Thread 8714671

Real edition

Previous Thread: >>8707841

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

>RESOURCES
Wiki: https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki | https://comfyanonymous.github.io/ComfyUI_examples
Training: https://rentry.org/59xed3 | https://github.com/derrian-distro/LoRA_Easy_Training_Scripts | https://github.com/bmaltais/kohya_ss | https://github.com/Nerogar/OneTrainer
Tags: https://danbooru.donmai.us/wiki_pages/tag_groups | https://danbooru.donmai.us/related_tag
ControlNet: https://rentry.org/dummycontrolnet | https://civitai.com/models/136070
IOPaint (LamaCleaner): https://www.iopaint.com/install
Upscalers: https://openmodeldb.info
Booru: https://aibooru.online
4chanX Catbox/NAI prompt userscript: https://rentry.org/hdgcb
Illustrious-related: https://rentry.org/illustrious_loras_n_stuff
Useful Nodes/Extensions: https://rentry.org/8csaevw5

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

Another AI garbage general? Just wonderful, thank you so much. Jannies, when are you going to purge the board from this cancer?

>>8714678
you'd be quicker to get a janny to look at the thread if you posted nagatoro lel

Oh for fuck's sake. If the inpainting area is too large the inpaint will be fucked fiercely, no matter what you try. Gotta inpaint, piece by piece, then correct as needed.

The more you know, I guess.

>>8714678
>>8714682
>actually true
Based.

If Miura had kept Griffith as Frikka as he originally made her, the manga would have been much shorter and we would have known the actual ending.

What are the best VAEs for sdxl? I know of the basic one but do others exist?

>>8717834
The best? NAIv3 VAE.

Just started using ComfyUI, where can I find workflows to generate text to image better?

>>8718283
you can only find missing nodes, shitty custom node shills and grifts

I've been annoyed at my gens the past 2 days and I thought that I had reached the limit of what the models and my pc could do. then I saw that the past 200+ gens have had an incorrect node link, and I feel retarded

>>8718283
read/watch a tutorial series and you'll be able to fine tune shit yourself and better understand shitty premade nodes you download

>>8719099
You're not gonna find S3 there anon.

>>8719116
i will keep looking

NAI GODS... how the fuck did this guy get something like this?

man some gens I get are really good, these ai models have completely outclassed 90% of artists, its insane

>>8714877
Box for original? Also what are you inpainting?

>>8719099
He lost a finger in 'Nam, poor bastard

>>8719810
He lost more than one finger anon, look at his hand

>>8719783
https://files.catbox.moe/29kfsa.png

It's not porn related, just people at different distances in the horizon in what I can only call a panorama.

Any suggestions on Wan2.1 or 2.2 uncensored anime models? Or good workflows. Since am seeing people post anime clips over here. In my case the results always has issues and anomalies while using the default template.

>>8719965
Thanks, I can now remake this but with flat chest!!

>>8719974
your legal name and home address?

>>8720182
must be this one https://danbooru.donmai.us/posts/7076610

Was everyone just waiting for the new thread? Or is /h/ only dead during euro hours?

>>8720196
ohhh it's the old duplicate, nvm

So, are we supposed to be using this thread or not?

yes

>>8720272
of course not just look at the OP

>>8720272
just move to /hdg/, let it die

>>8720282
you do that

Her proportions are all off, but this is only a for fun prompt, so I will leave it as is.

what's fun about that

>>8720321
Penis-sama finds it cute and funny.

>>8720272
boox?

>>8720278
>>8720282
Having taken a look at /hdg/, I'd rather post here with the other two persons that forgot to open the catalog.

>>8720336
Why?

>>8720336
based and same but the retard who said it was going to bake a new (real) thread better do it or I will

>>8720320
Bruh this picture is simply fucked up, literally just squish it

I wish Miku was real so I could install neuromorphic skin and tactile sensory arrays in her.

baking

>>8720332
https://files.catbox.moe/pnmxom.png

baked

>>8720272
>>8720320
>>8720336
>>8720388
Which checkpoints and LoRAs are you using?

