🍥｜anime

[wispy canopy]

yeah, you'll have to make an agent

a character that will interpret your prompts

7b models struggle at interpreting prompts at times though

[bleak cape]

which upscaler for anime?

[untold glacier]

Some say AnimeSharp. I say Fatal_Anime, aside from the darkening it does.

[wispy canopy]

depending on sampler. if using unipc - > ddim, ANY esrgan upscaler with denoise 0.5 will end up pretty similar

[untold glacier]

I often use UltraSharp or Anime6B though.

[bleak cape]

it was ddim

[wispy canopy]

any esrgan will do similar

still, i swear by ultrasharp

[untold glacier]

Ultrasharp is great.

[wispy canopy]

all my images are with ultrasharp

[bleak cape]

better than superscale? it has a long name so must be good

[untold glacier]

Lol

[wispy canopy]

using that logics, my real name would make me a god

which i'm not

xD

[bleak cape]

lol

is there any other upscalers you recommend?

[untold glacier]

I've found a bit too much banding on some gradients or fine meshes with Ultrasharp on anime gens, so if I want a more art-based gen I go with fatal_anime, if I want flatness I go with Anime6B.

[wispy canopy]

i know she's missing a hand... but she's looking down on me, isn't she?

[drowsy dove]

I suppose this is a success

[[USERNAME]]: Can you suggest me a prompt for an artificial intelligence image generator? I want it to be about an off road car in a beach  
[[AINAME]]: Certainly! Here's a possible prompt: "Generate an image of an off-road car cruising through the sand dunes on a beautiful beach, with the blue waters of the ocean in the background and seagulls flying overhead."

[wispy canopy]

unipc into ddim?

[bleak cape]

damn that never works for me

[untold glacier]

Sorry, should've specified--Just regular ol' DPM++ 2M Karras or SDE Karras

[wispy canopy]

use unipc when going for gradients and or softer colors

karras-kun likes to make stuff more flat-shaded

[wispy canopy]

if you are using oobabooga

[bleak cape]

whats the difference between ddim and unipc?

[untold glacier]

I've tried it before, and it does depend on the model, but I get better cohesion with SDE and if I want details I go 2M. I do need to do more testing with unipc though.

[wispy canopy]

ddim one of the oldest, but still very good sampler, unipc one of the newest, but unipc cannot do upscale, so is substituted for ddim

[drowsy dove]

I'm using llama.cpp, so I'll adjust the format. Thanks

[bleak cape]

ultra sharp worked well

[trim current]

https://tenor.com/view/hizzie-spongebob-gif-21331644

[bleak cape]

prompt?

[lyric zealot]

detailed background, masterpiece, best quality, highly detailed, hyper realistic, full body,

1girl, pink hair, twin braids, yellow detailed eyes, thighhighs, zettai ryouiki, sneakers, stylish, hands in pockets,
Negative prompt: (low quality, worst quality:1.4), (bad anatomy), extra digit, fewer digits, (extra arms:1.2), bad hands, by (bad-artist:0.6), bad-image-v2-39000
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 22, Seed: 930105857, Size: 512x768, Model hash: 4d651c7638, Denoising strength: 0.53, Clip skip: 2, ADetailer model: face_yolov8n.pt, ADetailer conf: 30, ADetailer dilate/erode: 32, ADetailer mask blur: 4, ADetailer denoising strength: 0.4, ADetailer inpaint full: True, ADetailer inpaint padding: 0, ADetailer version: 23.5.19, Hires upscale: 2, Hires steps: 10, Hires upscaler: 4x_fatal_Anime_500000_G, Dynamic thresholding enabled: True, Mimic scale: 7, Threshold percentile: 96, Mimic mode: Half Cosine Up, Mimic scale minimum: 4, CFG mode: Half Cosine Up, CFG scale minimum: 4, Version: v1.3.0

Used embeddings: bad-artist [2a38], bad-image-v2-39000 [b03e]

[wispy canopy]

ok, now i'm prompting this too 😄

freebie prompts are just there to be taken, right?

[lyric zealot]

sharing is caring

[wispy canopy]

oh shit. i just had to ignore the part where there is masterpiece high quality in there

[trim current]

iw was gonna say, needs more masterpiece

[wispy canopy]

i kinda killed that in my model lol

[lyric zealot]

do what you have to do

[trim current]

(masterpiece:111111111111111111)

[lyric zealot]

not enough

[wispy canopy]

https://tenor.com/view/100percent-gif-18871183

[trim current]

((((((((((((masterpiece111111200)))

i forgot to adjust my loras om comfyu earlier and had 3 of them on 2 lol

was extremely baked and horrible

[lyric zealot]

[wispy canopy]

horrible fail on my model lol xD

it just doesn't work well with booru prompts xD

[lyric zealot]

pretty cute tho, not gonna lie

[trim current]

then unbooru the prompt

[lyric zealot]

don't do it yourself, make an agent for that

[trim current]

kek

yes

[lyric zealot]

maybe even make a bot, then can detect whenever there's a prompt posted here, unbooru it, generate it, and self post it

[trim current]

kek

[bleak cape]

it cut the jeans lol

[lyric zealot]

lmao

weird fashion but whatever floats

[wispy canopy]

already did. waiting for gens 😄

[bleak cape]

cute car

[lyric zealot]

[lyric zealot]

toy car at carnival

[lyric zealot]

twin braids

[supple raptor]

I usually use loras at like 0.8 strength, but I have the dumb so I often forget to add the 0. and end up running them at 8 instead

[wispy canopy]

https://tenor.com/view/burn-elmo-pyro-burn-it-down-ashes-gif-5632946

burned images everywhere

[supple raptor]

They're not even burned at that point. It's just colorful blobs XD

[trim current]

i noticed that comfy applies them different than auto, i need to have it at above 1 for a single lora or below 1 for multiple different loras

for it to work properly

but 3 of the same at 2

is just

no lol

[supple raptor]

Wish me luck. I'm updating WebUI.

[trim current]

i just moved my models out and embeddings then fesh install lol

[trim current]

bumblebee transformer

[supple raptor]

why the fuck is Auto still installing xformers 0.0.17 by default?

