[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / r / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip / qa] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
Settings Mobile Home
/h/ - Hentai


Thread archived.
You cannot reply anymore.


[Advertise on 4chan]


File: wakanda3.png (2.9 MB, 1248x1824)
2.9 MB
2.9 MB PNG
Edition African princess

Previous Thread: >>7925070

>LOCAL FORGEUI **NEW**
https://github.com/lllyasviel/stable-diffusion-webui-forge

>LOCAL WEBUI SETUP *OLD*
https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki

>LOCAL COMFYUI *NOT RECOMMENDED*
https://github.com/comfyanonymous/ComfyUI
https://comfyanonymous.github.io/ComfyUI_examples/
https://openart.ai/workflows/templates

>LOCAL MODELS
https://civitai.com/models/257749/horsefucker-v6
https://civitai.com/models/288584/autismmix
https://rentry.org/5exa3 (hll+fluff) *OLD*
https://gitgud.io/gayshit/makesomefuckingporn (use wayback machine) *OLD*

>PAID
NovelAI: https://rentry.org/hdg-nai-v3

>RESOURCES
Features: https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features
Training: https://rentry.org/59xed3 | https://github.com/derrian-distro/LoRA_Easy_Training_Scripts
Tags: https://danbooru.donmai.us/wiki_pages/tag_groups | https://danbooru.donmai.us/related_tag
ControlNet: https://rentry.org/dummycontrolnet
IOPaint (LamaCleaner): https://huggingface.co/spaces/Sanster/iopaint-lama | https://www.iopaint.com/install
Animation: https://rentry.org/AnimAnon
Wildcards: https://rentry.org/NAIwildcards
Booru: https://aibooru.online
4chanX Catbox/NAI prompt userscript: https://rentry.org/hdgcb
NAI API/prompting in auto1111: https://github.com/Metachs/sdwebui-nai-api
Pony-related: https://rentry.org/ponyxl_loras_n_stuff | https://rentry.org/ponyxl_lora_previews | https://lite.framacalc.org/4ttgzvd0rx-a6jf

OP Template/Logo: https://rentry.org/hdgtemp/raw | https://files.catbox.moe/xk99gy.png | https://files.catbox.moe/35bdt5.gif

>Related
>>>/aco/sdg
>>>/g/sdg
>>>/d/ddg
>>>/e/edg
>>>/vt/vtai
>>>/vg/aids
>>>/b/degen
>>>/trash/sdg
>>
>no hdg logo
kys
>>
>>7927209
--------> /e/
>>
>>7927218
I will
>>
The fuck was this in the end?
>>7927089
>>7927114
>>7927121
>>7927129
>>7927143
Was it bait or something else? NovelAI doing shady advertising tactics?
>>
>>7927227
No, he was just joking!
>>
>no logo
>/e/ gen
>only one nipple piercing and it's got noclip on
unbased, i swear if you fuckers don't shape up i:m going to bake with my own gen in theop
>>
>>7927232
I wasn't
>>
>>7927218
it got stuck in her ass
>>
>>7927231
The guy is hacker/scraper.
>>
>>7927239
Scraper?
>>
>>7927231
>>ill try it out and PASS IT ON if you leave your (temp/alias) mail
aini clearly got bored of giving tossjob in the it department and came her eto amrket
>>
>>7927236
wtf that's rude, kys
>>
>>7927231
I got emailed a legit account with 3k anals (thanks!).
Dunno if it was an actual advertiser, though.
>>
>>7927231
it already happened before, someone gave out nai accounts with anals
>>
>>7927240
You can scrape the data off of websites that you can then use yourself.
>>
>>7927231
Wouldn't surprise me if 90% of the NAI schizos were 1 very dedicated shill samefagging. "They" have the same speech pattern
>>
>>7927248
How can someone scrape NAI logins?
>>
>no logo
>/e/ gen
>only one nipple piercing and it's got noclip on
>baked when 20 people were clearly sitting with thumbs up their asses
HOLY BASED
>>
File: 00017-2038835836.png (2.53 MB, 1792x1280)
2.53 MB
2.53 MB PNG
>>
pony knows princess peach but won't gen her without a fucking ponytail
im inpainting with long hair prompted and (ponytail:1.5) in negs at 0.85 and it won't budge
AAAAHHHHHHHHH
>>
>>7927244
wouldn't they have better things to do than come here like aren't they an actual company
>>
>>7927250
Based naitruther
>>
>>7927251
This why I also said hacker.
>>
File: 02269-1975926737.png (2.06 MB, 1552x1232)
2.06 MB
2.06 MB PNG
>>
>>7927254
a-anon?
>>
>>7927231
Kingdragon@live.de|TSyj@9qvR2L6nwf with 3000
ahhahahahahahahahahahahahaha@gmail.com|jinklebell1 with 3000
alan4friddle@gmail.com|Sonic101! with 3000
alex.ander14@hotmail.com|alexander1912 with 3000
alezampierif@gmail.com|molem2708 with 3000
allanimesubindo@gmail.com|opponeo3 with 3000
andrecruz662@gmail.com|Universo with 3000
mail|pass, first come, first served
>>
File: 00461-417417082.jpg (1.91 MB, 2560x2176)
1.91 MB
1.91 MB JPG
Ok anon, I hope you did your pushups. Take care of yourself
>>
File: file.png (21 KB, 1005x198)
21 KB
21 KB PNG
GUYS??
>>
>>7927262
nai gen, don't fall for the baiterino!!!
>>
>>7927256
mate it's been 2.5 years and they still haven't made v2 modules available
they spend 90% of time there masturbating in furry costumes
>>
>>7927223
Back when moot was still around, I saw someone ask him about making a futa containment board, and he said that the entire point of /d/ in the first place was to be futa containment, which I can accept. The part that irritates me is that despite being on /d/, there's almost never any sex, and the images have less imagination than /e/ even. It's so disappointing to see the same 1girl compositions over and over, it's disgusting.
>>
>>7927256
Unironically here seems like a good place to advertise.
A lot of normalfags are either still butthurt with AI or wouldn't know what to do with it.
>>
>>7927266
Firstworldies are done.
>>
>>7927266
>believing an extension literally with the word SAFE in it isnt safe
tech illiterate retard
>>
File: 1694330933914321.png (42 KB, 498x718)
42 KB
42 KB PNG
>>7927263
>isreal
this guy even has some eye bleach custom css theme
>>
>>7927265
Hot + can count to ten.
>>
>>7927276
lmao
>>
File: 00097-537665483.png (3.29 MB, 1536x2560)
3.29 MB
3.29 MB PNG
>>7927135
I ended up training it at 512 and it seemed to turn out okay. Incase anyone wants, Kirigakure Shura LoRA: https://pixeldrain com/l/QfQVVwZW
https://files.catbox.moe/9v7a32.png
>>
>>7927256
Companies are astroturfing all the time everywhere, welcome to the real world.
>>
>>7927276
Aww I bet he's just a little cutie, too bad he got hacked.
>>
>>7927265
Box please, I'll even do 20 more then.
>>
>>7927283
Those look like trial accounts, in the past NAI has given out steps to trial accounts (for LLM module training), but they carried over and became anlas
>>
File: s-14035504.png (3.18 MB, 1856x1280)
3.18 MB
3.18 MB PNG
My previous posts were a little /e/. I made sure this wouldn't be the case going forward.
>>
File: 00190-3509510065.jpg (325 KB, 2048x2048)
325 KB
325 KB JPG
>>
File: r16w5r bake 1 sample-s.jpg (2.69 MB, 4815x1895)
2.69 MB
2.69 MB JPG
>>7926042
>>7925445
5920 steps test (shexyo style lora settings), bake 1 of anon's r16w5r dataset
(sorry I don't know the artist's actual name because I am retarded)
>11 epoch saved
https://mega.nz/file/16kRkBTK#1e1-tw18V-274ZNF78-SpYyWIpsJ6uSI9lyMRx3wsrc
>If you try any of these, I recommend 0.8 weight as hands are coooked and this helps (a bit... theyre pretty bad)
>preview:
https://mega.nz/file/lyMz2bBK#0BLuT5c3Qvqe_B7UNZIIytOCAEULuusxlaHt2Sd5awE
I feel like it didn't really 'capture' the small details of the style and it is, of course, allergic to hands, so I'll be baking more with different settings. Shexyo lora json added to my mega folder of stolen jsons anyway. I think less than half the amount of epochs and a couple repeats would've been better with these settings for the dataset size (be warned beginning bakers, 5k+ steps is an all nighter for many)
>>7926961
some extensions don't work on forge.. but if you move them back to a1111 (assuming you still have it), you can just edit your user batch file for forge and in your command ARG, point folders towards the a1111 ones instead of moving it all over to forge.
>https://www.reddit.com/r/StableDiffusion/comments/1arfcp7/thoughts_and_query_on_forge/
(plebbit, I know, but the comment by 'Mutaclone' shows how to format it).
if you aren't using a 4090 or high vram setup, the speed increase of forge is worth sacrificing a few extensions for most users
>>7926966
bless you kindly, I truly love 1.5 sovl
>>
How do I use ControlNet openpose? Enabling it and putting in a control image doesn't seem to do anything. Any video guides don't really explain what the fuck is going wrong, also in the drop down menus for models there is nothing but people in the videos do? Where would one get these openpose controlnet models?
>>
>>7927256
Are you stupid? Actual companies have shills. Most of them anyway.
it's not an uncommon thing at all. They just hire a third party company to target different spots of interest.

You think profitable companies still buy cable advertising or something?
>>
>>7927275
You're telling me he got deleted for no reason? Its proof, bro.
>>
>>7927272
I mean it makes sense because /d/ in my mind was always just for the real abhorrent stuff like scat or guro. I like futa so I don't really see it as the same but I know other people vehemently despise it in a similar way as the former. It just ends up being way more of a popular fetish than others so it just overtakes the entire board.
/d/ kind of just seems like the the "it isn't vanilla so it ends up here" board which is weird. I'm relatively new to /h/ but even something like tentacles in my mind is quintessential "hentai" so to see that it doesn't belong here as well is just so odd to me.
>>
>>7927265
hot... can you make 10 rep of mating press her now?
>>
>>7927291
Wasn't Aini hired mostly to mock 4chinners and then send to Japan on NAIv1 model cash?
>>
>>7927287
shit joke and censored image, 10 points from hufflepuff and see me after class
>>
>>7927296
>see me after class
will you rape her? pls pls pls rape little girls children cunny rape kids
>>
>>7927297
Sir...
>>
>>7927287
>people complain 1girl is /e/
>demand a penis be added
>futanari, however, is for /d/
it's all so confusing. you'd think just slapping a dick on a 1girl standing would be all you'd have to do but then it just makes it even more confusing
>>
>>7927293
Scat and guro go on /trash/ only, not /d/. Tentacles is a weird exception and personally I'd allow them here but it is what it is.
>>
>>7927288
Slenderman armed bitch
>>
>>7927287
God bless you kind soul. Now anon can finally coom to this image.
>>
>>7927297
It's not rape if I mind control them first.
>>
>>7927293
Most board distinctions are kinda meh nowadays.
>>
>>7927300
>Scat and guro go on /trash/ only, not /d/.
Ah, shit. I guess I just have a super outdated or incorrect view of what /d/ even is because all I think of when I think of that board is that shitting dick nipples image. I don't go on that board often.
>>
>>7927285
MixDiffusers @ .4

Giving you pre-processed because it triggers hires setup otherwise and it's misleading
https://files.catbox.moe/izmyn1.png

Stay healthy
>>
>>7927301
Blind shizo anon is trying to speak
>>
>>7927305
There's plenty of gross stuff so it's warranted.
>>
File: catbox_jbzuik.png (2.66 MB, 1280x1792)
2.66 MB
2.66 MB PNG
>>
File: controlnet.png (698 KB, 1850x870)
698 KB
698 KB PNG
>>7927290
make sure if you're using SDXL you have downloaded the SDXL controlnet models:
>https://huggingface.co/lllyasviel/sd_control_collection/tree/main

I find thibaud works best for openpose
make sure you check the box at the top beside 'controlnet'
check all the things in pic rel that are squared off in red
use an actual photo for open pose, as it does not understand illustrations well. if you have no hands included, you can use the non _hands version
press the little exploding button thing the arrows are pointing at once you've input everything, so you can see a preview of what controlnet openpose has understood from your image (and check if its correct)
the preview will be beside the image in the big box
play with 'control mode' to see which results you like the best
play with ' control weight, starting control step, ending control step' if you want, usually not necessary
use a non ancestral sampler (example: euler instead of euler a) if it doesn't appear to be effecting the end result

controlnet depth also works well with pony if you don't want to use openpose

if you want to edit the preview image's pose, there should be a little icon beside it that opens up the openpose ui in another window
>>
File: file.png (5 KB, 294x82)
5 KB
5 KB PNG
Ah that's probably my issue then thanks. Yeah looks like i downloaded the SD 1.5 models and I kept getting this error but when I switched over the realisticvision to test it gave me body horror but at least it generated something.
>>
>>7927321
meant for >>7927319
>>
>>7927319
forgot to add:
once you download the models, put them in your 'webui->extensions->controlnet' folder
>>
>>7927321
yeah, it will give that nontype error if you mismatch 1.5 models with a sdxl check (or vice versa). should be good to go once you add the right ones
>>
Another thing I noticed is that SDXL loras fry worse than 1.5
1.5 would just get shitty clothing bleed, fried SDXL epochs are body horror
>>
I got one of the 9k anlas nai accounts but after playing with it for a bit I've decided it's worse than local. email me at
im6289aa@proton.me
and I'll pass it to someone else. one account so first response gets it.
>>
File: 2024-04-04 07_32_39-.png (454 KB, 1868x992)
454 KB
454 KB PNG
Any 4090 users know why some gens just flatline the GPU and basically freeze the PC when you try to do the hires passes?

Attached pic for reference; the settings of this image shouldn't be more than a 4090 can handle I'd think.

