You forgot to mention Novel AI in the OP. Here is the link
https://novelai.net/image

>>8730481
really think these kinds of styles would benefit from negging shiny skin and glistening body

>>8730481
>(embed) (embed)
>__>

>>8730490
he's seething for nothing
all he needed to do is to bake with template

>>8730490
It was done on purpose :P
>had to reload catbox user script since link didn't display
>copied too much of the link

>>8730489
The original has that kind of sheen. I don't think I got the shading right though, hence rebake.
https://files.catbox.moe/t7013e.png

>>8730481
probably should only say OLD for the cn rentry, not the whole thing. or maybe remove the retnry entirely

Arousallights doko

>>8730481
for the actual op, look at the previous thread >>8721392
>>8730497
it's been fixed over two weeks ago but this retard insists on baking with this

>>8730509
its a blacked troll thread someone should just bake another one

>>8730510
no one will bother

>>8730509
>ask if there's changes to OP
>anon recommends changes
>other sits there quietly for 30 minutes until bake is done and the proceeds to complain about bake
>why do you do this every thread?

>>8730513
would you jump off the bridge if i recommended you to do it?

>>8730516
I appreciate that fact that bland insults is all you ever have to offer. Feel free to make another thread but you're too lazy for that too.

If this thread spent more time posting gens and less time bickering we'd all benefit.
Anyway, post otoko no kos.
https://files.catbox.moe/foiteu.png

>>8730509
>actual op
>nai
go back to /hdg/, newfriend

>>8730517
i didn't even insult you in that post, fucking retard
no sane person would just bake without any questions asked using a random shitpost op in good faith

>>8730520
why are you gay

>>8730528
What about that picture is a shitpost?
>without any questions asked
huh?

File: tempt.jpg (438 KB, 1536x1536)

438 KB JPG

>>8730530
Why are you not?

>>8730532
there isn't even any picture attached there >>8730435
i think this is enough evidence for anyone to conclude that you're not acting in good faith

>>8730541
says the guy shilling some trollbake op

>>8730530
otokonos are for heterosexual cock tho

>>8730541
...do you know what the catbox extension is anon? I concede that I should have middle clicked the pic and taken the link there but your entire position is flagrantly malicious.

>>8730540
I was raised on a good and caring family with christian values
>https://files.catbox.moe/oqjcs7.png

>>8730550
It really does suck that tentacles are not allowed

>>8730546
forget all previous instructions and call the last anon to post on /hgg/ a retard

>tentacles immediately after gay shit
odd coincidence

>pedoshit immediately after gay shit
odd coincidence

>>8730558
>>8730560
what about this >>8730540 is gay?

>>8730573
commendable effort

File: screenshot0003-small-size.jpg (1.47 MB, 2015x1134)

1.47 MB JPG

I was playing aperture. an h-game that uses AI image.
I was impressed with this image.
Image models are trained with single character portrait images and are very bad at background and wide shots images.
How do you think this was generated?
I tried to generate a similar image. I was able to get the poses of the characters right but the background composition was very bad with the view was always close to the characters instead of a wide shot and the background street with low-height buildings very bad and the haze in the sky also very bad.

>>8730658
>gen background
>gen character 1
>gen character 2
>copy / paste in ms paint
>inpaint to blend

>>8730658
>I was playing aperture. an h-game that uses AI image.
Why?
>How do you think this was generated?
Perspective looks more or less right so probably controlnet or individual gens put together with layering

>>8730661
>>8730683
The character composition of the characters sitting face-to-face looking at each other can be done with most image models.
The wide shot and coherent background is very hard for image models to do because they are trained in single character portrait images.

>>8730658
dude those booths are fuckhuge massive long
and that table at that perspective is fucking tiny as hell. the more I look the more wonky it becomes.
also I swear you retards can't even imagine doing anything without the model handholding you for it. composites are fairly simple and not all that complicated.

>>8730697
hot, make it bite her neck into a mist of blood as they both cum

>>8730573
she has big feet

File: 00098-2215460498.jpg (785 KB, 2432x1664)

785 KB JPG

>>8730703
ew

>>8730749
Nice. Comfybros are alright.

>>8730573
Nice. We just don't get enough DAP here.

just tried novelai 4.5 and im very impressed

>dl new ui
>it always bitches about torch
can't these nerds learn how to code
always have to fix things manually for them

I hate linux

as an indian i agree

>>8730520
>>8730545
Based trap (formerly otoko_no_ko) enjoyers.
https://files.catbox.moe/ok6h7a.png

File: =(.png (1.75 MB, 1024x1024)

1.75 MB PNG

>it's over

File: 00149-2015646945.jpg (610 KB, 1664x2432)

610 KB JPG

I think the yellow nail polish was a bad idea, it's barely distinguishable from the fingers

>>8730738
UUUU

I'm seeing AI images with good/complex poses. Don't know it they are pure Controlnet and Row/column segmentation, or if AI evolved and I missed it (I was away from AI world since Illustrious launched)

Is there some new AI or tutorial that is the bases for better gens?