[cyan burrow]

he's slow and vlad is bloating his fork...annoying

[supple raptor]

Auto has been really active for a while now. He's just using dev and RC branches instead of breaking the main branch every other time he merges something.

[cyan burrow]

okay

but why not use a newer xformers then?

tsundere goth

[supple raptor]

No idea

[lyric zealot]

remind me again, is newer xformers better?

[cyan burrow]

works better with pytorch 2.0

[supple raptor]

0.0.20 is slightly faster and is deterministic at least within the same session.

[neon linden]

do you guys use online services or your own graphic card?
and if you use your own graphic card, which one do you guys have?

[cyan burrow]

I use my own local, a 3070 ti 8gb vram

[neon linden]

oh! and how is the performance? I have the same one

[supple raptor]

Most of us use our own. I have an RTX 3080 12GB. Ran it previously on an RTX 2070 before I upgraded.

[cyan burrow]

I am considering getting a 3090 24gb vram though. Second hand, but it will still work

the performance for generating images etc is alright

[neon linden]

does it get too hot?

[cyan burrow]

I can't train models, but I can train LORA's with limitation

[neon linden]

hm

[cyan burrow]

No but that might be because I have optimal airflow with meshes, 200mm fans and everything

never goes above 65c

I can merge models though

it's a much more convoluted process if you want to merge it effectively with only specific parts of other models

[stuck aurora]

Mine goes 75°C, should I be alarmed?

[cyan burrow]

no

[supple raptor]

image generation is generally less demanding than gaming as far as thermal load because you're only pushing the card for a few seconds to maybe a few minutes at a time compared to gaming which is basically a constant load.

[cyan burrow]

it can go up to 105 I think

[neon linden]

thanks a lot!!
do you guys believe that the tech will advance in a way that it requires less vram to work, or that we will just have to continue getting better graphic cards?

[cyan burrow]

if it remains under 83 C then you're fine

above that you need to consider upgrading cooling or undervolting

[neon linden]

as far as I know it shouldn't go higher than 70 or 80 before the component starts to degrade o.o

[lyric zealot]

i think going up to 105 is bad

like really bad

[supple raptor]

Ideally, keep it under 90C and your GPU is fine. Once it starts getting above 90C you'll likely see some thermal throttling.

[cyan burrow]

@lyric zealot it has been tested to not blow up and short circuit up to 105

doesn't mean it won't damage components

but yeah

keep it as low as you can

75 is fine

[lyric zealot]

[supple raptor]

Lower is always better, but all modern cards have thermal protections in place that will shut them down before damage occurs unless there is a malfunction.

[lyric zealot]

I remember there was a time, my pc was acting up, and the temperature is going over 100C

[cyan burrow]

clear out dust from the fans and heatsinks of the gpu, also peal up the sticker on the fans top and slightly loosen the screw and add oil certified for ball bearing use. If they have started to slow down or scratch

[lyric zealot]

and it blue screens

[cyan burrow]

J

[lyric zealot]

lucky now it's like that

[cyan burrow]

I have set mine to thermal throttle at 90

meaning it cuts performance in half

and temperatures drop

[shell timber]

only get a 3090 if u plan to train models, or generate massive images above 1024x1024. compared to a 3080, it generates 512x512 images at most a second or so faster.

[wispy canopy]

except for the part where you can run larger batches, do other things, and have a LLM loaded at the same time, i agree

source: i did that specific upgrade from 3080 to 3090

[shell timber]

never really noticed much of a difference with that when i did the same. although admittedly, i never really did batches more than 20 images or so back then

[wispy canopy]

i mainly got it so i can run 13b models with ease

[shell timber]

fair enough

[cyan burrow]

I would be using to run LLM and SD simultaneously and attempting to create a cross exchange between them. Or perhaps integrating the LLM into SD through an API. I would experiment a lot more, not just with AI but with 3d software that needs high vram. Oh and it would be great for games too and last at least two generations

[wispy canopy]

too lazy to api, just manually feed it XD

[untold glacier]

This

[supple raptor]

batch size is what you can increase, batch count is irrelevant. any GPU can generate with a batch count of 4 million if you wanted to since that just generates 4 million images one after another and is no more demanding that generating a single image. batch size increases the number of images that are generated simultaneously and therefore requires more vram.

[wispy canopy]

and usually slightly faster

[untold glacier]

0.19 xformers has been life changing ever since I started using it a month ago.

[cyan burrow]

I would use it so I can click a "reformat" button in SD, using the LLM to reformat my sentence

etc

[supple raptor]

0.0.20 is available now and has a small performance increase over 0.0.19

[untold glacier]

Seems stable?

[supple raptor]

Just installed it personally. It seems fine so far. It's a stable release, not a dev build.

[untold glacier]

🆒 🫘

[supple raptor]

To install it just activate venv and run:
pip uninstall xformers
then run
pip install -U xformers

[untold glacier]

They say that it uninstalls automatically when installing a new version though?

[supple raptor]

It should, but I'd always rather just spend the 10 seconds to make sure it was uninstalled.

[silk narwhal]

just in case they renamed stuff yea

[cyan burrow]

oh ffs, I couldn't even post my short story

said it contained harmful content, offensive content, bad bad content

which....I mean ok? Can't it point it out then, what I need to change 😦

[lyric zealot]

feed it to chatgpt, let it rewrite it

[wispy canopy]

"sorry but as an ai model... bla bla bla"

[lyric zealot]

[cyan burrow]

haha

[lyric zealot]

but isn't it trying to rewrite it without those stuff?

[shell timber]

want me to pass u a bypass prompt?

[wispy canopy]

DAMMIT EFACE STOP LYRICSPROMPTING

BRAIN SAYS NO

https://tenor.com/view/no-way-dude-no-oh-bugs-bunny-bugs-gif-22941840

[magic wave]

I say, lyric prompt for a big tiddy goth gf

[wispy canopy]

ugh... i know LLM model don't like that

[wispy canopy]

you had that one saved up? XD

[shell timber]

had it in backstock. figured now it would be good to post lol

[magic wave]

Too bad, do et

XD

[magic wave]

It is nice

[cyan burrow]

@wispy canopy grab some combichrist lyrics

see what it does

[wispy canopy]

it just accepted just like that?!