Catbox is the image I'm trying to replicate from an anon from a previous thread.
https://files.catbox.moe/w8s4y2.png
>>
I'm in actual disbelief how fucking stupid this general is
>>
>>7927330
>playing with it for a bit I've decided it's worse than local
How much images did you try? At what resolutions, settings? You know it still takes some getting used to, right? It's not a magic button.
>>
>>7927331
use Forge not A1111
>>
brainlet here, how can i set adetailer's default inpaint width/height to be 1024?
>>
>>7927332
then leave and raise the iq
>>
>>7927331
>4090
>light theme
This isn't a place for productive members of society. Please leave
>>
>>7927336
ui_config.json
>>
>>7927331
use tiled vae
>>
File: kanna.png (1.48 MB, 768x1280)
1.48 MB
1.48 MB PNG
Trying to prompt Kanna Kamui on Pony but it keeps giving her Ilulu's twintails, what do I put in the negs to fix it? "Twintails" doesn't work.
>>
>>7927338
Someone using a 4090 for anime titties isn't a productive member of society.
>>7927335
Forge is a different UI for Stable Diffusion? Does it resolve maxing out the GPU issues, or are we hating on A1111 (actual question, not sass - I'm unfamiliar with Forge).
>>
File: 1696541962707767.png (25 KB, 162x121)
25 KB
25 KB PNG
>>7927341
christ
>>
>>7927333
Made like 10 or 15 images. has really poor prompt comprehension. fantastic artist and character knowledge.
>>
>>7927341
Try putting the other character in negatives?
>>
>>7927345
it's nai
>>
>>7927346
>has really poor prompt comprehension
It just only understands Danbooru tags, and it understands them quite well. You didn't prompt it with natural language, did you? Can you share some prompts you did?
>>
>>7927209
>LOCAL COMFYUI *NOT RECOMMENDED*

I'm out of the loop. Why is comfyui not recommended? A while back I remember this board shilling this UI as the best.
>>
>181 full test grids
I'M XYYYYIGGG
>>
>>7927348
>768x1280
>overfried style (just look at the ass)
surely its nai and not local...
>>
>>7927350
its obsolete
>>
>>7927328
as >>7927289
who has fried their fair share of loras during setting tests, I think it's honestly just that we slap a super high step count on sdxl lora training when it isn't necessary. even the re-entry recommends 2.5-3k steps, where anons have shown you can get very good results anywhere from 600-1.5k steps. I can't really blame SDXL here, there isn't really agreed on golden bake settings, it takes longer to bake/reqs more vram/gpu, and so there is less testing.
>>
>>7927344
forge is an A1111 fork that isn't a pile of shit, and is explicitly better at sdxl stuff
>>
>>7927345
I'm not gonna bother fixing and touching up gens if I can't even get the hair right first.
>>7927347
Will do.
>>
>>7927330
yeah same kek, maybe im just a garbage prompter
thanks to the people sharing the accounts though, even if it really is marketing
>>
>>7927354
Oh for sure, I think most of my bakes ended up requiring below 60% of my steps. I have to recheck and adjust my settings later.
But SDXL lora behavior is definitely way different than 1.5
>>
>>7927339
thanks, found the settings with ctrl+f
>>
>>7927349
>It just only understands Danbooru tags, and it understands them quite well
'Member when in the first days of v3 there were attempts to unironically spread the meme that it's good at natural language? And some say there were no shills here kek. I actually subbed to my first month partially cause I believed the meme and was hoping I could do without the tag prompting for that month at least. Little did I know I was getting into tag hell once again.
>>
>>7927340
Ill give that a try
>>7927356
And this as well
Thanks for the help.
>>
>>7927341
She's pruned because it's a cuck model. You have to prompt all her features and not just her name. Even then good luck getting consistent results
>>
>>7927349
I know my danbooru tags. it did fine with 1girl but lost it's mind with 2girls, had no idea what a strap-on is, etc.

>>7927359
it's not marketing. they're stolen. if you don't enjoy it pass it on to someone else like I did. spread the joy.
>>
File: 00243-1230667523.png (2.83 MB, 1248x1824)
2.83 MB
2.83 MB PNG
>>
Any tips on training styles?
Mine always feels underbaked, and give some prompt bleed (maybe its a tagging issue)?
Like, one of the girls of the dataset has a braided bang, and when I prompt a girl with a braid, side braid, etc, she gets a braided bang when using the lora.
>>
>>7927360
yeah, I can't really deny that. I even miss 1.5 sovl a lot of the time desu, but when I remember fighting with anatomy every single gen and the horror of hands, I'm stuck with SDXL. I wonder if cascade will be any different
>>
>>7927350
It's just banter to make the ComfyUI developer seethe, since he is kind of an asshole.
Use whatever works best for you.
>>
>>7927370
this looks weird.
>>
>>7927373
Using a artist tag helped mine but the artists also differ a lot. Like, I have 1.5 loras baked perfectly on the same datasets that just refuse to work on SDXL even when others work good. Plus the quality tags have their own fried style.
>>7927374
The lora baking was easier for sure. Had it's weird logical quirks but SDXL is just way less consistent imho. The charas are pretty tight tho.
>>
>>7927371
I got your freebie (thanks once again) and have a bit of understanding how to use it, what were you trying to prompt?
Although it sure is less capable with smut concepts, but I enjoy it more as a 1girler desu.
>>
File: 00150-2972589237.png (2.46 MB, 1304x1952)
2.46 MB
2.46 MB PNG
Gobbussy sex at the beach
>>
>>7927373
that sounds overfitted desu. if you don't have it in your dataset already, try to add some images of what a braid, side braid, etc is and tag them so the AI can understand the difference better
(if not possible, add braided bang to negs I guess?)
if you want some jsons to try for style baking, I've been collecting/testing (stolen from metadata and posts). they're named after whatever I stole them from, generally (mostly from the re-entry but some from civit or posts, updating daily)
https://mega.nz/folder/429HjaaI#vel37O61SS1_mf2ZRNfSJA
>>
File: kanna.png (1.31 MB, 768x1280)
1.31 MB
1.31 MB PNG
>>7927347
I'm dumb, that worked, thanks. Back to the lab I go.
>>
>>7927333
NTA. The model feels pretty good, but the cloud service part is giga cucked, and it doesn't help that these accounts don't have unlimited gens to just mess around like an actual sub.
>>
File: 00024-654258301.png (2.67 MB, 1280x1792)
2.67 MB
2.67 MB PNG
>party member insists on inspecting your gear
>>
>>7927350
just some retard baker with an axe to grind, ignore
>>
>>7927371
yeah i sent it to the protonmail guy
>>
File: ComfyUI_01491_.png (3.57 MB, 1536x2048)
3.57 MB
3.57 MB PNG
>>
File: 00109-2072658544.png (2.62 MB, 1304x1952)
2.62 MB
2.62 MB PNG
Bnuuy
>>
>>7927289
didnt work mutaclone dudes comment. i tried but oh well. the only problem for me is for some reason multidiffusion on forge doesnt work, but on my a1111 setup it works good. that shit made my stuff really fast. The resolution doesnt become i what i set it to i hate my computer with 1050ti so much it's unreal :/
>>
>try nai
>don't try their inpainting
bravo
>>
>naisu its been an hour, bake should be done!
>go to check on it
>error on first epoch
>I forgot to specify the output folder
lmao fuck me
>>
File: 00554-3642763965.jpg (1.94 MB, 2816x2048)
1.94 MB
1.94 MB JPG
Squats are next anon, will you help?
>>
File: 00187-1070365140.png (2.65 MB, 1280x1792)
2.65 MB
2.65 MB PNG
>>7927379
don't tempt me
>>
>>7927394
On my face?
>>
File: wakanda4.png (2.74 MB, 1248x1824)
2.74 MB
2.74 MB PNG
you are so mean with me, I'm going to /b/
>>
>>7927401
this image made my eyes physically hurt
>>
>>7927263
>alex.ander14@hotmail.com|alexander1912
WTF that's my mail ?????
>>
>>7927401
Bye.
>>
File: 00073-2108482676.jpg (206 KB, 1152x864)
206 KB
206 KB JPG
>>
>>7927406
Thanks for the anlas, looser
>>
>>7927410
nooooo gib it back ?!
>>
>only three datasets i care enough to bake left right now
Peace.
>>
>>7927411
I coomed all over them, you sure you want it?
>>
Guess this is a good chance to try straight txt2img at their high resolutions.
>>
>>7927416
nice seibas
>>
File: highlights_507.png (2.69 MB, 1543x1133)
2.69 MB
2.69 MB PNG
>baker makes previous thread highlights

would this work or is /h/ too fast?

also, b4 u @ me, chill tf out it's not official just me being bored.
>>
>>7927418
where is my gen? why didn't you include it?
>>
File: 1700209245157325.png (2.27 MB, 1344x1728)
2.27 MB
2.27 MB PNG
>>7927413
>Ran out of datasets
>rebaked all my older stuff
>Haven't baked in 2 days
I'm jonesing since I haven't left my computer idle like this in a month or two now and I've built this autism that something has to be cooking while I'm not using the computer.
>>
>>7927418
if you have enough autism and free time, go for it. but also
>can't even get chosen for fucking highlights
GOD DAMN IT NIGGER SHIT DICK FUCK
>>
>>7927418
>the least /h/ gen is the best one
>>
>>7927425
>>7927422
Bad gen bros...

>>7927418
I like them their fun in /e/, but /h/ is pretty fast so ganbare if you're that autistic.
>>
File: 00123-36727998.png (2.15 MB, 1344x1632)
2.15 MB
2.15 MB PNG
>>
>>7927424
bake this anon's no one else has yet (I planned to but am running behind on my bakes)
>>7925585
>>
>>7927418
Nice selection of ABSOLUTE TRASH.
>>
File: luhv.png (2.69 MB, 1543x1133)
2.69 MB
2.69 MB PNG
>>7927418
these 3 give me feels
>>
>>7927424
I used to have that with 1.5 but SDXL baking is meeeh.
I have 10k images worth of dataset in my obsolete folder lmao.
>>
>>7927433
>not posting his 1.5 datasets so others can do bakes of them
but why
>>
why the ragerinos boys don't make prompts?
>>
>>7927417
Thanks.
>>
>>7927436
I wouldn't either, shit's like 55gb.
>>
>>7927418
Go for it if you have the time.
I know the anons that replied to you are joking, but I just hope it doesn't result in unironic petty drama.
>>
>>7927418
>3(three)[tres]/drei/ 1 girl,standing,looking at viewer
nice /h/ hightlight, retard
>>
>>7927441
I will heavily scrutinize it and hope it doesn't follow the trend of the same 4 anon's styles since I see bleed over from /e/.
>>
>>7927441
I wasn't joking, half of these are not even /h/ and not even good, just generic 1girlstanding. Are you guys baiting?
>>
last thread someone posted some hashed codes they liked, but I can't find the god damn post.
ctrl f
hash, code, tag
no luck.
>>
>>7927445
What, you missing your cocks?
>>
>>7927394
Fucking hate squats. Worst exercise by far.
>>
File: civitai.png (29 KB, 1571x470)
29 KB
29 KB PNG
civitai bros?
>>
>>7927435
heh... I saw that.
>>
>>7927447
So you're baiting, I see. Get a life.
>>
>>7927452
who?
>>
>>7927452
wut
who has a meltdown?
>>
>>7927452
who dis
>>
>>7927451
Heavy squats are soul crushing but deadlifts and OHP are king.
>>
>>7927452
hOLY BASED
>>
>>7927444
It's not a trend, there's like 5 anons tops posting quality stuff on /e/ and they mostly do same styles, everything else there is absolute slop. Why would you include slop in your le threaderino highlighterinos?
>>
>>7927452
>seething esl
kek
>>
>>7927452
I agree with him about censorship, I can't properly express my live for cunny on Civ. Such a shame.
>>
File: 00029-624423814.png (2.93 MB, 1408x1920)
2.93 MB
2.93 MB PNG
>>
File: file.png (13 KB, 873x224)
13 KB
13 KB PNG
>>7927446
found it by remembering sakimi funnily enough
>>
>>7927452
That last part was kinda rude...
>>
>>7927458
Bench and pull ups for me
>>
>>7927436
I can post it when I'm done but most of it is obsolete chara/concept stuff and now untagged.
>>
>>7927440
>pick favorites
>maek mega
>upload 20GB worth
>????
>>
File: anila 8.png (3.1 MB, 1496x2072)
3.1 MB
3.1 MB PNG
List of lora that need to be made, V2:

Paizuri invitation (concept) - Civitai has one. Might affect style too much.
Skintone slider (tool)
Age slider (tool)
Standing shota kneeling onee-san paizuri (concept)
Omizu_(h2o)
Motsuaki
Tomiokasena - Civitai has one under Tomioka Sena
Yuyu_(yuyuworks)
Suurin
UTTA/Utterangle
Rikose
Mamuru
Kaptivate - Civitai has one
Marui Maru
Kedamono Kangoku-tou
Anila (Granblue Fantasy) - Anon made one, mostly there. Civitai has a new one that seems better.
>>
>>7927455
>>7927456
>>7927457
https://civitai.com/user/iselestia/models
no idea honestly
>>
>>7927463
Isn't it enough to just keep your previews sfw in order to avoid removal?
>>
>>7927460
I mean that's the thing, shit you call "the 5 best" I could think a few are complete slop. That's why it's nicer as a tool to highlight different stuff that looked good. But go for it mang.

