Thread 8613148

Sovl edition

Previous Thread: >>8600493

>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

>>8613160
lil bwo has one joke between two generals >>8613158

>>8613090
What was your end control step for the tile upscale? I've noticed that CN only works well on artists the model knows but on my loras it does hallucinations below 0.95 end step. Idk what's happening.

>>8613176
0.8 denoise first pass, 0.7 denoise second pass
both passes at 0.35 strength and 0.8 guidance end

>>8613090
This is stellar, if it's not the same style would you mind posting a box for this one as well?

>>8613090
thanks, final result looks great

>used my image for bake
now if only I could figure out how to bake that lora again but less melty

https://files.catbox.moe/9o94qg.png
>>8603720
thought that style looked familiar as hell and it turns out i was right
https://www.youtube.com/watch?v=NPU0O5mUJbs
>>8613090
i need some new upscaling snake oil i should try that out

based wildcard enjoyer?

>>8613367
just started with this shit
how did i do

i have more but i'll wait for some feedback
don't even know if this is the correct board to post this

>>8613371
>>8613375
Well, if you want me to being honest
>>8613358
>>8613364
>>8613371
>>8613374
Those are """decent""" but way too similar from each other, the style is more or less good
This is the right board but avoid spamming your images this much, especially when they have the same setting

>>8613376
ok, will do anon

>>8613366
>>8613364
>>8613363
>>8613362
>>8613360
>>8613358
that's what cats get for always sticking their stupid buttholes in people's faces

>wonder what a big tit artist style would look like applied to a loli
>it looks hideous because the artist normally draws wide shoulders

>>8613375
They are pretty good, but posting a lot of variations of the same pic is considered spamming.

>full_body in the prompt
>gen is a cowboy shot

came doesn't want to work for me hm

>>8607584

close-up box please?

>>8613596
Based resolution ignorer.

Any Intel Arc B580 owners? Have you tried training?

>>8613224
Mission accomplished, I think? It's not quite as sharp as I'd want but whatever

>outlives /hdg/
We fucking did it, sisters.

guys where are the good gens?

Optimizers done, now only schedulers and loss (and huber schedule) and I should be done.

>>8613848
oh wtf lmao what happened
did the janny just archive it because of the /aco/spam?

>>8613852
i feel like it could be because op pic sorta looks like futa autofellatio kek

>>8613852
Seemed cursed from the getgo with the pruned related generals, highlightfag not highlight fagging, and I think some anon reported it early on as NAI shilling on irc.

/hgg/ also got archived because of the spam and reports one time, now that i think about it

>>8613800
Umm, sweaty? You can have wide shots in landscape resolution?

>>8614013
you still live?

>>8614033
Never left, I decided to not post at all on the last thread

>>8614035
the thread was that slow?

>>8614055
It has nothing to do with that, I was schizo testing ,uhm, things

>>8613850
>he tests one variable in equilibrium and thinks he's done

I like these panels but I need to get better at prooompting

>>8614416
dunnyo what the hell you're saying cuh but i shall continue testing

>>8614588
They're waiting for you Freeman, in the test chamber.

>>8614588
can you check if compass supports degenerated_to_sgd = true as an optional optimizer argument and test it?

>>8614602
Compass didn't even want to work on my end so no lol

>>8613153
>got another round of xyzs
>adamw8 looks like dogshit stylistically in comparison to adamw
lol
all these years of "it's exactly the same just faster"

>>8614588
what are you doing anon?

>>8614655
I just 1:1d every easyscripts setting.

>>8614657
I mean, waht is the final goal? are you training / fine tuning a model? loras? just larnign ai concepts?

>>8614671
A good lora config more or less

>>8614675
oh I see, good luck!

>>8614675
>A good lora config more or less
>>8614588
>dunnyo what the hell you're saying cuh
you're about to

it's nai

>>8614727
very nice angle/pose

>>8614749
What prompt can help me get that angle? I can't force it via img2img on CivitAI? Any specific prompt could help?

>>8614756
That's not your picture?

>>8614769
I did gen them + using bed invite lora. But i can't get angles right. It is random

>>8614756
you mean the pov?
I thought you'd knew yourself lol, it's quite nice

wait what, this IS /hgg/ what the fuck are those posts

>>8614775
Catbox??
Should just be lying on side, pov, under covers. Might need close-up too at 1.2 or so.
https://files.catbox.moe/54bsvc.png

>>8614784
https://files.catbox.moe/m2orli.jpeg does this help?

>>8614786
Woah this was made on civit itself?

>>8614791
indeed

>>8614792
nice. does the accidental thighjob gen often with this prompt?

>>8614791
So? Prompting and loras work the same way, it's the rest of the stuff that's lacking. Adetailer feels really weird, can't use controlnet or regional prompter, fewer upscaling options, etc.

>>8614802
I didn't say whatever you're insinuating, was just surprised by the metadata is all.

>>8614802
People still think you need their 10 layer of snake oil to get a gen

>>8614814
it's more important to apply the snakeoil during lora training

>>8614799
I'm pretty sure I got lucky, but I don't see it being that hard to nudge if you added thighjob to the tags.
https://files.catbox.moe/8ad0w4.png

>>8614793
>>>/y/

>starting to notice the weird unsymmetricality of higher batches
it's over

>>8615288
disgusting

>>8615288
kino
>series
what about idolmaster or persona?

>>8615288
>Just did this
It's clearly ai generated anon.

>>8615344
Proofs?

>>8615288
wrong thread

ermmm....

>>8615463
tomimi's
BIG
FAT
tail

>>8615422
>anal DP
B A S E D

>>8615422
More DAP

I was the anon saying v4.5 was shit. I figured it out, not so shitty anymore but I do prefer 4.0 aesthetics

>>8615527
>but I do prefer 4.0 aesthetics
you mean v3, right?

>>8615550
not when it hits right
v3 is just all over the place

composition wise, style wise is depends on many things

>>8615550
No, anon just loves artifacts and blur.

>>8615557
Me too but only if it's called sovl because it's on 2005 gay cgs

I'm confused, are you supposed to use the adafactor scheduler with the adafactor optimizer always, or it is just some specific one? I know it doesn't work on other optimizers.

>>8615569
Oh wait it automatically does use it, never mind.

>>8615463
Repost so early?

>>8615527
Post a picture, sweaty?

>>8615596
no

>>8615600
Mmmm nyes?

>came suddenly started working
ogey

>>8615594
ah, my bad
I just dump them in one folder, lose track sometimes of which ones I already posted

>>8615110
>went back to batch 1
>suddenly optimizers are completely different
predictable but i'm glad i went back to them

>hdg got archived again
oh rmao

hdgsissies just can't catch a break. kek

this is what you get for removing NovelAI from the OP~

can jannies not do that
i don't want a rapefugee flood here

>>8615678
They just tabbed over as we all do, anonie.

the fuck is going on with hdg, and what the fuck is the difference with this thread?

>>8615681
We exhibit self control here and laugh at the hdg purges.

you guys were posting so much trash hdg killed itself again...

>>8615681
i assume
>janny archives previous thread for acospam and assumedly reports
>janny sees scat spam and acospam take up half the new thread
>there now exists a calm alternative so he doesn't care if it gets archived again

If you don't have a post archived in /hgg/ #001, we don't want you. Shoo, refugees. Shoo.

whatever, I'll just post some hentai

Uh oh. That one got her boilin' n bubblin'.

janny should have taken this thread down, this one is the even lower quality hdg

>>8615678
Sorry...

adafactor's kyinda good

This thread is superior because it wasn't baked by highlightanon

As /hgg/'s first pillar, I should inform you all of the ground rules.
1. No shitposting
2. miquellas are encouraged
3. No catfag
Keep those in mind and you'll get extended residency until your jeet general returns.

>>8615692
Box?

>>8615697
>1. No shitposting

NYOOoOOOOooooOOOooOOO!!

>>8615700
I know it's tough, anonie, but we've found here that a great way to fight the urge is autistically discussing lora baking for days on end. I believe in you.

>>8615709
Bakeranon, you should write a rentry.

>>8615709
isn't this general actually dead

>>8615685
This but unironically. I was in the screencap. In fact if you didn't post a pic with the original metadata then you need to go.

what are you n-words doing this time

>>8615720
we dindu nuffin

>>8615720
Looks like mass reporting is the new black.

>>8615738
I should really learn to inpaint...

anyone tried wavelet loss yet for training?

