🍥｜anime

[gusty palm]

ok thats pre upscale

should contain metadata it wont let me post the prompt for some reaosn ><

(Illustration:1),(Evil:1.2) masterpiece, high resolution, highly detailed, best quality, 4k, ((cute vampire girl:1.4)), classical gothic Victorian sailor outfit, (extended vampire fangs:1.2), open mouth, ((huge fangs:1.5)), short hair, ((diva makeup:1.3)), sleepy, shy, petite, (small breasts:1.2), adult 1girl, creepy, haunted, chromatic abberation, (The Mountain of Fire:1.2), midnight, saturated colors, high contrast, abstract, foggy,,(Impressionism:1.3),(CineColor:1),close portrait,(manga:1.3),beautiful,attractive,handsome,trending on ArtStation,DeviantArt contest winner,CGSociety,ultrafine,detailed,studio lighting

stylepile adds all the artstation stuff but it seems to change the results alot

[dawn jasper]

Vampire Succubus hybrid?

[nova remnant]

is this a lora

[runic vault]

the style yes

[nova remnant]

the character

or how did you get those ice stuff >.>

[runic vault]

ice wings, char is cirno from touhou. does not need a lora for char only

very simple prompt cirno, solo, blue eyes, light blue hair, short hair, blue dress, puffy short sleeves, white sleeves, collared dress, red ribbon, hair bow, blue bow, bare legs, feet out of frame, ice wings, kneehighs, white socks, mary janes,

[nova remnant]

those wings remind me of a game character

[runic vault]

maybe because touhou is a bloody videogame

[nova remnant]

No I mean a really old game

(2000s stuff)

haha

I dunno what touhou is

[runic vault]

it has over 823k images alone on danbooru...
one of the most well known franchises in japan with a insanely large fan community

[nova remnant]

:P

there were many version of these

quick google find

[nova remnant]

I don't watch anime

just play my little games and do SD

[runic vault]

then why are in the anime tab?

[nova remnant]

I'm an impostor 👍

and should go to bed 💤

[past junco]

I need to fix my legs lmao

[past junco]

I love everything about this

Until I just realized they share the one scarf

Still great though

Don't mind if I just keep it?

[elfin bolt]

Go ahead, i claim no credit lol

[past junco]

Would you also mind sharing your prompt? We can DM and talk about our process and stuff

[elfin bolt]

Its dall e dude

[past junco]

Lmao

[elfin bolt]

No process just type

[lyric zealot]

[past junco]

Seems like a lot of people use it here even though this is a SD server lol

[elfin bolt]

😅 ive been here so long and idk where to share it haha

I plan to use some of these output as lora training tho

[past junco]

I have also been thinking of trying that

Especially for this art style I have

It's not like any Lora you can grab online

Pure homegrown Checkpoint Merging, and I mean some serious levels of Checkpoint Merging to the Nth degree

I kind of named all the stuff so I honestly cannot tell you what models I used for merging though

Especially when I accidentally deleted some of my models for space this one time...

Screwed me royally

[dawn jasper]

A witch.

[past junco]

These are all very nice

Still digging chicks and cars

Is that an Audi?

[native halo]

yes just with audi r8 on prompt

[past junco]

Dude you should totally try generating roadster convertible cars

Or perhaps the Mercades Benz AMG

I should definitely try doing some cars of my own

[native halo]

idk what happened here

[past junco]

What on earth

Transformers of a different variety

That really makes me think lmao

Shape shifter ladies who turn into cars and back into lady form

They also happen to be massive ladies

[past junco]

In the middle of a busy street with people waiting in cars lmao

I also find it interesting that the street is one way, with two lanes

There's so many interesting things I want to try

For instance, creating a collage by pulling different objects out of different images in Photoshop or Krita, then using Img2Img to create a full image out of the collage

Another idea is to just go google searching for a bunch of images, photos and etc and trying Img2Img on them

[wispy canopy]

first historic account of dancing catgirls, circa ~300 ad

[wispy canopy]

sdxl?

[native halo]

1.5

[wispy canopy]

damn, that's very consistent for 1.5

[native halo]

cuz i only post the good ones 🤓

[wispy canopy]

i post whatever trash i generate 😄

[native halo]

idk what happened here

[wispy canopy]

just your regular clipping ^^

[oak spoke]

so which lora is this

[wispy canopy]

none, just a prompt

[oak spoke]

nice

damn that's a very short prompt

[wispy canopy]

blueprint of a waifu, digital schematics, front view, parts list Negative prompt: nipple,

[oak spoke]

jet bird

[oak spoke]

what if......bp of futa

[wispy canopy]

good birb ^^

[wispy canopy]

internally hidden? XD

[oak spoke]

so like retractable?

here's the waifu one

[wispy canopy]

well, robot waifu would have some advantages, right? 😛

[oak spoke]

when I tried using futa inprompt instead it just give me this

I guess this's the augmented part

[wispy canopy]

i suppose it is

[rain valve]

is it sdxl? all this stuff

[cyan burrow]

back home! Cancer survived, again

[wispy canopy]

wb cryo ❤️

more model merging, want to get my mix even better 😄 (but what happened to v1.2?! -> that was a failure)

the middle creature tho, that's pretty cool

[nova remnant]

meowning people

[nova remnant]

Would let her run me over

[nova remnant]

Do you need a good GPU to mix?

