Thread 8699499

Ligne Claire Edition

Previous Thread: >>8693972

>LOCAL UI
reForge: https://github.com/Panchovix/stable-diffusion-webui-reForge
Comfy: https://github.com/comfyanonymous/ComfyUI | https://comfyanonymous.github.io/ComfyUI_examples

>RESOURCES
4chanX Catbox/NAI prompt userscript: https://rentry.org/hdgcb
Tags: https://danbooru.donmai.us/wiki_pages/tag_groups | https://danbooru.donmai.us/related_tag | https://tagexplorer.github.io/#/
Inpaint: https://files.catbox.moe/fbzsxb.jpg
Upscalers: https://openmodeldb.info
Booru: https://aibooru.online
ControlNet (OLD): https://rentry.org/dummycontrolnet

>TRAINING
Guide (WIP): https://rentry.org/yahbgtr
Anon's scripts: https://mega.nz/folder/VxYFhAYb#FQZn8iz_SxWV3x1BBaJGbw
Trainers: https://github.com/derrian-distro/LoRA_Easy_Training_Scripts | https://github.com/bmaltais/kohya_ss | https://github.com/Nerogar/OneTrainer

>NOOBAI-XL
https://civitai.com/models/833294/noobai-xl-nai-xl
https://huggingface.co/Laxhar

>NOOBAI SHITMIXES
102d custom: https://civitai.com/models/1201815?modelVersionId=1491533
291h: https://civitai.com/models/1301670/291h

OP Template/Logo: https://rentry.org/hggop/edit | https://files.catbox.moe/om5a99.png

>>8699499 (OP)
updated the op
you can try .co rentry domain since .org appears down

/hgg/enners, I come from /v/'s miquella thread and need only your strongest boxes.

>>8699499 (OP)
>Someone else baked
Nice I guess

>>8699499 (OP)
>291h
kys yourself you cattower easyfluff shitmix retard

tsmt

*crickets*

i'm doing cunny gens so I don't have anything to post here at the moment

avg_loss=0.0496

>>8699961
this desu

>>8700030
:(

>>8699771
zesty

>>8700044
whaddaya want

>>8700030
Extra finger.
Tits too low.
Nice topic though and catbox please.

>>8700046
Don't hurt her....

>>8700050
i'm out of practice gomenasai

i'm lazy

https://files.catbox.moe/5zyypy.png

>>8700083
Thanks, lazy-kun.

what's the magical simple solution for unfucking fishnet textures

>>8700187
pray or add an artist tag that does them really nice

Who are some ai artists that are actually worth checking out?

>>8700230
i only save ai from 4chan threads

>>8700230
me and him, perhaps also him, now that I remember there is also him

horse dildo or plain giant dildo

>>8700087
New pillarino?

what killed the /hgg/

/hdg/ put its foot down finally and dragged us back.
I requested a janny to lock this thread already.

>>8700310
This is a slow thread by nature.

>>8700310
answer the damn question >>8700275

>>8700310
I can't actually post half my shit I think I'm not sure where the line is drawn with hypno / hilichurls desu.

>>8700359
>hilichurls
those are fine, I posted some the other day

Yeah hilichurls is just stealth bbcfags

I ain't a fan of bbc but I cope hilichurls as like orcs / goblins / generic monsters cause they sometimes fuck up gens and every model just knows hilichurls as unique enough. I'd do bokoblins more but they're fucked as well to a 50/50 of working.

>>8700365
>I'd do bokoblins more but they're fucked as well to a 50/50 of working.
In what way? I merged my own model but it's primarily 291 and the only issue I ever have is sometimes I get a boko bigger than trap (formerly otoko_no_ko) Link in a 1 on 1 which some size tag prompting usually fixes. But if it's multiple bokos, it usually gets their size correct.

chat do I want to know what a boko is

>>8700373
I'll try some more it felt like it'd 50/50 get them right or just not work, the size issue wasn't as bothersome to me but yeah it'd make them way larger too.

>>8700374
no pico you don't

>>8700359
>I'm not sure where the line is drawn with hypno / hilichurls desu.
you don't know where the line is drawn on bestiality shit?
just post in /b/ and never come back

>>8700376
Only tags in one of my saved prompts relating to the boko itself would be: the legend of zelda, (bokoblin), interspecies, height difference.
Definitely one of those chars you might want to use the copyright tag if you normally do not.

>>8700381
Think that worked pretty well, thanks bwo

>>8699499 (OP)
what's with the bird

>>8700030
redo pls

>>8700393
havent you heard?

>>8700393
this is her canonical pimp

>>8700429
hey mister, can i have your style mix?

>>8700284
i've been here since october 2022

>>8700395
nyo

>>8700704
>i've been here since october 2022
Not many can be Sir Solange and establish pillardome immediately. Nothing wrong with coming to your own at a slower pace, lazy-kun.

Anybody tried Chroma yet?

>>8700725
it looks better generally but fingers are still bad, it's unusable atm. and you can't prompt artists so it would have always been unusable without a finetune.

>>8700725
>no artist tags
>barely knows any character
>terrible fingers
If I'm going to use a mid SDXL alternative that takes two minutes for a 1024x1024 gen, I'd rather use Neta, at least it doesn't have those issues.

netabros.. they're laughing at us, again.

>>8700726
>>8700731
>artists
I wonder why is that every other big finetune lacks artists. Lode even said they poured 100k H100 hours which is over 2x of noob (50k A100)

>>8700737
worried about liability / community complaints

>>8700737
It's missing danbooru knowledge in general. Guess saying it's trained on anime was just a marketing strategy, because photorealism and furries work well enough on it.

>>8700737
It's probably not in their priorities, it's also a furry model so >>8700744

>>8700748
>https://huggingface.co/lodestones/Chroma
>Training on a 5M dataset, curated from 20M samples including anime
>including anime
Including even 500k anime images would have went a really long way as pony v6 showed, yet seems like it just can't generate anything anime at all? The other anon might be right...

>>8700753
It can generate anime stuff, but it's extremely generic

>>8700757
>but it's extremely generic
SDXL 1.0 also could generate anime images, anon. It was extremely generic as well.

>>8700758
Yes, and extremely generic anime images are still anime images so it's not true that " it just can't generate anything anime at all"

>>8700759
>extremely generic anime images are still anime images
I have no idea what kind of mental gymnastics you have to go through to reach that conclusion. If it's worse at anime than flux-schnell, what's the point?

>>8700766
Can't you just say "I worded that poorly" instead of engaging in this kind of piss poor semantics?

What's causing it this time?

furGAWDS farting all over aniprey once again

the ani-mals in this thread gettin quite uppity.. need bestiality correction :3

At some point you find yourself so far down the social ladder that people don't even want to bully you anymore

>>8700758
>>8700759
guys calm down
point is even 2K image characters do not work, neither do artists
it was probably trained on some generic anime styles and captioned with NL like the rest of the dataset. Or just referred to e621 pics that happen to have an anime aesthetic.

>he doesn't use "very aesthetic" in his prompts
ngmi

Do we really need a separate general for this?

uhm.. we're discussing lora training, chuddie.

https://litter.catbox.moe/dz5thla0yot167rm.toml
Speaking of training, I'm trying to make this config work, but feel like i'm still overshooting a little. Fidelity is pretty good but background and hands are so so. I'm wondering if tweaking the edm2 lr or using dropout would help

>>8700805
what does edm2 do?

>>8700807
adaptive loss weighting

>>8700808
should i learn how to use it or is it snakeoil?