[cyan burrow]

I can't post most of them here, cause...well they're mostly profane

[lyric zealot]

where are the mummy milkers generation

[wispy canopy]

working on it!@

[lyric zealot]

[wispy canopy]

i hope clyde doesn't hate me though

[lyric zealot]

dont get ban tho

[shell timber]

any specific requests? lol

[lyric zealot]

just some mommy milkers

[shell timber]

fresh off the batch for ya

[wispy canopy]

please clyde, don't be angry

clyde seems to like them 😄

[lyric zealot]

BIGGER

[shell timber]

u sure?

[lyric zealot]

galaxy level

[wispy canopy]

... ok. i'm gonna fetch my galatic breast prompt again...

[untold glacier]

That's not something I thought I'd ever read.

[shell timber]

prompt? I made a whole model for that

[drowsy dove]

Local LLMs can do a lot of things, like... Enterprise Resource Planning. It seems to be same as Stable Diffusion models, the websites/services have restrictions while local models are limited by their training data

[lyric zealot]

[wispy canopy]

i... just need to find it

AGAIN

FUUU

@lyric zealot imma dm you whatever comes out, don't wanna risk getting a timeout here XD

[lyric zealot]

for sure for sure

[wispy canopy]

these are very much threading on the wrong side of the line xD

[shell timber]

prob gonna have to do the same lol

[teal pivot]

hey ppl :3

[wispy canopy]

hi and almost bye

this be my last generations, it's 2:15 am

[teal pivot]

3 am for me xd

just started 😄

[lyric zealot]

morning to you

[wispy canopy]

aight, imma go to sleep

buhbye ❤️

[teal pivot]

gn!

[drowsy dove]

Are the images straight from txt2img or there was editing involved? I ask because normally I feel like it takes so long to get good results (cursed hands, bad eyes, unrelated things blending)

[supple raptor]

Who's are you asking about?

[drowsy dove]

The people posting images, really

[supple raptor]

Mine are mostly straight from txt2img. Occasionally I do some manual editing and/or inpainting.

[drowsy dove]

Now I discovered AI-assisted prompt generation, I think I'll use it often. It seems to describe what I want so well compared to my emgrish pronts

[magic wave]

Late, but it here

[lyric zealot]

i need to use my super reaction more

[magic wave]

Same

XD

[supple raptor]

AI assisted prompting can be fun. I still like writing and refining my own prompts though

[lyric zealot]

I'm wasting it

[wispy shard]

where?

[magic wave]

Same xD

[supple raptor]

damn... darksushi goes hard

[drowsy dove]

I'm using Vicuna 7B locally via https://github.com/ggerganov/llama.cpp

[lyric zealot]

lmao

[magic wave]

honse :o

[wispy shard]

mommy

[lyric zealot]

[magic wave]

yes

[magic wave]

Very top heavy

[untold glacier]

Uh

[untold glacier]

Tried starting SD after the xformers upgrade

[untold glacier]

Just completely uninstalled and reinstalled xformers with the same error.

[supple raptor]

probably mismatched torchvision version. You could either rebuild venv, try --reinstall-torch, or try to fix manually. If you rebuild venv you will need to manually install xformers again, but that method will definitely work as long as you're on the current auto1111 commit.

[untold glacier]

Oh wait, I haven't updated Auto in a couple weeks. If I reinstall torch, should I update* first?

Or maybe I'd need to just wipe venv if I did that.

[supple raptor]

I would just go ahead and update auto then delete venv to let it install properly. That will put you on torch 20.1. There are some nice new features

[untold glacier]

Alright, will do that then. Thanks Nubby

Now I'll have to redo the sampler and SAG edits again though.

[magic wave]

Cow cat girl? :O

[lyric zealot]

yes

[magic wave]

XD

[supple raptor]

SAG edits are built in now so they're no longer needed.

[magic wave]

Whats that model/lora?

[nova remnant]

how do you update xformers nubby

[lucid rapids]

latest ghostmix model

and add detail lora

[supple raptor]

Looks like there may be some minor complications on previous WebUI versions and/or previous pytorch versions. So first question is when did you last update your WebUI and what torch version shows at the bottom of your WebUI?

[nova remnant]

1.3.0 and same torch as you

maybe i should update python just incase

[slow needle]

cyberpunk cowgirl cause why not

[supple raptor]

cool, that makes it easy then.

1. Go to your stable-diffusion-webui folder, click in the address bar and type cmd then hit enter to open a cmd window in that folder
2. Type venv\scripts\activate and hit enter in the cmd window to activate your venv
3. run pip uninstall xformers to uninstall the current version of xformers
4. run pip install -U xformers to install 0.0.20.

[nova remnant]

got it to work thank you 🤞

[supple raptor]

so we got the option to choose the hires sampler.... but it doesn't work with UniPC

[magic wave]

Ohhh, thats fair then

[supple raptor]

why you do this auto?

[slow needle]

just saying when you switched to tensorRT you likely wont be able to use hires at it max any more

and DDIM is fine for upscaling

[supple raptor]

tensorRT sounds like a headache that I don't have the patience to explore currently

[untold glacier]

Yep. SAG patch is broken again after this latest update.

[supple raptor]

The patch should be completely unnecessary now. The callbacks were added according to the patch notes.

[untold glacier]

Oh! Great news

[untold glacier]

Wow

[magic wave]

I might be cookin

[nova remnant]

when arent you

[untold glacier]

ᴛʀᴜᴇ

[nova remnant]

tfw no thic cow gf

[lyric zealot]

???

[terse ore]

Anyone tried rendering UE's Metahuman and feeding it into SD?

[pure oyster]

Generating an image in monochrome and coloring it yourself reduces that "AI look" by a decent amount.

