#🍥|anime
1 messages · Page 49 of 1
No game no life s2 when
and that'ts it. I don't even remember wwhat happened to her
Pre much

DAMN THESE ELVES HAVE BEEN POWERGRINDING FOR LEVELS!
what, I am pretty sure the ending was good or somethin
yeah it was
tbh I don't really remember lol
I now remember they were in a "hotel" at one point
one in like a desert or something?
ok I'll just rewatch man
more like post-apocalyptic/dystopian, in s2 at least
i liked it though, i didn't think they could re-do the magic from s1
do I need to rewatch s1 aswell?
it's a pretty good series, so go right ahead

I also forgot the whole of monster right after I claimed it as my favorite anime
idk anymore
didn't like the anime, loved the manga
oh boi. that's like... so many of the series
I am now buying it
the songs in the anime are good
this this armor has excellent features!
the back of the head is totally exposed
does it resist lightning
it is a helmet
masket!
EL PSY KONGROO
"my clock stopped working"
dude. that was so hard on okabe man
totally get why he broke down
"doesn't matter what you do. she dies". this is fun information to know
It's like she's trying to stop me from asking wtf is up with her friend's hood.
me vibing to the op rn:
yeah, those intro songs were bangers
idk what this is, but it looks cool
Randomly (well not really, I only took those that I liked) picked up from a store that offered free stickers
clean your damn dust filter and put it back on right
hehe
you should see my room lol
eery
not kids rly. just me
too messy
it's alright though I am gonna move houses in a couple of months
did you forget to save the version where I fixed her toes for you?
I think it's in the download folder
oh hell naw
bruh did the "super hax0r" in steins gate hack SERN (CERN) with an SQL injection?
Thanks
It's my civic duty to ensure toe quality.
I made it on the starboard lets gooooo!
feet enjoyer
Looks promising
bruh
either locon or epi
gross. epi is trash
know a fix ? 👀
Reload

Use lighting-locon from here. Usually it only needs around 0.4-0.6 weight. Then define the lighting in your prompt such as darkness or bright https://huggingface.co/closertodeath/ctdlora/tree/main/locon
both are also destroying detail
the locon too?
ye
By that logic, anything except the model itself is destroying details
what?
wat
what logic?
locon destroying details
by my testing, lowra destroys everything, locon has the best results
(tested different weights)
is lowra a new thing, or u mistyped lora
You'll want to do that with your prompt. When offset_noise is used properly it shouldn't inherently change the brightness much, it simply allows a wider range of brightness. If you want it to be darker, use words like darkness or night and increase the strength of those as you see fit.
so its a lora
i did add (darkness, dark, black:1.3) silly
yes
Not sure then. It works well for me. This is at only 0.3 weight.
that is not as dark as i want it to be
I get that, and to make it darker I would increase the emphasis of darkness associated tokens and/or the Lyco.
do you have dark background in the prompt? That can help. If it oversimplifies the background you can use prompt editing to have it dropped from the prompt after the first 7 or so steps. There's also the old trick of doing img2img on a completely black starting image with extremely high denoising if nothing else works.
first time seeing someone say "lyco"
i thought i invented that i2i trick
LOL
also I am waiting hard on someone to drop a LoLiCon
you better not be waiting hard
ayo, not what I meant
He's a child anyways so maybe it's ok? I'm not really sure tbh.
290 gens, good night
let's say it is
_not 100% sure if this falls under anime (correct me if I'm wrong) _
Anyway til that I can add black and white watercolour painting and black and white Japanese watercolour painting to my prompts and get things like these :)
this def anime
Cool thanks, I was like 90% certain :). The 10% uncertainty came from using style: Fantasy art for these.
is this intentional?
The left? Yes, absolutely.
translators trollers
Maybe you install it where it says install from url?
oh
You may also need to add .git to the end of the URL
for now I face problems with quality'
i don't what you add to SD so you get this quality
have you downloaded any "checkpoints"?
yes
can you send example?
meinamix
I showed you before .
It comes with experience. For now when you see an image you like, try clicking the image and selecting "Open in Browser" "Open Original". Once you have the image open in your browser, you can either download it or drag and drop the image into the the PNG Info tab. Depending on how the person uploaded the image, you will likely be able to see their prompt and settings.