>>8720415
What's it to you, buddy?

>>8720416
So I can make my own images of Kanna getting railed because Kaedeko has become fat and lazy with his lewding of her with the same basic and boring poses over and over. I don't need her to be taking 30 dicks but every picture for the last few months has been her sitting in some outfit and then her boobs are out. There is no creativity like the fan censor, the bathhouse locker room cover, the wet shirt from the rain, the hanging upside down, etc.

For all people complaining about AI, I sure get a low of views on pixiv for what are mere sfw images.

>>8720417
Nova anime. It does Kaedeko fairly ok. And you should be glad he doesn't lewd her, because all he does is smooth-brained ntr plots. They are really, really bad. He is also lazy as fuck with his writing.

>>8720423
Every month he does one image of Kanna, then throughout the month he takes off her top, then does a variant with wet nipples and other with cum on her tits. My point is his earlier images with her tits out are better than the more recent ones as those just follow a formula.

>>8720423
The only people truly complaining about AI are the low to mid artists, the good ones have the following and likely no issue with continued work it's the grifters and aspiring

>>8720427
Oh, I am painfully aware. He got on my nerves a long time ago. That's the main reason I make porn of Kanna and release it into the wild.

>>8720430
My thoughts exactly. If he kept doing these candid shots, maybe threw in a pussy or asshole every now and again, would be fine. But since he's been lazy, generate all the pics of her taking all the dicks.

>>8720388
thank you

Anyone here tried working with Kwen with less than 24GB of RAM?

Any tips to get ACTUAL micro bikinis?
micro_bikini and string_bikini tags barely do anything. They look like normal bikinis.

>>8720515
Use a LoRA?
https://civitai.com/models/818739/micro-bikini

>>8720511
?

>>8720515
you could always just prompt for nude and then doodle in some lines before the upscale and then append the prompt to reflect the swimsuit
example:
original lowres https://files.catbox.moe/cvvgfx.png
doodled lines
https://files.catbox.moe/5m5v28.png
https://files.catbox.moe/odabnh.png
upscales
https://files.catbox.moe/zue5fx.png
https://files.catbox.moe/d86rhc.png

or more well endowed version and in motion
original lowres
https://files.catbox.moe/pyzdxd.png
doodled lines
https://files.catbox.moe/o65hns.png
https://files.catbox.moe/g6yqag.png
upscales
<image related>
https://files.catbox.moe/thp06n.png

were the literal toddlers necessary?

absolutely

>>8720415
The catbox is right there you dumb fuck

where

deleted so fast I couldn't even make a move :(

>>8720635
https://files.catbox.moe/kfml7d.mp4

>>8720632
workflow

>>8720637
https://files.catbox.moe/b5lxpb.jpg
that's a first last frame to video

i forgot to put in landscape mode the gen so it cropped the first frame and transitioned easier by accident to the last frame <.<

>>8720515
An even easier method is to state which bits you want visible. Example:

micro bikini,areola slip,pussy peek

Will yield very reliably a very tiny bikini.

Not erotic enough to post in /h/.

https://files.catbox.moe/utvs8t.png

I swear, the day AI gets eyes right 100% of the time is the day real artists are buried and done.

>>8720648
subjective af since there are shit tons of eyes and levels of details

my point is, ask what is a good eye, no one will agree

>>8720648
>>8720657


He's talking about the retards who don’t use a 2nd-pass/Adetailer.

pfffff AHAHAHAHA

Wasn't anon gonna bake a new thread? What happened?

>>8720695
Nobody posted an image for op

>>8720695
Maybe I managed to convince him to merge back into /hdg/. Can't see any difference between the generals at this point.

>>8720713
No

>Illustrious and Noob shitmerges will be one whole year behind in its danbooru dataset next month
grim

>>8720730
Oh no think of all the FOTM gacha and seasonal anime whores!