>>7927469
If it's a favorite I already made the lora, rest are just sitting there. And still why would I upload 20gb of pictures, if someone wanted to make a lora of that chara / style just use img grabber, take 10mins to download 200 images would be faster.
>>
File: 00579-1235363860.jpg (1.9 MB, 2048x2816)
1.9 MB
1.9 MB JPG
>>7927400
Ok anon, since you asked nicely
You need to promise me 10 more pushups when we're done
>>
>>7927418
box for top left please?
>>
>>7927418
>>7927432
>>7927427
>>7927441

When making it, I rate it by

>style
stuff that looks unique/cool/new
>technical skill
fucked hands, anatomy, or out of frame dicks- bad
>anyone who makes it can't include their own gens
some neutrality

stuff i don't think should matter
>board rules
don't care if it's actually /h/ or not. obv no /d/

>>7927425
>>7927422

sorry boys :(
>>
>>7927471
Just some uoohhh uploader. Nothing of value lost
>>
>>7927472
Yeah but they banned the word loli, the fuckers.
>>
>>7927394
gymbox?
>>
>>7927418
these are always fun. people here have strong opinions so it might cause some butthurt but c'est la vie
>>
File: LIVEONTWITCH.png (3 KB, 218x72)
3 KB
3 KB PNG
I thought april fools day was over, civit thinks its all month?
>>
>>7927480
It's the same base as
>>7927265
>>7927306
>>
>>7927441
>I just hope it doesn't result in unironic petty drama
it will
although it might be a good idea to make people put more effort or post more interesting and varied gens
but you can bet on the drama
>>
>>7927473
I was talking about /e/, here in /hdg/ there's far greater variety of styles from way more posters being posted day in and day out. There, not so much. And I disagree that just cause someone applied a different style lora it instantly makes it a highlight, any retard can do that.
>>
File: penis_for_anon.png (3.09 MB, 1248x1824)
3.09 MB
3.09 MB PNG
>>7927454
here your cock
>>
>>7927482
What's the joke there?
>>
>>7927490
civitai
>>
>>7927445

Look through the last thread, there really wasn't much /h/ to begin with, and obviously it's my personal opinion on the matter.
>>
>>7927491
lmaoooooooo
>>
>>7927452
He's right objectively, why I don't post my stuff there.
>>
>>7927418
Gay popularity contest. Everyone will complain because those who know shit on those who don't know, and those who don't know, don't know they don't show, so they shit on those who know.
IE, some of your highlights are low effort garbage and you don't even realize.
Nah.
Keep this homo shit to a minimum.

Keep this social shit out of here. Post cool pics, enjoy cool pics. I am content with remaining anonymous
>>
>>7927495
>t. pedo
>>
File: catbox_cuqeev.png (2.52 MB, 1280x1792)
2.52 MB
2.52 MB PNG
>>
>>7927496
based
>>
>>7927496
Eh, I think it will actually be funny when attentionwhores will start getting at each other's throats over highlights, but in the end it'll probably the sane anons who will lose as always though.
>>
>>7927474
holy shit more please maybe give her some d
>>
>>7927488
hello overexposure-kun
>>
>>7927496

uh-oh somones gen wasn't included in the OFFICIAL HIGHLIGHTS /h/ world championship.
>>
>>7927506
It did, but you're still a faggot
>>
ebin updoots to the left
>>
File: 06578-00044-1610851136.png (2.29 MB, 2048x1536)
2.29 MB
2.29 MB PNG
>>
>>7927484
>7 loras
breh
>>
>>7927507

tell me which one

So that we know, to not show, to those, who know, but don't want to show, to those who know
>>
>>7927508
We are also starting a collage of highlighted posts and good news Sir yours will be the first one to make it!!!
>>
>>7927512
Read my name, queer.
I don't care about your social shit
now fuck off
>>
>>7927513
Isn't that already a thing?
>>
>>7927516
Text posts, Sir.
>>
>>7927418
Why is there more [Local/NAI] gens than [NAI/Local] gens???
Fuck off, you retarded [paypig shill/aco lover].
>>
File: catbox_sq0n3i.png (2.57 MB, 1280x1792)
2.57 MB
2.57 MB PNG
>>
>>7927517
Yeah.
>>
>>7927520
Okay you know what, your postings are now banned from our highlights, good day Sir.
>>
>>7927509
Ranma?
>>
>>7927499
>>7927519
Brown girls with braids is next level kino.
>>
>>7927523
Yes, using tottotenero style.
>>
>>7927514

very well, I know, to not show, the queer shit to those who wish to not know and those who shit on the knowers

>>7927518
The satirical criticism is understood, fellow knower.
>>
>gyaru manba lora on civit
I'm
ready
>>
>>7927524
great taste. I would love to do more but pony isn't that good at getting large voluptuous quad braids.
>>
>>7927527
post your gens
>>
>>7927526
you're a faggot on an imageboard playing at social organizer.
Nothing you say carries any weight, you dumb lolcow
>>
>>7927531
Holy bas'd.
>>
>>7927531
basadero
>>
File: shitbed.png (24 KB, 627x359)
24 KB
24 KB PNG
>>7927482
Joke's on us I suppose, cus it just shit the bed.
>>
>>7927531
pwnd that newb attentionwhore
>>
>>7927496
highlights have no effect on anonymity
>>
Do any of you use an online service like Paperspace or Vast.ai, something like that to proompt? My laptop's not powerful enough I am afraid so I have to resort to Vast.ai but the lora trainer there doesn't work for me neither does Google Colab anymore
>>
>>7927545
$25 -> NAI
Pay up.
>>
>>7927525
link to the tottotenero lora you're using?
>>
File: 1692575747230300.png (1.48 MB, 1248x1824)
1.48 MB
1.48 MB PNG
>>7927498
NTA but same, never posted my loras there never will
people upvote the most fried resolution and tag dependant shit ever (wagashi XL lora for example), there's no implicit or explicit quality control
>>
File: 00596-4119774264.jpg (1.83 MB, 2048x2816)
1.83 MB
1.83 MB JPG
>>7927503
Anon!
It's a bit... small.
No, don't look sad. In this Isekai, pushups make your manhood grow!
Let's do 10 more pushups
>>
afrobull https://gofile.io/d/Z7Jf2I
>>
>>7927547
>Significantly more expensive than the 10 bucks I put into Vast.ai which last me a month or more depending on my usage
>Can't make LoRAs of my characters
thanks but no thanks

>>7927551
>>7927549
>Discord
ehh I guess if I have to... if you can use Pony and make LoRAs aside from usual proompting then I'll hop to that one
>>
Surprised to see that shit doesn't fall apart at straight txt2img 1280x1856 without SMEA. Some oddities with proportions maybe, but it's coherent. Could they be using something like kohya fix for high res? I wonder.
>>
>>7927525
Nice.
>>
>>7927554
How'd you get that hand gesture?
>>
>>7927558
I hate them so much for giving us only base res for fucking $25. It's lacking the detail and refinement so much. Not as bad as local is at the same res, but local is also fucking free.

>b-but your 10k anlas
I'm 3 days into my resub and already spent 2k and I've enhance-upscaled like 5% of the gens I liked and saved max.
>>
>>7927555
So, fun fact, my 1.5 standard setup was 100 epochs, save each 10. I checked my saved epochs just now and
For the styles: 1x 10, 5x 20, 2x 30, 2x 40, 2x 50, 1x 60
For charas: 3x 40, 2x 60, 1x 70
Definitely gonna lower my max epochs to 70 on XL.
>>
>>7927564
Yeah, 10k of anlas isn't really that much
>>
File: 06567-00044-1610851136.png (3.33 MB, 2048x2048)
3.33 MB
3.33 MB PNG
>>7927548
It's the same on the rentry. I just rename them for my own cathegorization.

https://files.catbox.moe/30g21q.png

https://files.catbox.moe/sc7pis.safetensors
>>
>>7927563
I prayed to the lord of stones
>>
>>7927564
You can get something good at base res with the right prompt and settings after that just enhance at 1x for free and then 4x upscale for 7 anlas.
At least that's what I'm been doing now. Peviously I gen'd above base res and was burning through anlas.
>>
>>7927564
I agree, it seems like it's (with whatever the bandaids they apply) straight up better when you gen at 1024x1536 and higher, with smea especially. It could truly shine with local but alas.
I wonder if they are getting a lot of revenue from actual whales (the ones who'd buy tens of thousands of additional anlas).
>>7927574
Anon, 7 anlas upscale is just some crappy GAN upscaler, pretty much a scam. And 1x "enhance" is just img2img.
You can do a tile upscale for "free" though with the forge/automatic extension.
>>
File: 1704721211397956.png (1.81 MB, 1248x1824)
1.81 MB
1.81 MB PNG
>>7927582
I have freeu and khoya hires fix enabled
>>
>>7927549
>>7927551
actually hilarious how much assfleshlight has started trying to advertise here lmao. honestly, if anons are dumb enough to pay for his dicksword bot then maybe they deserve to part with their $$
>>
File: anon_enjoys.png (3.04 MB, 1248x1824)
3.04 MB
3.04 MB PNG
>>7927505
open wide
>>
>>7927416
>straight txt2img at their high resolutions.
this killed nai for me
>>
I wish the ADetailer only worked on a specific character, or specify which one is to be detailed. It keeps giving my dudes makeup.
>>
What was problem people said happened when you make a lora to overwrite an existing tag? I got so close with my style lora but something is fucking up the comprehension.
>>
File: 1697434062134396.png (1.81 MB, 1536x1536)
1.81 MB
1.81 MB PNG
what the fuck
>>
>>7927592
Click where it says "Detection" and set:
>Mask only the top k largest
to 1 or however many girls you have in your gen.
>>
File: 00139-2333809006.png (2.28 MB, 1344x1632)
2.28 MB
2.28 MB PNG
>>
>>7927592
Just use inpaint my guy.
>>
>>7927592
go to adetailer settings, make it do left to right. use the [SEP] tag to separate faces.
smirk [SEP] crying [SEP] laughing
>>
>combo of artists work well with base autismmix but shits itself with a character lora
pain
>>
>>7927441
>but I just hope it doesn't result in unironic petty drama.
I was gonna post this because that's exactly what it will become. It's nice when your gen is in the OP but it's literally 1/301 so people aren't salty if they don't get picked. If you decide to do highlights there will be people seething both in the thread and behind their screens which will worsen the vibe of this place.
>>
>>7927592
to prevent that let the gen go untouched then I use adetailer in img2img. Press 'skip img2img' and if it detects a dude you can simply use photoshop to undo it.
>>
>>7927568
In steps
style: low 800, avg 1850, hi 3300
chara: low 1240, avg 2635, hi 4560
for 1e-4
which is why i don't really like step targets
>>
man the guy that trained the truckkun lora actually did a good job, it is good even for nai genned images + prodigy
>>
https://civitai.com/models/321768/realistic-xl-indian-model?modelVersionId=360799
>Realistic XL Indian Model
>158 downloads
>10,000 online use
Kek.
>>
File: 02165-00044-1610851136.png (3.78 MB, 2048x2048)
3.78 MB
3.78 MB PNG
My years watching shoujo anime have influenced me.
>>
>>7927564
I haven't been able to run out of anlas but i also haven't had as much time to goon out the last couple of months
>>
>>7927615
good morning sirs
>>
>>7927615
the guy making pajeetas with hairy armpits is based at least
>>
feets
>>
File: 00639-3078987883.jpg (1.81 MB, 2048x2816)
1.81 MB
1.81 MB JPG
>>7927554
H-how many pushups did you do!?
10? no way.
It has to be at least 11 ..or even 12!
>>
>>7927593
Please respond. I'm so close but I can't figure out why this isn't working.
https://files.catbox.moe/pv0cfc.png
>>
>>7927591
Doesn't seem like it's THAT much better for smut but feels like backgrounds will benefit from higher res+smea. Gonna try it out later with some artists that do backgrounds I guess.
>>
>>7927617
more like katawa shoujo
>>
File: 1704282129502271.png (2.58 MB, 1344x1728)
2.58 MB
2.58 MB PNG
I am once again asking for your suggestions on artists / charas. Will do it if I like them, charas are more lenient though.
>>
>prompt "looking at penis"
>girl start looking at viewer
Am i really getting called a dick by my porn ai?
>>
>>7927629
Izumi from Blue Archive. Would love to gen all that cute meat
>>
>>7927617
drills ojou always loses to plain main girl, though
>>
Is it just me or pony can't into fantasy outfits?
>>
File: 02078-00044-1610851136.png (988 KB, 768x1024)
988 KB
988 KB PNG
>>7927636
NOT IN MY HEART
>>
File: 00013-2533300525.jpg (346 KB, 2304x1792)
346 KB
346 KB JPG
I don't want to go back to the wage cage tomorrow
>>
>>7927642
based ojou-sama avenger
>>
>>7927639
Seems... competent?

https://files.catbox.moe/ucyed1.png
>>
File: 00288-303121722-0000.jpg (2 MB, 1632x2456)
2 MB
2 MB JPG
>>
>>7927418
Too fast.
>>
>>7927629
Wanna do any fate chars?
>>
>>7927648
I meant more like mage robe swirlies type stuff.
>>
LEAKED
>>
>>7927639
>fantasy outfits
like what?
>>
File: 00039-68660291.png (2.02 MB, 1248x1824)
2.02 MB
2.02 MB PNG
>>
>>7927657
really nice outfit.
>>
>>7927662
(toned body:1.1), (removing pants:1.4), (bulletproof plates:1.2), (tactical vest:1.3), (skintight_fishnet_bodysuit:1.2)
>>
>>7927654
>swirlies
What do you even mean?