I mean you only sent img2img
im no betting man, but i bet he's using a resolution higher than 900x900
without highres
For better quality try using negative prompts, and also use prompts like "(masterpiece, top quality, best quality, official art, beautiful and aesthetic:1.2)", one that i use
best advice
altho, unless u are paying for internet, show the example lol
yes I face problems with img2img
only with that?
I use them and also
I don't see problems with text2img
Ah yes, giving the book a chirurgical operation
I think the resolution you put isn't close enough to the image you are trying to img2img on
right
so it stretches/squashes it
what happened to the other tighthight
She only wears one
Only needs to do laundry half as often
fashion
what negative promts i should use ?
you can go check images on Civitai and get a bunch i guess, its what i used to do
okay
civitai trash
man, i wrote the same thing twice 
oh, day2 picture, impressive
free prompts though
download like a thousand embeddings and then use a couple of em
sadly, no alternatives, also its doing pretty good these last days
for model hosting it is fine
but for images
trash
most people, or atleast from what i remember, edit the images or something so i stopped trying
umm, why ?
nah
not most
some do
I did what thezenith said and it's good
and some just don't have png info
the info you get there is not complete
I think I can continue with that rn
how ai destroy fingers
I hate this
One (1)
way better than before
YES But i want to make animation
so I should do it with animtor extention ?
Can't answer that to you. Never tried doing animations
in a way, yes
when u select a style, it will automatically gets added to the prompt (u can check after the generation, on generation details)
most cursed clip skip experiment
okay thanks
this is amazing
give up
I used photopea to remove eyes and edit some things
huh?
animation is currently bad with ai
I have animator extention and I just need to get it
howd you make this 😮
or that
I tried to make 2 animations and I get good reult
this one is not mine
yes It needs editing everyframe
wdym by good?
how do i create these ? 🙂
Look
I made this but not without stable
but I can make better if I spend more time
yes I know
how do you make them >
as I said you can do stuff like that
I used img2img
img2img
I forgot the name of the one that made them
maybe I can fix the frames and editing some things . but yes it won't be that wor
so to make the little short changing frames is img2img
the ones I sent are not img2img
what are they 😮
I forgot one of them but I think one of the techniques is prompt travel or something like that
my pc : some mercy
YES fizzledorf
Just don't look at her double thumb
THankyou
Prompt-Interpolation and deforum are like the main ones imo
but now there are some new stuff idk much about
i tried "thumbs up"
Beautiful lmao
:3
evenin'
#1072220168534642768 A blue bird
quiet tonight
using multiple (((()))) instead of weighting it (kk:1:3) can cause tokens to mismatch
and screw up prompts
even dwagons do karate
"to hell with stranger danger babe, hop in"
going sideways on a highway @65mph
Tried recreating one of my maplestory characters from txt2image only, think i nailed it? 
Yeahhh i tried... couldnt get it to work. Im still fairly new to stable diffusion. Could it be that the model i was using just couldnt understand what i was trying to do?
chest too huge
Its funny you post that cos im working on smth similar 
Nice! I dug that prompt up from a couple days ago because I needed a break from generating Kiki (my goblin OC)
dang my model is nicer than i thought, really shines through higher res
that's very pretty
why is it always the hands 😢
The AI experience
seems my model does great when you feed it stories
modern day elf?
hm?
black sweater, white undershirt but.... it blends?
i think it interpret it as shadowxd
hmmm... now i just need to make this LLM chat i have into an actual initiator prompt... i got it working in the style it actually collects the most important visual features at the 75 token marks, so it writes a coherent story with visual elements, and my model really seems to like it.
full of magic 🤩
cannonically, isn't it impossible to beat him?
the amount of plot armor that dude has
The power of plot armor. lol
xD
plot-armor, the dude.
aight, someone take over the daytime posting!
imma go make some breakfast
10/10
unfortunately got lucky and the only good one so far, all the rest have the rider fused with the dragon
add_detail lora is like the best lora that has ever existed!
i'm nearly only using it for the -1, not the +1 xD
add_detail + aDetailer extension is just something else
(zoom into my last pic)
or this
somehow, i really love the -1 power, something in this style just does it for me
now it's time for the women to thirst... although. dang. not homo but these be mighty fine looking ❤️
@merry tendon that's with add-detail on -1 lol.
lol, -1 from regular gigachads to 0.5 ultramega shiny-chads.