[nova remnant]

hot goth girl

[magic wave]

True :o

[near elbow]

product idea: washing machine with a side seat 🙂

how to clean your cat, part III

[wispy canopy]

damn, my my model likes high steps soooo much (35vs50vs100 low)

[rich crater]

who here likes goth chicks

spoiler alert: I do

and so will you

[rich crater]

exibit a

exhibit b

exhibit c

[rich crater]

YAYYYY! finally some hot men to look at

[gloomy scroll]

Maybe I should ask for progress report on that prompt inspector bot I heard this server had on development couple months ago so I don't need to save every image to check if they have prompt

[supple raptor]

[brittle jay]

Good morning every two

Does anybody know what is the max length of tokens possible in a1111?

[gloomy scroll]

+40k at least

[brittle jay]

Epico

[brittle jay]

Why is it not theoretically infinite? (If you had good enough specs)

[gloomy scroll]

I have no idea if it's not infinite I just have seen ~40k token prompt

[brittle jay]

Like it still just averages it all into a single vector or embedding or whatever it is called

Like 75 tokens it makes a new thingy then it averages all the thingys into one

So why no infinite?

@supple raptor

[gloomy scroll]

Is there reason why you want it to be infinite like I don't even see reason for that 40k token prompt?

[brittle jay]

I just wanna know if possible

[supple raptor]

No idea. There may be some limit to what python can handle at once I don't think there is a hard coded limit though. Honestly, once you get close to 1k none of the tokens are doing very much anyways unless you add a bunch of emphasis so it's kind of irrelevant.

[brittle jay]

Ye

Probably has to do with python limits

Idk

Also it took you very long to write

[supple raptor]

Typing on my phone while also watching anime

[brittle jay]

Oh then fair enough

What anime?

[supple raptor]

Monster

[brittle jay]

Nice one

What ep?

[supple raptor]

Just started ep 8

[brittle jay]

Nice

I am just here in math class

[supple raptor]

What are you doing? Get down from there!

[brittle jay]

My school be trippin

[gloomy scroll]

[stuck aurora]

How many Sampling Steps is recommended? I use 30, is it unnecessary?

[gloomy scroll]

I use 20 with 2M Karras

[stuck aurora]

Thank you!

[gloomy scroll]

She's preparing grand luna magic

[modest anvil]

made of my faviort images i made in a minute

[tawny helm]

Hey guys, if none of you have seen this yet this script is dope as hell.

https://civitai.com/models/23188/loopback-scaler

Considering how on top of the game a chunk of everyone in here is most of you probably might know. XD

[modest anvil]

check out https://civitai.com/models/58869/sd-prompt-enhancer

i think its probably one of the best scrips out there

[tawny helm]

Oooh, I heard about this just haven't checked it out yet.

Installed the extension, gonna check it out now before I turn in, been up all night. lol

[modest anvil]

check out llms to there super healpfull

[nova remnant]

...

[tawny helm]

Oh boy....

[nova remnant]

boy what in da

ima click it

maybe itll get him sum gift points

[tawny helm]

Please don't...lol

[nova remnant]

🤣

[compact warren]

Funny that i see this bc someone posted a handsome squidward lora on civitai yesterday

[brittle jay]

life imitates computer

computer imitates life

[compact warren]

not mine

[stuck aurora]

[trim current]

Try looking for the loopback + superimposer

[trim current]

https://www.reddit.com/r/StableDiffusion/comments/xpml78/loopback_and_superimpose_script_for_img2img/?utm_source=share&utm_medium=android_app&utm_name=androidcss&utm_term=1&utm_content=share_button

Found it wee

[brittle jay]

Yes

[lyric zealot]

weird shadow

[slate night]

Oh I am more than happy to give husbando if husbando is requested lmao

[trim current]

https://civitai.com/models/70248/color101-vae
New vae

Not mine

[brittle jay]

Gotta say though emo > goth

[lyric zealot]

[trim current]

2002 era scene/alt > goth/emo

[brittle jay]

Sussy baka

[lyric zealot]

is that too much?

[unreal ridge]

no

[brittle jay]

Could be argued

[unreal ridge]

backboob is basically cleavage but from a different angle

[trim current]

love this

prompt

[lyric zealot]

1girl, black hair, twin braids, yellow detailed eyes, back, tattoo, yakuza, gangster, 


Negative prompt: (low quality, worst quality:1.4), (bad anatomy), extra digit, fewer digits, (extra arms:1.2), bad hands, by (bad-artist:0.6), bad-image-v2-39000
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 22, Seed: 3906485748, Size: 512x768, Model hash: 4d651c7638, Denoising strength: 0.53, Clip skip: 2, Hires upscale: 2, Hires steps: 10, Hires upscaler: 4x_fatal_Anime_500000_G, Dynamic thresholding enabled: True, Mimic scale: 7, Threshold percentile: 96, Mimic mode: Half Cosine Up, Mimic scale minimum: 4, CFG mode: Half Cosine Up, CFG scale minimum: 4, Version: v1.3.0```

[trim current]

thankiess

[lyric zealot]

[slate night]

Oooh imma try this once my current batch is done

[unreal ridge]

Damn

I'm getting some really good ones

but they show nipple

[lyric zealot]

[unreal ridge]

https://tenor.com/view/arknights-seseren-cute-drinking-sip-gif-17757374

[lyric zealot]

put nsfw in negative?

[unreal ridge]

I never put nsfw in negative

Might use bra as a prompt

[lyric zealot]

[unreal ridge]

Decent tattoo but weird braids

my model hate braids

[lyric zealot]

[unreal ridge]

I'll switch to ponytail 😔

[lyric zealot]

that's good enough

[unreal ridge]

skull tattoo

[slate night]

I might get bonked for this but

[trim current]

https://tenor.com/view/family-guy-peter-griffin-weird-weirdo-side-boob-gif-12911316

All i can think about

[slate night]

I mean same though /hj

[trim current]

I should do more SD but TOTK

I need to fuse my shield to a minecart asap

Ree

[open valley]

the face is awesome !

[open valley]

because "small breast" maybe ? so try to put "large breast" "big breast" in negative

[unreal ridge]

already moved on to another prompt

[open valley]

np

[gloomy scroll]

I'll join this back tattoo gang

[gloomy scroll]

[unreal ridge]

Shit, that's

WIDE

[quick viper]

what is the checkpoint used please?