>local will never get a new model or move on from sdxl
grim............

All the rich kids moved on to wan tho. only poors remain in the imagen threads

>>8720735
This but unironically. I'm laughing at the poorfags who have to cope with their static porn lol

Man I realized I can't even bring myself to shitpost shill WAN videos, this shit is legit sickeningly terrible and I just felt bad praising it

>>8720747
The fact that 90% of i2v posted uses some uncanny 2.9d shitmix gen as base is telling of the kind of person who thinks wan is good.

>>8720747
i kind of want to try it just to see...

lol cope naishill

novelai gods completely own this space. localkeks just cannot stop seething

you wish :^)

the wish was already granted in 2023, localkekkyseether

naisisters, wangods keep laughing at us...

>>8720747
I didn't want to do it for a while but yeah filtering videos on /h/ and /e/ is the way to go.

Why do half of these civitai gens use underscores in tags and no spaces between them?
I don't get it.

>>8720735
y u say this? ;_;

I no longer get good gens, this sucks.

>>8720801
The question is: did your gens get worse or did your taste get better?

I no longer gen

>>8720711
Nigga are you for real, there was plenty of good images to use for OP already

>>8720837
recycling images for op is /hdg/ territory

>>8720851
those threads are fucking identical, fuck off with your cringe larp

>>8720858
Then go back

>>8720851
Well, too bad, there is no point on another new thread at this point, if we created another is more than likely going to get deleted so we are stuck on this one

Anacreontic Miku paizuri.

What caused the decline?

who gives a fuck about the OP image?

>>8720867
i do

Oh hell naw

good morning

How do I best merge loras? the supermerger extension for a1111 is crap and just throws errors

Hmm I was looking through some notes I made a couple years ago and nothing with NAI has changed. Even back at release, NAI's strength was its style (darkness/flat images) but the model itself was all they had. The interface was shit, you couldn't upscale into hires, you couldn't use CN, loras, it was all a real pain. Had that model been local the types of stuff we could with it would have been insane. Local had everything else but was really lacking in style. Now fast forward to today and style has significantly improved and we can even gen complex poses we couldn't do back then, meanwhile NAI's style is shit. Once again, the model is all they have since the ability to gen really complex stuff is insane but with shit style and no loras it's worthless. It's just so sad for the hentai scene. We're always so close to greatness but something always cockblocks us.

happy for you bro or sorry that happened

tldr?

nai [isn't] local and bad

>>8720975
Yeah, pretty much my opinion and why I stopped my subscription back then, I am waiting for a new model as everyone here does but honestly, if we don't get one I am fine with what we got

hello my favorite pajeet niggers, any new noob versions worth checking out? the ones I've tried are almost good.

>>8720989
https://civitai.com/models/1858821/noobai-v-pred-10-with-eq-vae
This is the latest "version" of noob that isn't just a shitmix

>>8720990
oh sweet I just dl'd that today but didn't think it would change anything much, I'll give it a go

>>8720990
I didn't like that one. "slight aesthetic finetune" my ass

>>8720990
>those previews
I'm gonan stick with ym shitmixes.

>>8720992
Well, you know...
>aesthetics
>bluvoll
LOL

>some anon gives a logical reason for high dropout 2 threads ago
>try it at 0.5 and compare for differences that dropout might affect
>it had a fairly significant positive effect
>didn't take that much longer to bake
neat.

>>8720990
looks bad.

>>8721028
Who are you quoting bro?

>>8721032
me

>>8720975
I just really don't get what did they expect without adding realism. On local, we have models like WAI to add a nice detailed style, NAI... just looks rough, like some troonime scribbles, looks really boring and plain. What kind of incel would find that appealing?

>>8721035
it's because nai lost

Can you use illustrious and noob checkpoints and loras interchangeably with each other?

>>8721053
to some degree yes, but loras trained on vpred will only work on vpred

>>8721054
man I don't understand all these terms, what's vpred? what cyberfix dude it's so complicated...what do they even do?