wait. why is it that the -1 chads can hold their swords properly, and have actual proper swords?! ULTRACHADS!
guten morgen everynyan
morgen 🥱
i am a happy prompter 😄
which model?
still using alfahentai
An illustration of a person standing confidently against a bright background filled with stars and galaxies. They appear determined and strong-willed, with their fists clenched at their sides and their head held high. Colors used include shades of blue and gold, representing strength and hope respectively. This the actual positive prompt 😄
now i have to be sure to save the LLM prompt though
whats a llm prompt
prompt you use for Large Language Models
guys wtf, this is AI generated (with use of controlnets ofc) https://twitter.com/emokakimasu/status/1632310859384127489
Basically, i'm running both a large language model and stable diffusion at the same time, i feed the LLM stuff, and it spits out something that i feel works well as a prompt for my SD model
a LOT of inpainting and or corrections were used tho, no way that's a single shot
ofc
and use of controlnets
thats very smart, llm generates faster ideas iguess?
you can see most of their process in the tweet
I mean if you schizo prompting then LLMs could probably help
but if you have something specific in mind then not really
okay makes sense 😄
man I always forget so much stuff that when I rewatch animes it almost feels like the first time
like how I am rewatching steins gate rn
too bad i can't unforget suikoden 2 😦
yeah. that's the one
one of the best, if not THE best old school rpg
imho
Anyone know any better settings when training on images that are like 5000x2000 give or take In resolution?

whu... does that even work...is that even possible?
This one artist makes really cool shiny images, but every time i train a lora, the details are never crisp or clear, in fact it looks blurry
for example
The texture of the image is very grainy
And im starting to believe its bc of how high the resolution is
Now i'm no expert, but shouldn't training be done one 512x512?
Either 512 or 768
iirc there's a setting called buckets which can bypass that
Yeah
(sorry, i'm a total noob when it comes to training and that stuff)
But idk what bucket settings i should set this for
The highest ive ever gone is 1024
Not even sure u can go higher
I mainly just follow holo's guide and colab to make loras https://civitai.com/models/22530
I use kohya unfortunately, too
I have better guide for you
This is focused on beginners. The training settings probably aren't 100% optimal but should do well.
Github repositories used in this guide:
1: Kohya Script: https://github.com/kohya-ss/sd-scripts
2: GUI to use Kohya Script : https://github.com/bmaltais/kohya_ss
3: Scrapper: https://github.com/sA...
best guide ever
No need for a guide anymore though if u already have a grasp on what youre doing
no no
because people don't use DAdaptation
which is best
And why is it best
I used it once the other day, seemed fine like my adamw trainings
I don't remember explanation by heart
let me search
Dadaptation (D-Adaptation, D is short for Distance) is an adaptive algorithm applied on top of an existing learning optimizer and will automatically find the ideal learning rate for your dataset.
For the Kohya trainer, the Dadaptation option applies on top of the beloved AdamW.
Here are the Pros for Dadaptation:
-Ideal LRs found automatically by optimizer
-Earlier epochs are usable compared to AdamW
-Use only one setting for all your Loras
Here are some drawbacks:
-Training speed is slower than AdamW
-higher risk of overtraining at earlier epochs
-No improvement if you have already found your favorite AdamW settings
-higher vram usage
"Tired of searching for the best learning rate and guessing cosine repeats? Have Dadaptation do it for you."
So word for word from the projectai guide page
Yes i have seen that
It looks like the cons outweigh the pros here by number
what how?
training speed - idc about that
higher risk of overtraining - just use earlier epochs (happened to me once so I just used epoch 1 and it came out good)
No improvement if you have already found your favorite AdamW settings - I don't wanna waste time finding the optimal one for every single lora I make if it doesn't even improve it by that much and also it would take overall less time probably to just use DAdaptation and would give practically the same results.
higher vram - cope
Good guide but I use colab because my vram is limited
I said by number
The cons outweight them
Not in usefulness
then you can try to match the json in the colab, I saw you can change the optimiser there
oh in numbers
bruh
that's what we do
sigh
I just really hate having to learn new settings for this stuff

Even if it is easy
you don't have to learn
just copy paste
it finds the optimal thingy for you automatically
Ill try it again after my current training
you can also just use this config file if you are using kohya ss gui
good morning
wtf
wait, png info is broken again?
only for me
what version of discord do you have?