[wispy canopy]

depends, it helps, but the merger takes CPU (it needs to combine weights and then send back to gpu)

[nova remnant]

But

Does that mean you still need a good GPU?

[wispy canopy]

if you can run the model, you can merge the model

you'll need a lot of RAM tho

(not vram)

hmmm... another win for my new mix?!

got the blueprint right

[nova remnant]

Like 32 or 64

I have 16

[wispy canopy]

32 is enough for regular sd

16 probably too 🙂

i run llm's at times -> then i need 64 just to get everything properly mapped

[nova remnant]

I dunno what LLM is

[nova remnant]

Which one is more correct

[wispy canopy]

i wanted a blueprint/schematics

so protovision dropped the ball by not making the waifu a blueprint

it's my to-go test to see if a model is overtuned/over stylized

overtuned models oftend tend to fail to make it a blueprint of a human subject

[past junco]

I probably have an extremely overturned model

Merged from a whole bunch of different models

[wispy canopy]

if you're using an anime model -> yeah.

[nova remnant]

What how do you proceed after you know it's over tuned

[past junco]

Lmao alright then

I honestly just did a bunch of random stuff and hope to god it worked

[nova remnant]

@past junco if you type girl does it throw big Booba at you?

[wispy canopy]

take a step back in the merging process, or try to find a model with weights to counteract the overtuning

[nova remnant]

[wispy canopy]

best way to do it. My merging adventure started with a drunk fluke model... so yeah 😄

[past junco]

Not that I am aware of, will have to check

My original theory was that it would blend the images in between the two merged models

[nova remnant]

This is alien language to me

[past junco]

No idea if that's even true

Oooh, I did do some countering by adding a realistic model to the anime model

Among other stuff

[nova remnant]

@past junco the model I like using is too

[past junco]

Idk

[wispy canopy]

basically, start over again, before it went wrong, or find a different model which can help counteract the over-styling

[past junco]

Did a whole bunch of random stuff

[wispy canopy]

this is often a good way 🙂

[nova remnant]

@wispy canopy would it be possible to make a model less without losing what it does right/to my liking

[wispy canopy]

yes, that's what i'm doing 🙂 then you enter the "merge block weights" type merging

[nova remnant]

What are block weights

[wispy canopy]

this scary thing 😄

basically, a SD model is various layers of neural network, all connected

with blockweight merging, you can try to keep the layers of the model you want, and swap them out with other models

for example, i know that for my base model, the lower IN and higher OUT are totally busted, so i'm having those replaced by the other model

however, it's mid intelligence and creativity are simply better, so i want to keep the creativity, but have the image quality output of the otherr

[nova remnant]

[wispy canopy]

new attempt went hwaywire, but i don't dislike that 🙂

[past junco]

Warning: Kissing

[nova remnant]

👍 bub

[nova remnant]

Careful with that

[past junco]

Nice censors

[wispy canopy]

welll.. uhh... kinda needed

[past junco]

Lmao

[wispy canopy]

i mean, model sucks at nipples, but can't have those here 😦

[past junco]

Yeah

[nova remnant]

Bad pixel blobs

[nova remnant]

This one is interesting

Floating pillow

You need skill to achieve that level of equilibrium

[wispy canopy]

first time i can't say "wow, it's certainly the best"

[past junco]

The floating arm aaaaaaaaa

And yeah lmao

I also can't stop SD from placing hands in the wrong place

First image has her hand on his back

And the second image has a weird ghost hand on the railing

[nova remnant]

Welp

Time to merge

[past junco]

"It's merging time"

I might be able to fix this in inpainting

[past junco]

How do you even tell which are good and bad?

[wispy canopy]

to quote @nova remnant -> i'm inspecting in closeup, and compare a lot of prompting styles and see if something is able to generate artifacts, and if it differs enough from the base model

just. keep. comparing. about 4 minutes per set of images

what i'm looking for now, is increased creativity of the middle vs protovision (the model from which i'm ripping the image quality, but am not fond of the middle weights that model has)

[past junco]

Problem is this is based on comparing different models

I'm still not sure what's going on

[wispy canopy]

it's not different models -> protovision is the model i'm ripping something frrom

brightprotonuke 1.1 is the previous model i posted

1.2 is a dud XD

[past junco]

Yeah so that makes it difficult for me to do

Because at this point

I have no idea what models I really uses to create my current model

And I'd dare say a lot of the other models I made are merged from other overtuned models lol

[wispy canopy]

1.3 is actually a re-mix using a similar recipe, but somet update block weightings

[past junco]

I wish I knew more about block weights

[wispy canopy]

i don't know a lot about them, just a "general gut feeling"

[past junco]

All that I really know is that I got a model I love and it works for me

And it's the model I want to focus on

[wispy canopy]

for example, here's a propmt where somethign disintegrrated in the middle model, now it's trying to figure out where it went wrong

[nova remnant]

[past junco]

The middle one does look a bit odd compared to the others

[wispy canopy]

so there's probably a couple of block weights that are not fitting well together

so, i go test this prompt with other weightings and see if i can reduce the artifacting

i use supermerger for this

[past junco]

Haven't heard of it

[nova remnant]

theres a SMOL person in the middle image

save him

[wispy canopy]

@nova remnant doing my best making more mergers!