https://files.catbox.moe/w5fufc.png
>>
>>7927642
based and drillpilled
>>
>>7927663
>skintight_fishnet_bodysuit
Neat. I'll keep this one in mind. btw nice style.
>>
File: 00019-1888105639.jpg (465 KB, 1664x2432)
465 KB
465 KB JPG
>>7927639
use loras, preferably of artists doing fantasy stuff like oda non
>>
File: 00233-3417883313.png (3.59 MB, 1248x1824)
3.59 MB
3.59 MB PNG
>>7927667
i made it in, every style (was testing loras)
>>
File: 00288-731169050.png (953 KB, 832x1216)
953 KB
953 KB PNG
>>
File: 00617-2980148197.jpg (1.77 MB, 1408x2400)
1.77 MB
1.77 MB JPG
>>7927623
>>
>>7927577
I wanna sodomize shiburin
>>
>>7927670
what style is that?
>>
>>7927673
nice
>>
>>7927613
>training a fucking lora on an AI style even though everyone already knows the secret sauce
>>
>>7927673
how do you manage so many views?
>>
>>7927681
I whisper to the stones
>>
>>7927681
((((((((((multiple views))))))))))
>>
>>7927679
those nai3 based loras work surprisingly well while mixing artist loras on pony is often really annoying
>>
>>7927677
Xenotrip
https://civitai.com/models/366990/pony-custom-styles

https://files.catbox.moe/2n5mww.png
>>
>>7927679
if everyone knows the secret sauce why is there retarded naigger coming here every thread asking about this mysterious undecryptable mix
>>
File: 00333-538175941.png (1008 KB, 832x1216)
1008 KB
1008 KB PNG
>>
File: 1695064298001790.png (2.6 MB, 1344x1728)
2.6 MB
2.6 MB PNG
>>7927653
Figured you'd tell me I'm just taking a list in notepad, most fgo's are promptable base in pony though.
>>
>>7927686
oh yeah i already had this lora so i thought it was this one, ty
>>
File: 00048-2215066032.png (1.44 MB, 1216x1024)
1.44 MB
1.44 MB PNG
Why some of my generations can take 40 seconds and others 11 fucking minutes?
Is my GPU dying? I have a 3060
Not using forge btw
>>
>>7927689
nta but my wife is promptable naked but her outfit are a mess and naked gen get boring, i should probably make a dataset for her considering the amount of picture i have of her honestly
>>
>>7927691
It's a memory choke with forged, I see it sometimes you cancel the gen or just reload the ui. Still don't know what causes it.
>>
>>7927691
You're probably going OOM and then falling back to system RAM.
Do you have "System Memory Fallback" enabled on your Nvidia control panel?
>>
https://www.ic3.gov/Media/Y2024/PSA240329
>>
>>7927691
Maybe running out of Vram on the one that takes forever. Use Forge.
>>
>>7927695
think of the latents, matrices and differential equations rights!
>>
>>7927689
>Figured you'd tell me I'm just taking a list in notepad, most fgo's are promptable
>base in pony though.
I wish. Don't know why, but it doesn't even know half of them.
>>
>>7927699
whoops meant for >>7927697
>>
>adetailer break faces instead of fixing faces
it happen when I use adetailer during img2img upscale (upscale at noise .3), anyone had the problem? how do I fix this?
>>
>>7927104
>>7927265
Based
>>
>>7927679
I mean it does a better job capturing the style vs other loras
>>
>>7927691
are the parameters the same and just the seed change? I noticed slow down with some lora and speed changing quite a lot gen to gen
>>
>>7927696
should I be concerned?
>>
>>7927691
Heh... sorry bro I was just borrowing a bit of RAM for my gens. Didn't think you would notice.
>>
>>7927703
Maybe the LoRA or artist tag you're using is too strong.
>>
ok but what is the best sampler?
>>
>>7927711
I tried with an old pic and I also have the problem, only parameter I changed is model from AutismMix pony to dpo, can it be the cause?
>>
forget highlights, make collages of the worst garbage in each thread instead
>>
>>7927715
Euler Karras DPS Ancestral C++ (No browns)
>>
>>7927623
Oh, that's Nice, can u share box of this one pls?
>>
>>7927718
When will the jews go mask off and release the no whites sampler?
>>
how many of (You) post here while drunk
>>
>>7927719
https://files.catbox.moe/jnhdep.png
>>
File: 06608-00044-1610851136.png (3.2 MB, 1536x2048)
3.2 MB
3.2 MB PNG
>>7927693
You need more samples of her clothing. Did you properly tag your lora? You need to be specific on colors and type of attire.
>>
>>7927716
Maybe. Try simplifying your prompt as much as you can (for the upscale only) and see if it helps at all.
>>
>>7927708
>>7927710
>>7927694
>>7927695
>>7927697

It seems that restarting the pc fixed the problem.
Don't leave your computer running for many days.

Thanks anons.
>>
File: 1709441135890469.png (2.37 MB, 1344x1728)
2.37 MB
2.37 MB PNG
>>7927701
>still didn't tell me who
Brother I
>>
>>7927696
cool pedo shit, pedo

die
>>
>he clicked
>>
File: 00009-2715809657[1].png (2.5 MB, 2048x2048)
2.5 MB
2.5 MB PNG
>average truckkun gen
>>
>>7927733
April fool's was yesterday though
>>
>>7927593
elaborate
>>
>>7927733
seems better than usual not sure why
>>
>>7927730
Alright, top of my head: atalante and alter version, martha, altera, medea, Brynhild and mata hari.
>>
File: 00289-826070038.png (2.44 MB, 1248x1824)
2.44 MB
2.44 MB PNG
>>
>>7927622
>>>/aco/8092173
>>
>generating thick milfs with veiny tits and sweaty feet
this shit's like crack
>>
>>7927306
>MixDiffusers @ .4
What's this?
>>
>>7927733
Is that even the right style? looks too clean. Also someone should go on his discord and tell him about the leaked passwords and the loras trained on his style.
>>
File: 00660-1293989396.png (1.11 MB, 1024x1024)
1.11 MB
1.11 MB PNG
>>7927744
For real, I haven't been able to stop hitting the generate button since this morning. Perhaps this will be the closest I ever get to the "press button, receive dopamine" experiment.
>>
>>7927744
Based
>>
>>7927747
perfect tool for knee pussy enjoyers
>>
>>7927752
Is it referring Mixture of Diffusers in the Tiled Diffusion extension?
>>
>>7927753
yes
>>
File: 1698370042479732.png (1.41 MB, 1024x1296)
1.41 MB
1.41 MB PNG
>spend 10 minutes just inpanting a gen because my autism demands it's as faithful to the original design as possible
>>
>>7927751
with all due respect, and I can't stress this enough: what in the cinnamon toast fuck is wrong with you
>>
>>7927754
And I guess the .4 has to be the denoise?
>>
wtf? linaqruf is "ani"?
>>
File: 00255-4016186491.jpg (778 KB, 1664x2432)
778 KB
778 KB JPG
it turns out the training lora i was using was just shit and masked loss may not necessarily take any/many more steps than usual. i guess this is what i get for bulk training overnight and not thoroughly testing the results out.

this is nice though since it'll make an oyakodon lora easier to train.
>>
>>7927762
crank it up higher for even more pussy
>>
>>7927756
Worth it, beautiful hag
>>
Fart lora doko?
>>
>>7927738
I baked a lora using game screenshots and as a result the dataset is filled with well known characters.

https://files.catbox.moe/d7dcu2.png
https://files.catbox.moe/2ilzlo.png
When prompting 1girls or characters pony already knows, the style comes out correctly and seems to be consistent so far (I might still need to change the dataset).

https://files.catbox.moe/42hmji.png
https://files.catbox.moe/rwl9l9.png
When prompting characters from the dataset AI fucks up completely, but its errors are limited mainly to the face and overall style. The outfits are pretty decent. Dataset size is 468 and I'm debating whether or not to cherry pick a few and keep it at 100.
>>
>>7927738
>>7927770
I vaguely remember someone saying something about an issue they had with a character lora fucking up because pony knew the character already but I don't remember. I also feel like a better dataset strategy would be to use a "sacrificial lamb" character that I will never prompt and get the style from mostly their pictures. I've made successful style loras before using pokemon characters so I feel like this should work.
>>
>>
>>7927770
damn that is so fucking weird, not sure i've ever seen results that look like that, it's probably because at least some of those have a shit ton of fanart already in the model. can you try pruning the character tags and then making a separate character lora for ff7r tifa or something?
>>
File: ComfyUI_01685_.png (2.93 MB, 1536x2048)
2.93 MB
2.93 MB PNG
>>7927654

https://files.catbox.moe/1xhl9g.png
https://files.catbox.moe/yfkx82.png
https://files.catbox.moe/dbthdn.png
https://files.catbox.moe/g2wroc.png
https://files.catbox.moe/roa2n8.png
https://files.catbox.moe/e78x3e.png
look I don't know how this thing came into existence but if I have to see it so do you
https://files.catbox.moe/0eugnw.png
>>
>>7927770
actually i take it back, i had a similar problem with trying to bake the salute tag, which in combination with my input images invariably converged on some very western, very thick lines, very muscular style, it hit something bad in the latent space
>>
>>7927775
>can you try pruning the character tags
I know from experience that this is a mistake. Pruning the character tags would force all 1girls to look like whatever character was in the dataset. The character name is the only thing that's making this lora work right now.
>>
>>7927778
multi character loras are fucking cursed anyway, just don't
>>
>>7927770
those looks so eerie and uncanny it's kinda amazing
>>
>>7927756
Your efforts are not wasted upon my eyes anon. Well done.
>>
File: 00000-2306060811.png (3.31 MB, 1882x1434)
3.31 MB
3.31 MB PNG
>>
>>7927778
use better captions
>>
>>7927784
i love big boobs so much bros
>>
>>7927779
In my experience this isn't the case. The multicharacter lora I made for a NAI style lora is filled evenly with different pokemon characters and everything works properly.
>>7927780
That's the last part I'm trying to fix. A major part of the problem is that, because it's from the game, a lot of the images have characters not looking directly at viewer. It's always off to the side or looking to the side. The problem is pony doesn't know the tag "looking ahead" which is what gives that uncanny look. I tagged those pics as "looking to the side" but it doesn't fully fix it.
>>
>>7927785
Please elaborate.
>>
>>7927756
still couldn't spend an extra 10 seconds to go to the right thread
>>
How tf do I start training loras?
>>
>>7927789
you shouldn't need to include the character names in a style lora if the training captions describe the content of the image. they're effectively doing the same thing
>>
>>7927756
need that butt on my face
>>
>>7927779
You just need the dataset to be properly balanced.
It's unironically easier to pull off than multiple costumes for one character.
>>
>>7927787
nig, the tag is looking at viewer
>>
>>7927784
(small breasts) as seen by AI
>>
>>7927288
Hot
>>
>>7927794
not this shit again
>>
What methods do people use for high quality pictures? Is HiRes the norm or do people use something else or additional things?
Would post something but nothing comapres to High quality stuff ITT
>>
>>7927778
>Pruning the character tags would force all 1girls to look like whatever character was in the dataset
I mean, that's not really how it works. As long as hair color, eye color, hair style, etc. are labeled it should generally be fine. If your dataset is largely dominated by specific characters that are labeled it's going to be treated as a style activation tag.
>>7927805
not that anon but this guy is talking about making a STYLE lora, not a character lora.
Tagging philosophies are slightly different between the two.
>>
>>7927733
gatobox, onegai?
>>
Any 8gb vram training configs?
>>
>>7927809
i don't think it's a style lora i think it's a mixed character style lora
>>
>>7927806
no need to upscale with pony imo, i don't need my porn in 4k
>>
what's better, prodigy, adamw or lion?
>>
>>7927794
This is incorrect. Take this pic for example.
https://files.catbox.moe/lpf030.png
The reason why this pic has these shorts and hairstyle is because Hilda is in the dataset, with all her features tagged, but her name is missing. As a result her tagged features get bound to "1girl" and forces 1girl to be an amalgam of Hilda and the other 9 characters I put in the dataset. Simply tagging her name completely solves this issue. What are you suggesting I tag in order to stop this from happening if not the name?
>>7927799
https://files.catbox.moe/v0ldg0.png
Let's take this pic for example. Tagger will interpret this as "looking at viewer" when in fact it's not looking at viewer, it's "looking to the side." That's because in game she's speaking to someone who is to the right of her. Tagging this as "looking at viewer" would fuck up your lora.

https://files.catbox.moe/ymswp1.png
This is a better example. Tagger will tag this as "looking at viewer" but a human can understand that although she's looking toward you, she's actually looking at something behind you. The tag for this is "looking ahead" but pony doesn't know this concept. The bulk of closeups from the game are like this and I was forced to tag them as "looking to the side." The only incorrect answer for both these pictures is "looking at viewer" which will fuck up the lora.
>>7927809
>If your dataset is largely dominated by specific characters that are labeled it's going to be treated as a style activation tag.
This is completely false since tags are shuffled. It would only kinda work that way if you put it as first tag and didn't shuffle that one.
>>
>>7927816
yes
>>
>>7927629
NTA but I found this list some time ago and I'd agree with pretty much everything here
ame_(amechan17391739)
manzai_sugar
senju_(snz0)
maruko_tooto
ed_(edemevas)
ezzydraws

Muscular babes section:
Yaki_apple
minew
Todding
sorahachi_midoro
1509virgoart
hanzo_(2929)
Putcher
bbc-chan
hataraki_ari
fumio_(rsqkr)
>>
>>7927814
Fair enough but i do wonder about the quality of the pictures. Is there a trick to that or is it all prompt work? Take this one for example (no disrespect to others, it's just that this one is the newest high quality gen) >>7927764
The quality is very high. Is it really just hires?
>>
>>7927820
Either hires fix or tiled upscale controlnet.
>>
>no good Gorobots lora style
its over
>>
>>7927816
Prodigy burns the TE so it's a meme. I tested it. AdamW or Lion both works.
>>
>>7927824
ykwtd
>>
>>7927812
It's a style lora made from multiple characters. This works because during training AI looks for similarities. If all the characters have the same style it will only learn the style, especially if it already knows those characters. If you don't tag their names the similarity between all pics in your dataset is 1girl so when you type "1girl" AI will pull from the tagged features in the dataset, ruining the result.
>>
File: abandonware.png (4 KB, 311x83)
4 KB
4 KB PNG
>>
>>7927829
cope
>>
File: ComfyUI_01850_.png (1.54 MB, 1536x2048)
1.54 MB
1.54 MB PNG
>>7927819
>ame_(amechan17391739)
https://mega.nz/folder/ou0GjTYI#Wxbh2WQrCpzZGFlqpECZdw
Maybe I never uploaded it, but its kinda weak, but is stronger with help from another lora like traditional style
>>
is it ok to train loras with 400 images or am I retarded?
>>
>>7927842
those better be 400 perfectly captioned, high quality images
>>
>>7927825
huh that explains why some concepts get fried with some tags.

