/vt/ - Virtual Youtubers

Request anchor

damn last thread died as i was posting the link to this one

>>45791345
guess I need to prompt lui with her eggs then

No one man should have all this power
https://litter.catbox.moe/ev4c6l.png

>>45791171
>>45791036
So I have a variant of my mix I used for holosuccubi and arabian holos where I've swapped protogen for basil in the pesto component of the model. While it's not super dark, it's certainly darker than what I was getting with the pesto variant.

>>45791534
Yeah, also note to self: prompt my own nerd Faunas so I don't have to steal someone else's

>>45791036
Most of mine are mixed with CreamyPesto412, but picrel is 0.5(luckyStrikeMix+hll3) + 0.5(nightSkyYOZORA+hll3)

+hll3 just means + (hll3vtubers-last-pruned - animefinal-full-pruned)

>>45791650
>cursed Fauna trying not to slap herself

>>45791467
Can I have a multiMoom image? Several Mumeis lying in bed together, inviting you to join them in a nice, snuggly Moompile? Preferably in some comfy PJs or casual wear.

>>45791467
I request your finest bottom heavy chuubas. The {{{bigger}}} the better.

>>45785541
maybe 7% is enough to mess up mixing, somehow?

>>45791623
i dont know what protogen is. i really should have used a dark girl instead of the great rosemi sama to test all these models
>>45791734
it creamypesto 412 on grugger's huggingface? (in collagb) i didn't see it in the rentry

captcha:MANNM

>>45791623
Abductable.
>>45791650
...she was supposed to be dancing. I grugged the lyrics to Nobody But Me (the famous cover by The Human Beinz , not the Isley Brothers original, they are VERY different).

well..thats alot of eggs

>>45791897
>>45791737
Tagged myself instead of you for some reason.

>>45791882
https://rentry.org/8nxtk#creamypesto412nuhard69a8d0d4b2

>>45792107
oh did you just trim the name on it?

>>45792148
ye

>>45791467
Realistic EN models please
it doesn't have to be the same as the pic, something similar to it is good enough

Let's see those prompts get {{{bigger}}}

>>45791467
Anyone know any other mixing strategies for the hll model? I am banging my head against the wall here to trying to keep character features while discarding EVERYTHING else.
(I am looking for MBW weights that people have successfully used with hll, or detailed information about where the character details reside in the hll U-Net.)

>>45792283
You cannot handle my biggest deer udders.

>>45791650
Was surprised to see one of my first gens again.

>>45792215
here you go pestophile. come back next week when the high wears off.

>>45791882
>i dont know what protogen is
https://civitai com/models/3666/protogen-x34-official-release
photorealism model. Pesto uses Basil (hehe, kitchen joke).

Dark skinned Coco as bonus

>>45791467
Fubuki and Mio after gaining weight, Fubuki disappointed in her still-lacking chest but Mio is big all over. Alternativly, some fubuki squeezing Mio's soft places

Also this >>45792283

>>45792673
im sorry i did not memorize every model on the civitai catalogue. i will not make this mistake again.

>>45792283
Why does she look so young? Not that I'm complaining...

>>45792863
This picture was taken before I put 4 kids into her anon

well the anal isn't doing so hot so back to this pose https://files.catbox.moe/siz405.png

>>45792759
>Provide explanation and link to resource so anons can understand what I'm talking about precisely because anon said he didn't know what Protogen is.
>Get some snarky comment about not knowing every model on civit.
Why?

>>45791881
Wait why do I have different percentages?
And different checkpoint hashes? Unless nai-full-conv is the same good old nai
Also share similarity script that makes seed random? Im still on 114514

>>45793464
because he thinks greentexting is only for epic snarky comments and not for actual quoting

>>45791738
Never tried multiple girls before, I doubt spamming the character you want 3 time is the way, but hey this is what I got.

>>45793574
> Nanashi Mumei, (brown eyes:1.2), (ponytail:1.1), feather hair ornament, brown hair, Nanashi Mumei, (brown eyes:1.2), (ponytail:1.1), feather hair ornament, brown hair, Nanashi Mumei, (brown eyes:1.2), (ponytail:1.1)
kek

>>45793574
Shockingly effective brute force method! That one on the left looks like it's got a little Moom in the oven...

>>45793574
Make base resolution bigger, multiple girls tag. This is 1536*768

I gotta sk. does she have a gorilla grip pussy? https://files.catbox.moe/e3aah6.png

>>45794865
Is my dick supposed to be sore three days later?