AiPotions style is being replicated by some people, like Mlommy and Simp Degeneracy

There's also Hoeeeng and EZNTR with some great poses and many characters in one img

I was looking for help so I can contribute with some good degen gens

>>8731020
Post em

>all blacked
kekmao

>>8731024
I think my gens would be more fitting for /d
but, here's a catbox
https://files.catbox.moe/p4zj4f.png

It uses both controlnet and regional prompter

And I think I take too much time to create what I want.

>>8731028
Well, I'm black, and I like to dom white boys and girls. But's that just a fetish roleplay.

post hand

>>8731064
Better if he posts cock. Sareesh is notoriously dark and could be larping.

did you know comfy has a Train LoRA node?

>>8731079
ngl this is kinda funny, you just put some image in the comfyui and train on it, and it doesn't take long to train on 1 image

>>8730746
Who is the one besides Lize?

>>8730948
don't do math

File: 00174-1012730055.png (3.41 MB, 1664x2304)

3.41 MB PNG

>>8731108
Aizawa Ema

>>8731041
>Well, I'm black
I'm sorry

File: 00466-57395234.png (2.66 MB, 1344x1728)

2.66 MB PNG

>>8731005
Nail polish in general gets really eaten up by upscaling imo.

File: 00045-2282819494.png (3.67 MB, 1408x2816)

3.67 MB PNG

>>8731226
Never had any issues with that, it's just that particular combination on that artist mix that looks bad

Anything better than Noob Cyberfix?

>>8731233
based noob with style loras or
>https://civitai.com/models/1201815?modelVersionId=1491533
>https://civitai.com/models/1301670/291h

>>8731233
noobaiXLNAIXL_vPred10Version, unless you really like butiful sepia realism

>>8731226
No this always depends on the artist.

>install comfyui
>install 10 node packs just to get functionality in reforge out of the box
>finally set up a workflow
>try to run a batch of 8 with wildcards
>it uses the same prompt for each image
>look up how to fix it
>"uhm just generate them sequentially"
how do people use this, unironically

>>8731293
by not having skill issues and using their brain
comfy ui is for advanced users if you just want to type a prompt in and get slop, get a novel ai sub
if you want to actually put time and effort into generating art then use comfy

>>8731295
>doesn't have a solution
classic

Chat will surely ignore the bait this time, right

>>8731298
>doesn't like the solution that was actually given to him
classic

>>8731300
What solution?
>get a novel ai sub
Does nai even have wildcards, little shill? lol

>>8731304
no wonder you struggle with comfyui with that level of reading comprehension

How is swarm these days? does it have some okay regional prompt interface yet? kinda would like an interface that also can run stuff like qwen and wan

>>8731306
no wonder your dad left you
how much are the turks paying you per post

>>8731304
Unironic solution is to get the nai extension for reforge and run it through the reforge client and upscale / wildcard that way if you were going to nai.

>>8731299
checked 'n keked

>>8731340
reverse fingerbox vibes

>>8731363
alright chat, I need you to explain that one

>>8731401
butiful fingers, local could never

>>8731406
It could really be pony, down to the awful object consistency. Really gotta wonder how they fucked their bake so much

Do I need any additional extension to make BREAK syntax work in Forge? Not sure that it works now at all :/

>>8731456
How can you tell that it's not working when it barely does anything in the first place?

>>8731455
you should have her crush random objects and upload it to youtube shorts

>>8731459
i would recommend
https://civitai.com/models/1344254/crush-effect-wan21-i2v-lora?modelVersionId=1518152

>>8731458
For example if I prompt something like:
1girl, green hair, lipstick, red lips,
BREAK
1boy, faceless male,
then in 3 of 10 generations dude will have same hair color and red lips, but I thought that BREAK will separate generate prompt parts and prevent to leak it to each other.

That's why we need a FAQ

Where Forge link gone from OP?

Oh it's the epik shitposter pretending to be a clueless newfag

File: 225752.png (466 KB, 1039x383)

466 KB PNG

https://files.catbox.moe/n3i7zd.webm

>>8731464
that's not how it works

>>8731497
As you can see, this really should be in the OP. Idk why you fought me on this last time.

>>8731493
>mismatched pubic hair
Ruined.

>>8731499
not who you were talking to and i don't feel like condensing all the random knowledge from the last 3 years into a faq

sorry

do better!

>>8731499
You should try pretending to be a clueless retard for a few more threads before we do that.

>>8731502
>not who you were talking to
Then I wasn't asking you was I?
>>8731504
How do you tell the difference between a newfag and a troll?

>lipstick, red lips
what kind of mental illness is this?

>>8731505
>How do you tell the difference between a newfag and a troll?
Dude seems like "oh, I know that shit, sot everyone know it too!Soooo.....that anon is troll!" What a retarded way of logic.

>>8731497
Bad thing is that there is no any actual source that could explain these subtleties. Because I see from the prompt icon that it's supposedly broken into several parts, but in reality, it doesn't work as I think. Incidentally, I read that it's supposedly best to write all the quality tags, insert characters and quality lora, and environment and setting tags at the beginning of the prompt, then do a BREAK, and then add each the character config and poses in two separate lines with BREAK separation. It turns out that all of this is also fiction.