it never worked for me
update discord
there's no button for that?
close discord completely and reopen
i started my pc like half an hour ago
Now... i'm not saying this party is fcked.. but i don't think this is a bossfight they can win.
open original > save as
right click save image
open in browswer -> drag picture to my "stolen SD pics" -> open that image in sd image
I think all these ways should work
the important part: OPEN IN BROWSER
the direct image from discord has pnginfo pruned
wrong
nah
i always did
try right clicking -> save image
straight on discord
should work
works
where prompt bot 
projectai
god i wish i had the smuggered emoji rn
nah I meant a promptbot to read pnginfo, we've been suggesting it since forever
they been planning to add it
they just wanna make it a bit better or something
there's a prompt inspector but only at touhou ai server
i hate how it dms you tho
it should be like those messages that only you can see
I quite like the dm, so you can always go back and see
something like the /pnginfo or stuff like that
RWBY - I burn (yagao song)
the LLM+SD really nailed the lyrics interpretation
fuck my close-up:1.2 in negative ig
omg your pfp is best girl fr
also best form of best girl
pixel hanekawa fr fr
ULTRA-KILL!!!
am trying. not sure if my prompts/model work well with taht though
cool
lol
dragon
DEE- [redacted]
i sized down and upscaled like 4 times to see what would happen
considering the og img was this i'd say it was a success

i nearly know the feel bro...
i don't like deleting data...
what language is that?
looks like a pile of blueberries
bye moon
What if Shinji just got in the robot? Michael Bay's Evangelion would make sure it happened.
Subscribe now for more Animation Domination High-Def clips: http://bit.ly/RXLBii
Watch Shows + Make GIFS at http://www.foxadhd.com
Like Animation Domination High-Def on Facebook: http://bit.ly/ADHD_FB
Follow Animation Domination High-Def on Twitter: htt...
why is there no lora for these badass fuckers
reject end of the world, embrace serinity
naw bro fuck that
i need my christ explosions
neon genesis evangelion, cross shaped explosion
lora needed...
Bro where can I see ai art nudes?
lemme try in my model
LOL
wrong server. go either to one of the discord channels that DOES do that, or 4chan, hdg channels
(bene there, done that)
once you're done with ai nude art, come back here ❤️
oo
Huh where da stable diffusion discord
here
i feel my model got closer with just those words already
i'd say it's almost there
that means it did grasp the concept
Does the nsfw tag on an model mean it's purposely made for nsfw?
Or it can make it and author indicates it through tag
Give ma a web to make nude art
it means it can do decent nsfw, not necessarily that it was built for it
mine is nsfw, i'm hardly using it for nsfw
Was wondering what was it
#🎥|animation of my picture
that, we can't do.
lol, her crotch exploded? XD
gonna do gens without prompt to get inspiration
it knows it should be making crosses/lines, but just can't grasp it
image assistant or professional philosopher my christ
that's why llm prompting is golden ^^
where should i start with those?
- have beefy gpu 2) have BEEFY gpu 3) in case pooint not clear: you need 12gb+ gpu
you can load lighter models, but the 13b models are just MILES ahead of quality of responses
dunno if you can run them on collab
i thought it was some site online
What GPU?
no, running this on my pc
ah, misinterpreted that
I'm asking him what GPU does he have 
6gb, probably 2060-70 or smth?
1660
probably one of (if not) the worst gpus in this chat
nah, i've seen worse
i remember when people made fun of 1080 ti for being an absolute monster
i've got the advantage being an older dude, and living alone -> can invest my money in my pc
gtx1060 6b
... i got a 3090 and it's not the most expensive single piece of equipment on my pc lol ...
tf is then
my headset, audeze lcd-gx
WHAT TEH FUCK
HOW
HOW DO YOU SPEND MORE ON YOUR HEADSET THAN YOUR GPU???
THATS A WASTEE
because i'm an audiophile at the same time
nah, once you hear good sound, and you enjoy that, you can hardly go back
Not the best choice, but you can run SD on that.