[sharp swan]

motrcleyner

[kind basalt]

So good

[brittle jay]

Epico

I am also back home, now I have to do homework already😭
I was just getting used to the freedom of vacation in another country

[brittle jay]

So @cyan burrow are you like totally physically fine rn then or are you still recovering?

[brittle jay]

[pure oyster]

@lyric zealot Am I doing this red eyed waifu thing correctly?

[glacial bane]

This looks like a character from splatz kitz

[sharp swan]

https://media.discordapp.net/attachments/936528714748030976/1101364382895702097/freeze.gif

[dawn jasper]

https://tenor.com/view/sponge-bob-old-man-alcohol-gif-25229103

[sharp swan]

Nvm

[dawn jasper]

https://encrypted-tbn0.gstatic.com/images?q=tbn:ANd9GcQAvwmshA4FLzygmCDvLH6yKzmHGznxYgrpiA&usqp=CAU

https://static.wikia.nocookie.net/fairy-fencer-f/images/b/b0/Alyn.png/revision/latest?cb=20140511045622

^For those of us familiar with Fairy Fencer F: The Series, did anyone do bikini images for Eryn by chance?

[astral drift]

Least horny anime channel poster

[dawn jasper]

@astral drift I had to ask, since Eryn rarely has any such images, really.

[dawn jasper]

@jagged wasp Is that a biker?

[jagged wasp]

ninja

and mech pilot

and a wizard

[astral drift]

Lol I'm not judging

Those kinds of pics are actually a lot easier to do than accurately recreating game outfits

Just gotta inpaint or manually paint their hair/face

[dawn jasper]

@astral drift True, but I'm rather shocked no one did those kinds of pics for Eryn here.

[astral drift]

Virus

Counterfeit XL is pretty goood

SDXL anime checkpoint

[wispy canopy]

new model/merge testing, i'm not dissapointed yet

[gusty palm]

im making rice and general tsos

[wispy canopy]

general tsos?

[nova remnant]

general TOS

[wispy canopy]

testing my model on 'bara'

got. "The BARA"

[gloomy scroll]

Totally same person

[jagged wasp]

trendy cute adult female anime wizard wearing supreme brand clothing bleak background

[wispy canopy]

torturing my new model mix, hoping it'll fall apart

[gloomy scroll]

Thed gaoy!!

[wispy canopy]

YOURO PRCOMP

[nova remnant]

who ? ifleso holl !!??

[wispy canopy]

Wizard with brand clothing. You made the prompt man XD

[wispy canopy]

regional prompting, or natural prompting?

[gloomy scroll]

RP mask

[wispy canopy]

gets the circle a lot better, i suppose?

[gloomy scroll]

Mask has worked best for me

I haven't gotten as good quality with anything else

But problem is that I would need to make new masks if I want some other position

[wispy canopy]

prompt/prompt-ex are really harrd to get right

you need a couple of tries to get the treshhold/trigger word right

[nova remnant]

@wispy canopy

hires 1.3

[gloomy scroll]

I don't get that magic circle as well with prompt/prompt-ex

[wispy canopy]

if you're on sd1.5 it's pretty much impossible

[jagged wasp]

Ya it’s a pretty ez prompt can change out brand and occupation for some decent images

[wispy canopy]

sd1.5 simpmly can't handle positional awareness

[gloomy scroll]

RP is just op

[nova remnant]

bit creepy of an image, what happened to the bird

[wispy canopy]

I believe that time will come
The days of the cherry blossom road trip illuminated by the moonlit night
Give me flowers in memory
Thank you thank you
Thank you
(best quality:1.3), (highres:1), (detailed:1.2), (incredible:1.3), (perfect:1.3), (perfection:1.3), (volume cover art:0), (illustration:1.3)```
Clean Tears - Sakura <- one of the best miku songs ever, fight me

[gloomy scroll]

Bird stealing a sip

[gloomy scroll]

Here's my brand clothing

[wispy canopy]

there's.... a brand

just not brand clothing XD

[nova remnant]

Is there any point in hires with 1.0 resolution ?

[wispy canopy]

1.0?

[nova remnant]

"upscale by 1.0"

[wispy canopy]

oh

uhh... i don't think so

[nova remnant]

👍

[gloomy scroll]

Yes it's just i2i

[wispy canopy]

https://tenor.com/view/that-is-blasphemy-satan-south-park-s4e11-probably-gif-21638722

[nova remnant]

@wispy canopy

[wispy canopy]

more. testing. more merging

[nova remnant]

Some AND_XXXX craziness

sadly I cant upscale so it looks bad

[wispy canopy]

this is fckn awesome!!!!

GIB PROMPT!

[nova remnant]

Its literally nothing lmao

standing, mountain background, clouds background, full body, small breasts, white hair, elf ears, 1girl, casual clothes
AND_PERP [ (magical circle behind her:1.1) :1
AND_SALT lightning ] :1
lora:boldline:1 lora:add_detail:0.3

[wispy canopy]

you say it's nothing

but do you realize how extreme 'latent space' is terms of dimensions?

[nova remnant]

now speak in a language that I understand

[wispy canopy]

basically, the only way that image could happen, is that model, with that prompt, you made up!

so don't say it's nothing!

[nova remnant]

ah

I'm a pro prompter!

[wispy canopy]

you are!

using your prompt to torture my model comparison tho 😄

[wispy canopy]

https://tenor.com/view/miku-gif-5451841

dall3 really being a thing man

[elfin bolt]

why nobody trained a lora on black and manga style with the screentone?

im going to try one soon

[wispy canopy]

like this?