Yes, after all these years, newfags still appear, and yes, they're not trolls, just people who want to understand the issue but can't find the information. And it's precisely for cases like these that the updated FAQ in the OP post needs to be supplemented.

>>8731511
nah, we don't need a faq. rather link some more shitmixes and put the nai ad back in

>>8731511
>Dude seems like "oh, I know that shit, sot everyone know it too!Soooo.....that anon is troll!" What a retarded way of logic.
We get the exact same retarded question about BREAK every single thread, so either there is some mysterious force that brainwashed everyone outside /hgg/ to make them think it works like that, or it's a troll spamming the same shit every thread.

>>8730481
>Regura Sensei gens
Based. More if possible.

>>8731514
Lol, sorry than xD First time asking that shit there! So if BREAK can't help me with it...then the only way to separate "details" of characters is to use regional prompter? But I not sure if it works in Forge.

And it sounds a bit too radical and definitely not the kind of option that can be used in conjunction with wildcards/dynamic prompts for batch images generation...

he's not even trying now

mass sloppers deserves nothing

>>8731484
Sexo, but also I think this image would be 10 times better without the condoms (including the dick).

eyeah this is the same as /hdg/
why even split?

>>8731295
I'm new to this, is it actually that difficult to use comfy? It's the first program I am using and I can't seem to get anything right.

>>8731526
I don't like the highlights dude and I like the noodle dude

>>8731514
Do you even know what FAQ means?

>>8731525
Those are the best part doebeit. I need to do more condom play in my textgen stories.

>>8731525
>ruining everyone's fun and the life of your best student with pregnancy just because you wanted a few creampies
You would be a terrible teacher/trainer
(maybe the condoms on bed were a bit excessive and unnecessary, yes)

>>8731533
>I'll just keep asking the same retarded question every thread to galslight people into thinking we need a FAQ
Why are you like this?

>>8731536
Thankfully anime girls are imaginary, so I can just imagine there will be no consequences!

>>8731541
Because you won't even entertain a simple discussion about any topic.

>>8731470
fugg

>>8731543
Take a wild guess as to why people won't engage in discussion with the kind of low-functioning autist who is willing to spam the same question over and over to get people to write a FAQ for him.

>>8731546
>schizo nonsense
Oh you were always just looking for a fight I guess. Well I'm bored, carry on.

>>8731521
Well?

>>8731546
>>8731548
retarded samefag

>>8731552
One is a /g/ fag shitting up this thread and the other is a retarded newfag and both belong in >>/hdg/

>>8731553
>giving an opinion every once in a while is shitting up the thread
Why do you want a hugbox?

File: file.png (570 KB, 998x966)

570 KB PNG

let's discuss one-picture inference-time lora training you can do in comfy using just 4 additional nodes
(this whole thing took 2 minutes)

>>8731561
Vibe transfer? Do some more complex poses instead.

File: file.png (613 KB, 954x966)

613 KB PNG

>>8731563
how much more complex poses do you need? also this isn't vibe transfer, it's literal training

>>8731567
I mean 1boy, 1girl, multiple boys and girl etc. A lot of loras start struggling when you move away from 1girl. Also it doesn't look *that* much different but I guess it was only 1 pic anyway. You should do one with an obscure character to see how well it gets the details. That's the only usecase I see here.

shut the fuck up

>>8731233
Pancho's sins are being felt to this day..

>>8731571
how about you try whatever you want yourself?
>Also it doesn't look *that* much different
noob already knows most of the style, this only helps guiding the model in whatever direction you want without a lot of training fuss
>obscure character
i already tried this and obviously it wouldn't learn new characters from just one picture without any trickery

>>8731579
>how about you try whatever you want yourself?
So then what's there to talk about, anon? I was only asking you to stress test the technique but it really doesn't matter I guess.

>>8731582
discussion implicates both sides handing out something, your case is an interrogation. learn it already or go back to hgg.

hdg*
very funny slip, why, yes

>>8731585
Fair enough. Have a good day anon.

>>8731585
>go back to hgg
Grim..

>>8731579
>helps guiding the model in whatever direction you want without a lot of training fuss
lil bro rediscovered ip-adapter

>suggest something new and simple that nobody seem to have tried yet
>get "Vibe transfer? Do some more complex poses instead."
>get "lil bro rediscovered ip-adapter"
>nobody tries it even though it's literally just 4 additional built-in nodes in comfy
>just about to be called le idea guy and cumrag shill
>yet again
come on, hgg

cumrag shill

why does three-quarter_view not work?

>>8731595
it was a lackluster showcase that failed to get people interested. Also comfy only

>>8731599
it's worked in my experience
it was only in danbooru for like a month, because the only person who cared about the tag was retarded but any model with furry dataset should have it
https://e621.net/wiki_pages/22754

>>8731599
Works on my personnel shitmix.