>>8615741
yes

>>8615746
nice

my old config was hilariously fucked up
>three times the lr
>takes twice as long to converge
glad i went and tested this shit

>>8615774
love sakura

What's the best way to translate a specific style from one model to another? My first approach would be to gen a lot of images on the first model with the style I want and then train a Lora with that for the second model but I'm wondering if anyone has a better Idea and I know that training AI on AI is bad.

>>8615827
>I know that training AI on AI is bad.
It's not if the images are properly selected

>>8615807
gonna share your findings?

>>8615827
There is plenty of loras trained on AI to replicate specific style mixes of NAI grifters on local and they don't have any specific issues, problems happen when you train whole models mainly on AIslop like LLMs do.

>>8615834
i am still doing it but i can write something up

>>8615851
oh, nah, if you're still testing then please continue

back from vacation for posting march 7th with small tits
Not sure if this function already exists somewhere, but I made a couple custom nodes for comfy since I was annoyed tard wrangling mixes with washed out colors if anyone is interested, just drop the files imto custom_nodes folder if you want to try them
Main useful one applies a luminosity s-curve to images, meant to be run on images right after vae decode (and then sent to hires pass nodes). Should be fine on defaults, just don't raise y3 unless you want to go blind. (or are doing some spot color/monochrome style) Defaults may not have good effect on images with backlighting.
luminosity s-curve file: https://files.catbox.moe/50x9cm.py
Other one was meant to do monkey color grading by a warm/cool axis but it also applies a multiplier on chroma which is some more snake oil to have fun with, use this node with low strength
color warmth grading file: https://files.catbox.moe/73o3uz.py

>>8615894
3 months ban?

>>8615901
kek'd

>three days of testing just to bake a single lora
worth

>>8615906
post lora

>>8615906
>three days of testing just to bake a single lora
My loras end in _XX, imperial. Can you say the same?

>>8616019
I gotta do it on the whole dataset first, but I'm starting to write up that "guide".

>>8615720
box please

>>8616175
it's just ixy m8

>>8615720
coomin my brains out because of the marvel announcement. we're so back.

>>8615827
>>8615833
>training AI on AI is bad
it's actually always bad if you're using plain eps gens to train a ztsnr model because you're overwriting weak low snr knowledge with the specific pattern that occurs in your dataset
training on nai or noob outputs is ""okay""
training on pony is bad

>>8616235
>The base model you are training on, it should be either Illustrious0.1 or NoobAI-Vpred 1.0 for most users.
Woah woah woah. What's wrong with baking on epred?

>>8616290
illustrious is eps
no real reason to bake on eps noob

>>8616290
its epred

>>8616235
nice but and the end of the day it's still some list of incomprehensible magic spells, there's too little evidence and even some incorrect stuff, still way better than whatever is in the op
>Scale V pred loss: Scales the loss to be in line with EDM
this is wrong, it just multiplies the loss with the snr curve, nothing super fancy, but this is mutually exclusive with min snr and it should not be used under any circumstances except if you're training the model using (neweset) v-pred debiased estimation (which is also mutually exclusive with minsnr and it's basically the same thing as min snr with gamma=1 but smooth)
>Flip Augment: Flips the latents of the image while training. Causes quality degradation. Keep on.
kek
>Max Grad Norm
it prevents too large gradients (basically gradients are the proposed changes to the network weights) from throwing off the training
>Noise offset
any kind, must be disabled for vpred/ztsnr
>Multiple batches allow for quicker training but also exhibit problems with symmetricality in the end bake. Keep at 1.
what? sounds like an issue with bucketing
>Keep Tokens Separator: Unsure of a practical use for it.
it's for nai-style training when you want to separate some meta tags (and keep it in place) from other tags, for example "1girl, artist, character ||| whatever", basically keep tokens but more flexible
>Cache Latents
this also increases vram usage unless you cache it to disk
>Dropout: Used to drop out parts of the model. Causes degradation of quality. Keep off.
most likely you've used way too large numbers here, it should be below 0.1
>Prior Loss Weight
>Regularization Images: For specifying regularization images, a method of supposedly reducing overfitting. Uncommon use.
these options are specifically for dreambooth-style datasets, prior loss weight is a multiplier of the regularization image loss

>>8616290
noobpoint5 bros?
oekakianon cultural erasure

>>8616298
this is why i didn't even want to make this rmao

>>8616300
are you scared of the truth?

>>8616300
Not him but do you see no benefit in discussion? Debate isn't supposed to be about making the other person look stupid or feeling superior but about putting our heads together to get closer to the truth. Some things you have there are accurate while others aren't. Why not update it to reflect his corrections so the whole rentry is correct?

>>8616299
I highly refuse to believe that anyone that isn't him is still using that version

>>8616305
pretty sure cathag uses it because all his images are washed out

>>8616294
>>8616295
>>8616299
Ok hold on bros I'm freaking out now and I have to make all my loras. What would be the difference between baking on base illu and vpred? Besides the settings. Shit my current bake just finished and btfo'd me too I'm so sick of this shit.

>>8616304
well more like everyone's an expert until it's time to sit down and write lol
i don't even have the code for it anymore, it's more of an AAR of my 1:1s than a proper guide, imho the most interesting parts of it are the fp8 stuff since you can say "well actually x works better for me than y" just like i did about actual parameters

>>8616300
What you can do is paste everything there into somewhere else so others can review and edit things around, have some overall revisions to then have the final "mostly correct/general agreed" current way to train a lora
Good initial effort tho

>>8616298
Nice,
I'm also surprised dropout causes quality degradation but maybe it's not as compatible with lora due to the low fidelity nature of how the lora values propagate into the resulting model on application
Another thing I'm surprised about it scale weight norms, I would've expected that to help counteract frying but I guess not
At the end of the day, I imagine the most impactful settings will be the dataset quality itself and sampler/optimizer settings anyways. What's the verdict on that and did your opinion on it change from whatever you had beforehand?
16/32 Locon AdamW cosine 1e-4?
How many steps?

>>8616311
>well more like everyone's an expert until it's time to sit down and write
Yeah you're not wrong either. I've had my fair share of this, specifically his brand of thing where everyone will say nothing for months only to show up and tell me I'm wrong and that "this has been known since 2008" when I finally do a write-up and post results.
>well actually x works better for me than y
Yep a lot of it comes down to just pure empiric testing which is why I'm saying your contribution is good. The more info we all have, the better and the less time we all collectively have to spend testing. Ideally if we knew *why* x worked better than y then every lora would improve, but just having some solid evidence is a good start.

>>8616300
If it's any consolation I appreciate you putting something together and having to collate info from the barrage of people claiming their way is correct way. I owe you a gobby.

>>8616298
>this is wrong, it just multiplies the loss with the snr curve
TBF I was going with the ezscripts definition.
Maybe it really should be put onto some doc for fixing up and reuploading.
>>8616321
I think the saddest part of this is that this field just doesn't do "why". Shit's a black box.
>>8616317
If you meant to @ me then yeah definitely, Locon was the straightest no subtlety improvement here, at least for a characters.
AdamW, like I said, instead of 8bit which logically (it is 8bit!) does make it slightly worse, I'm more surprised that nobody mentioned it before, but I bet a lot of it was just "waow less vram more good" during the early days when it became the "SOTA".
Cosine did surprise me because people recommend with restarts but yeah, the restarts version just kept more of a consistent "fry" in the images, especially relating to clothes.

>>8616298
>most likely you've used way too large numbers here, it should be below 0.1
Dropout is a method from ancient times. Only ML practicioners working on toy models or learning from old material actually use it these days.

>>8616331
or people experimenting
I find it makes the lora work better on further merges and finetunes
Also haven't noticed any quality degradation, and it wouldn't make sense for there to be any. Just slows down the training a bit.

>>8616328
oh yeah I replied to the wrong post lol
I guess it makes sense @ restarts since it cranks the learning rate back up when the chances you're at an actual minima during training in high dimensional space gradient descent is incredibly miniscule, meaning you're not even getting the advantage it's meant to provide in the first place.
How many steps do you usually shoot for?

>>8616338
I think I'm still team epoch but I have some diverse full bakes to do now to be sure.

>>8616317
>I'm also surprised dropout causes quality degradation
to be completely honest this is also my experience with loras
>scale weight norms
iirc it rolls back the weights that become larger than a specified threshold, this is basically only useful for stable training using high lr and small alpha, and even then you'd need to find an equilibrium
>I imagine the most impactful settings will be the dataset quality itself
as long as you're not training the model in the wrong way (by enabling noise offset for vpred/ztsnr model for example), this will always be the case, other things will only affect how quickly you train the model and possibly how much it will forget
>>8616328
>I think the saddest part of this is that this field just doesn't do "why". Shit's a black box.
projecting as is, it's just that educated people rarely visit 4chan's /h/.