well, time to change for adamw/lion
>>
>>7927822
Alright
>tiled upscale controlnet.
It's not the ControlNet included in Forge, is it?
>>
>>7927849
It is but the one in img2img.
>>
Is there an image recognition AI more accurate than the tagger in webui?
>>
File: stereogram33.png (2.08 MB, 1728x1632)
2.08 MB
2.08 MB PNG
is there a method to seperate layers on generation? like character from background etc?
>>
>>7927844
Uh its for a style lora.
>>
>>7927806
gen a picture until you kind of like it, inpaint/sketch it until you really like it. upscale in extras tab with your choice of upscaler. send it back to i2i for multi diffusion mixture of diffusors at .25 ~ .55 until its perfect. adetailer. post.
>>
>>7927856
forgot to mention, hires fix is not used by most here anymore. but whatever works for you is good.
>>
>>7927855
Then it's even more important.
>>
>>7927854
https://github.com/AUTOMATIC1111/stable-diffusion-webui-rembg
>>
Is it safe to presume my training speed will be my genning speed?
>>
>>7927851
Damn i need to include a picture there even though i haven't genned it yet? Lots of preprocessors and 0 models too. Wish it was explained somewhere what does what but i won't bother you anymore. Thank you!
>>7927856
>>7927857
Ooooh thanks for the tips too
>>
Fuckkkk, the artist draws some nice pics with thick lines but there's also some really good ones with a thin line
>>
>>7927859
I have this what I meant was a way to seperately gen character from background foreground etc, that way you can easily forego the erasing background step
>>
>>7927862
>Damn i need to include a picture there even though i haven't genned it yet?
That's why I said the one in img2img. Gen a pic, click the button that sends it to img2img then setup your controlnet settings from there.
>>
>>7927855
you better triple check that everything is in order
>>
>>7927864
I don't know if that's possible since AI starts with a pixelated image then denoises. You're asking to cut out parts of the image entirely which just isn't how it works.
>>
>>7927864
https://github.com/layerdiffusion/sd-forge-layerdiffuse
this gens images with a transparent background. idk if that helps you.
>>
>>7927867
how
>>
>>7927865
Oh i've misread. My mistake
https://github.com/lllyasviel/ControlNet/tree/main/models
There's only 2 models here for the controlnet. Am i looking at the wrong place? And what do the Preprocessors do if you don't mind me asking?
>>
File: controlnet.png (1.4 MB, 2560x2948)
1.4 MB
1.4 MB PNG
>>7927874
Use this anon's settings.
https://huggingface.co/bdsqlsz/qinglong_controlnet-lllite/tree/main
This is the model he uses. Just get the alpha and beta tile files.
>>
>>7927874
>>7927877
Actually no that guy is crazy and uses 0.75 denoise. Start with something lower like 0.3 and work your way up.
>>
>>7927877
>tilememe
>>
>>7927879
>everyone's a critic
>no gen to offer proof
/hdg/ never changes.
>>
>>7927874
Oh i was looking wrong. My bad again.
>>7927877
>>7927878
I think i know how to go from there now. Many thanks
>>
>>7927880
catbox?
>>
File: catbox_7o0efp.png (2.86 MB, 1952x1520)
2.86 MB
2.86 MB PNG
>>
>>7927289
>>7926042
>>7925445
bake 2 of anon's r16w5r dataset
testing ennuigrl lora settings
https://mega.nz/file/crdB0YzQ#IPVySlarTRIAQv11oPK5zDK1d8-rnPnQDWb3NpcU2uQ
>preview
https://mega.nz/file/V7F0lS6T#SEM7MOljW5mKgYnecRnyCCutFtkcWY_sqQhzRr2LA2s

small break, more bake tests in a few hours
>>
>>7927869

this seems to be it ! but I get weird errors with the Generating Foregrounds and Backgrounds Together (SD1.5) part, does it work for you?
>>
>>7927885
that artist looks generic as fuck
waste your time
>>
File: 00264-2598571294.png (2.1 MB, 1344x1632)
2.1 MB
2.1 MB PNG
>>
>>7927890
boxxu?
>>
>>7926327
based
>>
>>7927828
>>7927809
lol
>>7927820
it's just hires fix. ultrasharp at 0.4 denoise using adetailer for the face
>>
>>7927784
too flat
>>
>>7927889
it's probably just a meh bake on my part. I'm happy to have datasets to get to test on without having to make a bunch of my own
>>
File: 00271-1471753490.png (2.39 MB, 1344x1632)
2.39 MB
2.39 MB PNG
>>7927891
https://files.catbox.moe/pxgvs4.png
>>
>>7927897
<3 thanks
>>
>>7927896
bake a gorobots lora
>>
>>7927896
Ahem... how big of a dataset do you mind training on?
>>
>>7927896
There are resident schizos that complain whenever a baker does anything tangible. It looks good anon just ignore him. But as you can see, the real reason for his complaint is that you didn't do the thing *he* wanted, not that it's actually bad.
>>
>>7927770
Multiple characters in a lora seems really hard on pony, are you trying to have all of a characters' features tied to just one activation trigger? It seems like you are, I never had much luck with that, instead having to describe their features. Anyway there might even be an extra layer of strange interaction since pony already knows those characters. Maybe since it already knows them, training them even more is just frying the model's knowledge of that character.
>>
>>7927902
>the real reason for his complaint is that you didn't do the thing *he* wanted, not that it's actually bad
Pretty much. It's kind of hard to share loras in some ways because everybody uses different generation settings and models. I try to test my loras in various ways before I share them, but then I still get complaints that they look underbaked in some cases that I've just stopped caring. It isn't my problem if you are using Confetti, which requires far more training to get loras functioning on it or if you are using a ton of negatives when negatives really are not necessary on the Autism brand of models.
>>
>>7927899
I've got another anon's dataset I said I'd try baking first, but if you post a dataset for gorobots I'll bake it tomorrow (or tomorrow night, depending if I have to tag it or not)
>>7927901
what we talkin'? I've not tried many yuge or really tiny ones yet, so I'm happy to attempt either
>>7927902
Thanks anon, that's kind of you. I do think it feels a bit lacking compared to the dataset, but that's part of the trial and error of testing different settings
>>
Wait does cascade really require 16GB vram or does it just slow down when it swaps out models between stages?
Might have to consider buying a new gpu now than later if it's the former.
>>
Damn it's been a month since there was last a commit on forge now. Is someone else gonna take over the project ? Is it going to get merged into the original webui project? I really don't want comfy anon to win.
>>
>>7927909
+-400 images
>>
>>7927629
zenno rob roy from uma musume
>>
>>7927911
>UI
>winning and losing
What mentality is this when these are just tools for convenience?
>>
>>7927906
>are you trying to have all of a characters' features tied to just one activation trigger
No there's no activation trigger. I just tagged all the features of the images and their names.
>Anyway there might even be an extra layer of strange interaction since pony already knows those characters.
This is the thing I'm getting at. Someone once said there was some strange interaction when they mistakenly made a character lora of someone pony knew but he didn't elaborate further. I'm going to remove one of the characters, bake it again, and try to prompt that character to see if it works.
>>
>>7927912
Sure, I'll give it a shot. If it needs tagging though it might take me a day or two
>>
>>7927919
I'm finishing it, will probs do the tagging too
>>
Latent couple or regional prompter?
>>
>>7927916
I was being a bit facetious, really I'd just not have to use comfy for SD3 or cascade, I don't care that much about who "wins".
>>
>>7927921
Regional prompter is just the fucking name of the extension that can handle both "latent couple" method and "attention couple" method
>>
>>7927885
>lora
>zip
>3.5gb
wat?
>>
>>7927923
don't worry. maybe they'll add support in auto in a year or two ;)
>>
>>7927925
Ah ok, Im watching a video saying he prefers latent couple over regional prompter
>>
>>7927921
Both. Attention couple is better at splitting up the prompt per region and what you'll use most of the time. Regional prompter is better for masking specific parts of the image by hand.
>>
File: 00310-1346074334.png (2.03 MB, 1632x1344)
2.03 MB
2.03 MB PNG
>>
>>7927925
The guy doing the cascade finetune is making a gradio interface for it
>>
>>7927906
Ah you know what? I just remember I attempted to train scarlet with the dataset.
https://files.catbox.moe/88e3b2.png
This one at least looks like the dataset pictures but still isn't quite right. It's a real pain because you need a lot of pictures (more than the standard 100) for AI to learn properly and then you need to train for 10s of thousands of steps. This one was 46,000 steps. I'm confident this new strategy will work though but there's some sweet spot I need to land for this dataset.
>>
>>7927920
Sounds good, I'll keep an eye out for when you post it (or if you see one of my testing posts and quote and link where you posted, I should see it)
>>7927927
I zipped every epoch. annoying, I know, but I like to give anons the choice. if you think you want a specific epoch from the preview I can upload it alone
>>
>>7927934
Dealers choice. pick one, separate it from the rest, and let me play with it please and thank you.
>>
>>7927929
>watching a video
Just read the attention couple author's memo with a translator
https://note.com/gcem156/n/nb3d516e376d7
>>
File: 00322-2082737902.png (2.48 MB, 1632x1344)
2.48 MB
2.48 MB PNG
ok i'm liking lion
>>
>>7927933
>46,000 steps
based
>>
>>7927884
hot
>>
File: 00022-697234747.png (3.74 MB, 2048x2048)
3.74 MB
3.74 MB PNG
Multiple arms is such a pain to get looking right, maybe I have to swallow the monster anatomy cope
>>
>>7927943
wouldnt furry booru tags help with that
>>
>>7927935
Let's try 4 for 'I think after that is when hand coherency took a hit'
https://mega.nz/file/V2l12AwZ#8Ylq6xTKyqzZYBJtxJ8Q1XTmoQECe2FNDYXmhXSGV9E
>>7927937
mind sharing settings? I'd like to see what lion is all about
>>
>>7927629
An Ago LORA would be cool, I like a lot of his doujin work.
>>
File: file.png (1.3 MB, 1216x832)
1.3 MB
1.3 MB PNG
new fetish acquired
>>
>>7927948
what fetish is this
>>
>>7927948
I think no norm prevents this but I'm not 100%
>>
Any clue what went wrong here? I baked another lora with these settings earlier no problem. The only things different here are dataset and batch size (10->3). I'm using prodigy so LR is out of my hands. Judging by the samples it was learning well until it suddenly went to shit.
>>
File: 00344-3910581252.png (2.41 MB, 1632x1344)
2.41 MB
2.41 MB PNG
>>7927946
I yoinked a .toml from an anon that uploaded a ton of loras with Lion.

https://files.catbox.moe/xdhqpr.toml

Just adjusted batch size to higher values since I use gradient accumulation, and 2x the dims (so block_dims=16,etc and block_alphas=0.125,etc)
>>
>>7927629
Terufuu
>>
>>7927949
fried fetish
>>7927950
thats the twist, i'm using no norm
>>
>>7927942
Was kind of surprised how hard it was to get vaginal and anal fingering at the same time. Kept just giving me anal sex more often than not until I started using the double penetration tag but even then felt like placebo.
>>
>>7927951
are these just the samples it poops out during training? if so, test the epochs normally, they may be fine
>>7927952
thanks anon, will steal into json format and give it a test run
>>
File: GJ9QxTyacAAcv3E.jpg (224 KB, 1280x1856)
224 KB
224 KB JPG
1.5 looks like this?!? We got mogged so hard....
>>
File: 00232-2771429621.png (430 KB, 896x1216)
430 KB
430 KB PNG
>>7927948
I love fried girls bro.
>>
>>7927958
1.5 has sovl, but that it took 60 hours of gacha and inpainting, 50 embeds, 200 schizo negs and god's blessing just to get 10 fingers on a hoe was enough to make me swap to sdxl
>>
>>7927951
Prodigy blew up unet. You'll have to use a different optimizer or very different settings since you cant change your LR manually. I remember the OP lora guide prodigy settings was doing this for me.
>>
>>7927957
>are these just the samples it poops out during training?
Yes.
>test the epochs normally, they may be fine
Gives a completely black image.
>>7927961
Anything particular I could try to make prodigy work?
>>
File: file.png (1.36 MB, 832x1216)
1.36 MB
1.36 MB PNG
>>7927959
ikr
>>
File: 00021-3264128631-min.png (2.65 MB, 2496x4096)
2.65 MB
2.65 MB PNG
>>
>>7927962
I'm not really sure since I switched from prodigy a long time ago. You can give my old prodigy settings a try but I'd recommend adamw or lion.
https://files.catbox.moe/hwbq3p.json
>>
File: 00003-2327679760.png (3.67 MB, 1664x2432)
3.67 MB
3.67 MB PNG
More birb

>>7926186
>>7926751
The one on the OP's rentry
https://rentry.org/ponyxl_lora_previews
https://files.catbox.moe/swjqw4.safetensors