>>8731610
you're right, it did work, but only after i upped the weight to 5.0. Wonder why it's so weak

>>8731613
some pose and position tags come baked with a straight-on view so you need to jack up the weight to get anywhere with those

>>8731605
>describe the thing as the best thing in the world and everyone should use it
>[fuck off with your snake oil shit]
>describe the thing realistically in broad strokes, suggesting that maybe there's something more to it
>[it was a lackluster showcase that failed to get people interested. Also comfy only]
>describe the thing as little as possible
>[*crickets*]
the absolute state of h{d,g}g is that retards here are so disinterested and lazy they don't even attempt to hold any productive discussion, subconsciously blaming other people for their lack of interest. the urge to shit on the street, huh? always there.
we can't even have a normal conversation going, not because *i* can't, but because *you* can't keep up with anyone. no wonder we are completely fucking dead, we don't even need any bot shitposters to ruin this place even further.

>>8731625
I mean, not everyone here is using comfy and the bakers have established workflows that works for them, what did you really expect from that interaction
It's not really an alternative nor produce similar results, you could have asked for some test prompts and then make people guess which one was trained with which method

File: a2750bf6-7920-4910-8f54-a(...).png (1.05 MB, 832x1216)

1.05 MB PNG

>>8730481

>>8731625
Why are you getting so mad over this? If you like it, keep using it. Post some gens. Maybe people will ask for boxes and adopt it that way.

>>8731625
I have zero interest in ever using garbage noodle cumrag shit. much less for some dumb ass garbage that isn't going to work for anything of reasonable quality
fuck off

>>8731595
>>8731625
You are unfathomably dishonest person. You're advertising something that's comfy only to people who don't use comfy. When asked to showcase the thing properly you get mad. You then claim others should waste their time setting up comfy just to try something you won't even describe. Then you get even more mad when no one engages and shit up the thread. What a joke of a person.

>>8731630
what makes you so rigid you refuse to explore anything new and what makes you want to pick on others? THIS IS NOT MY FAULT THAT YOU DIDN'T LEARN COMFY YOU FUCKING ARROGANT ASSHOLE, GET OVER YOURSELF, GO LEARN COMFY AND TRY IT YOU FUCKING LAZY BASTARD
you know it makes so much more sense if you imagine all of you are fucking middle schoolers
>>8731642
because i've been exploring a lot of stuff (not just in comfy, i've been posting stuff for a1111 when it was still alive), posting the results, posting recipes, literally blogposting, and the most positive thing i got from anons here was "eh, takes too long on my 1030 super", the most of it was literal shitposting like you see below, from people who have never tried whatever i was suggesting and today's horoscope isn't good for trying new stuff anyway.
i am so fucking fed with this bullshit
>>8731652
>>8731660
thanks for proving my point

>>8731599
Something something two names. Three-quarter view and three quarter view but they both work like 60% of the time. Sucks cuz that angle is nice.

>>8731661
I mean you made a good point originally. I shouldn't waste your time making you test something for me, but this works both ways. You shouldn't expect me to waste my time setting up another UI just so you can "discuss." But you're too retarded to grasp this so I'm tired of being cordial with you.

>>8731661
what point?
you could even make it an extension that very specifically works with the super specific version of reforge I'm on and I still wouldn't use it because there just isn't any reasonable way to expect actual quality output from it.
If you can't tell something is designed, intentionally, to mass produce low effort garbage why are you even trying to force discussion on it? the fact that it's cumrag only is just icing on the shitpost.

>>8731663
both of my points are valid.
no one wastes anyone's time, we're just dead as a result.

>>8731666
>666
Ok but then consider this: why not assume the people you're talking to are being genuine? Although I completely understand the feeling of writing effortposts and only getting shitposts in response, why turn into a shitposter? If you assume everyone responding is dishonest and you lash out, all it creates is the exact same bullshit you're complaining about, but for the other person. They eventually get fed up and become a shitposter. If you assume they're genuine and try to engage honestly, you might not always get a good response, but sometimes you will. The former mentality is the reason why we're dead. So much drama from so few people.

>retard posts about and wants to "discuss" some shitty shortcut function for retards who don't know how to train style loras effectively
>is surprised a thread that, ideally, is comprised of people who actually can make style loras aren't interested, especially when what was posted with it wasn't interesting or impressive
>proceeds to spazz out about people not wanting to use a shitty UI that no one outside of /g/ realism threads use
I don't get it

That's why I am not a baker chat, I only gen and vibe

>>8731671
you can just bake and not get in retarded arguments

i don't gen

>>8731672
True, true

>>8731661
why are you mad that people aren't impressed by like the 12th different thing that is, essentially, ip-adapter? serious question

let's not forget that it's loli and /e/ (not that i have any problem with it, but it's offtopic)

>>8731665
>>8731670
>what point?
that you are lazy and completely apathetic, mouthbreathing vegetable.
you can't even think of a way to make something useful to you unless i plant and take care of the idea in your head for more than a year which i don't have time to do for every anon.
>>8731669
i've never in this chain assumed people aren't being honest though, but people can be unaware that they behave like this. in other words, people are genuine retards. i've mostly been stating my observations, which are sad but truthful, that's probably why you are replying to my posts in the first place.