>want a specific body type that a certain artist draws but want another artist's style
>use [artist1:artist2:0.6]
>it just werks
Man, I've been using these kinds of techniques for ages but never thought about applying it to this use case until now somehow.

>>8616349
sounds like me when I finally started blending hair colors. [red hair:dark brown hair:0.3] makes a nice auburn.

>>8616352
Oh yeah, I've been using it for color blending for quite a while. Also using negative sometimes to try and make it more consistent.

>>8616308
Are you baking for vpred though? That's the question.

>>8616359
So if I bake on illu and run on vpred will it really be that much worse? I'm running on 102d and all the epred bakes I did seem okay but after my latest bake I'm not so sure anymore.

>>8616342
I read up about it a little just now and the dropout making things worse makes sense. Due to the nature of adding multiple passes of gaussian noise in steps, diffusion models expects/requires a lot more consistency in terms of network output that dropout had no expectation of doing back when it was made and is counterproductive in regards to.
I'm inclined to agree with not using dropout now and just trusting that the layer/group norms in SDXL should be doing its job (not that I know for sure they're used in ezscripts lora training if anyone wants to confirm)

>>8616362
It's most visible in stuff like inbuilt styles on non-vpred character loras. You just get way less accuracy, and a lot of greyness because of the incompatibility. It's not bake ruining but it's worth comparing.

Actually, "OP" here, I was thinking, since I use full shuffling now, would adding an artist tag for a style bake still make sense?

>>8616369
Eh fuck it I'll bake twice and be empirical about it kek

>>8616369
No there's no point.

>>8616369
I'd imagine it wouldn't hurt, especially if there's some vestigial knowledge of the artist from danbooru so as to nudge that knowledge to the surface
If it wasn't in danbooru/after cutoff (or if the model native artist tag just fucking sucks), I'd say may as well leave it out since style loras seem to work fine without tags, why add another piece of overhead during runtime

>>8616369
I've done it with the "3D" tag once by accident and it still worked like a trigger despite shuffling. Tag order doesn't seem to matter as much on illu/noob anyway.

>>8616391
3d tag is a super special case imo.

>new config also converges styles way faster than old one
That old one needed like 10 epochs more than for characters, I guess it was all the dropouts.
It's kinda funny how much it can change depending on how you test it, but I think the bulk of the old setup I grandfathered from Pony.
I'll test it a bit more and probably post it tomorrow.

>>8616385
>>8616391
Hm yeah it's still needed.

>>8616328
>Cosine did surprise me because people recommend with restarts but yeah, the restarts version just kept more of a consistent "fry" in the images, especially relating to clothes.
ime restarts do help but it depends on dataset size and lr and probably a lot of other stuff. i did some tests a while ago and settled on one restart for every 2400 examples of a concept in the dataset at batch size 1.

Coolio, I'm gonna go on a bake binge soon.

any videos i can use to learn?

>>8616454
This is a reading hobby, sweaty?
Unironically people here would send you to the research papers back in the day.

>>8616451
エロ

>>8616454
nyot really
the people making vid tutorials are making them for the common jeet denominator

>>8616454
depends if you have a programming/maths baseline
if you don't, lmao, go follow karpathy course and good luck
if you do, having chatgpt break down the components of sdxl for you until you get to the level of technical detail you're satisfied with isn't a terrible idea.

>>8616454
learn what

>>8616454
If you're completely new, then go ahead and search for any video tutorial on YouTube. I started learning by watching those and then reading some guides from the OP a year and a half ago.

>>8616457
More like

where's the nai gens?
im tired of localslop

wrong thread, chud

where's the ironic posts?
im tired of genuineslop

Best way to improve face quality?

>>8616530
make it >30% of the image

>>8616530
>inpaint
Forget that. Use adetailer on a hires pass and type in the face details at 32 padding.

Weird, a couple months ago gens would make me diamonds. Now they barely do anything. Are the models getting worse, have all the good genners left or has my taste just drifted?

>>8616715
Unironically take a break, gooner.

Any other shitmix testing purveyor want to run this model through their saved prompts? I've been awfully impressed with some of the outputs albeit it's not without issue. Curious if it's just good gacha on my part or what.
https://civitai.com/models/832573?modelVersionId=1677841

>>8616298
>Multiple batches allow for quicker training but also exhibit problems with symmetricality in the end bake. Keep at 1.
>what? sounds like an issue with bucketing
How

>>8616817
I refuse to use shitmixes, if I want to fry my model with shitty loras I can do it myself without loading a whole new checkpoint.

k

>>8616817
What was it that impressed you? I threw a bunch of prompts at it and didn't see much of a difference. Didn't even include 1+29 or 102d custom because they're even more similar to base v-pred.

I think bottom line every merge is just some mix ratio of illu+noob with diluted knowledge and a different base style. As soon as you prompt an artist above 300 pics or add a lora, all the checkpoints sort of drift together.

>>8616715
>he was jerking to other people's gens
ishiggydiggy
>>8616817
I wouldn't say it's better than 102dcustom but it's better than most shitmixes I've tested, primarily regarding inbuilt artist replication.

>>8616912
kek

>>8616715
whose gens did you like most bro

>>8616715
You kidding me? Look at this shit >>8615738 its getting even more amazing, and apparently now you can animate them too? Things are getting better and better.

>>8616910
My go-to's are 291h and custom. Flip between the two as I find 291h to be soft (good or bad thing depending on the artist mix) but have more artist fidelity whereas custom has more color depth but can overpower the style of certain artists.
This shitmix I stumbled on is giving me a nice in-between of both those models which is ideal. But it wouldn't be the first time I find a model I think has potential and then dump a week later.
>>8616912
>primarily regarding inbuilt artist replication
Yeah that's what I liked about it. Gives me the fidelity of 291h but with more color depth but not diving into fried or blown contrast territory.

>>8616532
>and type in the face details at 32 padding
Sorry I'm retarded. Can you explain what this part means?

>>8616530
higher base res and inpaint

>>8616936
nta and he might not be too smart either since Adetailer is the same as inpainting.
But anyway when inpainting change your prompt to describe only the face (and style) not the whole scene. Adetailer gives you a separate prompt window to simplify this. Mask padding is in the settings.

>>8616936
In the adetailer prompt type the info for her face plus your quality tags, artists tags, lora etc. I organize my prompts so it's all easy to copy/paste. For your settings put the resolution to 1024x1024 and 32 mask padding at 0.4 denoise and faces should come out much better.
>>8616939
You're so retarded since I'm telling him to automate his inpainting rather than doing it manually. You've already conceded the fact that they're the same so what's the issue?

>>8616936
booru tags that describe said character's face.
>yellow eyes, short eyebrows, scar across eye, etc
32 padding is the area the model will derive info from to change your masked zone. 32 or 64 are usually safe bets along with soft inpainting to eliminate any inconsistencies in the masked area.

>>8616940
The post made it sounds like you thought adetailer was somehow superior to inpaint beyond just finding the face automatically. A misunderstanding then, thought at least we clarified for the newfag.

I would apologize for calling you "not too smart", but now you insulted me even worse so get fucked.

>>8616947
Based grudge holding anon.

>>8616947
Basic 4chan discourse I'm afraid, but I will apologize anyway since I see my insult was unnecessary.

>>8616949
Based bigger man anon.

>>8616235
>https://rentry.org/hgg-lora
were you the one who was empirically testing various settings and posting those grids?
if you are, could you share the collection of grids as well?
would like to see if i would draw the same conclusions

>>8616958
mmm nyo (i deleted them)

shitty nyogen impersonator...

How are you guys using BREAK?
I feel like the more I use it, the less I know. Are you supposed to organize your tokens by importance so that important things are early in each prompt segment (ie masterpiece, best quality are not in the same segment, but split into different segments as the starting tokens)? Are you supposed to group by concept, so like you put all the clothing tags together, all the background tags together, etc? Or perhaps do you group the tokens by visual proximity, so perhaps tags close to one subject get a group, tags close to each subject's faces get their own groups, etc.

>>8617086
i'm not

>>8617086
I don't. Some have said that it resets unet or something so that the very next token after BREAK gets full importance which can be helpful but I don't really use it. Organization is done with pressing enter rather than BREAK.

>>8617086
Specific use snakeoil which doesn't work the way you are intending to use it.

>>8617086
>How are you guys using BREAK?
The only legitimate use case for it is to prevent getting tags split between blocks. For example, if you are at 73/75 tokens and have "foreshortening" as your next tag, it would be good to use BREAK, so the "forshorten" and "ing" tokens don't end up in separate token blocks.