I want to eventually make my own YD LoRA as well, mostly to have the dataset available to make one for animagine. There's sfw stuff I want to make, but prompting YD in animagine turns too sketchy for my taste.
>>
What are the consequences of disabling text encoder learning for style loras?
>>
>>7927972
You won't get total convergence on learning the style. You need some level of TE training.
>>
>>7927976
What values would you recommend then?
>>
>can't find a decent guide on how to use latent couple with regional prompter anywhere
Suffering immensely.
>>
File: 116508415_p0.png (2.57 MB, 1280x1856)
2.57 MB
2.57 MB PNG
>>7927958
Superior in every way (except hands)
>>
>>7927958
>>7927979
wtf bros, why isn't pony this good...
>>
>>7927979
pony sideboob saggers could never compete with this...
>>
>>7927979
I wish pony could do skin like that
>>
>>7927958
>>7927979
we have to go back...
>>
it's over sdxl bros...
>>
>>7927972
You get an anon mad at you for not following his guide.
Seriously, it's all conjecture. You'll have to test it yourself.
https://github.com/bmaltais/kohya_ss/blob/master/docs/LoRA/top_level.md
>>
>>7927979
Look at that fucking lamp... christ. It's like I can touch it and feel its warmth...
>>
>>7927977
It depends on what optimizer you are using. I think a safe value to start with for AdamW is like 0.0001.
>>
>>7927979
79s of truth... we made a mistake somewhere along the way
>>
>>7927943
giwtwm
>>
File: 00272-2500826386.png (950 KB, 1024x1024)
950 KB
950 KB PNG
>>7927945
What kind of tags would make multiple arms more consistent? Right now it's just a gacha with "multi arm"
>>
>>7927979
wtf... time to redownload 150gb of concept loras...
>>
sd1.5 times were good. I remember using 8+ loras on a single gen.
>>
I can't believe 1.5 was actually as good as current NAI
>>
>>7927978
What's forcing you to use latent mode? It's kind of an old and brute force way of generating isolated couples which is why everyone moved to attention mode.
>>
Guys, I have acquired 200 images for an artist lora.
Should I decrease the number of images or what?
Whats a good step range for the style?
On my mind, if I train on 200 images, 200*1(repeat)*20(epoches) for 4000 steps, It seems like its worse training.
>>
File: 00146-20240403222611.png (1.09 MB, 1024x1024)
1.09 MB
1.09 MB PNG
>>7927999
Well for a scene like this for example. Will attention mode with rows work fine given that the regions aren't exactly clear cut? I can understand how to make someone stand side by side but how will it work for a sex scene?
>>
please somebody port loractl to forge, i beg you...
>>
>>7928002
Id say, give up. Even on comfy its shit.
>>
File: tmpbj0q1l9_.png (3.01 MB, 1704x1280)
3.01 MB
3.01 MB PNG
1.5bros ww@
>>
>>7928004
>background
SOVL
>>
>>7928004
holy shit.....
>>
>>7928000
it really depends on your overall settings. batch count also effects step count
(img count * repeats * epoch)/batch
200 images isn't bad by any means though, assuming they are quality
>>
>>7928004
oh god, i came.
time to install c*mfyui and gen foxgirl in snow field
>>
>>7928000
More images are always welcome if they vary in subjects/characters/concepts, but the style stays the same. This is VERY important, because some artists doesn't draw the same way they used to do in 2015 compared to today, for example. If you are training a Style, then the Style in the dataset should be as consistent as possible. That includes colors, shapes, line thickness, everything. Might want to check the dataset again to make sure, and prune accordingly.
>>
File: tmpzbd7dd2h.png (2.98 MB, 1664x1280)
2.98 MB
2.98 MB PNG
>>7928009
Jokes on you, I actually have something like that but I didn't use coomfeeh-ooaie
>>
File: 00169-3352520333.png (2.33 MB, 1536x1280)
2.33 MB
2.33 MB PNG
>>7928004
rait here bwo
>>
cascade please, save us from samelooking 2.5d slop.
>>
File: 00037-4137101325.png (3.57 MB, 1648x2432)
3.57 MB
3.57 MB PNG
>>
>>7928001
I wouldn't use attention couple(masks) in situations like this (unless it's 2girls/2boys + character lora) and just use BREAK for the general prompt, 1girl, and 1boy
>>
>>7928018
Do not worry friend, it will.
>>
who memed us into using a shitty horsefucker model? we must heal hdg...
>>
>>7928016
>1.5 does Ratia near perfect
>Pony can't even get close
bwos...
>>
>>7928026
porn
>>
File: 00096-20240403235723.png (1.12 MB, 1024x1024)
1.12 MB
1.12 MB PNG
>>7928022
I'm experimenting with just two individuals standing next to each other but now my problem is adetailer bleeding the prompts together for the faces so I end up with something like this. Any way to get adetailer to stop fucking it up?
>>
>>7928026
Must've been horsefucker himself
>>
i wish i genned coom /sdg/ sucks
>>
>>7927968
Thanks. After some fiddling around it looks like the situation was remedied by reducing d_coef to 1.
>>
>>7928013
big twin bens
>>
>>7928013
How do you make the subject not be in the center?
>>
>>7928031
I really don't want to recommend comfy because it's a real pain in the ass for a lot of things but you seem to be at the stage where you want full control of where your prompts/conditionings go.
>>
>>7928026
Honestly? Yourselves. The retards who, time after time, clung to models like Based64 and pretended they were le best model ever!!! even when they got outdone by other models. It was that retarded fixation that led to model development grinding to a halt.
>>
>>7928040
I don't necessarily want full control I just want a separation of character traits. Adetailer seems to have a function for separating face retouches, just need to figure out how to use it in practice.
>>
>>7928042
>t. Easyfluff knot taker
>>
>>7928038
Honestly idk, just 1.5 things I guess, I can offer you a catbox for those If you want to see the details about it
>>
>>7928046
>impyling I liked that piece of shit sepia shitfest
LMAAAAAAAAOOOOOOOOOOOOOO
>>
>>7928047
sovl...
>>
>>7928044
>Adetailer
If you don't need it to be part of an automated workflow then you can just manually inpaint
>>
Horse cock
>>
>>7928052
Based, horse-cock pilled.
>>
>>7927952
made a .json of this toml if any other bmaltais users want to try anon's settings:
https://mega.nz/file/xb813Z5B#0BL8JN0MyY9ul2YWtdGwhYu-Ci6dBK2GMO0MJfshFuk
(I think I copied it right, anyway)
max train steps were 1000 so I adjusted epochs to match that with the dataset I'm using, unsure if that is ideal or not, will see how this bake goes
>>
>>7928058
>Mo0mFuk
what is it telling me...
>>
Will be my day 4 of SD today after I get home from work.
Pretty much all I've been thinking about is generating content of my character and I feel like I'm starting to get better at it.
I'd say its been a refreshing break from all the other things, like watching mindlessly some yt videos.
>>
>>7928061
based newfren
is your character an oc, you're waifu or both?
>>
File: 00262-693867193.png (2.28 MB, 1280x1920)
2.28 MB
2.28 MB PNG
>>
>>7928065
Berry cute style
>>
>>7928061
Have fun bwo
>>
File: 00264-948028695.png (2.69 MB, 1280x1920)
2.69 MB
2.69 MB PNG
>>
Bros, can I get in trouble for AI? I genned some stuff based in an artist and it kinda spread out. I'm kinda nervous.
>>
>>7928073
oh no no no no...
>>
>>7928076
Bwos...
>>
>>7928073
prepare to get a scolding
>>
>>7928073
burn your computer now
also terminate your isp contract
better leave home for at least 9 months too
>>
>>7928073
It’s time to visit your local law enforcement and explain the situation truthfully
>>
>>7928073
delete all your profiles
>>
>>7928073
Show it
>>
>>7928064
She's oc and a new waifu. Takes some time to refine her assets for the model later. Might introduce SD to my older waifus as well later. Lots of projects would benefit from this.
>>
>>7928073
It's over
>>
>>7928092
I look forward to your creations, newfriend.
>>
>>7928092
homemade LORAs of OCs are great, it feel like making a mod for a game or something
>>
File: j6.jpg (92 KB, 1280x720)
92 KB
92 KB JPG
>>7928073
you gen'd, we came, we'll tear your soul apart!
>>
>>7928092
basado
>>
File: 00055-2289733061.png (344 KB, 512x512)
344 KB
344 KB PNG
>>
File: 00265-3310982905.png (2.2 MB, 1280x1920)
2.2 MB
2.2 MB PNG
>>
File: 00057-364084722.png (450 KB, 768x512)
450 KB
450 KB PNG
>>
File: 00007-4149472308.png (562 KB, 640x744)
562 KB
562 KB PNG
>be naibabby
>have no idea what local is
>start using webui via forge
>have no models/loras/etc.
>type shit in
>get this
>install models/loras/etc.
>type in basic ass tags
>generate
And get deviantart kino, I have reached my peak.
>>
Any bakers around who don't mind doing some uuuoooohhh esque content? I said I'd bake anon's dataset but I didn't realize what it was before hand (sorry anon...)
>>7925585
>>
File: 00034-383349961.png (1.52 MB, 1024x1280)
1.52 MB
1.52 MB PNG
>>
>>7928118
I am deviant who likes to age up lolis and give them massive tits. Give me the name of the loli so I can turn her into a lovly hag like Yukari.
>>
Neigh...
https://files.catbox.moe/efogzz.png
>>
>>7928121
>Eirin, what happened to your uvula?
>>
>>7928121
eirin lora??? or is it nai
>>
File: 00020-3174962796.png (1.44 MB, 1024x1280)
1.44 MB
1.44 MB PNG
>>7928127
I just baked it. Not sure if it's good enough to share.
>>
>>7928128
looks good enough to me
>>
>>7928118
It's alright, first baker said he was alright with loli. I could try to make it myself if you can give some pointers/config for style? Only got 12GB of VRAM though.
>>
>>7928128
based 2hu baker
>>
>>7928128
Speaking of 2hus, have you managed to make or do you know of a lora that gets the fairy brats' wings correctly? Daiyousei in particular is near impossible to get her wings right.
>>
File: 00266-3319364031.png (2.16 MB, 1280x1920)
2.16 MB
2.16 MB PNG
>>
>>7928134
not the 2hu baker but at this point if it's that hard I'd just wait for the cascade finetune, no pruned characters so 2hus will be freed
>>
>>7928130
I bake with 12GB 3060 so you likely can bake just fine. if you've got (or can install) bmaltais/kohya_ss, I've got a bunch of jsons here you can try:
https://mega.nz/folder/429HjaaI#vel37O61SS1_mf2ZRNfSJA
>open kohya GUI in web browser
>make sure you're on the "LORA" tab, NOT the dreambooth tab
>at the bottom there should be a 'configuration' section, if you click that it will let you load in the .json file which will update all the settings
>make sure you adjust, at the top:
>your folder pathway that leads to your pony XL checkpoint
>desired name
>your folder pathway that leads to your dataset*
>your output folder
>once you've done these, check the option that says 'SDXL'
>optional: change the # of saves per epoch
>if you want to calculate steps, its done like this: (# of images in dataset * number of repeats * epoch)/batch count (you should be able to see what these are in the updated settings if I haven't made a note in the training comment)
>click 'start training'
>watch in the cmd window and see if it starts loading stuff, or if it errors
>if it errors, post back here and I'll try to help troubleshoot what went wrong

*dataset folder: for it to work, you need to have 2 folders. 1 main folder, then the folder with the actual dataset inside of that. the one with the dataset inside of it will decide the number of repeats by #_name
so for example, if you're only doing 1 repeat, its:
1_foldername
and you select the folder that, that one is inside as your dataset folder

For 12GB, the following jsons should be around 1 hr in bake time, if I remember correctly, so starting with these is generally easy enough for a beginner:
>4chanjson-usethisone
>anonslionsettings (I'm testing this one now)
>pony bluethebone
>pony dominus medusa
>pony ennuigrl
>pony ushiyama ame

there is no perfect 'setting' for any single dataset, unfortunately, everything seems to need some fiddling to find out what works best for it
>>
I fucking hate that moment when I just finished cleaning images in krita/PS to use them for dataset after 3 hours, then look again at image databases and say "hmm those ones could be usable if I clean them + upscale"
Kill me please, this vicious cycle never ends
>>
>>7928135
>>7928107
based, looks like a good lora
>>
>>7928139
oh god, cleaning the images that have a shitload of text and sound effects are the absolute worst.
>>
>>7928129
I haven't fully tested it. It seems flexible with poses and outfits but it takes some gacha not to mess up her outfit and there's a bit of style baked in. If that's fine with you then here: https://files.catbox.moe/1ttwli.7z
>>7928134
Sorry, can't help you there. I'm still learning baking, I might tackle that challenge later.
>>
>>7928138
If you have any issues/confusion along the way feel free to post here, I should be around for the next 3-4 hours or so (and back tomorrow, I'm a neet kek). from my experience there's usually some kind of hidden tech angel lurking in the background ready to answer the most obscure baking questions, too.
>>7928139
>dataset is probably fine as long as I tag the text etc
>dataset ocd kicks in
>spend hours scrubbing every little signature/text from the images, sometimes even redrawing parts of the surrounding area by hand because llama/photoshop just doesn't clean it 'right'
its a sickness
>>
>>7928145
based 2hu baker
>>
just as a quick question for clarification, these hashed artists work on autism right?
>>
>>7928152
they should work on any pony based model (like autism), though in some cases their style may be diminished by merges
>>
File: 2340896745036.png (3.25 MB, 2064x1408)
3.25 MB
3.25 MB PNG
>>
File: 3059867905764.png (2.37 MB, 1080x1576)
2.37 MB
2.37 MB PNG
>>
>>7928154
>>7928156
<3
>>
>>7927979
Anon posts a very heavily edited pic and people loses their minds? It's like you're all new to AI.
>>
>>7928161
nah bro it was a single gen
>>
File: 1710562466997406.png (2.56 MB, 1344x1728)
2.56 MB
2.56 MB PNG
>>
>>7928139
>>7928142
What stops you from just tagging "sound effects" and "speech bubble" Imo the only reason this won't work is if pony doesn't know those tags.
>>
>>7928128
Do you also have 6 fingers? If that's your self insert
>>
>>7928161
look at how they're posting. all the replies are a single anon. add the common retard replies to the filter and the thread gets immediately less cancer
>>
>>7928168
Is there a list of tags used by Autism mix?
>>
File: wtf.png (1 KB, 161x49)
1 KB
1 KB PNG
>>7928145
anon, your eirin wants to give me the rabies
>>
File: file.png (1.81 MB, 1024x1024)
1.81 MB
1.81 MB PNG
uh my lora baking isnt looking good so far
>>
>>7928165
boxxy?
>>7928173
t..this is why I don't use the preview image option...
>>
File: file.png (1.51 MB, 1024x1024)
1.51 MB
1.51 MB PNG
>>7928174
wise decision
like bro no way its this shitty
>>
File: s-40325590.png (3.16 MB, 1856x1280)
3.16 MB
3.16 MB PNG
let me guess, you need more
>>
>>7928171
No he never released that info which was the cause of all the controversy recently.
>>
File: 02504-383349964.png (1.35 MB, 1280x1024)
1.35 MB
1.35 MB PNG
>>7928169
Counting fingers is against thread etiquette.
>>7928172
False positive. If it's even possible to make .safetensors files dangerous I sure as hell don't know how.
>>
>>7928178
n-no... not my .safetensors...
>>
>>7928154
>>7928156
which shirow-ish style is that?
>>
File: 1702596233421488.png (2.47 MB, 1344x1728)
2.47 MB
2.47 MB PNG
>>7928174
https://files.catbox.moe/hii8ax.png
>>
>>7928181
>spank mark has 6 fingers