>>8700810
I personally think it gives better results than other loss weighting implementations. But some people don't use any loss weighting at all, stick to plain adam and just endlessly tweak their dataset instead

>>8700807
if the loss is too high, it brings it down. if the loss is too low, it brings it up. extremely useful for pretraining, not necessarily so for loras.

https://civitai.com/models/1891663/fishnet-size-slider?modelVersionId=2141170
>amazing lora idea
>read notes
>first time lora baker
grim, exceedingly so

anyone here tried pollo ai?
is it actually good for video nsfw or is it trash?
i cant do local hence why.

>>8700852
>i cant do local
>>/hdg/

>>8700805
>edm2
shiggydiggy

>>8700805
What a weird config desu.

>>8700859
it's normal, average even

>now i want to fuck about with my config
ohh frick u

where did all the declines go?

Still doing cunny gens, sorry
i'll probably do that elf x human gen later tho

would normies pay patreon bucks for this

>>8700882
well
for 500 of those every two hours, if you're a 1% lucky guy

>>8700882
It looks good but it also looks like a lot of other AI generated stuff. I'm sure someone would pay for it in a vacuum but in a crowded market you've got to stand out.

>>8700888
>>8700889
what could I improve... being a 3rd world motherfucker even dolaroos extra a month for doing this stuff would go a long way

>>8700890
even 100 dolaroos extra*...

>>8700890
make him blacker
but yeah most of the sloppers upload literal thousands per day so u need a consistent no human pipeline if you want to do it

undersea internet cables were a mistake

>>8700890
you need atleast 1000 dollaroos extra for that 5090

>>8700890
Composition-wise this image is fine, even good. But in business you need an "edge", something that you do better than anyone else. Being able to consistently make better than average AI slop is nice but might not get you anywhere because a lot of people can inpaint away the flaws in their gens. You can try catering to a specific fetish if you think it's underserved in it's niche, but it might also be good to use a more unique style. Not sure if you're using any artist tags/LORAs but this looks very "generic Illustrious AI image" to me.

I've been out for a while. What's the best method for generating animations now?

>>8700904
wan 2.2

>>8700904
ungenuine

>>8700902
zhen_lu:1.3,karochii,riz:0.8 was a mix that I enjoyed with WAI but I think this model that I'm trying out (jankuv5nsfw) has a generic style cooked in

>>8700900
I guess that would make gens way faster

>>8700894
true the bbc pics I posted on pixiv are the most popular so far

>>8700882
you'd probably top out at about 50 subs even if you're very dedicated but yeah sure

>>8700914
hey 200 dolaroos a month is like a 30% increase to my current wage

>>8700187
What's the tag for that position? I was looking this up earlier and idk if it works. I know there's flying and a midair tag but idk.

>>8700230
Why are you asking for recommendations instead of asking for help to improve?

>>8700359
As you can see by those two anons, the rules don't matter. What matters is whether or not someone is angry at you and reports you.

>>8700919
well and the jannie bothering
you have those shitty ayacoon threads not even getting taken down nor persistent and constant /aco/ and loli

>>8700921
check this one out >>8683944
the rules really don't matter one bit

>>8700922
All the responses are positive. Unless someone reports it it's just gonna fly under the radar.

If only there was a very convenient blue looking alternative to this dogshit website

>>8700925
we want more posters to be banned, not less

>>8700882
You're better off finding niche gacha sluts to make pics of, but if it's niche it means small following so it's hard. Still the whole point of AI is about making stuff for cheap that artists would charge over $100 for so you could make a name for yourself doing high quality art of characters no one likes. But in the end you'll burn out like everyone else that's tried. This stuff should be for fun, not money.

>>8700911
Well if you're getting some traction on pixiv you can try to monetize it. The worst thing that can happen is that no one buys anything and you're back to where you are now.

grifter talk goes to /hdg/ btw

>>8700863
>did
>previous findings still hold true
Saved

>>8700928
I'm definitely having fun doing it, every time I open comfy UI I find myself staying in there for hours and hours testing stuff to the point of losing sleep over it.
Just looking into ways to automate the process to make tons of gens with different tags has me itching to try but I've also seen that there is potential to at least make a source of extra income if I get good enough it at it, so why not try.

>>8700929
I just started a twitter account and a pixiv and the twitter is dead as shit gets like 20 views a post if that but pixiv gets hundreds and is already getting me followers, it seems like it is definitely the better platform for discovery.

>>8700944
blud thinks he can test anything relevant in a day baka my head

>>8700949
it's a lora not a tune my dear lil zoomie

bf16 or f16 safe precision?

>>8700947
more like brazilian grift, and how about you post some gens if you're so good?

>>8700950
the eternal cope of second rates

>post some gens, if you're so great
https://files.catbox.moe/vtrj4u.jpg

>>8700951
training? bf16 is stabler
>>8700953
huh? i'm sure that sounded good in your eslie head

Where the fuck is Ayakon?

who do you think is making the slide threads?

>>8700955
>stabler
Lora training yeah. What does this mean?

>>8700958
it gens horses :3

>>8700952
Oh nyo nyo nyo you've awakened the amputee shitposter

more of her? >>>/b/938845397

it's the schizo schizo

>>8700962
The guy that made that rentry? I'm still convinced he was trolling

>>8700966
stfu pissbuttfag

>>8700966
what rentry?

>>8700961
>you've awakened the amputee shitposter
how and why.......

I'm afraid

>>8700969
check 10 threads ago or so. Some guy made a lora baking rentry and came to riveting conclusions such as adamw being better than adamw8bit (shocker)

>>8700917
no clue, it did that randomly. you could try sex from behind, arm grab, size different?

>he's still seething about the entry
jejerino

>>8700973
I'm surprised the one anon attempting to help after so long made you this mad. If it really was so insignificant you would have forgotten about it already.

no dudebro he was trolling ughh
he trolled so hard i still remember everything about it a month after ughhh

>>8700978
hard to forget when he pops up every thread and tells people to train in full bf16

it's the rentranny

>>8700975
It's "size difference". Anyway best one to use is "lifting person" but it's meant for loli and minigirl. That pic is weird, she's tall enough to reach the floor yet still floating somehow.

>it was almost three months ago
AIEEEE IT'S IN MY WAAAALLS

tfw

all me btw

>>8700979
Aren't trolls usually the most memorable posters?

>>8700980
I'm the one who asked that question and I don't even know what you're talking about. So f16 I guess? What's the difference?

>>8700986
thiss
i mean he brought a general to its knees with a single rentry

>>8700984
teach me thy secrets

>>8700989
absolutely not go fuck yourself

>>8700987
>So f16
mixed nigga. why can't you subhumans bother gpt with your inane questions? isn't that what it's for?

This general is only for Patreon grifting, fucking trainschizo.

>still got no serious feedback on my config
why even come to this place
>>8700996
this

>>8700990
at least post a gen

>>8701000
I would've considered it if you weren't flooding low effort garbage

>>8700997
I was serious though, EDM2 is fucking crap and all the random cropping and padding and weird ass optimizer/schedulers are just overcomplicated for not much reason. Why would you flip aug with 1 batch even?
32/64 is oversized for SDXL, imho, lower TE LR didn't really do good things on my end. You have insanely low max grad norm and then dump warmup on it. Plus training >1024 is a total meme for SDXL.

>>8700993
>why should would anyone bother talking to humans???
>They should only talk to machines like me!
I'm so sick of you dumb bots shitting up this place.

Why don't people post chroma gens?

>>8701002
show what high effort looks like then

>>8701005
Because no one is using that?

>>8700993
>isn't that what it's for?
what is this general for?

>>8701007
But it's the best anime model?

>>8701010
102d

>>8701010
Aside of your clearly attempt at shitpost, who said that

>>8701013
It was just released and it was of a better architecture.