[unreal ridge]

You can check the metadata

unless he removed it

[quick viper]

idk what metadata is honestly

[supple raptor]

darkSushiMixMix_colorful

[gloomy scroll]

darkSushiMixMix_colorful

[gloomy scroll]

I WAS FASTER

[supple raptor]

I didn't. I even added checkpoint names to my metadata to make it easier on people.

[quick viper]

thanks

[unreal ridge]

Just checked the metadata

how many extensions you're using?

[supple raptor]

Nope, I was.

[wispy canopy]

🍑

[unreal ridge]

F

[quick viper]

F Vil

[supple raptor]

Just adetailer for the initial gen and tiled diffusion + tiled vae for 2x upscale.

[unreal ridge]

No clue of what adetailer is

gonna search for it

[gloomy scroll]

https://github.com/Bing-su/adetailer

[supple raptor]

https://tenor.com/view/victory-gif-23899955

[supple raptor]

CBA to fix these hands so this one isn't getting upscaled.

[gloomy scroll]

@supple raptor Toes

[lyric zealot]

I'm going insane

welp

[untold glacier]

It seems like just the other day I was on a page that had a wealth of SD resources including a large assortment of different wildcard lists. Anybody know such a thing? I can't find it in my history.

[open valley]

[unreal ridge]

Damn, my luck ran out or some shit

I'm getting exactly what I wanted with my gens but the hands are beyond fucked

[unreal ridge]

https://tenor.com/view/kratos-god-of-war-rage-gif-25979591

[open valley]

do you have a lora or text inversion for that ?

[untold glacier]

Wildcard lists, anybody got?

[gloomy scroll]

[unreal ridge]

for hands? yes

just getting unlucky

[open valley]

sad

[untold glacier]

[unreal ridge]

You see this shit? Won't even bother inpainting or trying to fix it

just going to move on

lmao

[supple raptor]

Here you go https://www.sdcompendium.com/doku.php?id=start

[untold glacier]

I know the link, didn't think a list was on there but I'd be happy to be wrong.

[supple raptor]

https://www.sdcompendium.com/doku.php?id=wilcards:wildcards_0108

[open valley]

my gc is a 1050 so when i wait like 20 MINUTES (real time) for hands like pic 2, i'm VERY upset

[untold glacier]

I'm wrong, lol. Found it just as you did. Thanks Nubbs!

[ocean reef]

it's a miracle to get good hands in 1 try

[unreal ridge]

oof

[open valley]

i feel you

really

[compact warren]

Theres a lora for it now so it can happen more often

[supple raptor]

https://civitai.com/models/56519/negativehand-negative-embedding

[unreal ridge]

Ayo, the unrequested cleavage 👁️ 👄 👁️

[ocean reef]

Yeah, I know, use one of them but still, it's really hard

[untold glacier]

I'm still on the fence about that. I really haven't seen too many examples of it working properly.

[gloomy scroll]

Good hands cheat version

[open valley]

Very nice pic

[compact warren]

I have
I get better hands way more often than without

[gloomy scroll]

You are clearly missing this from negative

[untold glacier]

Then the way you prompt and/or models/samplers used varies a lot from mine, heh.

[gloomy scroll]

https://tenor.com/view/raccoon-steal-stealing-yoink-seeya-gif-10550534

[unreal ridge]

no

As long as I don't get bonked by clyde

[compact warren]

As does everyones. I prompt mostly with danbooru tagging techniques

[unreal ridge]

I'm a happy boi

[gloomy scroll]

Oh I didn't mean to reply

[open valley]

i just putted that on negative prompt and now... i have a bigger cleavage than before

[supple raptor]

cancelled a tiled diffusion upscale part way through and got a kind of neat result.

[open valley]

god blessed hands

[ocean reef]

1/50 😄

[gloomy scroll]

I meant to make counter for these cleavage images but I feel like my model can't do chibi with this in prompt

[open valley]

i never get cleavage you cursed mee !

[unreal ridge]

I see that as a win

[open valley]

no !

[open valley]

I don't want to see some boobs even covered for these ones i'm doing rn

[open valley]

wait why timer has diseappeared ?

i'll lose my image ?

[gloomy scroll]

When it doesn't disappear?

[unreal ridge]

Isn't that normal?

[gloomy scroll]

That thing in most buggy thing ever

[unreal ridge]

Happens 90% of the time with me

[slate night]

Mine just does that some times yeah

[open valley]

normally i have the timre

it's invisible only at the start for me

[gloomy scroll]

If you do something in other tabs it bugs often

[open valley]

oh

[open valley]

better than the previous one !

[gloomy scroll]

Previous one was older one I did with different model

[wispy canopy]

ex-hentai model has issue with flat chest

[gloomy scroll]

(Flat^2)

[wispy canopy]

lemme try neg prompting large breasts

[open valley]

idk if it work but you can maybe try putting a Cup

for the breasts

[supple raptor]

With the way CLIP works a cup is just going to give you an actual cup most likely.

[unreal ridge]

just put (small breasts:15)

https://tenor.com/view/clueless-gif-24395495

[gloomy scroll]

She's flat and proud of it

[wispy canopy]

large breasts in neg, that's better

not flat yet tho

[open valley]

"no breasts"

🫣

[wispy canopy]

no breasts has the word "breasts". lets not prompt that on my model

[unreal ridge]

no booba

[open valley]

just put breasts on negative maybe

[wispy canopy]

all varieties are now banned! XD

[open valley]

oh

[wispy canopy]

nope, can't get them smaller than this lol

[open valley]

put "chest binder" on normal prompt ?