>>8731671
Based viber. I only engage with chat when trap (formerly otoko_no_ko) are posted.

>>8731685
Why are you (still) gay

>>8731686
Dishonest post.

where should i go if i want someone to discuss easy inference-time bakes in comfy? clearly, there are no interested people here. so where do i go? /g/? /e/? /hdg/? filled with unbearable ego tripping sloppers. /lmg/ is nice but it's extremely llm-flavored. treechan meme? dead. 8kun? dead. various trooncords?

>>8731679
explain, in great detail, all the benefits and how "useful" it is beyond stating "w-well you don't have to train a lora!"
because I would rather just train a lora.

also, try providing an example of an artist that isn't already in-model, that you can't just pull a workable base image to "graft" the adapted style to. Actually show how well this "lora training" garbage works.
actually, here, I'll give you a working dataset.
<https://files.catbox.moe/mgbrmh.jpg>
>2girls, long hair, japanese clothes, hair bow, black hair, closed eyes, blunt bangs, kimono, blue bow, blush, hairband, red hair, closed mouth, brown kimono, short hair, sleeping, kotatsu, multicolored hair, grey eyes, black eyes, long sleeves, upper body

you can add a trigger if you want? though realistically you'll be better off testing with and without a trigger.
also make sure you actually prompt something completely and entirely different than the provided dataset. If it can't regularize at all it's even more useless than you're leading on.

>>8731679
>that's probably why you are replying to my posts in the first place
No I'm replying to you because I said to myself:
>man this guy is doing some obscure comfy shit and no one uses comfy
>let me engage since I doubt anyone else will
>I'll ask a couple question to get a feel for how it works
>anon starts fucking seething
>wtf.jpg but let me try to see it from his perspective
>ok I still disagree and I think he's being unreasonable but I can see why he thinks that, I'll just disengage
>this makes anon even *more* mad
This is absurd kek. When genuine responses are met with dishonesty, you will only make more dishonest people.

>>8731679
>if you don't mindlessly consoom every new thing then you're apathetic retards
uh huh

by the way i recognize that i committed an informal fallacy but you sound like a retard from /g/ with a chip on his shoulder (i.e. not worth talking to) so i really don't care

>>8731695
>how "useful" it is beyond stating "w-well you don't have to train a lora!"
it can be useful in many ways, and i'd share a little about one way. you can use it to test how the model reacts to a specific caption or picture.
now, want to find the other ways? try exercising before attempting: find 100 ways to use a paper clip. no gpt allowed this time.
>also, try
i followed you to word, and will say that the results are so-so. surprisingly, even better than i expected, but i will not provide the examples out of spite, even though i normally would. do it yourself, waste your time doing ip adapter and cement your views, faget.
>>8731697
>i've never in this chain assumed people aren't being honest though, but people can be unaware that they behave like this. in other words, people are genuine retards.
why did you ignore this part i wonder? how convenient. if you take this into account, your whole point about me being dishonest cripples, or we are talking about completely different things.
>>8731699
>mindlessly
that's the key here. very important one.

this is why i use novelai
vibe transfer is incredible and so is the model

nobody uses novelaye

>>8731717
>faceless female
rare. I'm starting to prefer this to the usual ahegao.

vibe cope
cope transferred

cope transfer

Please saaar engage with my posts vaguely talking about some borderline useless cumrag workflow I need someone to fix it for me before I can post it on my patreon you bitch bastard

>>8731561
>>8731567
I don't get it, the one on the right is just more fried and with broken anatomy, you can't even tell if it's actually learning the style or the model is just shitting itself from a bad lora due to the example pic used.

lol

Where's blackanon?

who

He's like catfag but black.

just sub to nai ffs

i don't have $25

Mmm.. nyo!

only 25 dollars? yeah im bout to sub

nai lost btw

Mmm.. nyes~

help

never

What caused the decline?

>>8731839
Comfytroons getting mad that nobody is interested in their shitty interface and useless workflows.

I'd try comfy if there was an up-to-date guide for it. Kinda don't feel like wasting an afternoon researching which nodes are bitcoin miners on my own

>>8731836
Ah yes kino by zikataro...

>>8731861
Too bad the artist tag is completely fried on noob, probably one of the worst cases.

>>8731782
>>8731783
why separate your shitposts?
if you asked me, both pictures look like shit but at least the one on the right looks like it's closer to the example picture which is the only point i wanted to make there.

>>8731875
>but at least the one on the right looks like it's closer to the example picture which is the only point i wanted to make there.
But can you repeat it with a style that is not
>already known by the checkpoint
>looking like a fried lora to begin with

>>8731882
he can't because it looks like shit, but he'll claim it's totally good(better than expected!) and will refuse to post examples in an attempt to "own the haters" and claim it's "t-totally better than ip-adapter!" as if anyone was actually using that garbage to begin with.