>>8701003
Most of that config is from the ezscripts fork guy and it simply gives me better results than plain adam.

>no one is using that
keep telling that to yourself schizo

>>8701016
why are you even asking here about it if you're just gonna say "well it works for me"
well then use it and don't bother me baka my head

>>8701005
It's not fine-tuned for anime yet and probably never will be

>>8701019
If you have a better config feel free to post it, but "your config sucks and is too complicated" isn't exactly helpful criticism

LAX is currently helping train this model https://huggingface.co/NewBie-AI/Newbie-Ai-Training

>>8701022
>DiT
it's dead in the water before it even began
sounds like something LAX would be retarded enough to bother with, though.

>>8701022
what is it

>>8701022
>>8701024

>>8701025
some custom lumina based arch

what about our favorite baker linaqruf and the exciting new animagine?

>>8701027
>lumina
lmao

!retard

>>8701022
>3B
it was over before it even began

>>8701047
but sdxl is 3b-oooooh

nai is 2b

who is nai

>>8700768
true

schizo thread where is the gens

>>8701054
me

>>8701054
This anon >>8701060

>>8700954
you know the rules

I hate females

>>8701068
you should love them

tsmt

>>8701063
Stream that movie if you're so great

>>8701079
wrong thread dumbass

WOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOW!!!!!!!!!!!1111 2

>>8701079
What's the prompt for this?

>>8701063
and so do i

WHat's the best wan model I can run on a 3090?

>>8701099
don't bother

Some of you guys complain a lot but don't even post anything

>>8701104
i do both

>>8701104
Don't you have a well to dig or something?

>>8701104
you seem to be in the wrong thread
maybe trash is more your speed

>>8701104
>you guys
go back

>>8701111
go back where? I've been in this godforsaken website for 16 years

>>8701111
wasted quads.
>>8701104
It's because you're posting browns. Here on /h/ the color of the dick matters more than the girl. They'd even fap to solo male if his dick is white. Who is this btw?

>>8701116
>It's because you're posting browns
It's because he's a third world grifter

>>8701113
yeah lmaooo nah

>2009 newfags itt
KWAB

>>8701116
I see... next time I'll be sure to neg prompt dark-skinned male so that people won't complain as much, girl is hysilens_(honkai:_star_rail) .

>>8701118
why would I larp as a 30+ loser brazilian making shitty AI gooner smut on a bangladeshi Kite testing forum

butiful vpred fry

the real money is in hags

>>8701132
if it's not grey it's fried so true sister

>>8701129
zootcoded

>>8701134
awwww lil niggette is on a 45% ntsc laptop monitor :3

this shitposter has been awakened too..

that's just a repost lil seething bottie kun

oh so these were /g/ shitposts the entire time...

AHHHH right fucking Nadeko exists, oh hell yeah more ideas to the backlog

>>8701144
well yeah he literally said he was from g and trying "other boards now"
now of course now he is seething because his sliding didn't work and spamming but that's another thing

Seethers and complainers I'll see yall tomorow

>>8701142
Utena is wife

can someone tell my monkey brain 1girl giant dildo is not as hot as it seems to be

>>8701158
Aww I was just about to post now that my lora is done. Goodbye anon, hopefully we get more real posters and less bots.

waoooooooooo the progress.... I can feel it....

>>8701162
I'll be lookin at the thread here and there but I've already done enough genning for today,maybe I'll try to figure out automation stuff later...

wait who is amputeeschizo

>he answered the call
Go back.

FUCKING KYS

K Y S
Y
S

listen i just want to know if i finally have a name and if it's for my love for katawas

to me you will always be lora schizo

>I want to be a namefag
Go. Back.

>>8701186
if i'm training a lora on a character that's only available in one art style or something how would i get rid of the style bias?

>>8701187
who isn't a namefag at this point doe?

>>8701189
copier method

>>8701189
I don't train characters but I would img2img a lot of the images with other styles

>>8701189
bake it and then gen with style loras for more images

>>8701187
>>8701190
can be finally use tripcodes so we can have proper communication?

>>8701189
1. Find other characters of the same style
2. Tag the style
If you can't do 1, it means the character IS the style.

>>8701190
Me
>>8701195
this

>>8701197
you are poopdickschizoust

>>8701189
just train a style lora of the art style and then merge it with the model or use --base_weights to do it during runtime. it's that easy tbqh

>>8701198
h-how did you know...

>>8701199
I've heard that said for like two and a half years and I don't think I've seen more than one irl example of it tbhdesu

>>8701202
It's bad advice schizo

>>8701199
like train two separate loras, one for style. merge with checkpoint and then do a lora extract or something?

/hgdg/ discord when?

uh oh nogen schizo is here

>he is in my patreon server already
>fucker needs more
jesus

>>8701219
teamwork kino

ha ha time to attempt to do a style lora for the first time in like two months
can't wait for the first one to be mediocre shit with a lot of weird shit because I'm completely guestimating the LR and I didn't even bother to curate tags at all teehee~

>>8701219
hiiii

>>8701219
I look like this

>>8701222
>bother to curate tags
gotta take the danbooru tags with automatic cleaning scripts pill

>>8701202
https://litter.catbox.moe/69b504hu5jx6yk9i.png
https://litter.catbox.moe/cu5utbck4anodpa1.png
https://litter.catbox.moe/t185wn57ahjafldu.png
https://litter.catbox.moe/1phh9lupifn34v4n.png
it just works
>>8701207
you just train the character lora on the merge. that's what the copier method really is, all the stuff about matching poses is superfluous. the style doesn't need to be all that accurate as far as i can tell, i typically shoot for around dim 8. you do probably want to train the TE though for whatever reason

>>8701221
yeah was looking for a pic for the lora
>>8701223
Nice. Mine is still coming out too smooth for my liking. Are you pixelating with prompt or in post? Oh I need to try that pixel extension.

>>8701228
Prompt, adding pixel art in the baking tags helps a lot.
But I also only kept the images that I could properly upscale with only nearest neighbor.

>>8701229
>adding pixel art in the baking tags helps a lot
Hmm I figured if I didn't add that it would help since it gets baked in? Idk man I'm tired of fighting with those one. Hmm I upscaled everything too. Maybe that's why. Idk how NAI did it.

>>8701232
>Idk how NAI did it.
16ch vae

>>8701232
It can definitely help. I also have all my oekaki artists with "oekaki, jaggy lines", just helps it not melt the lines away.
Oh yeah worth mentioning to disable bucket upscale if you have it on. Forcing images into artificial buckets with lanczos ain't good especially for pixel art.
>Idk how NAI did it.
it's called 16ch vae kek

>>8701219
almost good

>>8701238
You're right I fixed it.

>>8701239
Isn't she your avatar?

>>8701227
>you just train the character lora on the merge.
i see, i'll give that a try.
>that's what the copier method really is,
why is it called the copier method? i though that guy was shitposting

>>8701239
her arms are still fucked broski

>>8701242
I prompted for it

i'm about to ask gpt for a custom image viewer so i can more quickly sort artist images for potential baking

>>8701241
the name for it that took off was coined by a japanese guy. idk why he called it that and i could have sworn that the pr that introduced --base_weights called it difference training which makes a lot more sense

>>8701142
Still waiting for an uncensored version.

>>8701142
I pretty much look like this

oh damn i didnt realize a few image viewers have script support
nice

>set up qimgv with the script
>works
>but the viewer has some retarded quirk when clicking up first resizes the image despite it being bound to the script only
WHY CAN'T IT EVER BE EASY AND WITHOUT TINKERTROONING ALL DAY?!?!!?

because you waste too much time making loras you'll never use

>>8701265
doesn't mean it can't be not annoying
like i can bind it to anything else but fuck me it never is just like you want it because fuck you

>TINKERTROONING
why is anti-intellectualism so rampant among younger generations?

