🍥｜anime

[bitter latch]

in some countries, birds are also woman

women

😄

there thats better

I put the AI created songs in #🌶｜off-topic

[coral fox]

Any luck creating 3 distinct characters in 1 picture ?

With different clothes, eye color, hair

[supple raptor]

regional prompting can do it. Usually I find it easier just to gen 3 separate images, photobash, then inpaint/img2img if necessary.

If they're existing, well known characters, I've had pretty good luck doing 2 at a time with NAI.

[ivory pebble]

What prompts do you use to get 2 characters on NAI? I haven't had much luck with that

It usually just combines the 2 characters into 1, even after using tags like 2girls, 2boys, multiple girls...

I'm probably missing some prompts/settings so that's why I'm asking

[supple raptor]

that one was:
(artist:quasarcake:1.1), (ebifurya:0.7), (artist:wlop:0.9), (2girls:1.1), komeiji koishi and komeiji satori, standing, back to back, holding hands, touhou, black hat, hat ribbon, hat bow, yellow bow, green hair, green eyes, heart-shaped pupils, waving, yellow shirt, frilled shirt collar, green skirt, pink eyes, pink hair, black hairband, blue shirt, frilled sleeves, pink skirt, heart of string, (third eye:1.1), simple background, black background, best quality, amazing quality, very aesthetic, absurdres
I use NAI inside comfy via API and have a node that converts from Comfy prompt syntax to NAI so you'd want to change those prompt weights to NAI syntax if you want to prompt it on their site. each {} increases weight by 0.05 and each [] decreases weight by the same amount.

[ivory pebble]

Seems like adding "and" does the trick, that's the only thing I didn't try last time

[rough oxide]

2 characters should just work, i've done it a number of times

but

i always use landscape when using more than 1 character

...most of the time,. left one there is a square

works with creatures too easily lol

[ivory pebble]

It's working now, I added the word "and" so now it's consistent at the very least. Still very luck based when it comes to outfits however.

[rough oxide]

and then yea it just works if they're well-represented in data

if a char is like not very represented, the highly represented char will takeover their elements

is what i just saw

[ivory pebble]

Yeah that seems to be the case

[rough oxide]

this worked fine tho as an initial test

[ivory pebble]

Also apparently it wasn't the word "and" that changed everything. Just prompting their names without it seems to work as well, don't know why it didn't work for me the last time I tried it

[rough oxide]

ah yes, that isn't etna lmao.

although now it sort of is just wrong color

it works fine for generic characters at least.

eventually worked with same prompt tho

[green temple]

cool

[supple raptor]

yeah, landscape does make it easier since it naturally wants to fill the extra horizontal space.

[nova remnant]

not smol

[rough oxide]

i somehow clicked here right as that was being posted.

intuition.

[rough oxide]

i r addicted to shinx. any time i make a pic with a pokemon it's always shinx.

[bitter latch]

there thats better

hmm trying for panda pokemon

[bitter latch]

goes to bed.. thud.. ow

[coarse brook]

Sooo... When r we gonna see one of those smol things cry, cause they couldn't eat the entire normal (to us) sized cookie?

[coarse brook]

No bonk, only cookie

https://tenor.com/view/champagne-barbie-cookie-monster-gif-21025884

[bitter latch]

cackles

brain tired, brain creative, shut up brain. need sleep

[vital raptor]

time to setup a1111 on my new server and see what kind of generation speeds i get

[bitter latch]

random miku , just cause

[vital raptor]

should be interesting to see how cpu rendering does with 56 threads and 128GB ram 😄

[bitter latch]

slow?

[vital raptor]

possibly

[bitter latch]

hmmms

[vital raptor]

cpu is slower than gpu, but most comparisions i've seen are using normal single cpu consumer hardware

[bitter latch]

nods

[vital raptor]

this is a dual cpu server with 28c/56t total

[bitter latch]

BUT will it be limited to the one cpu or is it going to multi-thread?

[vital raptor]

not sure actually, i'd imagine it's multi thread

[bitter latch]

to be or not to .. crash.. burning smell

reboots "do you want to .. play .. a game?"

[vital raptor]

but, even if it's super slow single core, it has 128GB ram (upgradable to 3TB) and it's a server so it can take as long as it needs 😄

i can use it to render much larger resolutions at least

[bitter latch]

hehe

[vital raptor]

the next step might be selling my 4090 and getting a datacenter card to replace it..