So close.
>>
>>7927946
Lion is about convergence speed (which is faster than Adam for 2+ times as was proven empirically)
>>7927952
You should also try 8 bit lion
>>
File: 00043-1227846976.png (2.44 MB, 1248x1824)
2.44 MB
2.44 MB PNG
>>
File: 1703652557847675.png (2.59 MB, 1344x1728)
2.59 MB
2.59 MB PNG
>>7928182
We're gonna make it brah.
>>
>>7928178
>If it's even possible to make .safetensors files dangerous I sure as hell don't know how.
You'd want to somehow pass a commit to one of the UIs that reads files renamed from .ckpt to .safetensors as .ckpt.
>>
File: saya1686290867780854.jpg (2.41 MB, 3036x2472)
2.41 MB
2.41 MB JPG
>>7928004
I miss saya anon...
>>
>>7928189
This is giving me ces flashbacks...
>>
>>7928138
Thank you anon! Any tips on how I should tag the manga/doujinshi pages? Never tried training with those in a dataset before and not sure how to handle them.
Do I still tag stuff like "1girl, standing, etc" and then slap on "manga (medium)" along with "monochrome"?
>>
>>7928181
thank you
>>
File: 02163-1487345680.png (2.15 MB, 1584x1280)
2.15 MB
2.15 MB PNG
>>
>>7928192
"monochrome" & "greyscale" for black & white images, make sure you tag speech "bubble", "text", "sound effects", as they appear (or clean them out like I do), "comic" & "multiple views" if you dont want to split the page panels into multiple images
>>
>>7928179
ah, the lengths we go to to have the brain of the moon suck us dry...
>>
What's the preferred way to do region prompting with an XL model? It seems most 1.5 methods doesn't work.
>>
>>7928200
abandon xlmeme and come back home
>>
File: 00102-980576979.png (1.18 MB, 1024x1024)
1.18 MB
1.18 MB PNG
>>7927952
uhh wtf I might just be team lion now, thanks again anon
>>7928175
lmao I'm sorry anon but these are sending me... at least its free comedy while you wait?
>>7928192
I usually try to crop manga pages into singular focuses, but I say tag the best you can and experiment/see how it goes. tags like (where applicable) :
>multiple views, comic, speech bubble, text, english text, japanese text, sound effects, grayscale, monochrome
will all help
remember, you are trying to tag everything you can see, as you want the AI to learn just the art style. also keep in mind you have a 255 token limit, so try to prioritize if you're hitting it.

I usually use the tagger extension, ML-Danbooru Caformer option at around 0.35 weight, then remove redundant or incorrect tags and add things that will be important to neg (such as text, grayscale, etc)

if your tagging isn't perfect your first go, don't worry, you'll be able to see pretty clearly what is bleeding through where and it'll give you a much better visual learning lesson than any amount of nitpicking beforehand will do
>>
>>7928197
Alright, you think it's worth cleaning the pages? Been hearing conflicting things from anons, but maybe someone has baked the same dataset cleaned vs non-cleaned.
>>
>>7927424
i'd like to see ett style lora if you can bake it
>>
File: file.png (1.52 MB, 1024x1024)
1.52 MB
1.52 MB PNG
>>7928203
And no matter the epoch, it's not improving... it's almost done training...
>>
>>7928206
I'm rekt man, maybe it just needs to gen on an autism checkpoint.. with source_anime... and every other style in negs KEK
really though post results from an actual gen once you can, I don't trust the sampler
>>
File: file.png (1.43 MB, 1024x1024)
1.43 MB
1.43 MB PNG
>>7928207
Will do
>>
Hello anons, i am trying to figure out the best way to upscale a gen. Is it better to use mixture of diffusers or controlnet tile?

also for mixtures what are the best settings?
>>
File: 00040-4201356099.png (1.2 MB, 1024x1024)
1.2 MB
1.2 MB PNG
>>7928203
I'll post this final bake 3 tomorrow, I'm pretty sleepy and don't feel like fighting with megaupload's snail speeds at the moment. not even sure how to feel now, I really like the lion settings. do I even want to keep testing all the other lora metadata I was planning? shit, hakuna matata or whatever. guess I could harvest my 895 tabs of galleries I wanted to make datasets of..
>>7928211
there is at least one anon who swears religiously by mixture of diffusers and has posted in detail their process, if you search archives for 'diffusers' or what not I'm sure you could find their posts fairly recently. (If no one is able to answer you at this hour, at least)
>>
File: file.png (39 KB, 917x541)
39 KB
39 KB PNG
>>7928213
Me
>>7928211
Can't say for controlnet tile, but for mixture:
Upscale the image, then use the original size as the tile's height and width. TIle overlap doesn't seem to do anything with mixture, and tile batch is based on your gpu power. I also recommend Never OOM's tiled VAE

Denoise from 0 to 0.25 = no unwanted stuff
0.25 to 0.35 = very rarely unwanted stuff
0.35+ = can get unwanted stuff the more you're nearing 0.5 denoise, but a lucky seed or just photoshop editing helps.
>>
>>7928177
Guess I have to lurk more.
What I've tried so far gives a pretty good idea why it might be.
>>
File: 00000-1764253301.png (2.77 MB, 1920x1280)
2.77 MB
2.77 MB PNG
is there any way to combine region prompter with ultimate SD upscale? I want to do this image at 2k by 3k but keep individual expression/details, etc
>>
>>7928219
He meant pony creator btw not autisms
>>
>>7928213
Thank you!
>>7928216
Thank you! Does tiled VAE have any affect on the quality?

also when i gen pretty large on mixture it fails. is that bc of too high batch or is there a size limit based on you gpu memory?
>>
>>7928220
wow bro how did u do dat
also, I dont have an answer for u, sorry
>>
>>7928222
No effect on quality, and what do you mean by "fail"? out of memory error? if so, then yeah, tiled vae fixes it. Can also lowering batch size, but I think the batch would only be the cause if your tiles were too large, but even so, tiled vae can probs fix it.
>>
File: 00019-2336073556.png (1000 KB, 1020x680)
1000 KB
1000 KB PNG
>>7928223
control net, of this ratatatat74 image, regional prompter and photoshop to crudly fix things
>>
>>7928225
Sweet ill give tiled a try! and yes it ooms and automatically runs again with tiled vae. Ill run with lower batch as well probably.
>>
File: forge-screenshot[1].jpg (241 KB, 1814x736)
241 KB
241 KB JPG
https://github.com/gustproof/sd-webui-forge-style-components
Style Components is now in github
ComfyUI version
https://github.com/gustproof/ComfyUI_IPAdapter_plus_Style_Components
>>
>>7928226
That worked out insanely well bro holy
would you mind posting the box?
>>
man, i fucking love the results of Restart but its so fucking slow... any other samplers producing the quality of Restart but faster?
>>
>>7928222
>Does tiled VAE have any affect on the quality?
It makes the image very slightly more saturated in my experience
>>
>>7928231
DPM++ SDE No Brown is the only one I have found that can match it. 8 Steps.
>>
File: 00009-2844267853.jpg (2.21 MB, 2400x2400)
2.21 MB
2.21 MB JPG
>>7928229
i mean it tooks like a few iterations and inpainting, but if you use one of the pictures i posted as a control net at a low weight like .2-.4 and region prompt characters you want with tags like shackle, metal collar, chained etc you should get something pretty ok. I used ponydiffusion
>>
>>7928235
wot and how do I get that one
>>
>>7928236
Which controlnet are you using?
>>
>>7928238
sdxl canny, haven't tried out other ones they could be better. the only lora i used was the smooth style from civitai
>>
>>7928228
To add, it's ipadapter for anime for those that don't know what it is
>Style Components is an IP-Adapter model conditioned on anime styles. The style embeddings can either be extracted from images or created manually. This repo currently only supports the SDXL model trained on AutismmixPony.
>>
>>7928228
so let me get this straight... put 10 anime images of a style into this extension, and it applies it to the gen?
>>
I'm installing Forge. When I go to run i get "torch is unable to use gpu" What can I do here to get it working? I appreciate it :)
>>
>>7928237
I got you anon, someone helped me so lemme pass it on.

go to your webui foleder and go to modules and find sd_samplers_kdiffusion.py.

then paste this in there: ('DPM++ SDE No Brown', 'sample_dpmpp_sde', ['k_dpmpp_sde'], {"second_order": True, "brownian_noise": False}),

then save then relaunch and youll have it.
personally i find about 20 steps and 5 cfg is what works for me but i could be doing it wrong
>>
>>7928245
Did you use the package or git clone?
>>
>>7928248
the package, and i have git and python 3.10.6
>>
>>7928244
yeah supposedly it should extract the style in some sense from the images you input, I don't think you need to stick in exactly 10 of them
>>
>>7928246
are you bamboozling me
>>
>>7928249
Yeah you should just make a new venv and use git clone. Some people are experiencing problems with the package.
>>
>>7928251
no?
>>
>>7928246
20 steps sounds crazy on No Brown. I do 10-12 at most if I want perfect results, but even 6 is fine if you are rolling for a good seed. 5 cfg is right though.
>>
>>7927792
Get grabber, czkawka, wd14 tagger in auto, and easyscripts.
>>
>>7928246
>('DPM++ SDE No Brown', 'sample_dpmpp_sde', ['k_dpmpp_sde'], {"second_order": True, "brownian_noise": False})
Nta, but I paste this with the brackets?
>>
>>7928253
I'll try it. If you are bamboozling, know that you'll make me sad
>>
>>7928231
schedulers are more important than samplers, mixing schedulers is even better.
restart sampling isnt worth the extra time anyway.
>>
File: nobrown.png (38 KB, 936x296)
38 KB
38 KB PNG
>>7928256
Right in with the other ones like this
>>
>>7928261
i love you, thanks anon
>>
File: GKRX2haXEAAyURP.jpg (173 KB, 1536x1536)
173 KB
173 KB JPG
How would you prompt pic related?
>>
>>7928263
Just prompt pony regularly at wider res, half the time you'll get this!
>>
>>7927725
I meant with base pony since that was what was talked about
>>
File: 00539-2942249786.png (900 KB, 888x1184)
900 KB
900 KB PNG
>>7928263
deepbooru basically got it right
"score_9, score_8_up, 1girl, ahoge, ass, black footwear, blush, breasts, formal, from behind, hand on hip, high heels, jacket, jeanne d'arc \(alter\) \(fate\), jeanne d'arc \(fate\), looking at viewer, miniskirt, multiple views, panties, pantyhose, pantylines, pencil skirt, shirt, short hair, skirt, suit, underwear, yellow eyes"
>>
>>7928268
wotis deepbooru and how to use it
>>
>>7928253
i-its real...
>>
File: 00001-2091312080.jpg (1.4 MB, 4096x4096)
1.4 MB
1.4 MB JPG
>>7928216
what great timing. I finally tried it too, very nice. thanks anon. between this and trying lion I'm feeling pretty lucky tonight. my only complain now is 4chan's tiny file size limit kek
>>
>>7928261
exceptionally based, thanks anon
>>
>>7928272
>complain
complaint rip. bed time for me. see y'all tomorrow frens
>>
>>7928272
>uo denim
based. looking forward to it if you ever decide to post it. gn anon.
>>
File: 00089-2874212654.png (2.13 MB, 1200x1800)
2.13 MB
2.13 MB PNG
>>
>>7928268
Makes sense. I'll try.
>>
>>7928272
no problem. Also are u the same anon whos baking datasets? I got mine ready
>>
File: 00100-67906962.png (3.66 MB, 1664x2304)
3.66 MB
3.66 MB PNG
>>7928254
i find with 10-12 is a lil less detail then i get with 20

>>7928256
yes, including the comma at the end so you keep the list working properly.