>HEADCANON
strawman

>>8701227
so you would just go
>train character lora
>merge lora with model
>retrain on merged model
right? i never tried it

>>8701239
box please

>>8701270
Why can't you explain how he should use it?
>>8701272
Nice toes. I'm not a footfag but I love painted nails.

>>8701272
penises should be tiny and crushed by feets

>>8701227
How do I enable --base weights on ez scripts? Or are you using bmaltis?

>>8701275
Anon uploads are temporarily paused! Please wait...
https://litter.catbox.moe/omztc9y8khmp45v8.png
Litterbox works doe

>>8701274
you train the style lora, merge it, then train the character lora on that merge. or just use base_weights.
>>8701279
does ez-scripts not have a way to pass args? it's just --base_weights="E:/path/to/lora.safetensors" and then --base_weights_multiplier number to control the weight, which is where the real schizophrenia begins

>>8701283
>and then --base_weights_multiplier number to control the weight, which is where the real schizophrenia begins
Ok so it clearly isn't as easy as you're making it sound.

>>8701285
tweaking the style lora's weight or resizing it or changing its training parameters can consistently make a difference for how well the small details of the character lora turn out in shitmixes. i think that there's a weird goldilocks zone where you want the style to be powerful enough to curtail burn in but also weak enough so that the weights don't get too fucked up. and after around two years of messing around with this i can pretty confidently say that i don't know where it is.
i've just been running the styles through sv_fro resizing at 0.955 and using a weight of 1

>>8701116
well...

>>8701281
thank you

>>8701272
Asymmetrical number of pink ankle pads...
Anyways, can't believe my eyes, somebody ITT can still gen.

>>8701329
ah fuck I'm blind

>>8701329
>Anyways, can't believe my eyes, somebody ITT can still gen.
what does this mean

>>8701333
Means I saw something very rare.

>>8701335
You could just said you are a footfag and end it there, you know

>>8701336
No need to be so jealous bro, it's just one (you) and you are making a scene...

>>8701338
I always get triggered when someone doubts gens capabilities

NewBie AI is going to revolutionize local hentai

>>8701239
Can you reup your catbox please?

>>8701351
https://litter.catbox.moe/vl7ockp98vqshmyb.png

>>8701351
It's ", nigger slop".

>>8701356
Essentially yeah :^)

>>8701356
retard
it's , nigger girl overdose

I wonder what would be the top artists if you tabulated booru upvotes per image, excluding <15 pic artsts

>>8701360
i mean combining and then dividing by image count for a per capita score

>creatives unironically doing this to their images to destroy evil ai
https://danbooru.donmai.us/posts/7745909

>>8701372
that's nothing, the chinese go way harder
https://danbooru.donmai.us/posts/9749094

>>8701372
Ohhh... thanks for the recommendation. Imo watermarks would be our bane but I've only ever seen some of them do that. His pic is big enough so I don't see what the problem is.
>>8701373
As you can see, chinks will win in the end.

>>8701225
*booru tags are shit 4 out of 5 times
also they don't help when the artist doesn't exist on *booru to begin with

>>8701360
Probably nyantcha or some prolific western gooner artist.

>>8701374
the funny thing is that flux kontext is almost but not quite able to clear this stuff up, maybe a lora could do it or maybe qwen image can do it, haven't tested

>>8701377
>also they don't help when the artist doesn't exist on *booru to begin with
Stop jerking off to aco mayhaps?

>>8701356
>>8701355
is this lora still up anywhere? im having a hard time looking for it

Which one looks juicier and more attractive, Anon? Can't decide on one specific certain option by myself...

bait :3

>>8701383
https://x.com/Rotoi_1234/media
yes, oh so very aco
(*booru has 5 images from them but their twitter/pixiv has more)

>>8701385
Probably better off using the other anon's.
https://pixeldrain com/u/noG78sYu

>>8701391
Pick the one with the best background.

>white pupils
>oekaki
Of course... it's him!

>>8701405
Is that because the other anon isn't a nigger?

>ytboi thinks this isn't a black only general
lmao

>>8701222
actually works pretty okay for a v1. though there's a lot of stuff happening that's pretty obviously due to tags and I also forgot to manually correct a few things in a few images.
also probably need to increase the LR a small amount.

>>8701411
I could tell by how shitty it became.

>>8701413
Oh dat's the goy. Guess I should go make a dataset.

>tfw I have two different eencya loras with different hash and can't tell wich one is the good one
perhaps I should test this my way

>>8701408
To be fair - background is the least thing that I'll worry now. Want to move from WAI slop merge to something more creative and now checking and comparing new merges that were released not so much time ago and...turns out there are several really good, but select only one is very difficult mostly due the colors and shades difference.
>>8701392
Nah, you're wrong. Just want to take an account others opinions for subject.

mine is the good one
no i will not share it >:)

I already have it
and your bitcoins

>>8701421
>punctuation
baiter :33333

>>8701421
Ask /aco/.

>>8701424
I'm sorry if my manner of writing in a language that is foreign to me offended you so much~

>>8701431
nyyaaa bait-chan... n-not so hard!!

oh nyo nyo nyo nyo~ You'll get it nyow~

1anon, baiting, otoko no ko, crossdressing

>>8701355
>404
nyooooo

>>8701419
They are almost the same, they feel the same what the fuck are you n-words doing, you fight everyday yet you produce the same results

>he is unable to differentiate between melted, fucked up details and coherent details
uh oh

NAI is getting some tech for character referente using a few pics. Meanwhile, we get uhh... Ummm... Can someone point me out what are we getting?

>>8701443
NewBie-AI DIT

>>8701443
Is NAI getting better ways to upscale things or different (cheaper) costs for doing gens at native higher resolutions?

>>8701445
Nope :3 you get 1024x1024 & 1216x832 and you'll like it

ohh dis anon is back
hello

>>8701442
Alright, adding a proper background seems to really show the differences, the old one is winning

>>8701445
NTA but I have a NAI sub, it expires tomorrow and I've only wasted 500 anlas and still have 9.5k left because I don't really need to upscale the pics often. Eyes looks perfect on base res and hands usually are fine 70% of the time. Unlike current local. I guess not increasing base resolution but improving and finetuning a very good VAE was the answer to get good small details.

>>8701448
omg hiiiiii~

>>8701421
My favorite was plantmilk oat, what I'd go back too if I wasn't vpred snake oil'd at this point.

>>8701466
fuck I thought they were going to kiss

>>8701477
you mean like this one?

>>8701478
BASED, yeah just like that

>>8701437
>otoko no ko
Too soon, anon.

I bet bro is posting censored shit to red board because it's a PR campaign for his future grift. Uncensored stuff will go there once bro accumulates demand.

lil bro got em scouted no cap

>>8701482
Why is everything with AI always about making money? Why can't people just do something for the love of the craft? Why haven't we banned this faggot for off-topic?

>>8701482
anon he's just a schizo who NEEDS to SHARE (whether anyone wants it or not (and no one fucking wants it))
they're actively ban evading

>>8701481
couldn't they have gone with femboy at least? retarded admins

>>8701490
If he just needs to share, he can share without rectangles which he obviously added in post for his /a/ butt buddies. This feels like grift.

>>8701494
his workflow only works with censor bars
please understand. he's too retarded for anything else.

Yeah my final verdict is that the one I already had is the good one, thanks for following my blog.Next up is the blacked set for the last week poll winner, remember to check your emails

>>8701500
The one you already had probably isn't the best one either. If it's the one from anon who's periodically dumping his loras in bulk, he made a new version (picrel) afterwards.