[bitter latch]

cracks up due to lack of sleep

....

cackles

[vital raptor]

my roommate was doing images of mad max muppets

it was kind of hilarious

[bitter latch]

there better

oh mad maxx you say?

same seed turned the other way...

haha

[vital raptor]

lol

[bitter latch]

umm o....k

ok enough of thta

meanwhile in the world of fallout 76 neo tokyo

ok this new merge is doing stuff I didn't think it could oooh

I would watch thta anime

[bitter latch]

there.. now i'll go pass out now...

you ate the lsat cookie....?! prepare to die

[vital raptor]

she needs a banana

carrots are dumb

[green temple]

everyone is eating cookies today

[vital raptor]

eepy

[vital raptor]

ffs, a1111 won't start in cpu only mode if there's no gpu installed

because, why not

apparently i need to find a gpu and install the drivers so i can not use it anyway 🙄

[coarse brook]

HOLD ON A MINUTE! That is...

[gloomy scroll]

🍌

[green temple]

cute

cursed doll

?

[wooden coral]

HunYuan def has potential

[green temple]

bad legs nooooo 😢

[nova remnant]

I can fix them for you

Here

👍

[green temple]

thx

[dense geode]

I don't know the reason but many of the pictures I have it generate show too much skin.

The model is Animagine XL (version 3.1).
Anyone an idea?

[nova remnant]

I have an idea, yes!

Your model is

So put stuff like "nsfw" in negative

and don't prompt for things like "thighs, shoulders"

They will expose more skin than what you think xD

Positives like "SFW" should work on non-pony based models

On pony you can just do "rating safe" and it takes care of most of it

[dense geode]

I did not specify any other body part than the head.

[dense geode]

Thanks. That should work.

[nova remnant]

if theres still too much that you don't want showing, "revealing clothes" is a tag, it should also work on negative

[dense geode]

It's just that I don't want the picture to be 50% skin. That doesn't make for a good presentation.

[coarse brook]

Ok ok. I know its not Anime, but... I have these 😄

[coarse brook]

That first Cowboy chick... Thats an Ashe, if i read that figure correctly, ye?

[nova remnant]

yeah it is 😼

[coarse brook]

Knew it 😉

No Bob?

[nova remnant]

ult is charging

[nova remnant]

kot

bad pony

[nova remnant]

Is that mai

stupid character

im noob, shes too hard to play

[coarse brook]

IS SO SOFT THOUGH!

[supple raptor]

Cat with 4gb's suggestion is good. The simplest solution is probably to make sure you are prompting for what you want to see. Lack of clothes? Prompt for the clothes you want and that probably won't be a problem.

[nova remnant]

brings flamethrower

[green temple]

؟؟؟؟ 😂

[nova remnant]

Soon there won't be any limbs, just booba flying

[ivory pebble]

The most impressive thing about this image is the fact she's able to move like that with those melons tbh

[coarse brook]

Big Unicorn plush!

[green temple]

👍😂

[nova remnant]

pony

youre a brony now

[coarse brook]

God no.
Its a derpy unicorn, nothing else

[nova remnant]

its over

you'll slowly accept the fact

[coarse brook]

Is been over 4 yrs, and over 8 for my pfp.

Still haven't seen a single episode

[nova remnant]

no one other than clueless kids would actually watch my little pony right

Riiiiiiiiight

https://media.discordapp.net/attachments/733762994525306972/899670992023076926/image0.gif

anyway, a draenei

I like their tails

not sure what it is even lmao

[bitter latch]

demon

[nova remnant]

she didn't do anything wrong

[bitter latch]

didn't say she did

huh that came out pretty good...

[bitter latch]

anime van gogh

[fringe granite]

#🏞｜general-with-images

[bitter latch]

print this out and hang on wall and confuse people

cackles

Anime simpsons

[bitter latch]

it actually did it !

[bitter latch]

haha

[nova remnant]

Don't look at me like that

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

[nova remnant]

2 smol 2 eep

[versed jay]

What does checkpoint mean? because the lorax is a pony, I don't understand why that comes out, those are the requirements

[ornate echo]

Checkpoint = Model

[magic wave]

A Spooki

[bitter latch]

makes his own version

[magic wave]

XD

She charging up the super chomp

[bitter latch]

shivers that is.. covers eyes nooo

ok no more of that

Ramen comic?

[nova remnant]

Luna

[zinc flax]

https://tenor.com/view/rio-blue-archive-rio-pat-rio-tsukatsuki-gif-5097930668510294900

[magic wave]

Me :o

[nova remnant]

[green temple]

cool

[rough oxide]

shinx is most adowable pokemon evrawr.

[coarse brook]

Goat people

[green temple]

🤣

[bitter latch]

ah ha

[green temple]

Oh !

[bitter latch]

😄

ok not sure where the ars came from.. tho prob the model

changed models and hmm ok...

sigh, nice but not what I wanted

[bitter latch]

hmm I sorta got it down

[twilit portal]

Steampunk

[bitter latch]

hehe

[bitter latch]

little sister?