>>45791982
Wow where'd they come from :D

>>45793493
It will show different percentages with different seeds.
nai-full-conv is just old nai-full, re-saved in model toolkit.
It's not important. My point is:
- hll3 was similar to nai in style, so when you merge using "add difference" - difference is just what it learned about vtubers.
- hll4 have slightly different style compared to nai. So when you merge, difference is vtubers AND style.

Good evening /vtai/!

>bbw_hll3.1
This was merged using add difference at a 0.4 multiplier. I think this might work. This will need more testing, plus a comparison against an MBW version of this model merge.

>>45795136
after kanata yes

>>45791467
PIPPA WITH PIPPANTS

>>45795681
jesus christ anon your asses are so good. what's the biggest dump truck you can make?

>>45793574
I didn't like the left one so I ran it through some more, I satisfied with this one.
>>45794098
I'm working with a toaster oven, 720x720 is the biggest I can bake. I'm upgrading soon though.

>>45795826
I think this is the biggest I've made in the past couple days.

>>45795922
Im gonna go feral

>>45795922
Suisei sit on my face, please

>>45796014
Pretty Ewiwa

>>45795922
catbox?

>>45795681
how was that ass constructed i must know

>>45796255
dozo
https://files.catbox.moe/zu0ypa.png
its a depthnet
https://files.catbox.moe/48x26w.png

>>45795922
Wow, can I get this with subaru or owl?

>>45796488
mind if I ask what the original image was that was used?

>>45796488
>Depthnet
I'm disappointed.

>>45796488
what lora is dishwash? dishwasher1910 style?

>>45796669
https://files.catbox.moe/qkjh9d.jpg
>>45796371
https://files.catbox.moe/zqfwhu.png
>>45796688
I don't really understand the nets very well so I prompted the rest of the Suisei's without them anyways.

>>45796908
>so I prompted the rest of the Suisei's without them anyways.
I am no longer disappointed.

>>45796952
>I am no longer disappointed.
What's your hang up with depth nets anyways?
>>45796528
You can have Shuba. Moomers didn't come very well.

>>45795922
>>45797155
Since you've done these two, how about either chocolate elf or the holo with canonically the largest rear? (Marine)

>>45797155
I think generating the pose without help is more impressive than using a depth net, since it feels like cheating.
(Although for particularly complex poses involving multiple items such as a gun pose, I can understand using a controller model)
For something that simple I don't think it was required.

>>45797345
NTA but I feel like it's really just to help get the size they wanted. prompting for size has such varying results imo so I could understand using depth for that.

>generating the pose without help is more impressive

The prompt is simple, you just have to gacha till you get one you really like. The depth net helps with size and shape so you dont have to gen a ton.

>>45796901
Yes, it's in....... One of the megas. I'll upload a catbox after dinner.

>>45797310
Eh, I didn't like how Flare came on this. And Marine barely looks like Marine. Now I want to prompt some different stuff in the short time I'll be here tonight.
>>45797345
What >>45797484 and >>45797599 said. Prompting for size is a pain in the ass for AI. It just prevents a lot of time from being wasted especially with a simple pose.

>>45797985
Thank you for this

>>45791220
Ina, sana, and kronii im pretty sure are the only ones I haven't done yet

>>45797599
Can I get a catbox of this? I really like how you do these

>>45795681
Good to see you again, ass-man. That's some sublime, 2B level sui-ass you have there.

>>45798628
I would do unspeakable things to this rat.

Back to testing merunyaa LoRA. Looks best at 0.7 weight IMO. Picrel. It's Okayu for those who aren't form radicalists.

>>45798232
Beautiful background. Really nice. The Ina is lovely too.

More Merunyaa 0.7: Moomers.

>>45797783
ty anon, looking forward to it

>>45798628
Something about using sepia and the orange background makes me like it more. Not sure film grain really does anything, but Im just copy pasting the initial prompt.
https://files.catbox.moe/5pjc3r.png

>>45798094
Anytime my friend!

I'll be heading out now but I did manage to get this Suisquat just right first. I'll probably drop in again later tonight in a few hours.
>>45798691
Hey there Succubro!I wish I could styed for longer.
>2B level sui-ass you have there.
Funny mention that. I made some 2B Sui-ass on NovelAI back in teh day.
https://files.catbox.moe/cu84fo.png
Maybe i should revisit it sometime with how far the AI has come.

>>45799261
We all start somewhere. Pretty good ass even on NAI to be fair.

This Mimonel LoRA might be my favorite so far purely on account of the backgrounds. The Matsuri dumpy doesn't hurt either.