>>8617086
I'm not

>>8617098
this
everything else like "separating characters" is pure placebo

>>8617103
Isn't that also how you separate adetailer face prompts, or regional prompter areas? I don't use webui/forge, just thought I read that somewhere. Might be where their confusion comes from.

>>8617108
I mean technically yeah, but you're changing its function when you use it with extensions like regional prompter. By default it simply separates clip chunks which is only useful to prevent tag being split into separate tokens

>>8617108
Woah woah woah. You can separate adetailer face prompts? With regional prompter?

>>8617098
This isn't even happening in modern UIs though? If you don't put BREAKs, they only split by commas, never in the middle of a word. Well I only tested forge/reforge as they have convenient token counter, maybe comfy does it like you say.

>>8617111
>You can separate adetailer face prompts?
no, adetailer has its own syntax for prompt splitting and uses [SEP]

>>8617114
Right, sorry I mixed those up.
>>8617111
If you have multiple faces in the pic you can give them separate adetailer prompts and it goes through them all, left to right I think. It doesn't use regional prompter, just inpaints them one by one.

>>8617086
I haver never used BREAK under any circumstances

kusujinn https://files.catbox.moe/46yjxk.safetensors

>>8617150
Thanks? Can't image how it'll look, he's gone through like five different styles.

How many images do I need for a character lora?

>>8617154
>>8616468
>>8616451

>>8617157
depends but it can work with as little as like 10

>>8617157
Fewer the better to get a strong style, it just won't generalize well without a varied dataset.

>>8616715
Had the same thing. In my case I was trying to prompt increasingly complex stuff and never noticed the subtle decrease in quality and frying step-by-step. Going back to simpler less convoluted prompts fixed it for me.
https://files.catbox.moe/qqxvkl.png

>>8617188
Did you train this on a merge? It looks completely different on noob. Not in a bad way.

>>8617245
Also I can't tell if "kusujinn" is supposed to be a trigger prompt or it's just picking up the model's existing knowledge from 150 booru pics.

>>8617245
Nyo u can see it's baked on 1.0
>>8617248
It is

>>8617245
Is that a character? If not what is the prompt for that hairstyle? Parted hair, wavy hair?

>>8617251
>It is
working with noobs existing artist tags is inadvisable in my opinion. They are often misaligned and overtrained. It's possible to somewhat fix them with te training, but you're usually better off training a new tag from scratch or training the style into uncond.

>>8617251
Okay, that really brings out the /aco/face, thanks

>>8617254
Okumura Haru
I thought she was about as mainstream as it gets

>>8617255
Unless they're not misaligned. Then it's a huge benefit and you only need a little bit of training to drive it home.

>>8617266
I haven't played persona 5 nor any persona game because I was never interested. Worse, the mainstream nature of the fifth game made me lean out not in. Maybe I'd like it if I tried it.

Seems like the anon who was finetuning a vae for noob back then is now responsible for neta's lumina 2 bake? https://huggingface.co/heziiiii

>>8616349
Thanks for the tip, I should experiment more with prompt editing. You can probably replicate a lot of controlnet stuff with it.
https://files.catbox.moe/kkzluz.png

>>8617255
I mean I tested it 1:1 and it was better with, meh.

>>8617279
Is this good or bad news?

>>8617302
I dunno, it's interesting, maybe he could give us some cool insider info

>>8617335
>deadsd

oh yeah i was gonna post that toml
it converges roughly around the 11-13 epoch but it's safer to keep 15
https://files.catbox.moe/a22hr0.toml

>>8616349
>>8616352
What does [a:b:numerical value] mean?

>>8617370
>What does [a:b:numerical value] mean?
tag "a" is in the prompt for "numerical value" percentage of the steps, after which, tag "b" replaces "a".

i.e. [cat:dog:0.5] - 20 steps
for the first 10 steps, cat exists in the prompt, at step 11, cat is replaced with dog
hope this helps bwo

>>8617378
Oh, perfect, thanks anon!

>>8617356
is it for style or character?

>>8617397
welcome bwo

>>8617439
both

>>8617455
Hot. Catbox?

>>8617477
Don't worry I know.

>>8617477
stealth meta bwo
still testing the lora tho

>>8617455
i absolutely hate how this only looks crisp if you don't open it in full resolution

If that's fuzzy then my pics are hot garbage holy shit.

>>8617510
more like it's a vae problem

>>8617512
Which vae are you using?

>>8617510
some prefer their gens mushy and blurry while others want digital art level of sharpness, at this point it's just a matter of taste, (You)r taste

>>8617515
same blurry shit as you most likely

>>8617524
fixFP16ErrorsSDXLLowerMemoryUse_v10?

>>8617527
>fixFP16ErrorsSDXLLowerMemoryUse_v10
isn't that just the fp16 vae

>>8617271
you're not missing out on anything good, it's shit megami tensei for the bottom of the barrel r*dditors

>>8617464
I thought they were supposed to be different

>>8617534
that can happen but nah it works, i baked a style and a chara and it just werks
about to rebake derpixon and skuddbutt out of curiosity

>>8617356
oops i forgot to add a subset with the shuffle captions

>>8617529
vae_trainer_step_90000_1008?

>>8617543
oh ffs and i baked two loras including the kusujinn without it because i forgot
lmaoo
i need to rebake them

>>8617546
here's the fixed one lel https://files.catbox.moe/uvldis.toml

>>8617544
this one?
https://huggingface.co/heziiiii/noob_vae_test/tree/main

>>8617548
Yeah those are the only two weird VAEs I've seen people use. Otherwise I just use SDXL.

>>8617535
Sweet, i'll test then I need to rebake a character

Okay bakers how much dim should I use? I heard someone say something about overfitting on style with high dim or something. I don't remember this being a thing.

>>8617279
For fucks sake, can someone please tell him to finetune flux's VAE? I don't want lumina 2 to come out with the artifacted, washed out, biggest shit that flux's VAE has, I really want the next gen model to be as less fucked as possible
Please tell him

>>8617651
Post a VAE comparison. However bad Flux's might be, SDXL's is 100x worse.

>>8617657
You can't use SDXL's VAE on lumina, how would that comparison even work

>>8617651
This might be his civit account https://civitai.com/user/li_li/images all I did was search the noob discord for hezi and this name popped up

>>8617657
nta but
>However bad Flux's might be, SDXL's is 100x worse.
not an excuse to not try to mess with it just to get better details, and then a better model. in fact, this is the best moment to do itsince lumina 2 isnt finished so there is still time to improve the vae

>>8617661
https://huggingface.co/spaces/rizavelioglu/vae-comparison

>>8617672
Oh that's pretty cool, thanks for showing it to me.

>>8617455
interesting how this one came out with less paper/canvas texture on the outlines, I kinda liked that effect

>>8617647
I use 16
and I used 16 on my last config too

>>8617647
don't remember this being a thing with what? SD1.5?
Dims are much different because dims don't exist in a vacuum if so, it's proportionate to the actual model size. a 16 dim lora on SDXL is much more "powerful" than a 16 dim lora on SD1.5 hence more sensitive to frying as larger dim values are used.
You probably don't need such a big lora (could just also train/resize 32 dim down to 16 and see what method you like) unless you're training many concepts, at which point just finetune and extract.

>>8617631
just remember to shuffle the captions unlike me

>>8617736
>at which point just finetune and extract
Any rentrys for this?

>>8617546
>retrained two of them without it again
am i fucking retarded or something

>>8617797
that's literally me

>>8617810
I like your white hair.

>>8616715
I've been working on getting a mix for a simpler style but something about the faces/bodies just lack the "intensity" to get me going.
I may have conditioned myself to needing prominent toned belly/ribcage/hipbones or else I won't see it as erotic

>>8617820
>but something about the faces/bodies just lack the "intensity" to get me going.
it's called "context"

>>8617848
you're right chief, I should go back to just making unquestionably rape/ryona pics

/hgg/, slopposting yea or nay?
https://files.catbox.moe/s38rup.png

>>8617909
as long as you don't spam it should be fine, post whatever you want just don't over do it

>>8617820
Idk just seeing my cock slide into her wet pussy in pov always gets me. Bonus points if it's a waifu from one of my character cards.

Can someone help me bake this lora? I'm not sure what I'm doing wrong.
https://litter.catbox.moe/o0vbte3jla83xfx1.rar

>>8617944
>these two feet shots
https://files.catbox.moe/flhv1q.gif
but if you mean for a style lora, give it some tag, and prune all those style descriptors like "anime coloring" because it just dilutes the output

>>8617947
No I don't even care about feet but that anime has really good nails for both hands and feet which I wanted for the lora. I had added those tags because I thought it helped but I guess not? I'll try training it for longer.