[wispy canopy]

i would not be surprised my model kinda ignores flat chest

[gloomy scroll]

man's chest on female

[open valley]

yes, it's maybe trained to have big bobbas

[wispy canopy]

mixed. it was

it's an ex-hentai model

[gloomy scroll]

Time to mix few more anime models into it

[open valley]

my gen is cool but too many hair lol

[unreal ridge]

Thought it was branches for a sec

[open valley]

[open valley]

[supple raptor]

UFO? aliens confirmed

[open valley]

at the beginning my gen always show cleavage

[open valley]

indentified : leaf

[gloomy scroll]

https://tenor.com/view/eyes-burning-my-eyes-spongebob-gif-5183364

[open valley]

yes thats bc of my low vram

and my screen is not cropped

i'll fix your eyes with this

[gloomy scroll]

I meant your theme

[open valley]

idk how to change tbh

[gloomy scroll]

--theme dark

Or use browser with dark theme

[open valley]

I won't join siths

I'm the good side of the force

[compact warren]

boring

[supple raptor]

pretty sure you can also just add __theme=dark to the end of the url

[compact warren]

Theres an extension for theme editing

[wispy canopy]

lel

[gloomy scroll]

When I started using webui I thought there was only dark theme

[open valley]

i didn't put flowers on my prompt, only "flower tatoo"

and i removed "very long hair"

whyyy

[gloomy scroll]

Well flower is flower

[open valley]

i'll just put tatoo

but for the hair...

[unreal ridge]

bald
might fix it

[open valley]

[open valley]

need more sunglasses imo

[unreal ridge]

[compact warren]

What kind of stuff do i put in the hires t2i prompt box

[unreal ridge]

zentangle

[gloomy scroll]

In default prompt you put 1girl and in highres prompt you put 1boy

[wispy canopy]

@gloomy scroll LLM re-imaging of your girl

[compact warren]

Huh????

[supple raptor]

IMO it's only useful if you set it to save a copy of the low res version before upscaling or if you use live previews. Watch to see if things are being added where they shouldn't be, if so try removing whatever that is from your hires prompt. Not sure if Loras are loaded/unloaded in the transition between 1st and 2nd pass, but if so it could sometimes be helpful to remove some loras in the 2nd pass.

[gloomy scroll]

What you don't want to turn your waifus into guys?

[compact warren]

Remove or add? Cuz the negative prompt is there too

[compact warren]

[gloomy scroll]

Not an elf tho

[compact warren]

Downwards pointed ears make a difference?

[wispy canopy]

i know. LLM things pointy ear -> elf 😄

[supple raptor]

For example, the hires pass can sometimes add floating locks of hair. Removing tags that reference hair from your prompt can potentially help with that. You very likely don't need to mention long hair in the Hires prompt, because it should have already been established in the 1st pass.

[lucid rapids]

what VAEs do you guys like the most?

[gloomy scroll]

anime2

[unreal ridge]

kfl8 90% of the time

[lucid rapids]

I find that one very saturated, so I was looking for alternatives

[gloomy scroll]

Bless it

[compact warren]

So the og prompt is the first pass and the hires prompt is the second pass? And when i see something wrong in the second pass, i add or remove things in the hires section?

[lucid rapids]

wym?

[supple raptor]

I modified it to lower the contrast (which is the actual problem, not the saturation) https://huggingface.co/nubby/kl-f8-anime2-blessed

[gloomy scroll]

Don't know how to do it but here's some examples NoCrypt did
https://huggingface.co/NoCrypt/blessed_vae

[supple raptor]

bless09 is the version that I prefer. 08 is even lower contrast.

[supple raptor]

It's super easy to do and takes less than a minute to make one once you have it set up. https://github.com/sALTaccount/VAE-BlessUp

[gloomy scroll]

I guess these are for older models

[open valley]

the landscape is wow

[supple raptor]

the ones from NoCrypt are just the NAI vae with increased contrast.

[gloomy scroll]

Elden ring embedding?

[lucid rapids]

yea i see the contrast diff

[supple raptor]

It has both Nerf's elden_ring and lands_between embeddings.

[gloomy scroll]

I guess time to move those back to embedding folder

[supple raptor]

[open valley]

a second girl appeared during the upscale

wtf

[supple raptor]

what's your denoising set to?

[gloomy scroll]

I kinda moved all my embeddings to another folder since I sometimes noticed that I accidently used embeddings in prompts

[open valley]

0,7

[gloomy scroll]

Max .5

[open valley]

oh

[supple raptor]

yep. that's the problem. I use 0.51

[open valley]

thanks

i'll remember

[gloomy scroll]

if you use sd upscaler or some other tiled upscaler then I wouldn't go over 0.4

[supple raptor]

for sd upscale script you probably want more like 0.3 with 2M Karras.

[gloomy scroll]

I used part of your prompt and now I saw something I didn't want to

[gloomy scroll]

I use 0.1 myself

[supple raptor]

which part?

[gloomy scroll]

extremely detailed background, landscape, (grungy, dirty:1.5), (eldritch abomination:1.2), (horror:1.3), extremely detailed, intricate, elden_ring, (flower_style:0.7), bokeh, art by (fromsoftware:1.5), (dark souls, demon's souls:1.5), lands_between

[gloomy scroll]

I'm too lazy to check so here is mine

@supple raptor

[open valley]

i don't have other upscaler

only the default ones

[gloomy scroll]

Doesn't rly matter

I think SirVeggie uses default upscalers too

[open valley]

her arm is not even humanly possible lol

[supple raptor]

Here's where to get the other ones. They don't make a huge difference for Hires fix since it's running img2img on it after upscaling anyways https://upscale.wiki/wiki/Model_Database

[open valley]

thanks

i didn't know the process

[supple raptor]

Yeah, Hires fix is just normal txt2img at the resolution you set, upscale to the new resolution, then img2img on the upscaled result.

[gloomy scroll]

@elfin geode Is prompt inspector bot still in development?

[elfin geode]

The other <@&1060374302433738782> can answer that!

[gloomy scroll]

There's multiple of you

[rose tapir]

👀

[elfin geode]

He is summoned

[rose tapir]

Unfortunately not as Discord was being very wonky with exif/metadata so at this point I've basically shelved it

Not what you wanted to hear but that's the unfortunate side of it!

[gloomy scroll]

I was just wondering since there ready made bots in github

[elfin geode]

Is true! I think the metadata issue also affected the promptinspector bot?

[rose tapir]

Right right, however once those were out, Discord stripped exif from images afterwards, rendering it useless

[gloomy scroll]

But at the moment bots work fine

[rose tapir]

Ah, it's back up?

[gloomy scroll]

I believe Discord reverted it's change

[supple raptor]

Yeah, exif is not being stripped any more. PromptInspector works as intended even without the stealth_pnginfo extension.