--lowvram and it will work, slowly, but will.
so true
got a $200 soundcard too
i don't even use medvram
and it's totally worth for me
tiled vae carries me to >1k dimension
i have audio interface + studio monitors + studio headphones, but my pc is trash
takes me like 15 seconds to gen 800x450
i went from open -> closed -> open headsets... i'm having a REAL hard time adjusting to closed headsets 😦
they're great when it's noisy, but the audioscape 😦
That's cool, I got AMD 
use tiled vae and remove low vram stuff
i have to sit in 60 degrees exhaust air for a minute long doing a long generation xD
weren't you the guy who used over 200 steps total
not 200
i use a half open over ear headphones from german company. almost the exact same fidelity as my studio monitors, so thats a win for me. feels very natural
but yeah, i do tend to go high
I use medvram
tiled vae gives me gray tiles at the end of generation
beyerdynamics?
you german?
actually Teufel. im german yes
disable
als ob
and set encoder tile size to 512
😄 ja doch.
ich hab gedacht hier ist kein deutscher XD
I NEED THIS SHIRT
ok back to English now
waves to his country neighbour, hello from holland
I'm fine with my current settings, it works at least
istg tiled vae is a godsend and if you don't use it you'll regret
:p ist doch schön. also i love holland ❤️
@merry tendon megu2bro, this u?
who can make the best eyes at a distance that are still super cool to look at? here is my try:
maybe👀
society
do this at -1, i wanna see your model make lineart ❤️
k wait
the hair gives away that add-detail was used 😛
that allowed tho, all tools are allowed
1.2 weight tho
why?
i think i'm just a sucker for clean lineart
nice, i like super detailed or lineart, something in between not.. it just the most generic bums
@sullen hollow edge dude, it seems
wait, is mal a girl?
wait. is he/she?
"this is the internet, everything is male unless otherwise confirmed. even then, still male"
i swear i dont look like this in real life
i'd wish i'd look half as cool as that tho
no context eface, lets add my features
uhoh. llm made eface this 
try me please
i think my name is cool

how lewd, she's touching boobs
first prompt kinda failed imho
cocacolastic
Mefistic!
i knew the name was badass
these are my actual prompts i'm using in my model though 😄 try using stories in your own prmopts, discover how well they work!
negative prompt leotard, gets rid of the crotch grabbing from the dress
ohhh, that's big, goodsuggestion
didn't know
hello
from 1 to 10, what would you guys rate this pic(zoom in for more detail)
I even lost that guide I used before. How to install it? I have Automatic on local machine.
8.5/10
what's that?
a vae? or a model?
this image gives me that "tales of" feeling so much
oh.... inspiration!!!!!! images of the saga of tales
megu, what's this model?
webui > extensions > available > click load from > search "mu" > find "MultiDiffusion with Tiled VAE" > click install at the right > back to installed > click apply and restart ui
yep, absolutely
Where is that setting located? I couldn't find it.
the meshing of colors is perfect
!!! THERE IS A FP16 VERSION IN THE FILES DROPDOWN !!! !!! THE NEWEST VERSION IS v-obsidian-1.0 !!! Some of the images might use embeddings or Loras...
if you want the same style also use the detail tweaker lora
oh, it's Xynaptix
it's been in there since NAI, and it irked me from day 1. leotard fixed all those issues 😄
@wispy canopy I haven't done story based prompts since midjourney, maybe I should try again.
you should. darkgemini thrives in it!
nice
i've done the tests, your model is REALLY good on story prompts
you know it?
I wish the add detail lora didn't affect the face and hair as much, otherwise its a really good
bat can't make microphones XD
0.5 weight is perfect
It didn't want to add dead bodies though
i got triggered so much by it yesterday lol
the face is upscaled with aDetailer, and style lora is added, so not the fault of detail tweaker
i was there
uahahah
wtf is adetailer
Ah yes, indeed. It's inside the extension.
So, this is all what I need? Just like on your screenshot?
What about Tiled diffusion?
i tried it once and it worked, you just need to have the same prompting style as me(probably)
an extension
tiled diffusion is for upscaling in img2img
what did you wrote for the mic?
wow
idk was some time ago
detail tweaker at -1 sure is something
oh ok
what does it do
Is Move VAE to GPU needed?
no
it does an img2img inpaint fix automatically on far away faces so they become more detailed
inpaint+uscale the fac automatically
this. i feel it's way better at -1 than at +1
+1 is too much imo, i use +0.5
Okay, thanks. I'll try it.
So, what exactly this extension is supposed to enhance?