>>45792346
>(I am looking for MBW weights that people have successfully used with hll, or detailed information about where the character details reside in the hll U-Net.)
1 person here has used MBW. For the latter question, you'd probably need to ask hllanon
>>45791882
>it creamypesto 412 on grugger's huggingface? (in collagb) i didn't see it in the rentry
It is in the rentry, looks like someone already linked you. I made a separate repo for it in an organization - so I could "give" it to pestonuhard-anon, but they declined to make an account

>>45799547
catbox please?

think I'm done for the day https://files.catbox.moe/t0f4xw.png

>>45799261
>>45726948
Missed this post completely in the previous thread, but something like this was what I had in mind for the split:
https://danbooru.donmai.us/posts/4718642?q=split+ass_focus+

>>45799648
https://files.catbox.moe/hnlx2o.png

>>45799261
Thank you for the constant feed of quality ass pics the last couple of days.

>>45799074
Cute Oni. I'm always surprised by the AI's ability to effortlessly spit out sexy underwear.

>>45788296

>>45792215
Here you go
asked bing to show me "gawr gura", kek

>>45799632
I was specifically looking for creamypeso412, not orangecreamypesto412nuhardfusion. That's why anon had to clarify he renamed it.

>>45800312
Gomen, I don't know my head from my ass with pestomixes
>>45792215
I am...surprised. I thought Bing was just DALL-E?

>>45798628
I might go back to vintageposting. I like how it makes certain parts have more shine to them

>>45800689
requesting the uncensored version

>>45800484
One of the inaposters mixes a new pesto weekly

>>45800720
https://files.catbox.moe/j0idtd.png
Gonna warn u rq, they gave her a big asshole and very plump lips

>>45801248
i love both of those things so this is perfect, thank you

>>45796878
pretty! but too big
>>45798232
oooh I like the mix between her regular out fit and her fancy outfit

>>45796901
dishwasher lora, not to be confused with dickwasher hypernet
https://files.catbox.moe/mc3g7h.safetensors

>>45801248
oh my god that is pure sex

Don't think I know how to use dynamic thresholding. Or even what it's supposed to look like when I'm using it correctly. Mimic scale was 7, Threshold percentile was 99, Mimic mode and CFG mode were Half Cosine Up and scale minimum 3, if anyone cares

>>45791982
is she going to make more?

>>45791467
Your oshi doing the skeletor 'joke's on you, I'm into that shit' pose. Picrel. Do I have to spoiler skeletor?

>>45802414
If Claude Monet painted Nijis? I like.

More Mimonel LoRA testing.

>>45791467
I've had a rough day and could use a happy chubby kid Fauna to cheer me up.

>>45802559

>>45802864
Excuse me while I go tell myself it's a bad idea to train a lora on his art.

whydo the faggots on civitai not prune their shit?

>>45803140
woah thats cute. catbox?

>>45803177
Cus they are faggots
Leave a 1 star review

>>45803177
The majority of people only have 5 models or less that they ever actually use so they have no reason to care how big they are.

>>45803239
Using a currently unreleased model mix, quick warning. https://files.catbox.moe/jf8u0t.png

>>45803312
as long as the cute shit is in a lora i don't care

>>45803312
>>45803351 (me)
dammit.
okay. probably should have read the tags first

Wow, pippers is proompting AI waifus as we speak.

>>45803351
>>45803418
Not a single lora used, sorry. The main model in the mix is Crosskemonog, https://civitai.com/models/11888/crosskemonofurrymodelandhumanmodel, and you can probably mimic the pastel colors with the dynamic thresholding settings.

Good morning vtai
>>45773043
>>45773082
Nah they're not the same, Biker Lulu would never ride a bike while drunk.

>>45803177
>Unpruned models
>"Works best at 0.6" LoRAs
>(masterpiece:1.2), (ultra-detailed), (8k Unity Wallpaper:1.3), (best quality:1.4) example prompts
Place is a meme with a handful of quality posters

>>45803553
ok

>>45800111

>>45803564
Good morning, Lulu poster!

>>45803594
and then some gigachad rolls out a dark nipples lora where the weight of the lora is a direct slider for nipple darkness and has absolutely no side effects on your image at all. one star rating.

>>45803564
is this lulu out on another baby making mission?
I hope the twins are doing okay btw

>>45803594
like half these models have broken clip indexes too. how do you even fuck that up?