[rose tapir]

In any case, I can work on that again, however I'd have to make it from scratch as several other programs are being used outside of the A1111 inspector, so it'd have to differentiate the exif from each one

Would love to have it in as wide range as possible to cover as many image types as possible, from each one they're made in!

[supple raptor]

Here's the repo for PromptInspector if you want to take a look through how Salt built it https://github.com/sALTaccount/PromptInspectorBot

[gloomy scroll]

Maybe it's time for me to go read how bots work so I can join another server's bot take over

[rose tapir]

Saw that one before! But that one doesn't cover all programs used for generating images, and a few have different ways that they're storing it

It'll definitely be on my list to work on now, but got a few other priorities first, thanks for the heads-up though!

[gloomy scroll]

I only know a1111 and Comfyui formats

[supple raptor]

Yeah, just figured it may be a decent reference if you hadn't already seen it.

[rose tapir]

Appreciate it!

[rose tapir]

That one and have to have it check for SD exif too since it's in a different format. If only they all shared the same format then it'd make life easier!

[compact warren]

Wait i just noticed the two new samplers
Are they any good

[gloomy scroll]

2?

[compact warren]

Well ive always had my sampler list filtered, so if there are new ones I definitely would notice

But then again, i don’t update too often

Unless ik its safe

But anyway, its the 2m sde and the 2m sde karras

Those werent in my list before

[gloomy scroll]

Oh I didn't notice those

Well I don't rly change samplers so it's not surprise that I didn't know

[compact warren]

2m sde seems about as fast as normal 2m karras

[supple raptor]

Yeah, I like 2m SDE Karras. I ran and posted a few comparisons yesterday.

[gloomy scroll]

Reminder that hands don't exist

[gloomy scroll]

So this is what happens when I put lot of focus on background and have bokeh in prompt

[trim current]

i like it

[gloomy scroll]

I don't like how blurry that background is

[near elbow]

speed blur? 🙂

[trim current]

bokeh is like light effect that'll tend to do that

[lucid rapids]

maybe (bokeh:.7)?

[untold glacier]

Wow. Neutral Prompt looks great for people like me who use composable diffusion (AND) a lot in their prompts: https://github.com/ljleb/sd-webui-neutral-prompt

[sharp bluff]

do you still have to crop images to 512x512 or 768x768 for lora training

or is that a thing of the past yet

(im lazy)

[gloomy scroll]

This could help with your laziness.
https://www.birme.net/?target_width=512&target_height=512

[sharp bluff]

yeah i use that

im just still lazy with 200+ pics

[untold glacier]

[sharp bluff]

me ^

[supple raptor]

cropping is a thing of the past. just enable bucketing. you don't even need to resize images unless they're insanely big. If you happen to have any extremely wide, or tall images you may want to manually crop them down to a more standard aspect ratio.

[gloomy scroll]

Just hop in bed and get cosy

[sharp bluff]

ooo did not know that

tyvm

[supple raptor]

I think you'd have to modify how composable diffusion works to use it with this. AFAIK composable diffusion hijacks the normal AND function and uses that as the trigger to figure out the region that each prompt should be applied to. I dont think it would work out of the box with the AND_SALT or AND_PERP triggers that this uses. It does seem like a far more sane way to use AND in normal prompting though.

Or, maybe I should read more, because they imply that it does work with composable diffusion in the known issues section

[untold glacier]

"Neutral prompt is an a1111 webui extension that enhances the denoising algorithm of the webui. It replaces the original algorithm with a higher-quality implementation based on more recent research."
I think I'm in favor of this.

[gloomy scroll]

Reading

[untold glacier]

It seems to keep the original intention of the AND function, but adds AND_SALT and AND_PERP

There is a caveat for some of you though: "The webui doesn't support composable diffusion via AND for samplers DDIM, PLMS, and UniPC. As Perp-Neg relies on composable diffusion, the extension will revert to the unmodified sampler implementation when these are used."

I use 2M Karras Test though, so I'm good there.

[supple raptor]

Have you tried 2M SDE Karras yet?

[untold glacier]

I haven't but I've seen others tests of it and I'm not impressed. I'll try it soon though.

It seems to need more steps to work properly.

[supple raptor]

I've been using it at 20 steps just like I do for 2m Karras and I like it so far.

[untold glacier]

Hmm. Well I trust your opinion more than I do the average bozo on the interwebs so I'll have to give it a try.

[supple raptor]

Here's a few comparisons I ran. All at 20 steps.

[untold glacier]

Downloading to compare

[supple raptor]

at some point I'll redo the comparisons at 15, 20, 25, and 30 steps

[sharp bluff]

wait so should i change this then or leave it as is?

[supple raptor]

Some people say that training at 768x768 allows more details to be learned and/or better hands, but it will take more vram to do it. With bucketing, the resolution you set is treated more as the maximum number of pixels so 512 x 512 = 262144. Your images will be resized to have the same number of pixels, but they will retain their aspect ratio which could be something like 768x341.

[sharp bluff]

ohhh i get it ty

[untold glacier]

It's looking pretty good! With grid-0172, are the images scaled down compared to their originals since they're in a grid? I'm seeing a remarkable amount of detail in the small features: eyes, etc.

[supple raptor]

That's the magic of the adetailer extension. Automatic full res inpainting of faces.

[untold glacier]

I was just looking at that... I guess it's worth an install then. Is it difficult to work with?

[supple raptor]

It's incredibly easy. Most likely you can just turn it on and forget that it exists.

[untold glacier]

[supple raptor]

It allows entering a separate prompt for inpainting but I rarely do unless there's a specific effect I'm trying for. Depending on your sampler, you may need to adjust the denoising that it uses to avoid getting a visible rectangle around faces, but he the default of 0.4 works pretty well for 2M Karras and 2M SDE Karras.

[untold glacier]

Sweet Christmas. I can't wait to put it into action.

[supple raptor]

Just watch out if you're genning groups of people. It can make those take a long time because it will inpaint each face.

[untold glacier]

Ah gotcha. So in effect it's a much more advanced form of codeformer, but using the SD model to gen instead.