>>7928258
well im happy knowing that i wont make you sad

also if either of you two run into trouble lmk ill try to help but it should be a very simple copy paste
>>
File: deepbooru.png (10 KB, 674x75)
10 KB
10 KB PNG
>>7928270
The box on the img2img tab. After you paste something into the img2img you press it, then it will guess the tags from the image.
>>
>>7928281
Managed to make it work, but its twice as slow and I see no improvements(?), so, usage tips? steps cfg scale and stuff
>>
>>7928283
i use 5 cfg with 20 steps but another anon like 10-12, but just mess around until you find something you like
>>
>>7928276
>uo denim
AH, THAT's the artist name, damn, I feel dumb, I should've known it. It's anon's dataset so I can't take credit - will post it tomorrow though 100%
>>7928280
I'm at least one of them, post your dataset and I'll bake it when I wake up !
>>
>>7928252
yep that did it. Using the package was making it bug out but the gitclone worked just fine. Thanks :)
>>
>>7927816
i agree with other anon, i've done successful bakes with both adamw and lion.
>>
>>7927854
no. controlnet or manual compositing is your best bet here.
>>
>>7927860
No, lol.
>>
>>7928283
CFG 4-6, any more and it will burn.
FreeU - 1.25, 1.4, 0.95, 0.3
Since it's a second order method the steps evaluate twice (8 step No brown is actually doing 16 internal steps). Above 6 steps will be around 90% of the detail quality, 8-10 steps are 95% and work well enough for me since faces and minor details will be overwritten in the high-res pass later anyway. 8-10 steps is very quick for good output on my machine compared to any other sampler I've tried.
If you don't bother with img2img refinements etc, then you might want to increase the steps a little bit more. In most cases though, the resolution is more often than not the limiting factor for details, so I don't bother getting perfect details on base 1024x1024 gens.
>>
>>7928272
box?
>>
>>7928295
how do I know that
>>
>>7928296
Hmm what's FreeU?
>>
>>7928299
nta but this im guessing in forge. not sure what it does though
>>
>>7928216
i did some quick testing and mixture looks like the better option than controlnet. Control net maybe had very slightly better fine details but messed up things like adding a belly button. Mixture is def the way im going moving forward. Thank you again! (i did both on same settings at .4 denoise)
>>
>>7928299
>>7928300
https://github.com/ChenyangSi/FreeU
It's supposed to "improves diffusion model sample quality at no costs: no training, no additional parameter introduced, and no increase in memory or sampling time"
Never used it myself though
>>
>>7928301
You can, apparently, use both, btw.
>>
>>7928303
i tried it earlier and it didnt work, just threw an error every time.
>>
>>7928302
>>7928296
>FreeU - 1.25, 1.4, 0.95, 0.3
idk if its placebo, but i see a difference in quality with these numbers
>>
>>7928305
oh sorry for the misinformation then, the old tiled diffusion extension supported it.
>>
>>7928306
i tried it a couple days ago and havent turned it off since. Someone posted the numbers 1.3, .98, 1.5, .95 which is what ive been using, im gonna try the numbers the other anon posted to compare rq
>>
>>7928308
interesting, good idea. Share your results when you are done? :)
>>
>>7928307
youre good, its prob just a forge thing, i used to do something similar with ultimate upscale and controlnet back on 1.5 a1111
>>
>>7928268
Didn't quite work but attention couple can do an ok job of it. I imagine the one I want to copy has Torisan as one of the artists it is ripping off.
>>
File: 00020-467237990.png (2.37 MB, 1536x1536)
2.37 MB
2.37 MB PNG
>>7928313
Forgot image
>>
File: 06696-00044-1610851136.png (2.63 MB, 2048x1536)
2.63 MB
2.63 MB PNG
>>
File: 00028-3563426336.png (2.67 MB, 1920x1280)
2.67 MB
2.67 MB PNG
>and for my next trick... i'm going to make your cock... disappear!
>>
File: catbox_08rqd4.png (1.8 MB, 1248x1824)
1.8 MB
1.8 MB PNG
>>
>>7928309
How do i post multiple catboxes? I have the images but i cant post all of them lol
>>
>>7928319
uh edit them or post the links separately
>>
>>7928300
>>7928302
Basically they are scaling factors for different parts of the UNet (The thing that is denoising the image). The b values will smooth out high frequency stuff (finer detail in images), higher values means the network is less confused about what is a detail in the image vs what is random noise supposedly.
The s values no one has given a good explanation for really, but they work differently from the b ones.
>>
>>7928321
1.3 .98 1.5 .95
https://files.catbox.moe/2iiot7.png
1.25 1.4 .95 .3
https://files.catbox.moe/dktidr.png
non free u
https://files.catbox.moe/cduk1z.png

i personally like the style of the 1.3 numbers better but 1.25 followed the prompt a little better and was a little too saturated imo and the nonfreeu is kinda mid

all same seed and settings, ill prob test the number more tomorrow night
>>
>>7928323
is there a way to have the values saved on FreeU? seems like a pain to have to re-input them everytime
>>
You ARE baking right now, Anon, right?
>>
>>7928228
setting combine embeds to concat (the default) makes it crash
>>
>>7928316
box??
>>
File: 00031-1187814044.png (2.46 MB, 1280x1920)
2.46 MB
2.46 MB PNG
>>
File: 00033-3917987955.png (2.56 MB, 1920x1280)
2.56 MB
2.56 MB PNG
>>
>>7928325
huh, thats so interesting to me. Thanks, I'll try them out for myself
>>
>>7928327
its probably in ui-config
>>
File: catbox_pug6vk.png (2.72 MB, 2048x1536)
2.72 MB
2.72 MB PNG
>>7928330
>>
>>7928334
oh, you're right. Amazing, thanks!
>>
File: catbox_rqulco.png (1.23 MB, 832x1216)
1.23 MB
1.23 MB PNG
>>
File: file.png (914 KB, 2480x1069)
914 KB
914 KB PNG
I dunno about this one. Maybe I'm using it wrong? 1 or 10 images, it seems to make no difference. Also the fact it doesn't do faces or eyes is kind of retarded.
>>
>>7928004
i'm scrolling through my old gens now and i miss the schizo backgrounds so much bros...i miss -punks like you wouldn't believe...but the quality is just so, so bad compared to 1024x models, i can't go back
>>
>>7928298
empirical testing, let it run for 2 minutes and look at the it/s

training and genning are not the same and there are MANY options in training that affect performance as well as your specific hardware setup and your mem usage
>>
8pkyx8>>7928334
Yeah you can change the defaults in there,I did. I wish there were presets. Having multiple sets of values sounds nice.
>>
File: Duel_at_the_Forest.jpg (1.26 MB, 2400x1200)
1.26 MB
1.26 MB JPG
>>7928004
>>
File: catbox_hm1w27.png (1.02 MB, 896x1152)
1.02 MB
1.02 MB PNG
>>
>>7928299
It's great.
>>
>>7928339
Which extension is this?
>>
>>7928339
it's kind of a retarded extension, if you read the page you see it's actually 100 pre-trained embeddings with sliders and a method to grad slider settings from an image. it can't do everything.

>>7928347
style components, it's on civitai only no github lol
>>
>>7928348
it got a github recently, ctrl + f the thread
>>
>>7928349
it's still kinda stupid, if it had 10 times the number of embeddings it would be better i feel
>>
>>7928350
I agree, but you can get some pretty nice touch ups on the gen. Just lower the style strength.
>>
>>7928344
beautiful, but wrong board
>>
File: 00000-923626577.png (2.59 MB, 1824x1344)
2.59 MB
2.59 MB PNG
>>
>>7928332
isn't this the girl from that 3d platformer? isn't she a loli
>>
>>7928356
nice. but she's missing one of her piercings
>>
>>7928124
cute mare
>>
File: 00003-1618726275.png (916 KB, 1024x1024)
916 KB
916 KB PNG
>>
File: 00568-1325794597.png (2.26 MB, 1824x1248)
2.26 MB
2.26 MB PNG
>>
File: Untitled-1.jpg (2.78 MB, 2048x2048)
2.78 MB
2.78 MB JPG
>>
File: 00040-10741585.png (3.43 MB, 2048x1632)
3.43 MB
3.43 MB PNG
>when the roles are reversed
>>
>>7928356
as good as that is, wrong board.
/u/ is that way. their thread is actually somewhat active now.
>>
Holy shit, no brown is amazing. Racism wins again?
>>
>>7928374
gooooooooooooooooooooooooooooooooood morning sir
>>
>>7928374
wait till you hear about no yellow
>>
>getting gaslit about samplers again
Sigh, you win.
>>
>>7928378
NEVER MIND
I KNEEL
>>
https://files.catbox.moe/ldnhzf.png
>>
>>7928374
the fuck is no brown
>>
File: reimu.jpg (707 KB, 1280x1856)
707 KB
707 KB JPG
>>
So whos gifting me anlas or a rtx 4090?
>>
>>7928384
me, post cock and feet
>>
File: 117531633_p1.png (1.36 MB, 1344x768)
1.36 MB
1.36 MB PNG
Quick question, does anyone know if the novel AI extension for automatic 1111 allow you to combine with lora that you already have on your system?
>>
>>7928381
>>7928246
>>
>>7928386
No. That's not how it works.
>>
>>7928388
I'm asking because I've been using the extension because I have an AMD card and I can't get that shit to work on Windows
>>
>>7928389
that doesn't have anything to do with not being able to add loras to NAI
>>
>>7928389
Loras are local, NAI is on the cloud
>>
>>7928389
You can't use loras with nai3, you don't have access to the model itself, only API.
>>
File: catbox_1z9p3q.png (3.4 MB, 2304x1792)
3.4 MB
3.4 MB PNG
>>7928371
wtf my gen is on r34 LOL

https://rule34.xxx/index.php?page=post&s=list&tags=asuka_langley_sohryu+

holy shit I thought I was a degenerate but reading this comments damn
>>
>>7928392
>>7928391
>>7928390
So the problem then I am having is how to get the fucking AMD card to work. I have a 6700
>>
>>7928394
Rule34 comments are absolute cringekino.
>>
>>7928394
yeah dont read comments here, actual mentally ill schizo
>>
File: file.png (60 KB, 673x811)
60 KB
60 KB PNG
>>7928394
>>7928396
>>7928397

I want to end my life
>>
>>7928394
stop making bbc
>>
>>7928394
I'm the one who posted the one looking for the original. But I found it. Thanks. Also you made a flooxy Lora? Lol.
I bet the dude will be pissed seeing as how he goes out of his way to make sure no one finds his prompts
>>
>>7928394
This one is for blackgens right? I actually know the prompts he uses because me and him are actually friends
>>
>>7928394
I laugh because my absolute worst gens from 1.4 days are on there but nothing new.
>>
>>7928398
>no "limitless guy/lady" commenter
Less schizo than it could've been.
>>
Is pivix the place to post at? Even a possibility for getting some money?
I recall something about pivix and AI ages ago though..?
>>
FreeU was a placebo.
Influencing with Lora was much more effective.
>>
>>7928383
>jpg
nigger
>>
>>7928406
The most popular place i'd say, but be careful:
Censor penis/pussies
Tag your AI art as AI art
about getting some money, its hard nowadays... patreon banning people, booth does not allow uncensored stuff, idk where to go.
>>
>>7928406
shut the fuck up
>>
>>7928395
play stupid games win stupid prizes
>>
>>7928399
no

>>7928401
oh damn, I guess you pose the ones you find are kino?

>>7928402
NAI right? I wonder which artists his mix is based off of

>>7928405
there are so many horny 'women' on here, the comments are a fucking RP board
>>
>>7928408
https://files.catbox.moe/c744w3.png
File too beeg, calm down anon.
>>
>>7928318
looks nice, where the lora at?
>>
>>7928412
>NAI right? I wonder which artists his mix is based off of


artist:mossacannibalis, {artist:mamimi (mamamimi)}, artist:nekoshoko, artist:incase, artist:riz, [artist:kairunoburogu], {{realistic}},
>>
>>7928412
Do you have your stuff in a cat box catalog or mega file? Or something?
>>
>>7928412
Weird thing is that it wasn't particularly worse when they didn't require accounts. These people get off on exposure.
>>
>>7928414
Don't respond to idiots, just ignore.
>>
>>7928409
Censored? LOL
What are some other popular places?

>>7928410
No u.
>>
>>7928423
idk bro, wish I knew.
>>
>>7928423
I post on there all the time You don't really have to worry. The censorship is just like the most minuscule type of censorship You can easily slap on. And then link to Patreon and then you can rake all the cash. I'm currently at like $700 a month and I started posting in the beginning of March .

Of course you can also do popular places like rule 34 and then link to your Patreon

And Twitter.

Though in my case I've already crossed the 100K follower mark on Twitter
>>
File: blkpose03.jpg (716 KB, 1376x2072)
716 KB
716 KB JPG
BOC
>>
>>7928426
Same, but I'm making $3.5k, git gud
>>
>>7927877
It just made things more blurry? Doesn't seem like that's what you want
>>
>>7928428
Name?
>>
>>7928426
Hmm, yeah that's seems like a good incentive, have the censored images promoting a patreon.
But patreon/twitter doesn't allow for most /d/ material though? I'd like to make some fun stuff with that too.
>>
>>7928430
T****-k**
>>
>>7928431
Most material is allowed except pretty much loli
And Twitter really only ever cares about Loli
>>
>>7928432
Is the patreon about fucking trucks?
>>
File: catbox_0fksw3.png (1 MB, 832x1152)
1 MB
1 MB PNG
>>7928415
>>7928417
https://mega.nz/folder/70tQhTpJ#e1PEOIbeXq7Wfx0YYwk6kw/folder/LsMxRThQ

>>7928416
oh yeah I saw him posting his prompt on twitter, what a chad.


the lora is kinda overfitted or some shit, but I love the art style so w/e. Gen at 0.6-0.8 and you might need to put multiple_views in neg.
I photoshopped a bunch of white skin dudes with the black dudes so the lora doesn't give you bbc when you don't want it.
>>
>>7928427
wonder why the AI likes to make orc penis have pink tips
that is not how a real orc penis looks like
>>
>>7928432
No, it can't be you.
THE guy????

If that's the case I'm actually following you already and paying you. Fuck.

Hey by the way do you perhaps know what prompts floox is using?
>>
>>7928439
>I'm actually following you already and paying you
Based piggie.
>>
bakes are going good bros
>>
>People unironically pay for AI GENERATED CONTENT
>Pay for stolen content
>Even here
Lmaoooooooooooo what a fucking loser
>>
>>7928440
>>7928442

Just the $3 version. And I did it just to get the prompts. And that was only for like 1 month

In fact I started my whole Patreon journey because of this fucker
>>
File: image(2902с).jpg (2.79 MB, 3200x1984)
2.79 MB
2.79 MB JPG
>>
>>7928442
it's the attentionwhore collage schizo so it makes sense
>>
>>7928446
the desk phasing into the floor is very distracting
>>
File: 00006-2601808666.png (3.31 MB, 2048x1632)
3.31 MB
3.31 MB PNG
>Link vs Groose
>>
Guys, properly tag your dataset and choose only the best images. Really made the difference on my v3 to v4 lora.
>>
>>7928448
classrooms are absolutely schizo even on xl models sadly
>>
>>7928444
>And I did it just to get the prompts
Is he retarded enough to not delete metadata?
>>
>>7928451
you can do better.
>>
>>7928446
Beautiful artstyle
What Loras are you using?
>>
diplomacy with elves
>>
>>