[twilit portal]

[bitter latch]

wow alot of detail, now if it made sense that would be great

there more anime-ish

and with that i'm gonna lay back down

ok more safe

[green temple]

🤣

[magic wave]

Clean pic of one of my OCs with relatively high steps
that one guy that told me steps higher than 30 and 8 for hires fix dont matter lied

Takes a bit long with SAG (Self-Attention Guidance), PAG (Perturbed-Attention Guidance), Vectorscope, and LoRAs, but it looks very clean

[steady grail]

cfg matter a lot

but for the steps, the bigger the resolution, the more it matters

at 1024, after 30, the differences are trascurable

[magic wave]

I mean, that one is uhhh, 832x1216, 50 steps, 20 hires, 4 cfg (becuse PAG)

Also at 0.5 hires denoise

[steady grail]

guys, someone that played with deep cache?

i'm testing settings with it, but results are wildly different

[steady grail]

benefit from deeph chake looks really trascurable

[unkempt cape]

Anyone got a good colab for gelbooru dataset gathering? I want the tags associated with the images too

[steady grail]

1.5?

[magic wave]

SDXL :o

[steady grail]

with cfg 4 i suppose you are using some lora for low step gen like lcm, lightning ecc.

[magic wave]

I usually use DatasetProcessor cause it usually works better, some pics done have many tags, and also its only really worth having tags that describe the pics, artist tags or quality tags are bad to have
For LoRAs at least, for Checkpoints I got no clue

[magic wave]

PAG requires lower cause otherwise it fries the pic

[steady grail]

pag?

sag?

i didn't6 delve too much in it. i didn't like it for some reason i don't remember

[magic wave]

There is SAG and PAG, SAG is decent, I like cause its more contrasty, PAG does a few things, some things I don't like, but it improves how well small details are done, like how clean they look

Thats what I've noticed the most about it

[steady grail]

oh

i like that

suggestions in setting it?

[supple raptor]

https://stable-diffusion-art.com/perturbed-attention-guidance/

[magic wave]

The default settings work pretty well imo :o

[magic wave]

Hai nubby

[supple raptor]

Hai LUNA!

YuukaWave

[steady grail]

oh, itwas nubby

hi

confused you for smile XD

[supple raptor]

Hi Shaikorth

YuukaWave

[magic wave]

Haiii, how you been nubbers, been up to much? :O

[supple raptor]

I've been pretty good. Just heading out the door now to take a friend to the car dealership so he can hopefully buy car.

[magic wave]

Ayyy thats good, and thats pretty nice, hope it goes well :o

[bitter latch]

and for some random

[nova remnant]

❤️

[nova remnant]

I'll grab her ears

[nova remnant]

Needs thicker black bars

[vital raptor]

fixed

[nova remnant]

She looks stupid with star face

[ivory pebble]

No u

[nova remnant]

not cute cat

[nova remnant]

why do you want ugly cat prompt

[ornate echo]

Hi furrball.

[bitter latch]

Sith cat

yawns, ok haven't slept well going to bed. nite

[unkempt cape]

What is a DatasetProcessor can u give link

[magic wave]

This
https://github.com/Particle1904/DatasetHelpers

Its pretty good

[unkempt cape]

I want the tags of the images on the site not generate it with a tagger

[green temple]

which is better ?

[green temple]

i`m using tile controlnet + sd ultimate upscale

which preprocessor is better for keeping the same color

i used colorfix+sharp but i dunno how to use it

[vital raptor]

looks like the second ones have been 'smoothed' in a way that some of the smaller details get erased and some of the shading is changed, they're also smaller resolution, not sure if that's intentional

[vital raptor]

overall i like the first version more because the 'messy' details seem to match the art style better, but i like the colors on the second ones a bit more

[green temple]

yeah unintentional

[queen void]

Mhhhh interesting...

[magic wave]

https://github.com/Bionus/imgbrd-grabber If you really wanna do that you could try playing around with this :o

[magic wave]

Its pretty nice :o

[queen void]

I might look into it and upgrade my own scripts if it is better

[unkempt cape]

Yeah been trying to use that

understandably it doesnt work on my fucked up pc

[magic wave]

Fair :O

[magic wave]

Might be issue of the app itself, after a update it has been starting to crash a lot for me :o

[unkempt cape]

🤒

Ok its working

And then ill use whatever app u put to resize them (which is very kul)

What should be the resolution guys im think 768*768

[magic wave]

For 1.5 768x768 is the best yeah

But you don't need to resize them

Buckets already do that

[green temple]

bad inpainting 😭

[unkempt cape]

Xl

Reason i ask this cause i havent seen alot of 1024*1024 outputs

[unkempt cape]

Im training with colab so like is it an automatic process that just happens or something i have to configure?

[magic wave]

Oh, yeah for XL you wanna do 1024x1024

[magic wave]

It happens automatically

You may see somthing relating to this somewhere

That automatically resizes

[unkempt cape]

Oh ok

[magic wave]

Yeah

Does anyone know how to speed up ADetailer model loading? Its taking so long for no reason and taking 2 min gens to 4 mins
@supple raptor do you know?