>>8617949
i think styles generally do need "trigger" tags for "concentrating" them, never had much luck post 1.5 without them
and yeah those tags would just dilute that baked in tag and pull on the pretty strong inbuilt knowledge
i can bake it tomorrow out of curiosity, just say what trigger tag you want

>>8617953
>i think styles generally do need "trigger" tags for "concentrating" them
Oh yeah? What I noticed back in the pony days was that some datasets needed them but others didn't. However I haven't had much issue with styles at all until I attempted this and I've been having trouble with it for weeks.
>just say what trigger tag you want
I doesn't really matter I just need a proof of concept and metadata. Some faggot (who's probably here now) was showing off his lora but when he posted the lora he cut out the metadata so I can't even see what he did.
https://files.catbox.moe/z4p83a.png

>>8617738
oh yeah good call
no noise offset tho?

hello, I've been using NoobAI-XL-Vpred-v1.0+v29b-v2-perpendicular-cyberfixv2 as my daily driver, snake oil be damned, is there anything "better" that has been released since or maybe a paradigm shift that popped up overnight that I missed? Thank you.

>>8618044
Is that the one with the shit backgrounds? We've all moved on to 102d custom bro.
https://civitai.com/models/1201815?modelVersionId=1491533
>he recently release 2.5d boost
Interesting.

>>8618044
>NoobAI-XL-Vpred-v1.0+v29b-v2-perpendicular-cyberfixv2 as my daily driver
it's not that bad, you don't need any snake oil with that one, it does suffer of very noticeable artifacting with some styles tho
As >>8618045 post, we use 102d custom nowadays for a better and more easy time

>>8618045
>>8618048
Thank you, I'll try this model out.

>>8618052
the euler cfg++ kl optimal 28-32 step @ 1.5cfg settings they suggest give pretty neat outputs for hires pass, just don't use rescale cfg with that if you try it

Artist mix is settling out nicely. Though it has a tendency to make girls cute and funny, I guess I did sign up for that with the artists I chose and it's not like I have reason to post that often

>>8618117
Nai is looking good

it's local

>>8618138
luv me anime pubes

i hate pubes

>>8618165
fluffy pubes get a pass

tufts and fuzz are acceptable, it's those overexaggerated pussies with flaps and individual strands going everywhere trying so hard to be "realistic" that it shoots right past is where it gets appaling

>>8618161
yeah nah

>>8618193
I don't provide metadata so you just gotta believe me bwo (I guess I base gen at 896x1152 too if that even matters)

>>8618048
Cute Vibes. Cute Lize.

>>8618045
>Interesting.
>This 2.5d boost model provides a model that deviates from flat 2D to a slightly 2.5D orientation.
Why do they do this? Reminds me of chromayume, for whatever it's worth, started off as flat and then ventured into 2.5 slop.

>>8618270
hey look, two cakes
it's not much effort to just merge in more custom udon for another model. variety is the spice of life

>>8617700
>paper/canvas texture on the outlines
that effect was due to mistagging (lack thereof) motion line tags on some images in the set with finer, less noticable, motion lines drawn near the outlines.
missed out on tagging a couple again which were in another folder, so back to the kitchen with this

>>8618317
i like that the penises are small

>>8618208
Feels like wagashi but more generic.

>>8618330
yeah, it indeed started out with trying to find settings to make wagashi + a wagashi lora play nice with noob. For some reason, my setup would always fry body parts with it

>>8618040
hm? no

>>8618320
b-bro that's average...

>>8618363
oh nyo nyo nyo nyoooooooooooooooooooooo

what's the difference between this and hdg? I haven't been here in months, where do I post sex gens?

>>8618396
if you have to ask, you're meant for /hdg/

>>8618396
This one's constantly near-death, so shitposters see it as too much effort for no payoff. Not worth it imo, now I'm stuck having to refresh five threads instead of 4.

>>8618396
less shitposting, less gens and more technical talk

>>8618403
>>8618403
I just wanna use hentai gens to make e-girls perform things they can't or won't in real life.
like nigri, lyumos, katz and other cosplayers getting fucked out by tentacles.
it's challenging indeed for brainlet coomers like me.

>>8618412
>3dpd
may your journey to other boards be swift and final

>>8618396
Until a janny decides to start enforcing the shitposting rules on /hdg/ it's unusable to me. Maybe 1 out of 10-15 posts is genuine these days, if you exclude the spam. This thread might be slower, but it has 100 times less of the cancer that's now in /hdg/.

>>8618396
>this bait again
On the 1% chance that you're serious, just look at the previous /hdg/ thread

This worked surprisingly well. Although it does have a bit of a darker bias because of the 1st ep.
skuddbutt https://files.catbox.moe/nivhgd.safetensors

Ever since anons suggested to run without negatives my gens have noticeably improved on Noob models, so I'm pretty convinced on that front. But what about quality tags? I see some anons using [<quality tags>:x] or even [<quality tags>::y]. Anyone experimented with what works best?

"curvy" tag my beloved
>>8618438
I found them to have little to no actual effect. Maybe it has some meaning if you run no artist base model but like... why?

>>8618441
No effect? They have a crazy strong effect even when using artist tags for me on chromayumeNoobaiXLNAI_v40.

>>8617963
>when he posted the lora he cut out the metadata
I guess you're talking about my basedbinkie lora? I can give you the toml if you want but anon's config here: >>8617547 is probably better, I was gonna use it myself for my next lora

Reminder to never leave your prompts with a hanging tag. Always have a comma at the end.

>>8618451
wait why? is not having a comma at the end that impactful?

>>8618447
Chromayume is pretty close to base model.
Well, I guess I meant positive effect. Even in your pics they kinda screw up the look and anatomy imho. When I was testing it myself it was pretty minimal.

Rebaked Kusujinn
https://files.catbox.moe/ridycu.safetensors

>>8618460
as for the difference between shuffling captions or not, the rightmost one was not shuffled
it's pretty minimal but maybe the flatness of the shuffled ones corresponds to the style a bit more? who knows at this point
https://files.catbox.moe/cyhpct.png

>>8618438
It's not that negs are bad, although maybe they are. It's that there's a technical quirk with noobAI, where just having anything at all in negs causes quality degradation versus leaving them completely empty. Even if it's just a single letter, or an underscore. The effect does partially go away with merges but it's easily visible on base noob.

>>8618463
It's not NoobAI as much as reForge. Pony can get a similar effect. I think people just didn't see because everyone needed negs back then. Or maybe the dude changed the backend between Pony and Noob because it is caused by how it processes the uncond.

>>8618466
Made me load up Pony again and check. It also goes away with merges, and everyone used source_pony in negs so I guess we never noticed.

Thing is A11/Forge/reForge pass the uncond as empty, instead of encoding an empty string and passing the output of that.

>>8618474
>Thing is A11/Forge/reForge pass the uncond as empty, instead of encoding an empty string and passing the output of that.
Then, why does an empty uncond look better on SDXL models?

>>8618460
>>8618434
thanks anon(s)

>>8618493
we just don't know.gif

>>8618493
It would be kinda funny if there was some other weird quirk that was handicapping outputs. The reForge/Comfy/Classic output differences show that it's not always the exact same thing depending on how you run the model.

>>8618208
>I don't provide metadata
Why are you proud of this?

>>8618516
seethe prompt thief

>>8618450
Did you make the lora for the pic I posted? It's not really about the config, although that matters, it's about the number of pics used, the tags used, etc. Having all the information is best but since dataset is 60% of what makes a good lora, I'm still pretty much in the dark without any idea of how he did it. Also:
>>8617944
>>8617944
>>8617944
Any takers? It's a style lora that's already tagged. I thought this place had at least 10 bakers lurking around.

derpixon https://files.catbox.moe/9ws6sb.safetensors
i tried tagging the characters and herzha forms but it doesn't really want to work
>>8618531
i have it baked i need to test it

>>8618524
I'm seething that /hgg/ has started tolerating retarded attentionwhores not that I care about the metadata for artists I'm already using.

>>8618516
Consider post natal self abortion. Nobody owes you shit, mouthbreather

>>8618537
Go back faggot. I don't want you shitting up /hgg/ too. We're not your xitter fanclub.