>>8701504
Well, hopefully he would share it again (or not), it's not like I would use this style but I just wanted to compare the thread's loras on the same artist

>>8701504
Does he (You) prompt pixel art too? Pixel art always fucks up the pics and has always looked like shit on pony and on noob. The epoch that was able to get the pixels on the hair was fucking up the eyes so I trashed that one.

what is this censorfaggot game plan?

>>8701530
Bump limit, then slide the thread. (You) attract the worst schizos.

nai simply lost because it can't do video
let that sink in

wonder why so little people use this lora
https://huggingface.co/nblight/noob-ft/blob/main/noob-ft-1536x-extract.safetensors
it's pretty great, it lets you do stuff like this in one pass of the model consistently, especially combined with kohya deep shrink

>>8701540
How do I use it?

>>8701544
just load like a regular lora and assume your base resolution is 1536x now
use deepshrink if you want to go higher

>>8701550
And it doesn't affect your style and shit? Guess it's worth a try.

>>8701555
Not as much as you'd think, and also little details are somewhat lacking. But the anatomy improves considerably, this is what i get without the lora+deepshrink at this resolution

>>8701558
Yeah, but for me, simply deep shrink alone does the trick. At least at 1920x1088 resolution I like with 1.5 deep shrink scale factor. Do you really need the lora too?

>>8701558
>>8701559
>deep shrink
That's the kohya hires fix in reforge? What are your settings?

>>8701540
we already discussed this, it's great (considered the gen times) but also outputs very noisy gens that you will likely end up inpainting anyway

>>8701558
or
and try to follow me here
or
you can just gen multiple lower res images in a shorter amount of time and then just like
I don't know
manually upscale ???
you know, just doing things as if you're a sane, rational person instead of a retard.
you get absolutely no real advantage generating at a higher base resolution when you're still gacha'ing your shit.

>>8701569
His entire point and purpose for posting is that he isn't gachaing due to the lora. Retard.

>>8701559
well it's much better if you use both, deepshrink alone is somewhat lacking
>>8701563
I use it in comfy like this
>>8701569
just use sd 1.5 and enjoy your hiresfix retard
>you're still gacha'ing your shit.
and learn to read

>>8701573
he's not i2i'ing. It's still just t2i.
He's just trying to shortcut hiresfix. He's still going into shit blind.
>>8701574
>just use sd 1.5 and enjoy your hiresfix retard
That is literally what you are doing jesus christ you fucking retard.
>some dumb cumragguzzler
wonders never cease

>>8701577
what the fuck are you even talking about fucking schizo?

>>8701581
let me put this into very plain terms
what you're doing? You're trying to reinvent hiresfix.
hiresfix was already shit. This is also shit, for multiple reasons.
If your goal is just passible "look ma I'm gooooooooooning!" single shot stream of conscious garbage that you'll never look at again? I mean, I guess it's fine? It's pretty bad if you're actually trying to do anything with a marked level of quality though.
You may as well include one of those garbage turbo-mode step reduction loras for good measure if you just want to goooooooon to fast, endless trash.

why are both of you acting like hires fix is bad when it's not

that's right sissy we all enjoy 832x1216 melties, keep it up. who the fuck needs clear details?

>>8701596
So you're saying deepshrink is just hires fix on the backend but with extra steps on the frontend?

>>8701601
deepshrink was developed at an attempt of an upscaler wasn't it? If you're running it t2i with no curation then yeah it's being used in pretty much exactly the same way you'd use hiresfix. And in this case the "extra steps" is using a shitty extract lora to try and push base res of a model higher to help correct anatomical errors(it still has anatomical errors and will fuck up perspective).
in some ways it might technically be better than hiresfix but the only excuse to use hiresfix in the first place was if you're a lazy retard who needs SUPER HIRES PR0N and need it NOW, quality be damned.

Or, if you just want to know where the "difference in viewpoint" of all of this is, think of it as the difference between someone generating for themselves to get their rocks off without caring too much about what they're looking at vs someone making shit to share and wanting to actually have a level of control and quality of what they're making.

>>8701596
>shit
>trash
explain what's wrong with it you shitting ape
inb4: without babbling and sounding mad
>It's pretty bad if you're actually trying to do anything with a marked level of quality though.
i stopped posting anything i put my effort in long ago because losers like you will spoil all the fun, these are just random prompts grabbed off danbooru just to post itt

>>8701605
>and wanting to actually have a level of control
This is really the only issue for me. I don't see how doing an img2img base res to X res controlnet is that much better than hires fix. I think you're exaggerating how bad hires fix is and creating unnecessary elitism where there isn't one. Now as a second img2img pass? You shouldn't be doing "hires fix" but using CN. First pass? Meh it's not a big deal quality wise.

>>8701605
>deepshrink was developed at an attempt of an upscaler wasn't it?
no it was not you fucking imbecile, it was discovered accidentally by kohya that you can resize the outputs of a some certain unet blocks and the model with think it's generating at lower resolution than it actually does. the output of that is very blurry though so it's only usable if you apply it at early steps
>(it still has anatomical errors and will fuck up perspective)
just like literally any model
at this point i must conclude that you never actually attempted to use this lora, so kindly shut the fuck up

>>8701611
>of some certain unet blocks and the model will think it's
idk what happened there

>>8701611
please explain to the class the actual application and standard use of deepshrink and the purpose it served as it was introduced

>>8701618
so you never used deepshrink either and the only thing that caught your attention is that it's called "kohya's hiresfix" or whatever in the shitty gui of choice. i see, i see.
>actual application and standard use of deepshrink
t2i at higher resolution than base model was trained on

Why aren't people posting chroma gens?

>>8701638
h

>>8701638
>>8701005
>>8701007

>>8701622
>at higher resolution than base model was trained on
sure sounds like some "ummmm AKKKKSHUALLY" dumbass bullshit where you're intentionally trying to split microscopic hairs that don't really exist.

>>8701487
nobody here is making money off of a.i gens bwove.
it's all in ya 'ead.

>>8701540
>kohya anything
didn't read

>>8701652
basedbasedbased

>>8701599
bwo thinks he's generating mona lisa(s) ova 'ere.

>style mix looks way better (colours and fucking details) at lower CFG
the fuck, at this point I think I will run an x/y for each fucking mix for their optimal and particular CFG

>>8701682
Samplers make even bigger difference to me, so might as well run comparison in 2 dimensions.
And that's not even accounting for shit like genning base res with one sampler/cfg, and upscaling with another.

>>8701687
Thankfully on my case I have already consolidate a general sampler and scheduler (which I change on some very very specific artist mixes) but this whole CFG situation really made me concern, how is even possible that at lower CFG I am getting more details?
Did I get everything wrong? not I am not using any CFG++ sampler or snake oil

>>8701690
do you mind sharing what samplers you liked?

>>8701712
ER SDE (50 steps) and DPM++ 2S a (30 steps)

>>8701540
is you new

>>8701690
*What* lower CFG? Hard to say if you're defrying from 7 or what

Does anyone have a complete guide to generate nsfw anime girls? I keep seing different ways of upscaling and I feel like my pics never look as good as the one I find online

>>8701739
It's very easy
>head to https://novelai.net
>click on the login button
>create a new account
>subscribe to our Opus subscription plan for unlimited gens
>just write what you want to see
It's just that easy

>>8701739
It's very easy
>head to https://novelai.net
>click on the login button
>create a new account
>subscribe to their Opus subscription plan for unlimited gens
>just write what you want to see
It's just that easy

kurumuz isn't sending his best...

i mean 25 a month isnt that much

>>8701272
catbox?