[elfin bolt]

it doesnt have the screentone details

[wispy canopy]

yeah you're right about that

didn't prmopt for it tho

[lyric zealot]

https://tenor.com/view/miku-leek-anime-gif-11585653

[elfin bolt]

try prompting it

[wispy canopy]

not sure what to make of it... which model did best... i know i dislike protovision's style xD

[nova remnant]

heyy

it did what I thought my prompt would

replace the circle to lightning

[nova remnant]

big pokemon or something

[wispy canopy]

sdxl = crazy smart compared to sd1.5

[wispy canopy]

your prompt had church in it 😄

[elfin bolt]

yes it didnt work LOL

"church of leek"

[wispy canopy]

lol, i can even fuckn read stable diffussion words

[nova remnant]

lmao

[wispy canopy]

that's... bad

REAL bad

[elfin bolt]

my workflow soon will be using dall e 3 and then edit in sd

[wispy canopy]

then again... i've had the prompts that created inspirational images -> imagine trying to read a paragraph of sd texzt

[elfin bolt]

sadge

[nova remnant]

@elfin bolt

[wispy canopy]

first, as SDXL is not as smart in tokenizing as llm is -> find a suitable trigger word (this differs PER MODEL)

[elfin bolt]

hmm yeah, screentones is there for the background, but the hair and the clothes is still unaffected

[wispy canopy]

no worries, running second test 🙂

and, going to block grayscale in negatives

[nova remnant]

dalle is funny xd

legs a bit wonky

[wispy canopy]

that's dall3?

[nova remnant]

ye

[wispy canopy]

aight, dall3 -> make a meat/salad gundam 😄

[elfin bolt]

damn

[wispy canopy]

i fear it's going to be ezpz for dall3

[elfin bolt]

yeah but not easily replicable per output

[wispy canopy]

cuz that thing took some serrious effort in prompting

[gundam::4], [mecha:(android:0.1):1] made from (steak:1.3), (armor:0.2) made of salad, meat, leafy greens, carrot joints, tomato shield, carrot sword, cheese

[elfin bolt]

this is likely impossible in sdxl

[nova remnant]

err

[wispy canopy]

this IS sdxl

[elfin bolt]

WHAT

nice

[nova remnant]

not quite

Probably have to specify more than "Meat and salad gundam" lmao

[wispy canopy]

that prompt is VERY reliant on sampler/model quirks

[nova remnant]

BEHOLD @wispy canopy

[wispy canopy]

it's really all about replacing the prompt parts at exactly the right part so it substitutes the parts prroperly

[nova remnant]

I CHALLENGE YOU

Hahaha

[wispy canopy]

dang

😮

that took me A WHOLE FUCKIGN EVENING

you did use my prompt, i suppose

[nova remnant]

gun to this sword fight

I'm playing unfair

this one has a gunsword

[wispy canopy]

oof ... this might be hard for a generic model

[wispy canopy]

even better than i got it!

[nova remnant]

are screentones those little dots like paper drawn ?

[wispy canopy]

yes

[elfin bolt]

its more like the training caption didnt capture the screentoned clothes and hair

[wispy canopy]

dude, i'm but a humble model mixer 😄

[elfin bolt]

👍🏻 yeah not ur fault, everyone is fixated on one style in anime back then

and now sdxl anime isn't really taking off

[wispy canopy]

and my model mix IS a generic model

[nova remnant]

is this kinda it or does it need to be black and white ?

[wispy canopy]

i'm trying to compete with a probably a 100b+ (layered models) image model with a just 2b model

this is going to be very harrd

[elfin bolt]

we use lora

[wispy canopy]

i don't use lora

i like sufferirng

i play path of exile

[nova remnant]

ew stop that

literally do something else

[lyric zealot]

i have no idea how to use RP mask mode

[wispy canopy]

well, my latest build isn't suffering

it's ez mode 😄

[wispy canopy]

literally just facetank the game

[wispy canopy]

can i has prompt to punish my model test?

[wispy canopy]

screentone effect seems the closest 🙂

[lyric zealot]

here with metadata, but it's not working properly, so the prompt doesn't reflect the generation

i was trying to get RP mask mode to work

[elfin bolt]

yep the background is atleast on par with manga

[gloomy scroll]

Each mask layer needs it's own BREAK and masks sometimes need to be little bit bigger than what you want

[nova remnant]

when I inpaint a larger image with controlnet it resizes the image to like 512x512 or whatever, is there a way to make it resize the inpainted area back to the larger image ?

[wispy canopy]

aight, about region prompt masks : they "just draw" latent attention

[lyric zealot]

and we just need to remember which mask was drawn first so that we can know which break refers to which?