>>8721057
You’re not alone — this stuff *is* complicated, especially when people throw around terms like “vpred,” “illustrious,” “noob checkpoints,” and “LoRAs” without explaining them. Let’s break it down in plain English.

---

### First: What are these things?

#### **Checkpoints (like "illustrious" or "noob")**
These are full AI models — the big files (usually 2GB–7GB) you load into Stable Diffusion to generate images. Think of them as different “brains” trained on different styles.

- **“Illustrious”** = usually a checkpoint fine-tuned for anime/illustration styles.
- **“Noob”** = often refers to beginner-friendly checkpoints, sometimes trained on broad datasets for general use.

They’re your main model — the foundation.

---

#### **LoRAs (Low-Rank Adaptations)**
These are *small add-ons* (usually 5MB–200MB) that tweak the behavior of a checkpoint. Like plugins or mods.

Example: You load “Illustrious” as your base model, then add a LoRA called “Cyberpunk Glow Eyes” to make characters’ eyes glow in a cyberpunk style.

LoRAs are meant to be *layered* on top of checkpoints — but they don’t always work with every checkpoint.

---

### What is “vpred”?

This is short for **“v-prediction”** — it’s a *training setting* used when the model was originally created.

There are two main types of prediction modes in Stable Diffusion:

1. **epsilon (ε)** — older, more common
2. **v-prediction (v)** — newer, sometimes produces better contrast/dynamic range

> **Crucial point**: If a LoRA was trained using **v-prediction**, it will only work properly if the *checkpoint you’re using also uses v-prediction*. Otherwise, the image might look weird, broken, or just… off.

Same goes for training settings like resolution, dataset, etc. — mismatched settings = bad results.

:^)

>>8721058
So can you use illustrious/noob checkpoints and LoRAs interchangeably?

Sort of — but with BIG caveats:

You can try any LoRA on any checkpoint — nothing will explode.
But it may not work well (or at all) if they were trained with incompatible settings (like vpred vs epsilon).
Best practice: Use LoRAs that were trained on the same base model or compatible architecture.

Example:

A LoRA trained on “Illustrious v2 (vpred)” should be used with other vpred models, not an old epsilon-based noob checkpoint.

Pro tip: How to know what’s what?

Check the model card or CivitAI page — good creators list:
Base model used
Training settings (vpred? epsilon?)
Recommended checkpoints

Look for tags like “SD 1.5”, “SDXL”, “vpred”, “anime-trained”, etc.

And “cyberfix dude”?

That sounds like someone frustrated with how complex this ecosystem has become — and honestly, fair. There’s a LOT of jargon, incompatible formats, and tribal knowledge.
TL;DR — Simple Summary:
Term What it is Can you mix freely?
Checkpoint Main model (e.g. Illustrious, Noob) Mostly yes, but style varies
LoRA Style/plugin/mod for a checkpoint Only if trained compatibly!
vpred Training mode — affects compatibility Match it or risk bad results

Think of it like car parts: You can bolt a Ferrari spoiler onto a Honda Civic… but it won’t work right unless the mounting points match.

Let me know if you want help picking compatible models or decoding a specific LoRA/checkpoint combo — happy to help!

skip the line
sub to nai

im already subbed

>thread getting discussions again?
>restart the NAI spambot

its that good and only 25 dollars its a no brainer honestly

>they discus?
>spam eet

Wait NAI (novelai.com) is ONLY 25 dollars?!?
I thought it was at least 100 dollars from how good it looks!

>>8721074
hard to believe theyre so generous!

>trollbake with 22 images
still gets spammed

is nais :--D

I tried the 30 freebies, and wasn't very impressed. I'll maybe consider buying an NAI sub when they add better samplers and noise schedule. These being the only options is laughable. They're ahead in many things, but behind in some aspects. I do like the character prompts feature and being able to insert a character reference. The character reference feature doesn't work as well as I'd hoped and there's no real way to adjust the strength or how it is being applied. I'd recommend just sticking to local shitmixes and coping with character and style loras. Thanks for reading my blog.