>>45803653
Hello FubuMioposter
>>45803761
That's just how she is, getting heated in the outdoor, doing inappropriate stuffs in public

In other news, Twin Lulu are doing fine

okay at this point they're just troling

>>45791467
>(Uoh)/Chibi Bae receiving Milk training
She must be corrected and do her DB reps
How do you not know what "OVER 9000" means

>>45776900
Phew lad these are fucking amazing, Wolf Lumi doesn’t seem to get enough lewd art, AI or otherwise
I actually wanted Kaneko Lumi but frankly this works even better because it’s of a chuuba who is woefully under-represented

>>45791467
Ina using Takodachi as chest padding for a nice bra

>>45804563
>>45791467
>Ina picrel with Takos in place of the balls
>Ina/Gooba/Pettan picrel with the balls proper

>>45804896
are you planning on uploading this model or should i just try and reverse engineer it from the furry model you showed earlier?

>>45804986
Lulu's Lulus...

>>45803564
When are we going to see drunk driving Lulu?

>>45804365
>A model with worse clip positions than Elysium v3
Now I've seen it all

>>45805007
It's not finished, because I wanted to continue testing out merging a certain model in to improve it's NSFW capabilities, but I can upload it to pixeldrain if you really wanted it now.

>>45802414
what does the prompt look like?
for the middle option - since that'll be the easiest one to explain how dynamic thresholding does: it adjusts scale over time
>spoiler
So since you used half cosine up, with cfg set to 7 and mimic set to 7 - it started at 3, your minimum, and went up to 3.5 - half of 7. the "half cosine" scheduler uses a cosine curve and goes up to half of your set real cfg by the last step
sir, I am in AWE of your Comic Sans xyz grid, please let me know how you did this
>>44919471
quoting and comparing some examples I made
>>44879740
big comparison picture
>>44918943
another backlink/quote list
4chanx if you aren't already using it...

99 is way too high of a threshold percentile even for cfg 30. go lower.

>>45803177
>>45803273
>>45803298
>>45803594
>>45803744
>>45803874
Unfathomably based

>>45804365
holy fucking shit, does it even have a functional CLIP encoder at this point?
post x/y/z of a detailed prompt before/after the fix, kek

Glaze-Namie-LoRA-anon, I'm packaging this turd up for you at the moment.

>>45805192
im sorry i already deleted the orignal, give me a second to redownload it, drive to the store, and come back.

>>45805192
oh sorry, didn't notice you also called me about the lora
I'll cook that after i make the comparison
after i come back from the store

pvc tako not sure if you are around, but I extracted a LoRA from the PVC checkpoint you use. I used NAI as the base model for the LoRA (not sure if that was right - I'm new to this). Anyways, it seems to do alright with grapefruit juice & hll3.1. I tried it with a 20/80 PVC/grapefruit model merge, slightly modified your 2-shot prompt and it seems to yield okay results.

https://files.catbox.moe/5zb206.png
https://mega.nz/folder/IfhWWJoC#kI-DbIIs7N960N345Gb1Nw

>>45804986
comfy family

>>45804986
>I was there...
Anon better be a good father to those little Lulus >:0

>>45805377
ooooh I'm currently making my own ina lora dealing with python dependency hell
I'll try it out! One problem I had with just merging models was that it watered down the pvc look, hopefully it's not an issue with your solution!

>>45805146
The comic sans x/y grid was just made in paint.net, I'm sorry. If you go to User Interface in settings there's a font option, but I've never used it. Here's a catbox of one of the images. https://files.catbox.moe/eptbn1.png
I'll definitely be testing it out more, it seems pretty interesting, if highly confusing.

>>45791604
holy shit, how?

>>45805236
>>45805299
get me a cornetto

I'm doing something a little autistic only a little with the RAR I am making. Originally, the plan was for "1, 2, 3" folders for each set of 20 images I ran. Now:
The raw runs at intensity 2
>1
>2
>3
Folders with "completed" versions of each of the 20 images, mixed runs
>1_F
>2_F
>3_F
Picrel did not successfully Glaze even at the highest fucking setting, so folder 2_F is one image short
The runs of the "stragglers" at intensity 3, 4, and 5:
>4
>5
>6 - One image, still failed
All the source images in 20 image folders
>1
>2
>3

RAR lets you deduplicate while the archive is being made, no wasted space in the archive. Will probably waste an extra dozen megabytes when you extract it if it can't into making hard/sym-links at extraction time.

>>45805715
>1girl, solo, elira pendora, blue hair, blue eyes, head wings, hair over one eye, hair ornament, squatting, peace sign, head tilt, closed mouth, smile, white sundress, spaghetti strap, flower field
You're going to have to prompt a bit harder or schizoprompt a little even - no offense intended with the prior comment, I can only see room for dynamic thresholding to give you interesting fuckery in the background here. Add something ridiculous like (fine fabric emphasis, floral print:1.4) and watch what it does.

>>45805902
Original image

>>45805902
The glaze source code was leaked/cracked
The incomplete part is it running a CLIP test on the glazed image to see if it still holds up to the old one.