[vital raptor]

adetailer does a gen on top of your gen

it generally makes it takes twice as long

on the other hand, i'm guessing you're using forge since it's moving models in and out of the vram, so that adds more time

[unkempt cape]

Bruh i misremembered the number of good images for lora, it was like 200->400

[vital raptor]

you can try turning that off so it doesn't move the model, but you'll need enough vram for it

[unkempt cape]

Yeah i misremembered and did 600

[zinc flax]

20 is also enough lol

[vital raptor]

i trained my lora on like 8, but it was a very specific face, nothing more

[magic wave]

O, oooof

[magic wave]

How do I do that? I do have 24 GB of VRAM

So I don't need to be too economic on my VRAM spce xD

[magic wave]

Also what are you training with that much? :o

[vital raptor]

i think it's --always-gpu in the commandline args

[magic wave]

Imma try that

[vital raptor]

unless they changed something since i last used it

i had a similar problem with my 4090 with forge, it was horribly slow

[magic wave]

It seems like its that, and yeah, when it load fasts it does like, 2 mins with high steps and a bunch of stuff that increases the time, but one model swap took like 188 seconds, literally 3 minutes on the model swap

[vital raptor]

yeah

in my experience, forge is basically hot garbage for anything over 8gb vram (last i used it) they put so much effort into lowend cards they basically broken highend cards

[magic wave]

Oh really? What do you use then? Normal A1111? :o

[vital raptor]

yeah

[magic wave]

Fair, I haven't tried it since I got my 3090, maybe I should :o

[vital raptor]

I use TensorRT to make my gens faster, and forge straight up won't work with that either

[magic wave]

TensorRT? :o

[vital raptor]

it's a pain in the ass to install properly, plus generating the onnx engines for your models, but if you only use a couple models it's not so bad

reduces my base XL gen times from ~6s to 3.5

[magic wave]

I mean I usually use uhhh, one model if you mean like just Checkpoint :o

[vital raptor]

yeah

some people change between models a lot, i only use 2-3, so i only had to build 2-3 engines, they take about the same amount of space as a normal checkpoint

[magic wave]

Now my gens are 40 secs, god dammet forge, the swaps were taking way longer than I thought

[vital raptor]

also loras 'might' work while using it

[magic wave]

"Might"? xD

[vital raptor]

with TRT you have to build the onnx engine for each model/checkpoint, and then 'convert' loras to work with TRT, the extension is buggy though so sometimes it says it converted but it didnt actually do anything, so that lora won't work with TRT turned on

[magic wave]

Oh, so you gotta like just convert it a few times to make sure? xD

[vital raptor]

no, it either works or it doesn't

[magic wave]

O

[vital raptor]

something about how the lora was created/saved that is basically wrong or non-standard, so it just wont work

[magic wave]

I mean, though that would mean if one of my LoRAs work, all of them should be good

[vital raptor]

i rarely use loras, so it doesn't bother me much, and if i need a lora i just turn off TRT for those gens

[magic wave]

Fair :o

I think now I'll be swapping over to A1111 and see how that goes, what are good args for that? I assume --high-vram is one :o

[vital raptor]

not needed

i use --xformers so i can use it when i'm not using TRT, but otherwise nothing else performance wise

[magic wave]

Fair, thats the only thing I had so I guess I'll leave that there :o

[unkempt cape]

Idk i just download a fuckton of gelbooru inages

I was going to download an amount somewhere in the limit, but i pretty much misremembered and doubled the amount

[bitter latch]

your morning does of random

MUKE?

[magic wave]

MPICS

[magic wave]

Fair :o

[zinc flax]

Lowest I tried on character was 15

Some madlad tried 1, I should also do that lol

[vital raptor]

well, if you don't do enough then your angles are limited

at the very least

if i generate a girl looking away from the viewer so you only see just a bit of a side view of the face, it looks really bad

because i didnt train that angle

[unkempt cape]

I mean look if theres anything we learned from llama 3, its probably that recommended training dataset size is bullsh

):

[magic wave]

I mean, 150 works pretty well for me, I usually do less, but when I can do 150 I do 150

[magic wave]

@vital raptor Do you know anything about OOMing like dat?

[vital raptor]

yeah, it usually happens when you're generating too big

generate smaller and upscale after if you want to

[magic wave]

I'm generating 832x1216, that isn't that big

[vital raptor]

no hires?