[supple raptor]

Sort of. All it really is, is face detection (it also has hand detection and full body detection). It masks the full rectangle that the face, hand, or body are detected in based on which adetailer model you select and then runs inpainting at full resolution on that area.

[untold glacier]

I do wonder if it'll work properly for me though, as more than half the time I've had problems inpainting. It is the heel to my SD Achilles.

[supple raptor]

Johny Bravo has a gun! Everyone get down!

[untold glacier]

[supple raptor]

[sharp bluff]

alright lets hope this one goes better than the last

[gloomy scroll]

[untold glacier]

OOOooohhhhh I love it so much! This is the magic I was looking for!

LOOK AT THE DETAILS ON THAT SMALL FACE

[sand geode]

Anyone have some obscure extensions/models/Lora's for anime? I'm bored and want to have some fun

[sharp bluff]

is it bad if i have to use my lora at 1.2 emphasis for it to look good?

like it looks perfect at that strength but idk if thats a bad thing

since most peoples are at like 0.8-1.0

[untold glacier]

Weight can depend on the model you're using it with. If it looks best at that for you, then great.

[sharp bluff]

Oh true I haven't actually tested it with other models yet

I feel bad having it on such high emphasis 😩

Gotta fit in 💀

[untold glacier]

Don't listen to anybody telling you that having things set one way is a bad thing. What this all boils down to is what looks best to you. I've worked with LoRAs like the This Is Fine lora set to 1.5 or higher before. If it looks good, then it works for your purposes.

[sharp bluff]

yeah i just bumped it up to 1.4 and it looks even better

this seriously turned out so good

[supple raptor]

The only concern I would have about it needing that high of strength is that it might effect how well it works with other loras. That's just a guess on my part so it may not even be true.

[sharp bluff]

ive only tried with 1 other lora so far and not so great :/

[supple raptor]

A little more training might help with that. But either way, it doesn't have to mix well with other loras, that's just nice for making them more versatile.

[sharp bluff]

yeah i think i wanna retrain it but im not sure what id change?

[supple raptor]

probably either just a little longer training or very slightly higher unet LR.

[sharp bluff]

so more epochs?

[supple raptor]

more epochs or more repeats would both work for making it train longer. how many repeats and how many epochs did you use for this version?

[sharp bluff]

3 repeats and 10 epochs

[supple raptor]

Assuming you have it set to save checkpoints at every epoch, I would maybe try training like 12 epochs. that way you can try both 11 and 12 to see what you like better.

[lucid rapids]

has anyone figured out how to make a manga panel with ai yet

[supple raptor]

Generating separate images and manually compositing them together and adding text is the best way. With the right tags you can get it to generate something that fits the format of a manga panel, but good luck trying to get it to make a story one that actually makes sense.

[lucid rapids]

lmao

yea i was just wondering if the tech was that advanced

[stuck aurora]

I drew the eyes, is it good?

[supple raptor]

Not even remotely close. The AI doesn't have any understanding of what is going on in the image, it's basically just matching noise patterns to the words that you put into it and when everything goes right, that noise turns into a decent looking image.

[lucid rapids]

just made a thing

lmao

[lucid rapids]

yea thats what i figured

wonder if itll eventually be possible 🤔

[supple raptor]

Not likely with diffusion models like we're using now. It would probably be doable with a multimodal ai model. Combine a decent LLM, vision, and diffusion model all together and it seems very plausible to me. Unfortunately something like that is going to take significantly more vram than most of us have available with current tech.

[stuck aurora]

what model did you use?

[nova remnant]

https://tenor.com/view/anime-anime-girl-head-nod-head-nodding-head-nod-yes-gif-24603507

[pale monolith]

feel like you could do it with a regional prompter and a manga base prompt

[stuck aurora]

How did you make it into manga? Did you add "manga panel" or something in prompt?

[lucid rapids]

yeah basically

[stuck aurora]

oh, thanks!

[lucid rapids]

try manga panel, black and white

[pale monolith]

it also works in color

[stuck aurora]

Wonder if img2img works with prompts too, or it does not consider the prompt?

[supple raptor]

Yeah, I suppose that could work. I feel like the success rate would still be really low though.

[lucid rapids]

ooo

[nova remnant]

well it did what you wanted

now all you gotta do is erase the text in photoshop and replace it

[lucid rapids]

lmao

[crimson stirrup]

cursed

[nova remnant]

..

[shell timber]

the tags 2koma, 3koma, 4koma, or 5koma works

[lucid rapids]

wat those tags mean

[shell timber]

comic with 2,3,4 or 5 panels. although the exact number of panels never seems to work,

[lucid rapids]

ah ok

[lucid rapids]

anyone got some niche checkpoints i could try on civi

[pale monolith]

nah but i got a long list of loras

[sage bluff]

Tbis cane out really good

[lucid rapids]

👁️

@dull jackal

[pale monolith]

but can we do better?

[slate night]

Lmao why do people still fall for these scams

Bless their hearts

[nova remnant]

thanks for free nitr0 😄

[lyric zealot]

because people always want free stuff

[nova remnant]

true

[lyric zealot]

anyway, right click > app > Report to staff

but best to straight ping staff first if possible

[nova remnant]

[lyric zealot]

then if no response then do that

so that they can get to it asap

[sage bluff]

So@good

[nova remnant]

Fair :0

[lyric zealot]

what makes these bots, join so early

just to do it like months months months later

[runic vault]

i guess it's a way for them to try to get past automods. just sitting there like a ticking time bomb

[lyric zealot]

likely

or might be an already hijacked account

[slate night]

I'm leaning toward hijacked account

[lyric zealot]

yup, so don't click on that link, if not, you might be next

[keen hollow]

evenin'

[pale monolith]

👋

heya

[keen hollow]

what's up?

heard there was a big auto update. interesting

[magic wave]

Was there? :o

[keen hollow]

Think there was recently

[magic wave]

Oh, is it the added new live preview method or soemthing else? :o

[weary dune]

need fix hand -_-

[hardy tree]

saw this on pixiv from ReikArts does anyone know what model makes Illustrations like this?

[weary dune]

u can try find it on civitai
there lot of models and loras with images

[wispy canopy]

i'm going to get a lot of things done today