[wispy canopy]

correct

[lyric zealot]

[gloomy scroll]

Yes

[wispy canopy]

your masks/prompts need to align perfectly

if they don't -> it won't even work

all before ADDCOM -> is not considered as a part for regional prompts

[lyric zealot]

okay im getting it

[wispy canopy]

unless your prompt had 2 masks, this would not workr

[lyric zealot]

it's kinda working now

just the prompt i was testing with need heavier weight

[wispy canopy]

to break down your prompt, you had "common prompt"+ mask 1 + mask 2

[lyric zealot]

yup yup

now i removed common prompt

but added a background mask as first region

[wispy canopy]

you can also use that as "addbase"

ADDBASE

[lyric zealot]

reason i did that, was because i didn't want the background prompt to apply to the whole region

just part of it, and some contrasting prompt to be added like the portal

[wispy canopy]

&&

^^

regional prompting is fun, isn't it? 🙂

[lyric zealot]

I didn't pay much attention to mask prompting style

[wispy canopy]

aight. lets see if i can get this prompt worrking on a gerneric model and sdxl, using my prompting skills ^^

[wispy canopy]

pretty sure i won't be able to get even close to this

[elfin bolt]

yeah look at the screentones.. its so fuckin good

-_-

im sick of not having the option to generate nsfw on something good

[wispy canopy]

you'd be stuck in an endless fap. like all of us, if you could

[elfin bolt]

[wispy canopy]

you know it's true XD

[nova remnant]

where does that portal lead to

[lyric zealot]

my house

[nova remnant]

Nice

gonna grab your gpu for me

[lyric zealot]

here you go

[lyric zealot]

[gusty palm]

anyone tried the fabric extension

https://github.com/dvruette/sd-webui-fabric

it looks like a thing where u generate some images and tell it which ones u like and it uses your prefrences to guide future generation

[wispy canopy]

have it installed, didn't try it

[nova remnant]

likes every booba image

[gloomy scroll]

Here's big booba image for your

[wispy canopy]

@lyric zealot @elfin bolt trying to combine both your stuff (screentones + regional promptin) at the same time in SDXL

[nova remnant]

https://tenor.com/view/raccoon-disappear-trying-to-save-my-salary-melt-gif-9874114

me trying to find the booba

[wispy canopy]

replaced door with gateway 🙂 that's a lot btter

[lyric zealot]

[nova remnant]

:oh:

[lyric zealot]

gateway might be a better word

[nova remnant]

[wispy canopy]

screentone effect ADD_TOPK black and white, manga screentone effect ADDCOM field, sunset, gateway, cinematic, horizon, cloud, sky, upper body, girl, ADDBASE magical dimension arch , glowing aura, gateway BREAK 1girl, pink hair, forehead, high ponytail, yellow eyes, looking at viewer, girl:0.5

[nova remnant]

why topk

also

shouldn't it go before the prompt

or is being on the same line fine

[wispy canopy]

xD

[lyric zealot]

guys behold

the magical fence gate

[wispy canopy]

trying to figure out the relations, how to get them both working nicely, this being the same image as my previous image

[nova remnant]

Is this the gate you want me to go through ? Seriously ?

[lyric zealot]

you might get stuck going through that

https://tenor.com/view/winnie-the-pooh-stuck-coince-gif-8896709

[nova remnant]

I've seen that before in a video

wait

[lyric zealot]

i resisted posting even more ||horni|| gif

[wispy canopy]

still same seed... oh boy... interactions between regional and neutral prompt. ANOTHER can of worms we can open

[gusty palm]

apparently AND works even without neural prompt

but i dont really understand how it works

[wispy canopy]

AND has worked for a very long time

[wispy canopy]

basically, it's submitting 2 prompts at the same time, and then asking the diffusion process to converge on a image that settles for the both of the prompts seperated by AND

[rustic jetty]

Been working on SDXL loras to improve cute anime expressions/flustered faces:

[gusty palm]

hmmm

[wispy canopy]

neutral prompt is a bit smarter about it, what goes where

(regional activation)

[rustic jetty]

This is my version 4 so far, gonna keep refining it tho

[wispy canopy]

fuuu... another failurer

ah. found it, comma

[gusty palm]

woa

[wispy canopy]

shit. i was afraid so, even from earlier testing -> things in regional prompt CAN escape the neutral prompt

well, i suppose that's a good and bad thing

more testing required

(still on the same seed as #🍥｜anime message)

[gusty palm]

be cool if more words were incorporated like AND,
BUT, .. i cant think of any more

[nova remnant]

discord - a new way to chat with friends & communities

[gusty palm]

have u guys heard of discord

[wispy canopy]

AND and BREAK are just things trying to get composable diffusion going

OH..... NOW i see the pink hair on the girl is on purpose (Didn't even inspect hte prompt)

@lyric zealot i'm using prompt-ex

[lyric zealot]

sdxl model understand context better

[wispy canopy]

with prompt-ex, you ask it to make a mask after a couple of steps, where the "activation region" is for the word you specified

[lyric zealot]

yup yup

I was using that previously

what's the difference compared to prompt?

[wispy canopy]

prompt-ex has a bit better region-activation iirc

[lyric zealot]

[gusty palm]

is this a ok use of AND

it seems if i dont put it at the end of my prompt, it basically interprets anything after AND under the AND

i saw the neural prompt did a thing putting different sectiosn of the prompt in brackets but i couldnt get that to work

[wispy canopy]

prompt VS prompt-ex

[wispy canopy]

that's correct. you'd have to consider both prompts as seperate entities when using AND

and from there, the diffusion will try to converge to an image satisfying both

[nova remnant]

what in the cursed ui is this

[gusty palm]

waaa

[nova remnant]

[gusty palm]

its so neat and colorful

poor text, instead o f rich text

[wispy canopy]

neat and colorfull... pfffff

[nova remnant]

one hand stole a finger from the other

[wispy canopy]

DUAL WIELDING COFFEE!!!!!