[magic wave]

With x2 hires :o

Though I think its OOMing at the VAE stage

[vital raptor]

so you're generating twice that

[magic wave]

I usually do on Forge :o

[vital raptor]

i can usually do 1024x1024 with 2x hires without a problem

[magic wave]

Fair, I guess maybe A1111 isnt too good with 832x1216 2x hires :o

[vital raptor]

it does fine with that

[magic wave]

Huh, thats weird then, dunno why its OOMing :o

[vital raptor]

no xformers, or you're using the old broken vae maybe

not sure what else it could be

[magic wave]

I do have xformers on, but lemme try with another VAE :o

[vital raptor]

that res is also pushing my vram to 23.3GB, so if you have anything else using vram at the same time it's gonna go over

[magic wave]

OOMed again, and ohhh, ooof, guess I'll stick to Forge then, cause thats like my usual res :o

[vital raptor]

realistically it should just overflow to the system ram, it'll take a bit longer but it'll finish

dunno why your setup isn't doing that

[magic wave]

I mean, I got 16 GB RAM, could that be why? :o

[vital raptor]

if you're using your vram for other things AND only have 8gb system ram (it only uses half) then yeah, that would be it

[magic wave]

Fair, I do plan to upgrade to 32 at some point xD

[vital raptor]

I have 64GB, so i rarely hit that unless i'm pushing resolution really high

[magic wave]

Fair, I would go for 64 but I think 32 x2 is too expensive to buy here :o

[vital raptor]

why that resolution though? seems a bit odd

[magic wave]

Well, I saw that the NAI "normal" res is 832x1216, and I was like "I wonder why", I tried it and it works better, specially with styles that have clean lineart, like anime screenshot type styles

[vital raptor]

but you're not using nai

also, the current nai model is just sdxl, 1024x1024 is normal base res

others are supported, but it's based on the training material mostly

[magic wave]

I mean, I'm not but, it looks cleaner for styles like that so I kept using it :o

Tis 1024 vs 1216 :o
I mean its also a higher res but imo the lines look cleaner :o
Though its also ArtiWaifu so its not the best example, but I've been playing around with it and I'm lazy to swap xD

[vital raptor]

yeah, whatever makes it work, i had a 1.5 model i could only use in 512x768 otherwise it didnt look good

[magic wave]

Fair, yeah on 1.5 I think most are limited to 512x768, though SDXL is a bit more flexible

Fo some wide pics I sometimes do 832x1536

Like wallpaper style

[vital raptor]

it wouldn't even do 512x512 properly, but yeah i've done some interesting sizes with sdxl

[magic wave]

Oh damn, now that is funky :O
And yeee, can't wait for SD 3 though, but I'm kinda cynical on it so I'm skeptical on if it will perform as good as it seems :o

[vital raptor]

ultrawide is always fun

[magic wave]

Yeee it is, ArtiWaifu seems good t it and specially backgrounds, but I always do white background xD

[unkempt cape]

got 600+ prompts to fuck around with now lol

[nova remnant]

looouna

you appear here agane

[magic wave]

I do :o

[nova remnant]

Grabby gen

👍

[zinc flax]

Gotta love twitter trend lol

[nova remnant]

twitter

https://tenor.com/view/cringe-gif-24107071

[green temple]

coolf👍

[nova remnant]

We're lacking smols today in chat

[green temple]

[bitter latch]

oh?

[green temple]

i want to eat her face 🥹

[bitter latch]

tried to do hands out grabby but didn't work

hmm wonder if I specify the park will it know... lets see

I watch a live streamer who visits some of them so i've learned the names of alot of parks around Tokyo

and yeah I think this is the back side of Ueno park

hmm it doens't wanna do ninja run

an effort was made

so how am I doing?

ok off to work be back in like 3 or so

[vital raptor]

nom nom

[magic wave]

Well, I'd say :o

[nova remnant]

wtf

[bitter latch]

wow

[stray jay]

generate a random picture

[radiant flower]

thats good

[unkempt cape]

💀 i was today year old when i found out dall e 3 already accepts image input

img2img was already a thing from the start and not a 4o thing

[supple raptor]

CirnoPopcorn

[autumn violet]

Filo

[supple raptor]

https://tenor.com/view/the-rising-of-the-shield-hero-shield-hero-filo-god-bird-gif-18523081

[queen void]

https://tenor.com/view/filo-tate-no-yusha-anime-the-rising-of-the-shield-hero-gif-21376260

[bitter latch]

The adventurers finally find the dragon king on his thron

there's a flash and.... .

and

and i'm off

[green temple]

😍

[nova remnant]

grabs

[unkempt cape]

doing some gens with Kaggle siu

[unkempt cape]

any wd news?

[paper sluice]

I'm new to SD and a1111. My prime focus is I want to colorize webtoons/manga. What models/approach you can recommend to avoid loosing consistency?(and make the model only colorize) Here are a few tries that were the most successful

[steady grail]

it's hard to say

i attempted to do the same and failed miserably

the most problematic part is the bg

[rough oxide]

you could probly do what you want with controlnet

but i'm not super well-versed in controlnet stuff

[steady grail]

well, there is a contronet for colors

there is even region prompting

but setting everything is very time consuming

testing new workflow XD

i'm looking if i can smoothly run xl on a complex workflow

i spend more time making workflows than genning XD

[slender barn]

trying a new workflow, I haven't really gotten crowds down 100 percent, might have to make a body diversity detailer, so I can get more variations in the crowds.\

[steady grail]

you can make a detailer for every body?

like hooking an entire body on the bg?