>>8721077
Lol you think they care now? They just start their shill spambot at any /h/ thread and laugh when people get angry at their bullshit, they don't check the thread stats. Our only hope is for mods to rangeban Israel but you know that is never going to happen.

>>8721080
yeah for me as a local user the thing I like doing the most is making comprehensive plots comparing different samplers and schedulers. with all that shit missing on nai i literally have no idea what the fuck to do lol

>>8721080
if i only got 30 test images of local id give up on AI completely

>>8721086
shut the fuck up retard nai sucks

>>8721080
I only used 12 of my 30 images before getting bored, maybe I'll use the rest for shitposts

>>8721083
Why does she look so korean?

>>8721089
what does that even mean

>>8721083
hmmm, catbox?

>>8721089
it's aegyo sal

>>8721083
https://files.catbox.moe/oxoefh.png

>>8721086
Retard

>>8721092
bless you!

catbox?

>>8721094
thanks bwo.

chattini, act 3 bosses are really cracked up

Bateman opens his laptop with deliberate precision. A faint synthetic glow washes over the table as he clicks a tab.

PRICE (suddenly enthused)
What’s that, a new checkpoint?

BATEMAN
New LoRA. What do you think?

McDermott leans forward, narrowing his eyes. On the screen: a suspiciously glossy “anime waifu” with blinding plastic skin.

McDERMOTT
Whoa. Very nice. Take a look.

He gestures, and Van Patten peers over his shoulder.

BATEMAN
Uploaded it to Civitai yesterday.

VAN PATTEN
Good shininess.

BATEMAN
That’s FlexiGloss v3, with DreamBooth residual training.

McDERMOTT (envious)
FlexiGloss v3?

VAN PATTEN
It is very cool, Bateman. But that’s nothing.

He flicks his own laptop open with a smug snap and pulls up a LoRA gallery.

VAN PATTEN
Look at this.

The table leans in. A plastic idol stares back, its legs so smooth they could deflect laser fire.

PRICE
That’s really nice.

Bateman clenches his fists beneath the table, fighting to maintain composure.

VAN PATTEN
ShrinkWrappedGradient, trained on eight RTX 4090s.

(Turning to Bateman)
What do you think?

BATEMAN (barely able to breathe, his voice a rasp):
Nice.

PRICE (lifting the screen up to catch the glare)
Jesus. This is really super. How’d a nitwit like you get reflections that sharp?

Bateman stares at his own LoRA gallery. His jaw trembles as he looks back at McDermott’s upload—another plastic waifu, shinier still.

BATEMAN (V.O.)
I can’t believe that Price prefers McDermott’s LoRA to mine.

PRICE (with relish)
But wait. You ain’t seen nothin’ yet.

He clicks open his profile.

>>8721163
PRICE
Raised specular highlights. Pale nimbus skin…
like she’s carved from Tupperware.

BATEMAN (choking on sheer anxiety)
Impressive. Very nice… let’s see Paul Allen’s LoRA.

Price smirks and, with dramatic cruelty, scrolls. The name glows across the screen:
“PAUL ALLEN – ANIME_DOLL_SUPREME_v5.”

Bateman swallows. The room dies into silence. A faint, rhythmic heartbeat surges in his ears as his eyes devour the hideous brilliance.

BATEMAN (V.O.)
Look at that subtle wax finish. The lifeless smoothness of it. Oh my God… it even has an artificial watermark baked into the model output…

His hand trembles violently as he grips the table edge. The LoRA fills his vision. Drops of sweat bead on his forehead. He can’t breathe.

CARRUTHERS (from across the table, concerned)
Patrick… you’re sweating.

chat, can't you stop losing your mind already, I am almost done with silksong I swear

should have ended it with nai

>>8721163
catbox if local?

why would you put a cat in a box?

because thats where my dick is

>>8721054
False. Loras trained on vpred work just fine on Wai plus improved colors.