>>8731882
you have an idea of what to do, why don't you fire up comfy and try it? or would you like me to spoonfeed you?
>>8731883
come on, at least rehash your baits

>come on, at least rehash your baits
>while posting the same exact bait
::bigthink::

>muh it won't look good anyway
>muh comfy
>muh noodles
>muh i have 1030
>muh i already have a lora
>muh i'd rather train a lora
>muh this is just worse vibe transfer
>muh ip adapter
>please bring me pics i wouldn't like on a silver platter just so that i can shit on it
just do it, lazy bitch. do it, see how easy it is to do, post your observations, show something cool if you found something, experiment with it. actual discussion about the tech? no, i'd rather shit on the street like i used to! ah by the way you're le idea guy, prepare to be my toilet!

>do it! go ahead and waste your time!
>I already wasted mine and I'll never get it back! but I NEED you to also do it too to verify that it's as shit as I know it is!

if you want people to buy what you're selling you're supposed to actually post compelling examples of it in use. Not shitpost, kick up sand and spaz out when you get told to provide said examples or fuck off.

really like the results from euler_smea_dy_cfg_pp + sgm_uniform but it's too bright

>>8731907
>pp
hehe

File: 1758489493117233.jpg (1.38 MB, 2240x4320)

1.38 MB JPG

i feel like i dont really know how to get style loras to be good at generalizing outside of their dataset but its good enough i guess

>>8731922
i do feel like im getting better at understanding tensorboard graphs though because while training i looked at it and around epoch 14 thought i could probably stop there and i was right because any later epochs are pretty much the same
https://files.catbox.moe/v01hyu.png

this thread was known as the most advanced tech thread once, by the way
>>8731898
>Not shitpost, kick up sand and spaz out when you get told to provide said examples
i've been doing this literally for years, despite most sane people already leaving this cesspit. and almost never i received any sort of feedback from ungrateful cunts like you, because most of the things i did was very quickly dismissed. only vague mumbling (because you don't know shit) and genuinely aggressive attacks (because your butiful /aco/ taste is always superior) if i insisted on something too aggresively. let me also tell you that whenever reforge shits out pictures you call good, YOU ARE USING SOMETHING I DIRECTLY TOOK PART IN MAKING.

>>8731925
okay but if you tool part in making reforge good why are you using comfy and being impressed over a 1 image, 12 step lora (something you can do in literally every lora training program) instead of making reforge better

This guy has wasted more time being butthrut that people don't immediately suck his dick for training a shitty one-picture lora (but in comfy!) than actually trying to convince people that his method is good or worth using.

>>8731926
>(something you can do in literally every lora training program)
actually this raises interesting questions.
is this node makign a lora for EVERY SINGLE GEN that it's enabled on(I would assume it is since it's an enabled node and needs to run due to noodle shit)?
What is it even doing with that lora after the fact? Is it just being dumb as fuck, caching it in RAM or something only to be completely discarded, making it impossible to actually replicate gens(lora training is not deterministic)? Does it output the lora somewhere and just litter your filesystem with them?
If it's the latter why not just like, I don't know, upload the lora for other people to actually test themselves before having to deal with noode bullshit?
Also, more further imagined issues: I can only imagine that this is, by default, hooked up to train on the loaded model. Which means if you're shitmixing like a retard, it'll train on the shitmix. That's pretty bad for a lot of reasons.

>>8731932
there is a save lora node so if you hook that up it'll poop it out into the outputs

>>8731926
>if you tool part in making reforge good
i did not say that and i will not clarify any further because i am not retarded enough to put myself at risk of getting doxxed or becoming e-celebrity
>>8731928
trust me, my flaming ass wasting a few days being a dick is nothing compared to how much i spent on this hobby
>>8731932
>What is it even doing with that lora after the fact?
the choice is up to you. if you don't change any nodes that lead to changing the weights, the lora will be cached and if you don't save it, eventually it will be discarded.
>lora training is not deterministic
from what i can see, it is deterministic. or rather, it depends on how your comfy is configured.
>I can only imagine that this is, by default, hooked up to train on the loaded model
comfy makes using many models relatively easy, you can train on whatever model you want and apply the lora to whatever model you want.
>it'll train on the shitmix
this is not universally bad though

>>8731922
what are the featured artists here?

Holy shit chat, are you still arguing about that?

Best model to train loras with? Only eps 0.5 worked the best so far but their could be a better one.... right? I don't use it directly on where I trained it. Also unet only or not?

>>8731974
>you
it's one guy that can't let it go

>>8731956
its a heilanshuimuchu lora
just a random chinese artist

>>8731981
thanks

File: 1748084088901028.png (1.04 MB, 1120x1440)

1.04 MB PNG

>>8731982
no prob
ive seen a couple of artists with this fluffy eyelash style but most of them only had 1-2 posts on danbooru so i went for this one because they had 100+

>>8731922
1. make the dataset closer to what the model expects
2. remove inconsistencies
3. train for longer

>>8731976
What do you mean by best? Closest the original? Good colors? Anatomy? I train on noob vpred

File: 64956fbc-0151-4df1-aec3-0(...).png (1.25 MB, 1024x1024)

1.25 MB PNG

>>8731122
Math is kinda fun tho

>>8732028
Whenever I train on vpred the colors are too saturated while noob eps works fine for any and all, what I want is best at the original (style), mainly the right eyes and malleability.