[slender barn]

yea, you can do all of different parts of the face, the body, and hands(I think there was a a feet/leg model too)

[steady grail]

wow, didn't know there were so many adetailer models

interesting

[slender barn]

I think this once OOM before finishing the Detailer past, still looks kinda cool.

[steady grail]

anyway, the idea is cool

using region?

[slender barn]

yea I am using area composition to try to get some more complex scenes, compare to regular prompts.

[steady grail]

yeah, makes sense

i would have done the same

i still didn't attempt anything like that

[slender barn]

I am still working on it myself, still hoping to get a little more out it.

[steady grail]

i'm more involved in adding depth to the image

i feel that 2d feeling plain

[slender barn]

that sounds awesome, how does that work, what does it add to the image?

[steady grail]

i first gen a bg and a chartacter, feed them to depth controlnet and mix them to the final ksampler

there is an exaple with the workflow

it works amazingly well, but it's very slow

so, now i'm working on a workflow without any cn to speed up stuff

[slender barn]

aawesome stuff, yea alot of this stuff is super gpu heavy. stuff can get slow.

[steady grail]

as predicted, only depth cn can gen that effect

[steady grail]

i wonder if i can work directly with noises to obtain such an effect

[steady grail]

how much region impacts on the gpu?

[supple raptor]

yeah, adetailer comes with person_yolov8n-seg and person_yolo8s-seg, both of which do bbox + segmentation masking for entire people. it's not quite as effective as just face detailer is in most cases since it's typically not upscaling the image as much, but you can always do person detailing first and then face detailing second.

[steady grail]

oh, that's very interesting

just testing the workflow that prompt bg and character separately. the result is promising, but without the different depths, the result is very boring 😦

[steady grail]

new workflow for xl 🙂

no idea what's that stuff on the bg XD

never prompted something like that 😛

[steady grail]

cool

[green temple]

which is better ?

[vital raptor]

first one

[bitter latch]

anime smurf

add's some mroe color

ok goes back to watching anime

[nova remnant]

Grabby

[bitter latch]

ok what kind of dresss is that?

[bitter latch]

hmm ok nothing special in prompt interesting

[steady grail]

cool, what't that model?

song of the day

https://www.youtube.com/watch?v=oXAhf1lUFzk

[bitter latch]

oh just a mix I made 😄

[steady grail]

it's cute lol

how do you mix stuff?

mine always end with horrible mutations

[bitter latch]

carefully....?

and sometimes it's down to the prompt or even if you add a artist style

[steady grail]

yeah, i press the merge button very carefully, but they still looks rigid and very mutagenic lol

[bitter latch]

but here's a anime song for anime peeps 😄 https://suno.com/song/7270aef6-ca1d-46a5-8462-b06d6309d9d7

[nova remnant]

@wooden coral gift for you!

[steady grail]

good, i just needed another steam gift, ty

as usual, don't click it, we are joking

it's a scam

[steady grail]

another scam

[zinc flax]

This is a rare one, it's not steam giftcard lol

[nova remnant]

no mods available to take these free gifts

[steady grail]

yeah, it's a stright up virus lol

[steady grail]

so sad for them XD

[versed jay]

I am following this tutorial to create a lora, but in the beginning when the txt is made automatically with the promt, it makes all the txt the same for me, only the name of the prefix that I put appears... why does that happen to me? https://www.youtube.com/watch?v=05sxY_bLI8c

helpp???

@vital raptor

@steady grail

[steady grail]

looking the video

i'm not an expert in merging, but i'm pretty sure the tagger is not sensitive enough

[nova remnant]

im bored

[steady grail]

read or study something

[green temple]

Sleep

[bitter latch]

or play some games

[bitter latch]

or make some music like this https://suno.com/song/45263c26-3f52-4746-903f-e12b0400eefe

@nova remnant for you https://suno.com/song/f7499295-7944-4df7-aea4-47ad3d00c383

ok i'll stop

https://suno.com/song/13a95c53-166b-4ed5-95f1-bdc9b0529aa4

odd how sometimes you can play in window and others it's a regular link. shrugs

[nova remnant]

[supple raptor]

For anyone who doesn't know. When these sort of spam messages come up, you can react to the message with ⚠️ and it will put it into a queue for the mod team to review. Same applies to images that violate the rules. Your reaction will disappear after about 1 second, this means that it worked and will be reviewed.