[nova remnant]

thats so funny

Im angry and I'll drink 2 coffees angry as hell

[gusty palm]

i wan coffee

[nova remnant]

[wispy canopy]

adding "tired of your bullshit" helped the prompt a lot tho 😄

TWO HANDED COFFEE!!

[nova remnant]

[gusty palm]

lolll

[wispy canopy]

ignore double thumb -> i have collected all coffee wielders from POE

[wispy canopy]

don't as me to make a coffee magic caster

... ugh

coffee caster it is

[nova remnant]

whatchu trying

smol

[lyric zealot]

https://tenor.com/view/tokyo-ghoul-gif-22882608

tbh

[lyric zealot]

i was trying something like this

just sky, and reflection water

but getting too much details on my generation

[nova remnant]

your model

[wispy canopy]

COFFEEMANCER!

[nova remnant]

sad coffee

[wispy canopy]

this better @nova remnant?

[gusty palm]

is this as slow as it looks?

[nova remnant]

what resolution you doing

[nova remnant]

[gusty palm]

384x688

[wispy canopy]

that a 1.5 model?

[lyric zealot]

sure is

[wispy canopy]

damn good reflections for a 1.5

[lyric zealot]

https://www.youtube.com/watch?v=XvHF1MzO25Q im trying to get something like this scene, but it's hard. PS it's a chinese "anime"

[nova remnant]

What is your gpu

[gusty palm]

gtx 1060 6gb

[lyric zealot]

i just want mirror reflection water

and sky

[nova remnant]

looks slower than it should

[astral drift]

Bro sleeping on 1.5

[wispy canopy]

FUCK. i think my model merge is a succes (1.3reversesmoothstepdiv3base)

[gusty palm]

using xformers and tiled vae/diffusion

[nova remnant]

[nova remnant]

is this image new

[lyric zealot]

sure is

I used the moredetail lora at -1 weight

[nova remnant]

but are you doing anything special like AND or regional prompt ?

[lyric zealot]

[gusty palm]

i was using AND in those i think

[nova remnant]

@lyric zealot I swear I've seen that image before maybe im crazy

[lyric zealot]

lmao really

[gusty palm]

with no ANDs im getting 1.42/it so its faster for sure i think

[lyric zealot]

i just generate stuff that's all similar

[wispy canopy]

AND is slower

(double image/diffusion)

@gusty palm, unless you reallly specific need it, i recommend BREAK instead of AND

and if you need and, neutral prompt is better...

[lyric zealot]

something close

[wispy canopy]

(AND_PERP, AND_SALT, AND_TOPK)

[nova remnant]

not in frame

:<

[lyric zealot]

what about this, in frame?

[nova remnant]

https://tenor.com/view/cat-lava-cat-minecraft-cat-gif-13440479435726322249

ded

[lyric zealot]

that makes me thought of something

[nova remnant]

what did you prompt for the background specifically

[lyric zealot]

maybe we can gen something like this

[lyric zealot]

here

[nova remnant]

faces upside down ? good luck

[lyric zealot]

^this with metadata

[wispy canopy]

ugh. this crushes my confidence for a new model 😦

[nova remnant]

explain

[wispy canopy]

the new model fails the prrompt

(just a single seed, but still)

ah... recovery

success! seems my prompt simply failed

[lyric zealot]

this isn't what i wanted

[nova remnant]

@lyric zealot

can't seem to be able to get rid of the little mountains

[gloomy scroll]

@wispy canopy
dreamy, abstract painting, sketch, fantastically absurd saffron, ancient japan,

[lyric zealot]

yup, that's kinda the hard part

or just random village/building

ground, land, tree

[nova remnant]

@gusty palm are you using medvram ?

[wispy canopy]

i was trying to go to bed xD but sure, lets run it 😄

[gloomy scroll]

But it's only 1am

[wispy canopy]

i have to get up at 6...

[nova remnant]

@lyric zealot oh I got it

[nova remnant]

(SD Trolls me by pitching the camera up)

[gloomy scroll]

I can sleep till 8:30

[nova remnant]

Kek

[lyric zealot]

[nova remnant]

o:

its not great but a fair deal of negatives are working

[wispy canopy]

@gloomy scroll these images are gorgeous tho 😮 #3 tho ❤️

[nova remnant]

image for you little vil with metadata

( My model is extremely so those negatives are necessary )

[gloomy scroll]

I think I stole that from Veggie's prompt long time ago

[wispy canopy]

lets see what it does without the rest of my prompt 😛

[wispy canopy]

as pure prompt on my model 😮

[wispy canopy]

aight, imma go to bed for real tho. see ya lovelies later ❤️

[nova remnant]

smol

[primal wharf]

Holy, nice

[nova remnant]

those aren't from dalle3 :p

he was asking if dalle3 could do those, I then did some on dalle

[nova remnant]

lmao

you have the same problem as me

girls going inside the water

[lyric zealot]

i have no idea what's going imma be honest

[nova remnant]

looks like when an app breaks and stuff is lagging out

then you start to see your background image

@lyric zealot

tried my best

[gusty palm]

after tweaking tiled vae settings and resolution got up to 1.6it/s

[nova remnant]

katieeeeeee

Are you on medvram or lowvram

[gusty palm]

no

[nova remnant]

none ?