@cyan burrow
i use 1.2 lol
yeah, +0.5 is plenty enough for most anime models
vram usage is MUCH lower for bigger images
what model?
eh
a link to detail tweaker
I think yours
ok 👍
ty
I think there was a steins gate style lora some while ago, can someone pls link it?
am i the only one who really dislikes the style here?
Usually I can't generate anything more than 632x632 with hires.fix and when I upscale I can't use more than, well the same 632x632, I got not enough memory.
So with this extension I potentially can get more?
https://www.pixiv.net/en/artworks/107292962
img2img from the pixiv link above
DONT use highres fix
that's my take
OMG, in another server everyone were convincing me to use it.
too much detail hurt eyes
i recommend only upscaling on multidiffusion
???
hires fix is great
and you should use it
i dislike it a lot
its purpose is not to upscale but to fix the problems high resolutions make
finally someone who appreciates it!
because models are mainly trained on 512x512
I upscale later in img2img with SD upscale script and 4x_Valar_v1
what is that upscaler
and that's why something specific for upscaling is better than something simple like that
so it fixes stuff like duplicated heads and weird compositions
which are the effect of models being trained on 512x512
no
I have no Idea, found it somewhere on Reddit, they said it's great, and yeah, I found it's better than R-ESRGAN 4x+ Anime6B, sharper I'd say.
am i the only one who hates multidiffusion upscaler?
you can't use hires fix at 1 and than upscale with extension?
upscaling makes too little changes to the composition and stuff imo
weird. i personally use 4x-ultrasharp
nah, same here
though yeah it might be good
you on my side bro🫂
(got confused with cn tiles), I had a hard time understanding how to use it
I don't even have that one, could you share a link please? I wanna try.
i dont want to read an entire book to understand upscaling, i want a good pic.. thats all
why is everyone recommmending this all of a sudden? is it that great?
I also don't get people using ultimate upscale script
but how much do I have to divide it?
it's better than a finetune for anime made with it
like there is literally no reason
.pth to \models\ESRGAN right?
yes
I will try it now
how do you mean?
also for all other upscalers https://upscale.wiki/wiki/Model_Database
oh my gooberness
without using chat gpt or something
this is basically my LLM prompt, which i gave a couple of examples how to create those storry teeller prompts
Thank you. I'll try it.
So it's the same, img2img and SD upscale scrip but with 4x-Ultrasharp?
not sd upscale script
so the only thing i gave it is your username, and the rest is all LLM going on a hallucination trip
on img2img use tiled diffusion, default settings except change the upscaler to 4x-ultrasharp and the size multiplier to what you want
hmmm so it establishes subject, it establishes location of subject, it details subject, it details location, it explains event, it details event
Ok good, time to try it out. Thanks.
I'm gonna do this with existing picture to compare results.
@brittle jay
literally me
well it got half of it
LOL
kinda reminds me of this character
her ass IS crying, DO NOT listen to her!!!
her ass? you mean her eyes
hehe
sauce?
lol these birds
ok I tested and ultrasharp is fine
it all comes down to preference
it is like around the level of nmkd
my fav upscaler is lollypop
can you make a comparison between it and ultrasharp
no
i just said "no", whats so bad about it man?
its failing to follow my story prompts though....Are you using auto or vlad? are you using vanilla prompt or some extension? are you doing 1000000 images and then cherrypicking?
it was something fast and simple and not much reason for you to refuse, it was just unexpected
it can't even get half right on mine
what's ultrasharp?
I think I have a better upscaler than ultrasharp though
this actually might be better than ultrasharp
here is the reason: i cant because i am doing something else atm
this is the finetune i talked about
nah, i'm just usting the first one that come up
nothing much just splash arties
.
vladmandic, vanilla prompt
is it better than animesharp?
wtf is a vladmandic
I disagree
yes
a fork of a1111
what's the difference?
@wispy canopy vladmandic...is that the popular vlad or another vlad?
its more up-to-date
the opinionated vlad?