>>8721083
>>8721094
>solo focus
Hmm... maybe that's the issue? I want to do pov but from the side like that rather than the traditional straight-on. Maybe female focus will help too.

>>8721193
grey should not be a color anymore.

>>8721196
greyscale is cool

https://www.pixiv.net/en/users/105587572
>this guy somehow has 2k patreon supporters
>cutesexyrobutts has 1.5k which makes this absurdly unlikely to be legit
>his pixiv/twitter posts barely ever have posts going over 1k likes

I've gone down a bit of a rabbithole and have realized there are creators out there that are inflating their counts with gift memberships that they can send out for free at no cost. what even is the benefit of this? if you look at their kimono there are a bunch of bot like comments on their posts.

>>8721272
>there are a bunch of bot like comments on their posts
People who pay for AI porn might as well be bots anyway.

I'm not sure whether having the appearance of bigger numbers gets you more actual numbers, but clearly he at least believes it to be true. Well, assuming you're right.

>>8721272
I think it's just a cabal of ai sloppers that do exactly the same thing to the T. I would not be surprised if they're all based in India.
They share the same exact format, have sizeable followings on Twitter, all have the blue checkmark and all have a patreon where they post 50+ 1girl sex gens every day.

>>8721272
>there are creators out there that are inflating their counts with gift memberships that they can send out for free at no cost.
and im assuming you have proof of this right

>>8721312
His proof will be calling you a shill for defending a known grifter.

>>8721326
damn he got me

>>8721080
You get 30 gens per account or IP?

gimme your lora training settings sars

>>8721058
>>8721060
Chat is this real?

jelly belly :3

>>8720871
Never forget, anons.

Wow. An actual schizo haunts this place. Very grim.

>>8721516
bwo did you miss last 1.5 years of hdg or something

It was a simpler time.

>>8721524
Bro this schizo wasn't there before.

Been out of the loop for a while, any Illu / noob checkpoints you would recommend now?

>>8721542
NoobAIXL vpred 1.0 is all you need.

>>8721028
Hand over the .toml old man

>>8721586
ppl dont use it anymore though
stuff like wai-nsfw 15/14 is more common

>>8721596
if you are retarded shit-eating civitai slopper then yeah sure switch to wai like everyone else.

>>8721598
Using a bad outdated model doesn't make you smart or cool.

>>8721605
WAI is just as "outdated" as noob.

>how do i make local good
you cant lil bro.......

>>8720423
At least on Pixiv, people got sick of all the slop and started filtering AI-gens out. I used to get 10k+ bookmarks and now I'm lucky to get 1k.

>>8721589
i haven't done a lot of testing but i literally just added 0.5 tag dropout to the end

>>8721589
also what are you trying to do?

>>8721609
But it's not dogshit like noob.

>>8721648
Totally agreed, it's much more dogshit.

>>8721643
Steal everyone's .toml's and bruteforce something

>>8721689
no i mean like, bake what

It's not that I don't want to post more, it's that I am too busy gathering datasets and experimenting, so I barely gen anything /h/ worthy.

>Best fidelity loras are actually merges of finetune extracts and snakeoil high dim loras
grim

>>8721786
prooves?

>>8721796
I wouldn't lie, it'd make noodleanon sad

>>8721798
I actually tried some lora extraction shit but it was so insanely time-consuming to do compared to my normal workflow of comparing lots of different step counts of what I bake that I can't justify it.

stop posting on this thread you retards

>>8721813
nobody here is coming back to /hdg/ retard

>>8721816
no I mean, use the other new /hgg/ thread

>>8721392

>>8721820
ok, but not because you told me to

>>8721820
i post wherever i want, if schizos spam threads i just have them open in tabs

>>8721616
box please

>>8721822
stop being tsundere or i'll breed you