>>8701741
I'm talking about local gen though...

>>8701713
lil nigga is me. SGM and Beta?

>beta
shiggydiggy genuinely

>>8701391
The white outline makes her look out of place, even if the background looks more realistic. Retry the XY with "no lineart".

Also, her outfit is not exactly pool-practical.

Why is it talking to itself?

Is she ok?

>>8701739
>I keep seing different ways of upscaling and I feel like my pics never look as good as the one I find online
You should be comparing the upscaled pic to the original size if you want to blame upscaling. When compared to other finished gens it's more likely your prompt and checkpoint, CFG, etc.

How many times a day do you cum?

>>8701782
way less than i used to :(

>>8701782
Once, twice very rarely. Some days, I just don't fap.

>he cums
ngmi

>>8701787
based and dao-pilled

>>8701787
based and caged yboi pilled

>>8701505
>censored ecchi on /h/

what does yboi mean? ya boi? yaoi boy? yikes boy?

I LOVE AUTOMATIC MY BAKING WORKLOADS WITH CHATGPT SCRIPTS!!!

is this the real life

>>8701756
anon's eencya lora from several threads ago + gogalking

>>8701824
aww man I missed that
we could've talked about the futility of baking loras for styles that already work

>>8701824
just post the box faggot

no styles *really* work on v-fried

>>8701827
hmmmmmmmm nyo

>>8701827
seethe

>>8701828
nothing works on base v-fried, that's why nobody but oekakibros use it

>>8701832
it's even worse on shitmixes because even if the eps didn't fuck up the artist themselves it also fries the vpred
people who unironically use inbuilt artists are fucking retards
t. tested like 70k inbuilt artists before

>>8701834
That's why you use loras and seethe about shit hands on base and shine on mixes

>>8701836
well yeah that's what i do

just bake a lora.

it is still funny that once we got a model that can actually reproduce styles via loras people stopped baking them and started using v-melted inbuilt artists

>>8701842
>that can actually reproduce styles via loras
16ch vae please...

>>8701843
for me it's >1MP gens
you can't reproduce lines people drawn at 7MP no matter how many channels you put in

All these reasons AND MORE is why I just pay 25 a month.It's THAT simple.

why do lora bakers have that school shooter vibe

>>8701846
Nai can't do any artists proper thoever

>>8701850
I just baked a vibe transfer thoughbeit

>>8701850
I'm glad they put out that trial because it made me realize how fucking shit at artists it really is. V3 was unironically better going by other people's gens.

>zero shot?
more like zero shot at ever getting pussy after using this garbage lmaooo

>>8701852
It's pretty good at limited amount of artists/styles ("painterly", sketchy, oekaki etc) but anything with smooth shading and it's pretty bad.

who is she talking to

shes talking to your chances of getting pussy

#pussyfree

what caused the decline?

>>8701887
novelai

>>8701888
this

>>8701832
>oekakibros
oekakischizo's other nickname is EPS 0.5 schizo
he also bakes loras on 0.5

>is called novelai
>is only used for image gen
why did they do it?

>>8697602
Decided to get off my ass and test this out.
What I have learned is that I need to have sageattention disabled or I can't load Q8. After disabling sageattention I can finish the first pass but ComfyUI crashes on starting the second pass, I still haven't fixed this yet.

>>8701892
Their 70B LLM was shit.

>bakers meltdown
What caused it (this:1.8) time?
>>8701892
>>8701894
You WILL coom within 8k context and you WILL like it.

>>8701894
why haven't they finetuned deepseek or something yet?

>>8701896
instant bankruptcy

>>8701896
Out of their capability
DS being a "small" player still has several thousand cards at their disposal

rurormaomao!!!!!!!!!!!

quant firm vs poor turk
yeah the poor turk ain't finetuning a giant moe for sure

>>8701897
>>8701898
damn you would think with the hundreds of bbc slop patreons with opus subs they could afford it

>>8701910
>they could afford it
LLM makers are burning through money faster than they earn it and that's despite
Claude Max - $200 subscription
ChatGPT pro - $200 subscription
Google AI Ultra - $249 subscription
NovelAI's opus tier is poverty tier
DeepSeek doesn't care to turn this into a profitable business and is funded by a quant firm
You don't train LLMs without real deep pockets and a willingness to see most of the sweet dollar go up in flame
once the bubble bursts, the losers in this competition will die with their knees deep in debt

cny

yay tanoshii!

Chinese New Year? isn't that in February or something

every day is a chinese new year for me if you know what i'm saying

>>8701918
my gf

mods we have a problem here

>>8701905
>>8701906
>>8701908
I nooticed

>>8701926
To ban pedos is up to the mods, but to send them to him, is up to us.

Hey all! Whenever I boot up reforge, I get picrel, which is weird since it seems to always install requirements? I would not mind it but it does not start without internet, and that is pretty annoying. Any ideas?

>>8701283
ah thank you i switched to rawdogging sd-scripts some time ago so i will try this

my use case is i have an artist that has a doujin series with some cyborgy features that i want to make a lora for without including the art style

hear me out: decreasing batch size during training

>>8701934
switching to comfyui solves the issue of using jank software

why are you scared of drawings?

what are some good settings for whatever the smallest lycoris network is?

>>8701735
I was using 5-6 until not much ago

>>8701763
Mostly Beta but with some special mixes (pic related), Karras Dynamic and AYS

>>8701949
>smallest
Go big or go home. You have the disc space, don't you?

>>8701953
i'm interested in the weights not the disk space

>>8701622
>>8701646
So what was the conclusion of this debate? Isn't the purpose of a debate to learn? You're claiming that he's splitting hairs but you never actually said what was wrong.
>what is deepshrink/Kohya hires fix and how should we use it?
You don't have to be civil, but you should at least be able to answer this question productively so we can all learn.

What should I buy for 25 doollers?

>>8701964
novelai opus tier

>>8701739
Rather than ask this question, post a /h/ gen and we will workshop it with you. It better be good or we will dogpile.

in cumfy what's the best way to save an image with the prompt in the filename?
right now I'm using a node to combine my prompt with the folder prefix like
"folder/[image]" and it works until I gen something using an artist tag that needs escape /. If I do it thinks I want to save it in ANOTHER folder
so it becomes "folder/whatever/artist tag" and then it breaks because it can't find the path

>>8701826
It doesn't work, that's why you bake it with keep tokens and the artist name as an activation tag. Then it works better than either alone.

>>8701972
I usually default to new unique token. Don't really want to bother with laxs fried vectors

>>8701973
I never saw much benefit in activation tags unless I was overwriting the existing tag.

>>8701969
>until I gen something using an artist tag that needs escape /.
I use the regex replacer node from the D2 suite (https://github.com/da2el-ai/D2-nodes-ComfyUI) to automatically filter out characters that aren't valid in filenames and the String Function node from comfyui-custom-scripts to separate the prompt into multiple parts, with the artist mix part being sent to regex -> filename

>>8701891
based loremaster flexing on newfaggots

I was going to post some gens in the /d/ thread but maybe not

>>8701985
Rather keep them to yourself? Why not?

>>8701989
That's what I have doing this whole time yes, someday I will hopefully find a proper bestiality thread

>>8701977
I'll take a look at those, thanks

>>8701892
>running shitty unlimited 20b at 16bit
v
>shitty 0.9b imagegen with limits
cash

>>8701992
oh
yeah I guess only /trash/ accepts those
/d/ if it's a half beast half person

>>8701985
>>8701992
Post 'em here, but catbox. You'd think pokephilia would be more common but it isn't.