[gloomy scroll]

But @ mods

[bitter latch]

boots up webui forge

waits

ponders art

hmm cyber shredder?

switches it up

grinz

sure is quiet

there we go

Dragon transforomation?

better

I would watch this anime

sighs

pokes people randomly with a stick

[vital raptor]

oi

[bitter latch]

yes?

[slender barn]

gonna mix up the style a tad

[nova remnant]

I do both

twice as helpful!

[bitter latch]

sleepy (damn storms)

[radiant flower]

thats so cool

are u able to change the color of the hair?

[nova remnant]

hello march 7th miss intertwined fate!

[unkempt cape]

anyone knows a good free bulk content aware cropping stuff?

datasethelpers did not infact content aware crop ):

[unkempt cape]

nvm screw content aware crop

[unkempt cape]

training started i have 0 expectations but i think my dataset is about as good as i can ever care

[bitter latch]

there we go

now for my own spin

that took a turn

[bitter latch]

goes back to bed

[jagged field]

nice <3

[steady grail]

guys, in your opinion, dpo is worth it?

[unkempt cape]

its... alright i guess

fire outfit tho

pleaseeeee colab make it to 9th epoch 🙏🙏🙏🙏🙏🙏🙏🙏🙏

[jagged field]

What does this mean ?

[unkempt cape]

training lora

siuuuu 9th epoch

[wheat falcon]

nope

[steady grail]

reasons, if i can ask?

the plastic effect?

[wheat falcon]

i dont like most of the technologys which help to get faster images but lowering the quality. Like Turbo, LCM, Lightning, DPO,
The only thing that now seems good in quality and speed is this Hyper model thing.

if someone merged the dpo into a model its okay to use it like with lcm models, but i wouldnt add the dpo lora to my gens

[steady grail]

dpo is used to lower the steps? to me it looked like his only function was to add details to image

and adding as a lora looked having no effect at all

[wheat falcon]

yea its for adding details right

i tested the dpo model for 1.5 and sdxl and the result wasnt that great.

but i havent looked for any good dpo community models

[steady grail]

i had this impression it jellified all the fluids

[wheat falcon]

which one did you tried?

[steady grail]

https://civitai.com/models/242825/dpo-direct-preference-optimization-lora-for-xl-and-15-openrail

this one

[wheat falcon]

but thats a lora, i thought you said it had no effect

[steady grail]

you merge it with the model

and you get some improvement on details

a 1, you get a fake 3d plasticized effect and solidification of fluids

under 85%, you get weird mutations

[vital raptor]

like with most things, it likely only works with very specific models and not much else

[steady grail]

the ideal is between 85 and 95

i suppose so

there is a full checkpoint of this stuff? maybe is more worth testing it

[vital raptor]

it's basically impossible to make a lora (or anything else) that works 100% with any model except the one it was trained with

[wheat falcon]

i tried one model which has DPO included, and i think its really good at details:
https://huggingface.co/ND911/EclecticEuphoria_5x_HDTML

[steady grail]

i should test it

i'm terrible at merging

maybe some dpo may improve my results XD

[wheat falcon]

it has everything in it xD

[steady grail]

full models i found are https://civitai.com/models/237681/dpo-sdxl-fp16 https://civitai.com/models/240850/sd15-direct-preference-optimization-dpo https://civitai.com/models/239836/dpoxlplus

i will test merges with these 4

rigorously on anime models XD

so counter intuitive eheheeh

[steady grail]

cute

reading blame makes me wonder

oh, forge got updated

[sly girder]

good morning

[steady grail]

ohi

i just remembered something, merging models with dpo is pretty meaningless

it's better to train the base model

[unkempt cape]

Oh fuck my dumbass

I trained the lora it was doing really shitty, so i went and checked

I changed the images all to jpeg BUT did not change the caption arcodingly

omfg i wanna die

[steady grail]

eheheh

[unkempt cape]

😢😢😢😰😰😰😰😰😰😰😰 😭😭😭😭😭😭😭😭😭😭😭

[steady grail]

stuff like that happen to me very often too

[unkempt cape]

Just gonna redo

Who thinks its a good thing to do to not throw when a prompt file is not avaiable 😭

omfg almost did it again. Note to self: abc.png matches with abc.txt not abc.png.txt

[steady grail]

a suggestion, make a check list

[supple raptor]

I haven't heard of any. Afaik all of the "DPO" community models just have the DPO lora merged in rather than actually being trained with DPO techniques.

[bitter latch]

how do you merge in the lora to a model I keep getting a dict can't be be converted to pystring error ,

oh nm I see

DPO does seem to make a diff, sorta like the more detail lora

L: before R: after DPO added

[bitter latch]

before and after with DPO and More detail

[oak spoke]

after look terrifying

[bitter latch]

It is from last of us. so yeah sorta

[oak spoke]

oh

that's ellie?