[gusty palm]

:> none

[lyric zealot]

looking nice

[nova remnant]

huh

@lyric zealot tried -1 on add detail

it deleted the elf

[gusty palm]

seems to fit within my stuff usually

[lyric zealot]

first world problem

[astral drift]

Improved the image tbh

[nova remnant]

👍definitely what I was going for

my model is a bit more realistic than plain 2d

I have (anime:1.4) to make it look more 2d

smolsus was trying to get some anime looking scene

[dawn jasper]

Neat visual there, is that for a video game?

[lyric zealot]

not really, it was just rng

[nova remnant]

o:

pretty cool effect

[lyric zealot]

[dawn jasper]

I like the expressions there.

[nova remnant]

what did I come back to

cute elfangel save me from those aberrations

[lyric zealot]

taking a bite

[gusty palm]

millions of teeth

[nova remnant]

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

[dawn jasper]

Why is she in the waters?

[native halo]

cause shes wet

[lyric zealot]

why not

[dawn jasper]

@lyric zealot I feel curious is mainly why I asked.

[nova remnant]

@lyric zealot I'm gonna eat you aaaa

[lyric zealot]

https://tenor.com/view/anime-gif-6216499

[dawn jasper]

https://tenor.com/view/kuroyukihime-shock-accel-world-surprised-omg-gif-9441659

[native halo]

https://tenor.com/view/chiyo-chichi-chiyo_chichi-cat-everynyan-gif-9439922

[nova remnant]

@lyric zealot

https://tenor.com/view/cat-blind-flashbang-light-mode-gif-21281004

[sharp swan]

leaked image of my father

[lyric zealot]

https://tenor.com/view/how-not-to-summon-a-demon-lord-shera-collar-slave-collar-elf-gif-16543989

[nova remnant]

@native halo You changed picture I literally couldn't see your message, your weird picture was everything that told me it was you typing, now im lost

[nova remnant]

!!!

[lyric zealot]

[nova remnant]

what are those called

Cough cough might try some for research purposes

[lyric zealot]

choker?

[nova remnant]

yeah but

choker will give me those smol black ones

[lyric zealot]

but what

[nova remnant]

I want like that gif

[lyric zealot]

i guess you can add chains

or even slave?

or something similar

no idea what's that actually called

[nova remnant]

[lyric zealot]

Im trying something new

[nova remnant]

What is new there that you're trying

many things going on that I dont understand

[lyric zealot]

trying this new thing someone told me

basically use sdxl model as generation

then convert to sd1.5 model for hires fix

[nova remnant]

o:

[nova remnant]

that is smort

[lyric zealot]

it's lacking vae atm

but im lazy to baked in vae

yup

but generation is taking forever

[nova remnant]

how many it/s on sdxl compared to 1.5

[lyric zealot]

2mins and 10 sec for this gen

whereas i can do it normally in >40 sec

in sdxl, im getting s/it

[nova remnant]

@lyric zealot nhac

[lyric zealot]

but sd1.5 is it/s

[nova remnant]

oh

I'm in the wrong model

I was so confused why the images were like that

very anime lmao
And not booba

[lyric zealot]

might be worth doing in comfyui or if you have a lot of vram

PS no vae

[lyric zealot]

[nova remnant]

is it possible to grab a sdxl gen seed

for 1.5

to compare

or are they completely different

[lyric zealot]

let me see

im changing model now

[lyric zealot]

same seed

so very different

[nova remnant]

https://tenor.com/view/cat-cat-watchin-cat-watch-cats-of-internet-cute-cat-gif-17661352

Just a bit more...

My PC crashed

[gusty palm]

i have to go get groceries

[nova remnant]

[gusty palm]

used cuztom hirez fix for this one

[nova remnant]

What is happening

[dawn jasper]

Very beautiful display, there.

[nova remnant]

@lyric zealot

[lyric zealot]

how many ears?
yes.

[nova remnant]

damn you sneaky animal ears

[nova remnant]

[nova remnant]

put fox ears too :<

[nova remnant]

many ears

fluffe

[gusty palm]

i cooked rice earlier and its like .. crunchy?

[nova remnant]

@lyric zealot

shes definitely nice and friendly

[gusty palm]

shes definitely nice and friendly

is there any way to reset every single seetting to default without fully reinstalling the ui

hirez fix on is outputting the resolution i have set as the main generation instead of the upscaled version

so i think i broke something

[nova remnant]

My PC is crashing

Doesn't wanna do cute images

[gusty palm]

nuu

[dawn jasper]

She works at a Maid Cafe?

[lyric zealot]

maybe, maybe not

[dawn jasper]

She's wearing those ears and a tail to work.

[gusty palm]

is this 384*768? it looks bigger

[nova remnant]

in settings it should be on the very bottom?

[zinc flax]

the newest bing AI thingy is scary good

[nova remnant]

It iss

Must be some insane computing

[zinc flax]

Very impressed with the text

[lyric zealot]

i have no idea how i get this

[nova remnant]

He gives me robot vibes

[lyric zealot]

trueeeeee

[nova remnant]

Bretty

I like the realistic details

[lyric zealot]

i like it as well

kiss

[nova remnant]

Is this regional prompting magic

[lyric zealot]

nope

normal prompting

[dawn jasper]

Very beautiful display with a beautiful woman.

[wispy canopy]

good morning

[lyric zealot]

[brittle rampart]

Guys, tell me if it makes sense to use the word BREAK in the prompts line and how, other than roughly finding out where 75 tokens end in order to put this word?