>>8701977
>automatically filter out characters that aren't valid in filenames
I installed the node but I can't figure out how to set up the regex replace
just sticking \ in it gives me
>bad escape (end of pattern) at position 0
what does yours look like?

>>8701999
>>8702003
It's not bestiality per se (but I think it counts for the most part), it's not hyper either (well, technically...), I don't know how to classify this but my monkey brain decided for the past couple of days that this is peak erotism (for now ¯\_(ツ)_/¯)
>https://litter.catbox.moe/599oggjojuzyt9vd.png
>https://litter.catbox.moe/4ffikzn7tgesbilc.png
>https://litter.catbox.moe/1hvlm0psb6zooc32.png

>>8702007
nevermind I figured it out, I had to escape the \ with another \

>>8702008
I think that's fine here, until it bulges her belly up to her tits
but they look loli-adjacent which might be a bigger problem. younger than in their pedo appeal source material even

>>8702008
Oh you're the giant dildo anon. I've been genning stuff like this. Might do a whole pokemon roleplay in a bit.
https://litter.catbox.moe/pmai0pvasy6ssyb0.png

what causes this?

>>8702012
zestiest nigga there ever was

unc thinks cause and effect still hold true in the big 2025 :skull:

what did he do to make them niggas so mad?

>>8702012
zesty ahh houndoom

>>8702009
run this one for maximum peace of mind (not many but there's a few booru tags with annoying characters)
['\"\<\>\-\+\\/\*\?\!\:\|]
I probably have too much extra escaping going on but I never remember which special character can be used without escape in char groups and escaping the ones you don't need to escape doesn't cause issues so wutever
some are valid in filenames but I really don't want to see shit like ' in a filename

>>8702012
pokepihilia is great but yeah it's that weird spot of where to put, just been on a hypno / bokoblin binge

>>8702026
I should do more hypno and hypnosis in general. Hilichurls are still the best if you don't want orcs. Genshin did a lot of good for AI.

>>8702028
orcs is fine sometimes they just come up weird but yeah those are just my go to for monster male hilichurl, hypno, bokoblins. Size play gets screwy cause there's poor ways to enforce it and I've had female orcs sometimes just on orc so those 3 somehow are strong enough for visual distinction with rare bleed over.

>>8702017
>>8702021
>staring at the dog instead of the girl
That's real gay my nigga

>>8702031
Serena is the worst pokegirl MC by miles tbf.

samefag status?

>>8702032
No that's plain jane Gloria.

>>8702031
>makes the dog stare at you in the gayest way possible
>"stop looking at him!"

>>8702036
>stare at you
Do you even know what eyes are? Just come out of the closet and go back to /y/

>>8702035
0/10 gloria is great.

>zoomer shitfus
it's kasumi and may

For me it's Iono

>>8702039
>old is better
Go sit down jii-chan. Everyone knows Rosa is the hottest MC.

>>8702042
>shit tastes even within zoomerfus
if anything it's made for anal hilda

>pokemon
I'd beat you up with my duel disc if we ever met in person

>>8702044
>le ebin contrarian
Both BW girls are equally hot for different reasons but May has always been trash. Misty was far superior even if she wasn't that good. Dawn is good doe.

Can you schedule bakes in ezscripts or should I finally take the raw sd_scripts pill?

>>8702048
There's a queue feature but i never used it

All protags are objectively inferior to Chicken.

>>8702047
>n-no you don't get it sweaty bike spats are actually bad

>>8702049
Oh shit yeah I didn't realize, it blends in well.

>>8702051
Ok you got me there actually. But that's why Rosa isn't trash. Hair buns, long twintails (not shitty short ones), and white pupils are peak.

feet

>fucking around with LRs either brick or it doesn't cook
Losing my gourd with this shit.

>>8702048
That's the only reason I've stuck with it.

why is prodigy always so shit when I use it? How was this shilled here in the past?

>>8702083
her canonical age?

>>8702084
I was about to say this shit kek.
>>8702083
This is illegal content in the UK.

>>8702083
if it's hot and dry enough her bush could have a forest fire

>>8702083
You really should post such things on /e/. Not that it's out of place here, but there's at least a chance you might get a positive response over there.

>>8702083
if she were to travel back in time and get fingered by a caveman they would discover fire

>>8702091
is the blud from her hymen or the bird?

>>8702093
kek

>>8702093
The bird can't hurt her, it's technically a part of her subconscious that's just manifested into reality. So it's likely her hymen, or she's just on the rag and still letting him hit.

>>8702094
Or if you want a bird to go spelunking...

>>8702095
>just on the rag
Australian?
In other news, how did you do her eyes? Inpaint?

pedo stuff goes to /b/, sir

>>8702101
you have the opportunity to introduce a new 3rd character using her butthole

wawaweewa...

>>8701963
>You're claiming that he's splitting hairs but you never actually said what was wrong.
if you read
>t2i at higher resolution than base model was trained on
you can read that as their argument hinging heavily on the "t2i" portion, meaning it was originally developed to be done blindly(...just like hiresfix) and that it very technically isn't "upscaling" despite upscaling also intending to push output to a higher resolution than the model is actually trained on.

it's just dumbass semantics to claim deepshrink isn't used(and wasn't originally released) primarily as an upscale technique. It's an argument you make to try and nitpick someone else into being "wrong."

>>8702071
it was good when you were a retard running shit adamw lrs ig

>>8702102
She's just a naga'd Eileen the crow from bloodborne. No idea why she turned out looking younger, it wasn't instructed.

>>8702098
Yeah, I inpaint anything I want good details on once I have a good overall generation.

>>8702109
Is she in prison? Also it's just raw inpaint? Do you sketch or anything?

>>8702110
I sketch major items, but for those eyes specifically I cleared the entire prompt and rewrote a new one just for the eyes and then inpainted. Works pretty well IMHO.

Also, it's her imaginary castle.

>>8702112
Fish cosplaying a fish.

>>8702110
Here's something that required sketching.

why is it so grey blurry and shiny

>>8702120
I prompted it that way. Wanted it to be a kind of fade out splash image. Here's another like it, this is in an incomplete state, so it will be more refined when I get back to it.

>>8702107
the important thing about hiresfix is that it's an *actual* upscale, it requires you to re-encode image twice using vae and sandwitch an upscaler model inbetween (unless you're doing latent upscale in which case you are retarded)
deepshrink is NOT an upscale technique you dumbass

>>8702116
Fish cosplays a wise leaf.

what caused it THIS time?

>>8702143
What caused what?

>>8702146
>>8702143
Probably the blood. These images don't exist within the same canon. It's not like everything she gets drawn doing happens to her all at once. That's only the case if you make a comic or something. If you need an excuse, you could say that these are all just robust hypotheticals or dream states.

>>8702143
Bad parenting.

>>8702149
ligne claire effect is kinda cool, but the proportions are nasty and the lora looks kinda fucked on later timesteps. I'd consider rebaking and getting better taste

>>8702122
Plenty of ways to dress up an idea.

>>8702152
I think this series turned out well. At any rate, it was just an experiment.

>>8702152
>>8702154
Ah, before I move on I should share the principal forge.

>>8701918
Decent, but NAI would have done belt and other clothing details better:3

+1 anals have been deposited in your account

why is this retard spamming this garbage that no one actually wants to see
I get that he hijacked the OP to try and strongarm this shit but what's even the goal here

it's like watching a movie frame by frame. i wonder what happens next!

>>8702173
>>8702176
I didn't post OP. Someone else used that image.

What is it about these images you dislike so strongly?