Negative prompt: negative_hand, (low quality:1.3), (worst quality:1.3), (lowres:1.3), (error:1.3), (fault:1.3), (undefined:1.3), (crude:1.3), jagged, spiky, ugly hands, extra arm, extra hand, split arm, missing finger, extra finger, three fingers, four fingers, six fingers, (disfigured:1.3), (unclear:1.3), (indistinct:1.3), merged fingers, (bad anatomy,:1.3) misplaced hand, misplaced foot,, cleavage, sex, nsfw, ecchi, slutty, leotard, (text:1.3), (signature:1.3), (title:1.3),
Steps: 75, Sampler: DPM++ 2M Karras, CFG scale: 5, Seed: 700778892, Size: 640x640, Model hash: c20d1e3f21, Model: alfahentai_v11, VAE: kl-f8-anime2, Denoising strength: 0.55, Clip skip: 1, Token merging merge attention: False, Cutoff enabled: True, Cutoff targets: "['One girl has long curly hair', 'the other short straight hair']", Cutoff padding: _</w>, Cutoff weight: 0.5, Cutoff disable_for_neg: True, Cutoff strong: False, Cutoff interpolation: lerp, Hires upscale: 2, Hires steps: 25, Hires upscaler: 4x-UltraSharp, Dynamic thresholding enabled: True, Mimic scale: 2, Threshold percentile: 100, Mimic mode: Half Cosine Up, Mimic scale minimum: 4, CFG mode: Half Cosine Up, CFG scale minimum: 4 Negative Template: negative_hand```
yeah, the popular vlad build
is getting it worth it?
hes slower tho
depends
yes, is faster(on my side), has better support
umh @wispy canopy what's all the rest? I don't understand most of it
oh lol, i even kept the cutoff targets
scratch those
it's an extension?
@cyan burrow on ur model, darkgemini
like cutoff, mimic mode
extensions are entire folders, this is a pth, upscaler
these are upscalers
so what did you use? I dunno what is cut-off and what is something else
so, a script
you use it inside a script
cutoff is not used, there's nothign in the prompt that's it's using actually
ohhhh, it's like R-ESRGAN 2x+
the other is dynamic cfg treshhold, which lowers/raises CFG during generation
An ethereal figure floats within a glass chamber, encased in a state of suspended animation. Crystalline structures surround the container, pulsing with bioluminescent energy. The figure's translucent form glows softly, its contours shifting and morphing in time with the light show. Tendrils of data weave around the container, connecting it to a larger network of systems and machines. A holographic interface hovers nearby, beckoning the user to enter and explore the digital realm. The atmosphere is charged with anticipation, as if something momentous is about to unfold.
Negative prompt: negative_hand, (low quality:1.3), (worst quality:1.3), (lowres:1.3), (error:1.3), (fault:1.3), (undefined:1.3), (crude:1.3), jagged, spiky, ugly hands, extra arm, extra hand, split arm, missing finger, extra finger, three fingers, four fingers, six fingers, (disfigured:1.3), (unclear:1.3), (indistinct:1.3), merged fingers, (bad anatomy,:1.3) misplaced hand, misplaced foot,, cleavage, sex, nsfw, ecchi, slutty, leotard, (text:1.3), (signature:1.3), (title:1.3),
Steps: 75, Sampler: DPM++ 2M Karras, CFG scale: 5, Seed: 700778892, Size: 640x640, Model hash: c20d1e3f21, Model: alfahentai_v11, VAE: kl-f8-anime2
so just that?
yeah, pretty much(if you omit dynamic treshhold)
yeah
might explain the higher accuracy
O.O
There is not enough GPU video memory available!
Every time I use it.
I even dropped WxH to 64x64 it dies anyway.
try tiled vae?
idk never used that extension rly
you're not using my model there though @wispy canopy
I tried to upscale
what dimensions?
yeah but try it with tiled vae
with all respect leave him to me, you don't know all the context
was wanting to mention that, i copied from my image browser, but the later iamges were made on your model
I heard tiled vae helps with vram
Of source image? 768x512
he is already using it
oh
are you sure you have 6GB vram?
guysm where do i place the upscaler file?
ah ok
check at task manager > performance > GPU > dedicated GPU memory
No I didn't turn it on when upscaling. I thought it's only for txt2img. Do I have to turn on both?
oh
you're supposed to yeah
I have RX6600XT with 8 GBs of vRAM
just make sure please
@wispy canopy
That's pretty cool
very nice ❤️
the correct is DEDICATED* gpu memory
send one like mine
ah
@cyan burrow if anything, i feel your model does this kind of shit better than mine sniff 😢
its 8gb
oh yeah i see
7.7/8.0
75 steps?
8GB dedicated and 8GB shared
tf
not mine originally
yeah, that's me XD
did u use --medvram