[bitter latch]

yep 😄

[oak spoke]

oh boy I accidentally made smt so scary I can't even post it here

if I have trypophobia I would've died

[nova remnant]

Gib money

[steady grail]

charming 🙂

[nova remnant]

She's on duty

[steady grail]

a mission to acquire cookies!

[nova remnant]

Cookies for sure

[steady grail]

she is trying to acquire the chocolate factory?

she must be the next willy wonka XD

[nova remnant]

[steady grail]

naturally, the cookies need chocolate <.<

me stupid lol

[nova remnant]

You're making me want sweets

[nova remnant]

Donuts are low tier

[autumn violet]

I have no idea what DPO is

[wooden delta]

draw a cat

[ornate echo]

No

[nova remnant]

yes

[autumn violet]

I can see you are a true artist

[nova remnant]

squish

[zinc flax]

she thic

[nova remnant]

so thick it looks like her thigh ate her hand

[bitter latch]

with DPO

without

so maybe helps. hmmms

grumpy

very grumpy

hehe

she's not had coffee.. incoming!

[autumn violet]

I know the best way to get coffee fast

Coffee cannon

[unkempt cape]

man lora training failed asf

):

[zinc flax]

What are you trying to make lol

[oak spoke]

average souls series enemy body scaling

🤣

[hexed ermine]

Which model did you use to get this image ?

[gloomy scroll]

Nai

[hexed ermine]

I thought sdxl can do it

[vital raptor]

nai is sdxl

just a paid trained version

[unkempt cape]

Is there a colab or kaggle for full model finetune sdxl

[oak spoke]

wtf happened?

someone come and killed all my pics

[oak spoke]

is that....... a POG FACE??????

🤣

[steady grail]

i wouldn't put the finger, but something is weird, anyway, what's better? XD

[oak spoke]

in term of what?

hands? or everything except hands?

[steady grail]

hands? what with the hands? oh...

XD

in general

maybe the skirt is a little short

[oak spoke]

I like 2nd pic's coloring better

it just got way worse hand

[steady grail]

let's not look ad them

i should use adetailer or inpaint XD

[green temple]

2nd

more smol

[green temple]

guys , why whenever i download multi diffusion extension the tiled vae and tiled diffusion dosen`t show up ?

[unkempt cape]

epic shit bg moment

[green temple]

🥹

[unkempt cape]

artifact heaven but still look pretty cool

[green temple]

🥹

[unkempt cape]

wtfffffffff

its like my 20000th attempt at this point 😭

[green temple]

??? 🤣

[lime patio]

I cannot get over the outfit and her hair.

[tropic sentinel]

Ok

[zinc flax]

huuge ahoge

[bitter latch]

hmm no

she says "MINE!"

[bitter latch]

there done

[bitter latch]

a past version

hmm not a white haired elf getting headpats.. ok...

[bitter latch]

😛

[supple raptor]

Are you using forge by any chance?

[green temple]

Yes

Tried it on A1111 and it worked but not forge

[supple raptor]

Forge has them built in so the extension doesn't work. The built in implementation lacks the regional prompting feature of the extension though. Tiled Diffusion is under Multi Diffusion and Tiled VAE is part of NeverOOM (bottom checkbox, it's not adjustable though).

[green temple]

Oh never knew tiled vae was on OOM

[supple raptor]

Yeah, it took me a little while to find it there.

[green temple]

Ok thx

[supple raptor]

Coke

[primal wharf]

Ghosts.... rituals.... decapitated head, wonder what she's doing.

[green temple]

she summoned smol ofc

[primal wharf]

@lyric zealot she summoned you

[green temple]

🤣

[primal wharf]

[nova remnant]

[bitter latch]

first one I tried, didn'work so well.

bbl

[green temple]

wow

[green temple]

wow

[dim spire]

I really need to learn how to fix hands with something other than paint, heh.

[nova remnant]

gen x900 times one will have good hands

or use sdxl

[unkempt cape]

outfit looks normal to me

hair is not

May be ill do some kohya ss gui stuff over on kaggle

maybe shit'll work and i'll be finetuning instead of this lora crap

if kaggle allows that is

[supple raptor]

lora is a lot easier than a full finetune. if you can't get good results out of lora training then you aren't going to have luck with a full finetune either and will have spent a hell of a lot more time to figure that out.

[vital raptor]

hmm, i wonder if comfy works on a network share, a1111 seems to be having issues

might be enough to make me switch if it will work

[vital raptor]

of course comfy works instantly after i spent hours trying to get a1111 working

i guess i need to find some fun nodes now and get everything figured out

[green temple]

(Badhandv4 : 1.5) 🌚

[vital raptor]

funny thing about hands, autismmix model actually has pretty bad hands by default, but some of the artist styles tend to do them just fine for the most part

[zinc flax]

v prompt cheat code lol

[vital raptor]

"reaching" or "waving" helps too 😄