[lyric zealot]

why do you wanna put STOP in prompt line?

if you are using a1111 there should be a token counter at the top right

[wispy canopy]

which ui are you using?

[lyric zealot]

that's probably the token counter

[wispy canopy]

that's the token counter, yeah

[brittle rampart]

a1111 with extentions

[wispy canopy]

you can use BREAK it's just specific in use

for me it's worth a lot, but outside of my freak prompting, i'mnot sure what the uses cases would be, except cleaning up formatting and making sure "red eyes" would not get seperated at token 75 and 76

[brittle rampart]

As far as I understand, using this makes it possible to correctly process very long prompts

[wispy canopy]

correct

[brittle rampart]

but how to calculate or see these 75 tokens?

[gusty palm]

ah ok i ll keep that in mind. i been scared to try it . i think i was just mistaken about my problem with hirez fix

[wispy canopy]

you'll have to trust the token counter

[brittle rampart]

Does it matter where the word BREAK appears? at the end of a line or at the beginning of a new one?

[wispy canopy]

nope, it applies BREAK wherever ryo uuse it

newlines aren't tokenized

[lyric zealot]

it's only dependant on the token count

[brittle rampart]

oki-doki. Thanks!

[gusty palm]

does BREAK automatically happen after the 75th token

like is it the same as manually adding BREAK

[lyric zealot]

no

BREAK is to manually fill up the remaining unused token in batches of 75

so example, if your token count is 34, using BREAK, your token count will go up to 75

[lyric zealot]

if you have a token count of 120, using BREAK at that time, your token count will go to 150

[wispy canopy]

and if you BREAK BREAK BREAK you're already at 225 tokens 😄

[lyric zealot]

[bitter latch]

woof

[lyric zealot]

double glasses

[bitter latch]

woof woof

[bitter latch]

😄

[brittle rampart]

By the way... is everything said about BREAK true when used in a negative prompt? if so, then how to separate the prompt when using Embedding

[bitter latch]

dur

[dawn jasper]

I like her expressions.

[wispy canopy]

no, negative doesn't require break

then again, you can't target either with the negative prompt

[brittle rampart]

oki

[bitter latch]

Corgi's in spaacceeee

[gusty palm]

stardog

[gusty palm]

theres a new starfox clone i wanna play

i been wanting a modern starfox clone for ages

[gusty palm]

its not multiplayer though so bleh

[brittle rampart]

and another question. How correctly will the prompt be processed depending on the model and the selected resolution of the initial image? does aspect ratio affect? How, then, can you find out what aspect ratios the model was trained for?

[zinc flax]

Love the eyes

[lyric zealot]

i'm trying RP to add more depth to the eyes

https://tenor.com/view/アイドル-idol-alexis-oshi-no-ko-gif-17708032838448296645

almost at this level ^

[zinc flax]

Cuuute Ai

[astral drift]

You're gonna have to check the documentation, but most models have similar resolutions depending on if they're SDXL or 1.5

You can usually just use the base model resolutions on any derivative models

[brittle rampart]

you mean 512x512?

[astral drift]

Yeah, for 1.5

or 512x768

[brittle rampart]

but is there a workflow that will allow me to get a further 4:3 from a 1:1 ratio?

[lyric zealot]

what do you mean? If you wanted a 4:3 ratio, then just go for it straight away?

[astral drift]

I think there are LoRas for it, I usually just keep to the resolutions the model is based on to keep the distortions to a minimum then crop to my desired ratio

[gusty palm]

or to turn a 1:1 into 4:3 by expanding the edges

[native halo]

where

[brittle rampart]

for example, I have a model trained on images with a 1:1 ratio. Will such a model correctly generate Images based on prompts at a ratio of 4:3?

[native halo]

why dont u try instead

[lyric zealot]

well, it might not be the best, but it's better than nothing, and using hires fix, should help

if the model is trained on 512x512, don't go too high up like over 1024, without using hires fix, because subject might get duplicated

[brittle rampart]

I tried it. and the result was more correct for 1:1 ratios

[native halo]

🤷‍♂️

[past junco]

Maybe try outpainting instead of cropping?

[brittle rampart]

That’s why a question arose in the workflow of obtaining an image with a 4:3 ratio (for example)

[gusty palm]

i generate in 2:1 mostly

most models dont seemto care

[native halo]

sorry cant help u only way to know your model capabilities is by generating a bunch of images at different resolutions

some can and some cant so either try or stick with default 512x512/ 512x768

[brittle rampart]

oki. thanks

[past junco]

Models are specifically trained for specific resolutions, the standard would be 512x512 for SD1.5 and 1024x1024 for SDXL

[past junco]

One of the things I've noticed is that trying to generate images in 1024x1024 using models designed with 512x512 in mind have defects, often adding more characters and objects into the scene than what was prompted

This is usually the case for my "four people group shot" prompts, but this could also be SD being SD, for the same reason extra fingers are added to hands

[brittle rampart]

it true!

It was precisely because of these incomprehensible additions that the question arose about in what ratio to receive the original image before HiRes.fix

[past junco]

The absolute best advice I could give is to play around with the settings, experiment, and see what works and what doesn't

[past junco]

Yeah

One thing I did find that helped a lot with this was ControlNet and Img2Img

[gusty palm]

384*768 is my go-to for wide pics