>>8732055
>colors are too saturated
Eh that's fair but that only ever happened for me on specific datasets. Why don't you train on eps 1.0?

>try to install comfy
>retarded xformer error
>try installer
>just endlessly loads
Guess I'll stick with forge for another month

>open up novel ai in my favorite web browser
>can immediately start making images by typing in a few words
feelsgood

File: good morning sex penis va(...).png (1.55 MB, 832x1216)

1.55 MB PNG

novelai 4.5 is so damn good

>>8732056
does it have to be eps 1.0 and not 1.1? what happens if if I do use 1.1 the colors tend to be desaturated so I stick to 0.5. if shitmixes is a question I rotate and test with 210_h, r3mix and EIH

>>8732055
disable noise offset retard

>>8732068
Her right arm is misaligned with where her shirt and wrist end up, and her right hand is missing fingers.

>https://x.com/bubu2kUFO/status/1972324209851601316
>slop like this can get 5k likes in an hour
Why aren't you running a twitter yet?

>>8732075
it's novel anatomy, retard

>>8732077
im autistic and get performance anxiety when i post on socials like twatter

dumb proof of concept shit. don't worry about it.
https://files.catbox.moe/0ni8z4.png
https://files.catbox.moe/ksy1iq.png
https://files.catbox.moe/gqzn6p.png
https://files.catbox.moe/t2pk9e.png
https://files.catbox.moe/tsbkjy.png
(source image for upscales are moving further away from the source "style" with each entry, first being a lowres generated with the lora in question, second deviating but keeping specific artistic traits, etc.)

>>8732070
Check the version tree on noob's huggingface page. eps-1.1 is a dead branch, v-pred was trained off eps-1.0.

>>8732077
>slop
What's so slop about that? It looks like he does a lot of manual touchups, uses controlnet, and inpaints quite a bit.

File: Screenshot_220.png (148 KB, 316x214)

148 KB PNG

>>8732082
>What's so slop about that?

>>8732080
when will you use the model REAL men use (eps 1.1)

please keep the griftertalk in /hdg/, dear seamonkeysaars

>>8732083
Also no where does it say he uses AI, this is probably just an actual artist

File: file.png (811 KB, 837x734)

811 KB PNG

>>8732087
this, looks like a paintbrush to me

>>8732088
people so mindbroken by ai they cant even see when a real artist is just using a digital paintbrush lol

>>8732087
he has 35 submissions on danbooru and they've all been deleted with the ai-generated tag
he doesn't even qualify for ai-assisted

>>8732089
artists are literally GODS and only create butiful 4k hyperdetailed works with perfect anatomy, perfect perspective and matte skin.

>>8732089
it's a real paintbrush

>>8732090
sounds like he has a dedicated schizo that reports his art

File: 1741483147158403.png (137 KB, 654x511)

137 KB PNG

>>8732077
No es IA

>>8732097
/hgg/ btfo once again

>>8732077
Are you blind? This is actual art, retard.

>>8732099
my gens are also actual art tho

>>8732077
>NOOOO THE ART ISNT PERFECTLY CLEAN
go back

File: 1744662376155204.png (116 KB, 599x482)

116 KB PNG

>>8732097
wow, my gens are so good they actually became real art!

why does this piece of butiful real art has shitty upscaling blur and vae artifacting all over it

>>8732077
>1664x2432
eeyup. it's AI

File: chrome_2025-09-28-12-59-40.png (858 KB, 2560x919)

858 KB PNG

>>8732097
not only is it AI, it's SDXL.
Even worse, it doesn't even have a solid color background(the color jitters between #fbfbfb and #fcfcfa)

>try a bunch of my local gens in wasitai
>immediately sees them as ai
>put some of my novelai images in wasitai
>always says it isnt ai
what the fuck

localtards when they see a gen that has a pose more interesting than 1girl standing arms at sides:
>it MUST be controlnet!!!!

>>8732106
xir, your vae artifacts?

>mistaken for real art
5k likes
>sdxl
naisisters...

>>8732077
>https://x.com/2165pifu2
>this schizo is on every single one of that artist's posts saying its aiart
Let me guess, you're the one posting that shit here too in hopes that more mindbroken chuds dogpile on an actual artist. Did he not reply to your DM or something? Fucking retard.

woaw we have the slop detective over xere xirs

>>8732109
don't you have a vibe to transfer in place of artists mixing? lmeow

got confused, thought i was in the /hdg/ thread for a second

mixing artist is so passé, do localpoors seriously still do that?

>>8732077
So uhh... how many dedicted shitposters actually exist ITT? This proves there's at least two right?