>>8618539
just quickly posts some tired bait in the other thread and he'll be occupied for a while

>>8618531
>Did you make the lora for the pic I posted?
Nope, in that case I have no idea what you're talking about

>>8618539
Fuck yourself you worthless cunt. If you don't like people posting gens that's your problem. Again, consider suicide you fucking dipshit.

na hours

>avatarfagging doesn't exist
>desperately trying to look cool in front of strangers isn't a thing
>don't mind me I'm just posting gens

i like asking for boxes bwos
sometimes i find some nice artists to train a lora for

>>8618550
it's one of few things this general is good for still

>>8617547
>RuntimeError: quantile() input tensor must be either float or double dtype
just errors out for me on the fork lol

>>8618560
eh i kinda like the technical discussions
but yeah sharing boxes is nice and comfy, who cares about the grifters

>>8618573
If you don't want to share it's fine, but then why post here? Just keep it to yourself and enjoy your super secret recipe. In reality every other intelligent person from the SDXL creators to the forge/comfy coders could have said "I don't owe you anything" and used the tech privately but they didn't, yet your insignificant contribution is the thing that belongs to you? Retarded ladder pushers should in fact be shamed.

>>8618343
Did some test and all the output were fried, so idk what was wrong with mine

>>8618577
wrong anon bwo, i do share... all my gens have stealth like >>8618550

>YOU MUST SHARE OR DONT POST AT ALL
what the hell? lmao

Yes, still not sharing metadata of my insignificant gens. If they're so insignificant why do you care

>>8618580
I was speaking generally not to you specifically.
>>8618582
>>8618583
Yes faggot. Why do you have no argument? Because your position is completely indefensible.

bros why are the hdg rapefugees still here the thread got rebaked
and i know it's you because you posted the same style in hdg too

>begger thinks he's contributing by whining and shaming image posters
it's time to go back

>attentionwhore thinks he's contributing by posting pictures where he's not wanted
Go back. /hdg/ is the perfectly place for you. You don't need to be here.

ywnbaj, this isn't your thread chud. you can't police what people post

>47:27
>49:04
>50:43
jej

>>8618531
retagged and used kotonoha for a trigger https://files.catbox.moe/n125j8.safetensors
i don't know if this is what you wants but it is a style lol
i think ai always smooths out the bloomy atmosphere if that makes sense

>>8618598
Can you post a catbox so I can post a comparison pic? Thanks for baking it btw.

>>8618601
https://files.catbox.moe/sprs7g.png

>if you post a picture without metadata youre attenionwhoring
rofl

Any other anons obsessed with horror?
https://files.catbox.moe/tnmkhn.png

>>8618623
Not obsessed nor I do really like terror but I do like to gen unsettling gens from time to time

>>8618598
>>8618609
I had to change some stuff because the lighting wasn't coming out good for some reason. Either way I think these pics showcase the differences.

This one is the original lora. You can see how the overall style is very close to the screencaps but also how detailed the desks/walls/window are. The buildings too. However his lora always has that backlighting/side lighting effect on all pics which makes me think he just used like 20 pics he found on a wallpaper website and called it a day.
https://files.catbox.moe/l1ljla.png

This one is yours which is close but doesn't quite have the detail of his, especially with the skin and the way the desk/window/buildings look.
https://files.catbox.moe/blrns0.png

This one is mine. Idk why it's doing this lighting thing. The desk/window is a bit closer than yours to the screencaps but mine looks overbaked. I don't understand. My config works for 90% of datasets but now I'm repeatedly have issues.
https://files.catbox.moe/oiq1v5.png

Is this your config? >>8617547

>>8618623
Oh thank you for reminding me. I enjoy horror gens with nakamura regura but yeah I should make some more of them.

>>8618634
>Is this your config
ye

>>8618634
I wouldn't be surprised if he overbaked on some shitmerge and that somehow made it better
I know the old N64 lora that was bretty kino was baked on AOM3a or something

https://files.catbox.moe/c4v4oe.png
>>8618629
I think stumbling upon a the ring porn parody when I was young was what did me in.

>>8618637
Regura is nice, should make another mix with them included. Karasu raven also makes some real nice monster girls, but they desperately need a lora for noob.

>>8618643
Hmm I guess this is worth a try. He was using this model for his gens.
https://civitai.com/models/1442151?modelVersionId=1732221

the urge to bake a lora on those weird ass old mugen hentai animations

oh yeah i also gotta retry that cursed game cg lora on this config

Whats the difference between Hentai Generation General and Hentai Diffusion General

>>8618663
waow, the 2006 was awesome!!! https://files.catbox.moe/cadhtw.jpg
>>8618673
the middle name

>>8618623
>be looking at scraped cgs
>stumble upon this https://files.catbox.moe/sfqyrf.jpg
uh i think you'd like this

>>8618692
So uhh... I guess that's semen on the walls then?

>>8618462
catbox for some of the images pls? i can't get the exact same style as yours with the kusujinn lora

Fugtrup https://files.catbox.moe/an57lq.safetensors
It can kinda work natively but it's better with
>>8618730
Semen On The Walls is my new band name
>>8618740
Don't have those but I prompted it like this https://files.catbox.moe/0bpfaq.png

>>8618516
not necessarily proud of it and was just preemptively addressing it if it was thinly veiled metadata bait. I'm not sure where you'd get the idea in the first place besides projecting/boogeyman but I can make up many reasons to not provide metadata
1) Spent hours figuring out what slight modifications to add to mix to get it to play nice and I'd prefer to not see my efforts end up being used in questionable subject matter by someone with what I would consider abhorrent tastes.
2) Workflow is incredibly schizo and identifiably mine. Ironically enough this reason is because I specifically want less attention because it's low hanging fruit to pick at that could easily follow me across styles if I always posted metadata.
3a) To spite you specifically
3b) It's closing in on a plausibly deniable style concerning which artists it's copying from, so it's a candidate for maybe reviving my xitter/pixiv accounts :^)
4) Mercury is almost in retrograde

gay

based

kek so it was to shill xis twitter

don't worry skinny-tits-from-above-kun, your gens are already extremely identifiable even if you don't post any workflows :)

>>8618771
My point is that a general should be a collaborative environment and there is no collaboration without sharing info whether that's pic metadata, lora metadata, configs, controlnet settings, etc. Outside of the social aspect, I don't see any purpose in having a general. We don't have themes, contests, challenges, requests, or anything else so that just leaves the typical attention whoring you'd see on /trash/ doubly so if you're posting porn of all things. Regardless I don't want to shit up the thread more than I already have but at least you're reasonable.

>>8618784
Uh no, seethe more style thief

>>8618784
Yeah that's a fair place to be coming from. I figure if there was a sanitized /b/ thread without all the wack shit/toddlercon, I'd be inclined to post there.
Mostly I've just been throwing in 2cents @ the lora training stuff here since I have a basic understanding of the underlying math and may have something to add beyond "empirically, this is observed" like the discussion over dropout. I figured I'd post some images to add activity to the thread since the most common comment concerning the thread is that it's "too slow"

>>8618782
>skinny-tits-from-above-kun
underboobless-kun ;)

https://files.catbox.moe/ccpe58.png
>>8618692
Heh, funny since I did try to gen some fatal frame girls, sadly they don't work natively. I tried making a shiragiku lora that never turned out well, but maybe I should try to rebake.

>>8618782
>>8618801
valid even if potentially inorganic, I wonder if something I'm using is overfit for that.
I put bouncing breasts in the negatives earlier because something overfit for breasts looking like there's an unnatural pressure on them on an earlier version and this reminded me to remove it at least, thanks

>>8618663
i still don't think i can get it to work lmao

>>8618827
Yeah it's bullshit but the civitbros might have discovered something that I need to test.

>>8618771
just curious but is this a wagashi with worldsaboten mix?
the hair highlights and lineart reminds me of wagashi but the eyes and face in general remind me of that worldsaboten lora

did you know when you're training with batch size below 32 you're fucking up every other batchnorm layer

>>8618623
dark theme, bleak ambience, dark persona, evil smile is a powerful combination

>>8618835
funnily enough, I also thought I saw cactusman in it so I added the lora at one point and it instantly made it worse, so it's not in anything I've posted.
eyes in that one are from the kindatsu lora on civitai at low weight interacting with the other artists I'm using

>>8618834
well i am the guy who made that config
the dataset is just cursed

why are zoomers so obsessed with fish? are catgirls seen as a boomer thing now and this is their attempt at counter culture?

weird windmill to fight desu

Ellen Joe did nothing wrong.

think it's a coincidence there were two highly visible highly produced sharks in the past few years, the latter may be caused by the former via usual trend chasing
zoomers would've bought into any theme'd girl if it garnered enough social media attention

>>8618879
what about the kraut orca?