>>8702177
I hope it eventually stops being 480p. hd should have buffered by now

>>8702179
>fried to shit
>ugly as sin
>emphasis on grossout shit
you'd unironically get better reception on /d/

>>8702177
That was the general idea, it came out kind of sloppy though. This is actually a pretty old gen, I wasn't as good at getting what I wanted out of the effort.

>>8702184
Those are all pretty subjective. What's an example of this gross out shit? What do you want to see?

>>8702185
>{{{{{{photorealistic, realistic, real life}}}}}}
I don't think it's working, bro

>>8702186
He's just some schizo or maybe the usual disgruntled pedo

>>8702181
Works on my machine. Like I said just a minute ago, these are all pretty old. Never went back and upscaled them. You're free to do so if you like!

>>8702186
>What's an example of this gross out shit?
Emphasis on grotesques anatomy, clit-penis shit, "odd piercings," period blood, etc. is all "gross out shit," retard.
>What do you want to see?
Not 10 of nearly the same exact image back to back, that's for sure.

>>8702187
Those are just counterweights. I wasn't trying for literal photorealism. I didn't know how to properly mix styles at the time, so that was how I compensated.

>>8702191
Hmmm, I was alright with the period blood and everything, but color palette in this one made me vomit:(

>works on my machine
>is posting fried v4.5 slop
what did he even mean by this

>>8701963
More importantly, what does all of this mean for RAUnet? Achieves the same efect but seems to limit composition and seed variance, similar to that failed SMEA-DYN implementation which was also more stable at high res. Any clear winners here, and can they be used together?

>>8702193
I've posted quite a bit of diversity I think, but I can go further I suppose. I happen to think the clitoris is pretty attractive, and it's not period blood, but defloration. After the fact shots, if you will.

I still don't understand why retards posting western slop gravitate to the JAPANESE porn board
what makes them see the sticky and go "hrm. Yes. My western themed, hyper realism garbage definitely belongs HERE!" ?
I just don't understand

>>8702195
Can't please everyone hahaha.
>>8702202
There's no board for Chinese characters like these, or are you saying the style isn't Japanese enough for you?

>>8702205
why are you reposting garbage you didn't even make and trying to pass it off as your own, as if it's somehow relevant to you posting westoid hyper-realism shit?

>>8702197
Hm, can you elaborate on what you mean by fried? I'm all ears.

>>8702207
Who says I didn't make it? And if the comment isn't directed I can only assume it's meant for everyone -- which includes me.

>>8702049
it worked nice

>>8702210
>Who says I didn't make it?
Me, because I made it you fucking retarded faggot.
https://litter.catbox.moe/3iscyxm9nd4fdysb.png

GEG

>>8702210
>>8702213
What if I didn't make any of this stuff?! What if it was all you?!

how did /g/ attract such mentally deranged schizos
and why the fuck did they come to the fucking japanese porn SUB-thread(not even the "real" hdg)

>>8702217
People who were liable to be on /g/ before are the least path of resistance to trannying and culture war.

>>8702208
nta but I can try. It's used to refer to either overblown contrast, like from using way too high CFG, or it's many little things. Like in >>8702162 if you look at the sweatdrops they're like a broken crystal, way too many sharp lines where there's supposed to be a clean gradient. Her right nipple is a square. There's a random streak of purple in her hair. Her eye is melted despite having enough resolution for a clear one. Some of the lineart is red-brown or green instead of black, hair highlights are sharp geometric shapes that make no sense, pussy looks like an unidentifiable slab of meat with too many skin folds and too wide at the top, there are green veins on the underside of her breasts, etc.

On local these are signs of using a burnt lora, or pushing prompt emphasis too high. Easiest way to tell is to try inpainting them or upscaling, they won't go away if something's burning the result.

>>8702216
>>8702213
>>8702217
Sorry, I'm just gaslighting myself to bring attention to how much effort I spend on these gens. I literally live on these threads and attention is my lifeblood.

Seriously why would you try to claim as yours stuff you saved from here when it's a "community" of 8 people who are immediately going to call you out?

>>8702210
>Who says I didn't make it?
it's a default 4chan-saved image filename without any metadata that looks nothing like your trash.
while your slop is all default nai webUI shit with metadata attached.
doesn't take a genius to rub two twigs together to figure it out.

>NAI
>effort
good one

>>8702222
Nice quads. Those are interesting details, but they seem irrespective of the overall mosaic being presented. Those are definitely derivative of the model being used and lack of care for the minutiae but the generator, but there are similar mistakes and oversights in every type of media.

Look for a problem and you'll find one, so to speak.

>entire thread is two loratards arguing with themselves interspersed by obvious bots talking to each other for the rest of the day
comfy :3

>>8702227
You've never deleted something by accident and had to redownload it from your own post?

I'm probably retarded though, so I get it.

WHAT CAUSED IT THIIIIIIIS TIME????????

i like that it ignored the box lmao

>>8702232
Point is those problems do not show up frequently when using reasonable settings. There's maybe one or two in a gen, but never such a crazy amount unless something is wrong.

>>8702234
I wonder if this may be the most harmful post in the whole thread. True or not.

I mean once you have to suspect every post of being a bot, is there any reason left to even read them?

tldr? i don't read bot posts

>>8702242
Yeah I get it, that one in particular was done on a phone using novelai on a trial account and a dozen inpaints. It's also pretty old, so the eye melt wasn't something I knew how to fix.

Reasonable criticisms all in all, thank you. AI generated art supremacists aren't worth talking seriously, but I appreciate good insight.

>>8702122
>>8702153
What's the workflow for something like this? Are these loras or do you just prompt and inpaint as you go? I always have a bunch of ideas for scenes based on my textgen stories but I don't even know where to start for a lot of the complex positions. And for those gacha sluts did you sketch the details like the interlocking fingers on the sash or the "shark" on her tail? AI never gets those correctly.

nothing of value was lost

>>8702248
why are you posting more shit you didn't make unless you're trying to pretend to be noodleschizo now, too?

>just realized i could automate my dataset upscaling even more
maybe nodes ain't that bad

>>8702244
I don't think having a human being the post makes it any more likely to have something of value.

I'm real and watching dexter right now

>>8702179
>Someone else used that image.
i now regret what i did and sincerely ask for forgiveness from everyone.
i can only offer a filter rule for 4chanx (put it into filename field)
>/^.*-....-....-....-.*\.(jpg|png)$/i;boards:h;op:no;

>>8702248
uh, sure
I didn't really meant to criticize, just that anon was asking about what fried means and I needed an example pic

I hereby declare the decline.

>>8702254
I don't think LLMs can get erect and even have such a thing as taste in hentai to discuss

>>8702249
Careful prompt structure does a lot of the heavy lifting. Making the character details first and with emphasis seems to help let the groundwork better than anything else. For complex scenes like the Ellen/Miyabi one, I just try to get the general location and orientation set up then inpaint or img2img forward. Sketching is usually something I need with on final passes, like if my descriptions are just not working. Sketch + inpaint + fresh prompt.

Another muddy one I need to detail, but it took all of these elements I just mentioned.

>>8702265
Post catbox for this.>>8702153

>>8702259
I was the one who asked and the poster of the example, haha.

This nigger acting like we can't see the metadata

sure wish "moderation" was a thing that fucking existed

>>8702261
Does that matter to you? What if the boot just says it got hard? Is that different from when a human intones that? How do you know?

>>8702268
That's not the original unfortunately, so I don't think there's any metadata for you to glean. When I saved those NAI had one of its customary captcha conniptions and the only way I could salvage them was to pull them out of temporary files. It's why that set never got refined.

>>8702274
cool it with the anti-pedophilic remarks

>>8702261
>I don't think LLMs can get erect

Do I wanna know what is going on this time