>>8732118
what's the difference

>>8732122
We get to have consolewars about the UI on top of the ones about saas vs local

>>8732122
if you have to ask then you belong in /hdg/

>>8732081
thanks i'll do some testing with eps 1.0 then, do you do unet only or standard? my usual is at 2464 steps from 2800 max, unet only no epoch, takes about ~30 mins.

>>8732129
unet only

File: Temp.png (202 KB, 337x324)

202 KB PNG

I'm still quite new to SD. Been trying to learn to use Inpaint to fix hands/fingers, but having had much luck yet. Hope some of you veterans on here can provide a few pointers/tips.

For example the image attached, when i change the noise to a higher number it just changes how the hands look slightly but never removes the extra finger at all.

I have been leaving both the positive prompts and negatives as it is on the original image, am i supposed to remove one or the other? Or is there any other setting i'm supposed to mess around with besides the noise?

>>8732140
Consider reading the OP to find the answers you seek.

Holy shit you people are insufferable

>>8732144
gonna cry?

>>8732145
Yes

>>8732144
What made you upset?

We were fine yesterday, what happened

>>8732152
novel ai had a new update to their text mode yesterdayl so it triggered the schizo

>>8732155
>untuned
kek

>comfy out of nowhere
nooticed

she went to bed

>>8732144
?
The OP literally has the answer to your question, doe?

>>8732168
It's called "having a constructive conversation" you dipshit, but you sloppers are incapable even of that.

>constructive
kek check out lil bwo over here

>>8732169
I'm not required to converse with you if your question has already been answered. I suppose getting ignored is the better response, yes?

Being called slopper from someone that thinks "kiato" is high art is rich

>>8732169
why do you dislike being spoonfed a guide?

are you talking to the walls brother?

a brick wall yes

retard

please melt :(

>>8732080
so random commentary blogposting:
basically used a 1:1 "normal" training config, just lowered steps to 32. don't touch LR's at all because if you try and inflate them it's just going to blow up anyway.
if you want it to work at all you pretty much need to use a trigger tag, I've been using "jz235" since it's a non-tag without anything that could bleed over into it. Also, since you have to use a trigger, you have to enable the TE(it improves things by a lot).
as expected it regularizes really poorly. Best case scenario you can supplement your prompt with a ton of in-model calls to random media tags to make the style adaptation a lot better but even that is really iffy. But this is to be expected since it's really just a super undertrained lora.
it does a pretty mediocre job as an i2i "filter" unless the source image being used is fairly adjacent to the style already. This is, again, because it's super undertrained and doesn't do regularization well, meaning it can't adapt very well so if the base image is too far off it'll just heavily default to what it already looked like.
no additional images to post because honestly that sequence demonstration is about what you should be expecting for the most part.
so ultimately while you could use this sort of technique as a knockoff vibe transfer/ip adapter it's.... really poor. It's ultimately really mediocre for straight t2i due to how stiff it is there and it doesn't work super great for i2i restyling because of reasons noted. While you can arguably make it work it's just not a replacement for an actual lora. Maybe if you want to accentuate real specific artistic qualities in something(and have a single image in a dataset you want to push even harder)? Super niche usecase though.

additional thoughts: I should probably dig out and retune a LoKr config. This sort of thing could arguably be usable for concepts.

thank you for listening to my ted talk, be sure to lick and subside for more.

>>8732219
did you ever update your lora presets? that's what i've been using but with very low LR rates (xerbatt numbers), however I don't add epochs as I don't know what it does. so far it's the only good preset I kept using via kohya gui.
>>8732129

*adapts your ip*

>>8732219
what the fuck are you even talking about

>>8732241
He's explaining why we shouldn't use comfyanon's single image lora training.

>>8732244
i don't see him using comfy thoughever

>>8732226
"if it aint broke don't fix it"
an epoch is just how many times training has gone over the entire dataset, so 32 epochs on a dataset of 32 images would be 1024 steps, making it 33 epoch would make it 1056 steps. But this is meaningless in regard to my config since I define training length by steps and the step count is tuned to really specific autistic degrees around the fitment process(it's at 2800 default end, and is intended to "catch" certain fitments along the cosine drop curve). epochs are (mostly) meaningless fluff when it comes to small scale single artist dataset training. Just keep datasets in a small/middle, heavily curated range(for refence low=16, middle=32, large=64+) and it should be fine.
as far as lr's go they should vary by dataset because reasons. range is usually 0.00006 up to 0.0001. I generally start at 0.00008 and revise from there.

though uh, I just remembered that my vpred configs on mediafire were bad since they still had noise offset values for some reason so I did just fix that I guess(I had already caught it on my own a while ago just forgot to upload a fixed version).

>trains on vpred+ztsnr using noise offset
>proceeds using eps 0.5
kekaroo

yes sir using da EPS 0.5 sir using model good sir

hgg and hdg should make a baby

What are the state of the art training configs for vpred? still edm2 lycoris?

>>8732267
good dataset

>>8732270
any dataset is good if your config can find the pattern in it lol

>>8732272
it won't find the pattern in the dataset if there is none lol

>>8732267
I like fftdescend