>>8618771
>3a) To spite you specifically
based

>>8618867
they should like crocs instead

Anyone know if there's a tag for two-tone clothing in which the front and back are different colors, rather than say the bottom being different from the top, or the colors being in a more striped pattern. In biology this happens and is called "countershading". This does exist as a tag but not with many examples. Is there a term for it in clothing?

>>8618652
Funny that. I like to test rando shitmixes and recently tried the v3 one of this. Wasn't impressed.

>>8618747
>Fugtrup
I'll give this one a try but feel fugtrup stuff works best with Pony or those 2.5/3D focused noob shitmixes.

>>8618523
She is cute. Do some non-/h/ with her sometime

>>8619004
All 2hoes are whores, I can't really picture myself doing something non-h with any of them

Does base sd-scripts have any annealing schedules?

>>8619010
The internets has damaged your mind.

>>8619022
The real reason is that I don't think the internet needs more 2hoe images, I would rather do more of my cute obscure gacha wives tbqh

>>8619023
OK that's a good one.

so the adafactor finetune config on the rentry leaves me with around 5gb of spare vram. How do i snakeoilmaxx with that?

>>8618747
>>8618973
I thought fugtrup works inherently?

>>8619023
>I would rather do more of my cute obscure gacha wives
Unfathomably based.

>>8619052
Nah. You can kinda prompt engineer a bit with tags like 3d, realistic, blender, etc but unless your model is already slopped, it's hard to replicate faithfully.

>>8619023
the true worth of AI gen
found out someone did a good train for my VN waifu on civitai with all outfits god bless

>>8619023
>The real reason is that I don't think the internet needs more 2hoe images
this is loser talk.
you are a loser.
the internet needs more 2hu not less.

>>8619052
>>>>It can kinda work natively but it's better with

>>8619023
What about not-so-obscure Vtuubas?

>>8619111
Those as well of course

>>8619111
>not .gif
missed opportunity

fun
not something i'm gonna use every day but fun

So like, does anyone have a sense for why exactly the model doesn't simply just perfectly know how to render anal_tail? It can definitely do it, but sometimes it wants to make the tail more of a plug or vibrator, sometimes it doesn't even place one, sometimes you get one tail and one object in the anus instead of them being the same thing. There should be way more than enough samples in the dataset. Is this just a really hard concept to get for the current amount of parameters?

>assless panties have 700 posts on danbooru
>surely it must work
>prompt for it
>it works but makes the character topless
>prompt for the character's clothing explicitly
>the panties turn into normal panties
It's all so tiresome.

>>8619158
have you tried "backless panties" instead?

>>8619158
Meant backless panties not assless. It's an alias so I was thinking it while making my post.

Ohhh is *that* who this shitposter is. I should have known.

schizo gen is over there if you're into pointlessly prolonging that kind of activity

what is lil bro talking about

all me btw

Why are we so dead tonight, /hdg/?

geg?

/hdg/?
and it's the weekend, go outside

>>8619271
no more than usual
if you want action go to the shitposting thread

>>8618550
I assume you are the fabled bwoposter I was referred to, could you help? >>8619527 Thank you.

>>8619038
increase batch size or train text encoder

>>8619579
hi bwo, will post it in a day or two, currently baking so i've not got the vram to make previews

>>8619640
Can't wait, thanks again 'bwo.

>>8618048
>102d custom
Could u suggest other models that are equally good as this but for realistic (dandon fuga, sakimichan, zumi) and 3d (fugtrup, slash-soft) style?

>>8619688
I wouldn't say custom is that bad at them, picrel
A lot of styles just need a higher base res and or lora to get them fully correct, that goes for any artist

holy mother of /aco/

>>8619605
both of these don't really increase quality and it converges fast enough already.

>>8619695
>>8618840

blackpill on this? no jeets as authors for once.

>>8619696
why?

>>8619699
the inputs become way too noisy, batchnorm layers will fit to your dataset extremely quickly. actually, they will even if you use large bs, so the best solution is to freeze the batchnorm layers, this way you even you would even have more free vram

>>8619697
It doesn't really do anything good for the model, at least on my end.

>>8619707
>this way you would even have more free vram
shit

>>8619697
>the virgin wavelet

>>8619707
>best solution is to freeze the batchnorm layers
so is there an argument for this or do i have to go on a vibecoding adventure? are batchnorm layers included in the usual train norm?

>actually listening to 4chan advice without images
ishiggydiggy

>>8619711
some trainers may let you select the trainable layers, but i don't think it's implemented in sd-scripts, or at least in the finetune script
you can try hacking it into https://github.com/kohya-ss/sd-scripts/blob/a21b6a917e8ca2d0392f5861da2dddb510e389ad/sdxl_train.py#L52

>>8619714
worst case I waste some electricity, best case I get better loras

Fuck my lora didn't bake last night. How do I install both python 3.9 and 3.10? Installing one always breaks the other.

>>8619720
Getting uv is the easiest.

>>8619718
So I didn't find any references to batch norm in the unet library and gpt tells me that the groupnorm32 layers that are used instead aren't dependent on batch size.

>>8618840
i dont think sdxl uses batchnorm since their vae uses groupnorm

>>8619727
>>8619729
>groupnorm32 layers that are used instead aren't dependent on batch size.
yeah that seems to be the case, actually. sdxl uses groupnorm
i swear i've seen batchnorm somewhere in sd though, maybe it was during early lora development back in sd 1 days..?

>>8619720
you'll need a virtual environment.

https://www.freecodecamp.org/news/how-to-setup-virtual-environments-in-python/

https://files.catbox.moe/sd0srl.png
>>8618841
dark theme and bleak ambience are working tags? I've used black theme before but never dark theme. I usually only use the "horror (theme)" + "dark" combination.

should I just not bother with activation tags for full finetunes? it looks like most of the style makes it into uncond anyway

>>8619813
Would you say that all style loras need activation tags or only some of them?

>>8619814
I feel like it's mostly a preference thing and you can get good results with both, regardless of dataset.

>>8619815
I feel like some datasets need activation tags to work while others are fine without one. It's all a black box which was my attempt to answer your question.

>>8619816
been my experience too
I think it depends on whether the model already recognizes similar styles or not

>>8619790
e621 tags, bleak ambience doesn't seem to have many images, so it may not work well
i personally use theme since I'm specifically trying to not get the effects horror gives, opting for more of a "good girl acting aggressive" lean. there seems to be dark aura too fro danbooru, which I might try

>>8619711
>>8619729
>>8619733
can confirm that I looked into a model's state dict out of curiosity just now and could not find any running_mean/running_var as would expect to be seen from a pytorch batchnorm layer

>lora hell
I should have stuck to my old config...

certain "people" here should put a cannon to their head

wrong thread?

>>8620088
such as?

How do you add vpred keys to a model again? I forgor :skull:

>>8620113
If the .py I saved back when noob vpred was still new is correct:

from safetensors.torch import load_file, save_file
import torch
state_dict = load_file("foo.safetensors")
state_dict['v_pred'] = torch.tensor([])
state_dict['ztsnr'] = torch.tensor([])
save_file(state_dict, "bar.safetensors")

>>8620117
thx

https://files.catbox.moe/uosrqa.png
>>8619903
Makes sense, but are you sure dark "theme" does anything different from "dark"? Might as well save some tokens. Dark aura is usually purple/black glow around a character.

>>8620140
i honestly don't really know or care if it doesn't have much difference since it gets the job done used with dark persona/evil smile for my purposes. it's an e621 tag that autocomplete gives that I just take.
dark/night tags are usually what I also reach for in combination when I do low light settings and there are loras/color grading techniques if I ever wanted more (I usually don't, as I like skin color rather than everything ending up dark blue)

>>8620140
>>8620152
also, I don't believe in "saving" tokens being a worthwhile effort for the most part. The TE will get what it gets and I've always doubted that the range of outputs will be just that much different because of the amount of tokens it gets, as lo g as the prompt's general meaning is within the same ballpark.
I just don't think it's that sensitive in the underlying math (CLIP input -> latent space mappings) and that most people got psyop'd into caring too much about it during yhe early phase where they lost their minds over calling it "prompt engineering" with the mental framing that came along with it

>tag doesn't get generated in every image
>up the weight
>now other tags get fucked over
>give them weight
>still other things get fucked
>download a lora
>it interferes with some parts of the image, lowering the weight makes it interfere less but also work less effectively
>there is no solution that doesn't fuck something else up or demand more manual labor (in the form of inpainting, or browsing through dozens of gens for the perfect cherrypick
God.
Is there any hope for a new good model on the horizon?

>>8620176
(Dark:1.2) doesn't work for me neither

>>8620176
Sounds like skill issue to be honest. But you could try raising CFG, it was meant for these cases originally before anime finetunes turned it into a blur/burn slider.

