

Just wait

[green temple]

[modest fog]

@green temple This is it

Smol!

[vestal ferry]

https://tenor.com/view/doro-nikke-gif-5383253755470222749

[green temple]

smol died !

[vestal ferry]

always

https://tenor.com/view/doro-dorothy-nikke-gif-6198201261302376921

[green temple]

[modest fog]

He bad person

[green temple]

help

[modest fog]

I'm calling the police

[green temple]

finally bloo

[vestal ferry]

https://tenor.com/view/nikke-doro-dorothy-pipe-bomb-gif-4643514375044855612

[vital raptor]

ffs, my chair just broke and I can't get a warranty fix because the company went out of business

Just as a heads up to anyone that might be shopping for a chair, don't buy AKRacing stuff 😄

Super nice chairs, especially for tall people, but they're gone now

[modest fog]

Mine has been broken for years

[vital raptor]

the welding that holds the back on the chair snapped, basically just a stool now

second time it's happened, but they replaced the back last time so I have been happy with it

$500 chair, didnt even last a year

[modest fog]

@green temple gaming smol

[modest fog]

Gaming stool kek

[vital raptor]

I wish i was smol so i could buy cheaper stuff that doesn't break

[modest fog]

I'm fat

I think I'm at the weight limit for my chair

So it just got weaker and weaker

But it's a cheap one

[vital raptor]

I'm a bit fat, the bigger problem is my height, normal chairs aren't tall enough

so i spent a lot on a big tall chair, and it's broken twice

[modest fog]

Mine is not a stool but without shenanigans it basically is

I have to put like a pillow and rest it on something so it's possible for me to rest my back on it lmao

[vital raptor]

lol

well, looks like i'm getting amazon to give me most of the money back if i ship the chair to them

now i have to find a new chair though

[vestal ferry]

Smol

[vital raptor]

ugh, $569 for a new chair... merry christmas to me i guess

[modest fog]

expensive

[vital raptor]

yeah

[terse ore]

No craigslist?

[vestal ferry]

smol

https://tenor.com/view/doro-nikke-gif-5383253755470222749

[modest fog]

@green temple he attacking smol again

[vestal ferry]

https://tenor.com/view/non-non-biyori-renge-walking-anime-baby-gif-16138179

[modest fog]

get away from me

[supple raptor]

https://tenor.com/view/loli-eat-you-loli-kawaii-anime-you-die-gif-16032459

[vestal ferry]

https://tenor.com/view/die-hachikuji-monogatari-punching-anime-gif-26529250

[worthy basin]

hi

im new

[bitter latch]

it's cold

[humble current]

I saw what you did

[humble current]

she looks familiar...

[green temple]

4 mins x1 upscale

[humble current]

let me try

[green temple]

30 step 5 cfg

[humble current]

I just finished this one at 1.8x upscale...

[green temple]

for me 20 step with 5cfg takes 3 mins

[humble current]

1x upscale 20 steps 5 cfg takes me 44.7 sec...

but the details aren't as good

[green temple]

yeah ofcourse

[green temple]

👍

[green temple]

how much time this took ?

[humble current]

2 mins 44 secs

[green temple]

👍

[green temple]

this is my generation time with hires off 😂

[humble current]

what gpu do you have?

[green temple]

gtx 1660 super

[humble current]

ah I see

[humble current]

I decided to do 1.8x upscale 40 steps and 40 highres steps and it took me 4 minutes

[green temple]

noice

[humble current]

I have a 4060. (8 gigs of vram....help me). I might upgrade when the 50xx comes out

or not if they turn out to not be a big leap

[green temple]

mine is 6gb vram 🥲

[humble current]

damn

I swear if the 5060 card has 8 gigs I'm gonna flip

[lapis orchid]

so i just started diving generation so im a noob and i was just curious on how many times to apply the hires fix. i was experimenting and noticed that it made some improvements but made the hands worse and worse. i was also using 30 steps in each. i also keep the height and width the same. if that makes a difference.

[vital raptor]

hires fix is for upscaling, if you aren't changing the resolution then it's not needed

although it does help with refining the image a little

[gloomy scroll]

Highres. fix is img2img

[lapis orchid]

ah so it really isnt giving a return doing it more then once

if at all

[vital raptor]

it could, but probably not worth the time, especially with that many steps

I used like 20-30 steps for the initial generation, then if i want to keep the image, i turn on hires fix x2 for 7-10 steps, anything more is mostly useless

[lapis orchid]

im using xlpony if that helps

[lapis orchid]

ah so perhaps i was using too many steps aswell. i noticed when i turned the sample steps to 50 i got alot of extra random stuff.

i was thinking more was better xD

[vital raptor]

depends on the model and what you want out of it, there's no 'right' answer

but, hires fix specifically, there's very little difference beyond 10 steps, and given the extra time for a higher resolution it's not worth it, it's faster to just generate more images instead

20-30 steps is most common for most XL models, changing the sample and cfg also changes the results, so finding which you like best just takes a lot of testing

[lapis orchid]

so 30 steps is just too much for hi res

[vital raptor]

I use autismmix_confetti, a ponyXL based model, and usually generate at 25 steps, my upscaling/hires step is different though since I don't use auto1111/forge

[lapis orchid]

yes i am using auto1111

[vital raptor]

but also, i'm doing some funky things with prompt interpolating so i needed a few extra steps, 20 steps is what i used to use

[lapis orchid]

i got into this because i play at the cardshop alot and buy alot of proxies so i thought it would be a cool hobby to make my own. but now im just having fun making my own anime art.

[vital raptor]

CFG:6, Steps: 25, Original Resolution: 896x1152
HiRes Fix would be equivelent to 10 steps with 0.2 denoise with 2x upscale

[lapis orchid]

the models i been using mentioned 1080/1360 im guessing i could make the images smaller? because what im also wanting to do is make them more wallpaper sized. but i was getting mutations to say the least.

[vital raptor]

XL models are trained on 1024x1024 base image size, certain ratios are generally equivalent like 896x1152, 832x1216, 768x1344, 640x1536 should all be fine to use

anything outside those might have problems, if you want something like a 1080p wallpaper then something like 768x1344 then upscale 1.4x, you can also try 540x960 and x2, but it might not generate right

[lapis orchid]

oh thats a good point.

768x1344 is pretty close

[vital raptor]

i use 896x1152 because it's standard 3:4 ratio which is the same as a lot of cameras, and i started doing this as an alternative to photography so it's just a ratio I like

[vital raptor]

I mostly do portrait style images anyway

[lapis orchid]

oh thats a cute one. hopefully i can make ones like that after some practice.

right now im just learning prompts and how everything functions. i been using alot of references for help aswell

controlnets been super helpful if im wanting something more specific

[lapis orchid]

so one more thing. is it better to generate a sampler then hi res. or do them together? it gives me the option to toggle it on and off

[vital raptor]

i usually generate with it off, then turn it on and re-generate the image if i want it bigger, but i'm impatient and want to generate through as many images as fast as possible

[lapis orchid]

yup! thats me aswell xD probably wasted alot of time. i been generating that image alot trying to get a good one.

[vital raptor]

it shouldn't make any difference results wise, just whether or not you want to wait for the full resolution version if you arent keeping it anyway

I generate a lot of images some days, and for me the difference is 3.5s or 10.7s, which is nothing like what some of you have to deal with

waiting ~11 seconds for a single image when i could have done 3 in about the same time just doesn't make sense unless i know i want it, at which point you just click the upscale button and wait the extra time

if i remember right, a1111's upscale button just regenerates the current image with the hiresfix settings so you don't have to toggle them