>>8620179
It's actually a dangerous tag on noob v-pred, with how well it works. But merges fuck up lighting very quickly.

If I add anything besides the trigger word to the prompt if comes out deformed. Did I overtrain? For reference
Scheduler: Cosine with startup
Lr cycles:3
Lr rate: 1e-4
Unet lr:1e-4
3000 steps
text encoder:0
Alpha=Dim
Adam8bit

>>8620179
Doing a different thing. IIRC dark did work for me on normal vpred when I tried it in the past. It's totally possible it gets fucked on a merge.

>>8620180
My CFG is already higher than normal. At this point I've tried everything except the snake oils. You say it's a skill issue but it's a known architecture issue that the more stuff you try to pack into an image, even if it all makes sense and there isn't conflicting tags, the model will simply just choke. It likely doesn't help that SDXL has the 75 token limit and does the chunking concatenation hack.

>>8620189
>3000 steps
for what size dataset

>>8620189
>Lr cycles:3
KEEEEEK

>>8620189
alpha should be twice dim?

>>8620189
The original SDXL guides and our early Pony bakes used this LR with alpha=half dim, and 2K steps. By that logic, yes you did. But without examples and metadata it's just a wild guess.

>>8620200
The original purpose of alpha was that it would always be lower than dim, at most equal. That training tools even allowed a higher value was lazy on their part.

>>8620189
3k steps has always been fine for me but yes post a picture. I notice you didn't post batch size tho.

>>8620189
60 img
50 reg img
>>8620196
was told it affects the scheduler, not overall learning rate.
Now I feel kind of stupid :{

is there any way to fit pagedadamw8bit into 24gb without fullbf16?

just set alpha=1 and let rngsus take the wheel

>>8620218
use adamw4bit

if you arent using edm2 loss you dont know shit about lora training

>>8620218
use adamw2bit

>>8620230
edm is dead

>>8620211
Batch size:2
and here's one
https://files.catbox.moe/ckdru0.png

>>8620236
post one of the deformed ones

>>8620242
https://files.catbox.moe/wcleqe.png

>>8620249
And you tagged all her features and stuff? If so then yeah just lower the steps.

>>8620236
Isn't batch size 2 basically like doubling your step count? That's how I treat it anyway.

>>8620258
I removed tags that are in every picture like hairstyle and glasses so it's tied to the trigger word.
And Ill try 2k steps with one lr cycle and lr1e-5.

>>8620269
So you never want to take her glasses off?

>>8620269
anyon why aren't you keeping pierodic epochs/steps

>>8620263
It's not really anything, if anything it's halving, since you gotta bump the LR a bit.

>>8620276
nope
>>8620284
I trained 5 epochs and I tried all of them. they had the same issues. I'll just have to try again

New config, https://files.catbox.moe/kg3ivs.toml what do you think?

Oh no not another one

>>8620323
Don't use this, it fries your GPU

>>8620323
I thought anon said above 1024 resolution caused issues? Does this work for you?
>DORA
?

>>8620339
you can go a little higher without issue, but there won't be any benefits in detail. I just have it set to 1152*1152 here to somewhat combat the over sharpening that lanczos causes when downscaling very large images.

Why do artist mixes sometimes make those weird ass fucking gremlin things in the background

>>8620376
less common than you think since I have no idea what you're talking about
post gen

>>8620378
sometimes i get this typa shit and it's always in mixes lmao https://files.catbox.moe/5m0iyi.png

>>8620379
yeah never seen that before
try prompting your artists like artist:zankuro to avoid their names leaking into something else

Is 102d better than vpred10? I tried it out and feel like it's a bit more coherent but also less capable of some art styles and concepts.

>>8620387
It's a shitmix, so it will be more stable and will generete more "aesthetic" (ie slopped up) images at the expense of being able to replicate styles.

>>8620387
Sounds about right. Like every merge, it dilutes the base model's knowledge somewhat in exchange for a style bias. Long as that bias doesn't conflict with what you're trying to do; and you aren't relying on 100pic danbooru tags knowledge, it's pretty good.

Trying to fix hands so I'm using meshgraph hand refiner but I keep getting
>ModuleNotFoundError: No module named 'mediapipe'
I install it with pip install mediapipe --user in my comfui folder directory and it still gives me an error after a restart. Any idea how I can fix this?

anyone know how i can get a plain text of all booru artist and character tags?

>>8620428
Ask our ai overlords to write you a script that scrapes them from the API

>>8620430
wont i get blocked from too many api calls?

>>8620424
>Trying to fix hands so I'm using meshgraph hand refiner but I keep getting
The fuck is this? some comfyui node or something?

>>8620434
Yes. Do you guys use something else to fix hands?

>>8620443
Yeah I use cyberfix/wai/102d. Otherwise I just inpaint sketch.

>>8620387
Give 291h a shot. I flip between it and 102dcustom.

>>8619645
hi bwos, posted the bakes
RadishKek: civitai.com/models/1662074
Aza/Manglifer: civitai.com/models/1662450

>>8620497
>sdxl_vae_mod_adapttest_01.safetensors
huh, what's this?

>>8620452
Best inpainting tutorial?

>>8620501
its a modified sdxl vae from an anon in /hdg/ >>8618360
its a little sharper than the noob vae and doesn't have the slight blue tint of xlvaec_c0.

>>8620505
hmm, did you train those loras with it? sounds like it shouldn't work otherwise, or it otherwise needs comfy to work and anon's custom node

>>8620502
https://rentry.org/fluffscaler-inpaint

>>8620513
nope i did not, i just tried using it for genning and kinda liked the effect.
it gave me the same results as the c0 vae but without the tint, so i'm happy with using it as-is

anyone here still have oekaki anon's
>slantedsouichirousep26-step00000132.safetensors
lora?
I accidentally deleted mine awhile back. tried looking in the rentry and archives but couldn't find mine

vibin'

what is lil bro vibin' about :skull: :skull: :thinkingemoji:

>>8620497
nice bwo,
by any chance would you be willing to share a "fail"/overbake of the aza loras? Wanted to see what outputs I could get from one of them

>>8620505
if it works for you thats fine but im still trying things out and wouldnt recommend using it, the trained model actually produced worse results now that i tested it more in enc+dec (very blurry rather than overcontrasted), and its only the encoder that i trained so it should actually be the same as original sdxl vae when used for just making gens

>>8620570
hi bwo, do you have any specific version in mind? i'm not sure if i've retained the original fails (the more recent ones might still be in the recycle bin, but i'm not sure)

>>8620585
alright, thanks for the heads up.
i only tested it against the noob vae and the c0 one i was using. Found it sharper than the noob vae (might just be that the noob vae is ever so slightly blurrier, will have to test against the original sdxl vae as a control). looking forward to the results of your vae training experiment!

>>8620599
I think I liked how >>8615606 seemed to turn out. If there are ones with more steps from that attempt, that would be cool instead too.
Also, as an aside observation on the lora, it's hilariously capable at getting legible english text moans to come out.

>>8620599
>slightly blurrier
thats probably the case because the one tuned decoder vae i tried from civit clearly didnt use a perceptual loss like lpips and was blurrier

my idea might not go anywhere, mainly wanted to demonstrate that there might be a way to upgrade from the 8x compressed vae, to a 4x compressed one without that much training
encoder needs to be adapted, decoder also shits itself a little for some reason with having its 2x upscaling removed (the vae actually outputs coherent stuff even without training it, just a little artifacted), and there needs to be a hopefully minor finetune for the higher res since its the equivalent as generating at 2x higher res than usual

Can a kind anon prompt a blowjob where you're sitting in the car (maybe driving) and the girl leans over from the side to give you a blowjob? Can't seem to figure this out.

>>8620532
apparently even I don't have it anymore(and I nuked it off mediafire since it's a standard lyco and I don't think it was actually super great to begin with). And I never really tried doing a fatter/more "recent" config run of it since I still had PTSD of running that dataset on pdxl.
I did still have the config and I'm pretty sure the dataset has been untouched since then, so here's a (sort of) reprint
pixeldrain com/u/quBr9bVC
if someone else has the original still that one is technically still going to be different due to the whole process being random because ML is fun like that, but I did at least check through a few output steps and 154 was (still) the best, though it's also a bit fried and bleeding at the edges. though that's probably fine-ish when used as a mix at lower weight and/or on derivative models instead of illustrious 0.1

>>8620428
Get one of the csv that people already scraped like from the autofill extension and then ask gpt to get you a conversion command.
I posted an updated one more fit for Noob, either here or in hdg, I don't remember.