>>45805902
Oh, I misspoke - 3_F would be the same folder as raw_run_3 all 20 images succeeded at intensity-2 - so there's no need to make both

>>45806195
fucking kek. no wonder that step is dogshit slow, it's running CLIP on a CPU

>>45805377
oh also according to civitai, pvc is based on SD 1.4
--
I just went to check civitai and there's a brand new pvc lora https://civitai.com/models/15509/aoaoko-pvc-style-model released ~2 days ago.
looks like I wasn't the only one who wanted it.
Thanks for your work in any case, I'll try both.

>>45806440
NAI based models on Civitai say they're based on SD1.4 because it's the closest model available.

>>45805656
Yeah you should be able to turn up the LoRA strength to get the PVC look. However I find it tends to burn things easily especially if you used a PVC mixed-model.

>>45806526
oh I see. still kinda new to this, thanks for the info
>>45806540
if you tested with my exact prompt from the other day it might be over schizo'd because I had to really increase the weights for them to be used at all.

>>45806287
It defaults all system settings to none while also not providing a windows user the means to change them. So no wonder it is slow

>>45805105

>>45805192
>>45805299
>>45805902
>>45806287
Huh, interesting. WinRAR will actually re-create hard links when you extract an archive that contains them. Neat. Dozo: https://files.catbox.moe/zd1c3k.rar
If anyone wants to peruse what images look like at various levels of Glaze, there you go.
>>45806996
It's dogshit for many more reasons.

>>45791467
Bimbo Danchou

Thats it

>>45807339
windows xp my beloved

>>45807606
3.1 actually, kek

Well I didn't get as much done today as I would have liked but I do have a bigger pool of sleepies to say good night with.
Also good night.

Good morning.

>>45805902
I'll definitely be testing this more, but now all of a sudden I'm preoccupied with getting the "Font for image grids that have text" setting working. Neither Comic Sans nor Comic Sans MS worked. Ah, the many distractions and novel ways to procrastinate in one program.

>>45808078
Are you running it locally? Try "comic.ttf"

>>45808078
>>45808160
and if that doesn't work, just throw
C:\Windows\Fonts\comic.ttf
in that box

>>45808160
>>45808371
comic.ttf worked! I will no longer have to throw my plots into outside editing programs for stupid jokes.

>>45808545
>insert smug image of a chuba here
nice

>>45807976
Wawa

>>45791467
(You) looking into a mirror and suddenly realizing...
(You) are Lamy

>>45808609
The model is trying its best. I do not think it understands the concept of smug.

>>45809078
is that the same one with the busted clip encoder?

>>45809237
Just checked with toolkit, completely fine clip encoder.

>>45807339
Home, will compare clip now
>>45809237
You got your anons mixed up

>>45809537
oops

>>45808968

>>45809828
AIIIIIIIIIIIIIIIIIIIIIIEEEEEEEEEEEEEEEEEEEEE

>>45808968

>>45809828

>>45809828
Holy shit I'm cute

>>45810011
You know, if it wasn't for the sink, I would've thought this was a funeral or memorial.

>>45810123
Less fucked finger

>>45810334
I wonder if the AI knows what an Iei is. Kind of doubt it.

>>45810445
MY BROTHER AND I BECAME YUKIHANA LAMY!

>>45805377
based pvc lora. i trust any lora here more than one posted on civit after seeing the garbage checkpoints I've been downloading.
>>45807976
morning. that lora looks amazing.
>>45805135
i'd love a version now that's cute and sfw. not every model needs to be a porn generator. >>45806195
has anyone unfucked the source and uploaded a good program so we can actually test this shit in a human time frame
>>45808545
god this shit is cute
>>45805192
setting up xy now

>>45802016
thanks anon
have a chigusa
https://files.catbox.moe/th2sfb.png

>>45805007
>>45810786
Alright, MoeMixVtubers uploading to pixeldrain now, it's giving me an ETA of 7 hours, so I guess it'll take a while.

>>45811007
pixeldrain moment.
it actually speeds up after a bit.
the HDD i put the checkpoints on is being retarded. give it a moment.

>>45811089
It's bumped down 2 hours to a 5 hours eta, I'm just going to go make some food and hope it's finished by then.

>>45809828
Could you catbox that please?

¿¿¿¿¿does clip even matter?????
hires: https://files.catbox.moe/wlfoul.png
>>45811363
it'll take like half an hour to an hour.
>>45810802
yw

>>45811492
https://files.catbox.moe/ostuhe.png

>>45811818 (me)
>>45805192
ohhhhhhhhhhhhhhhhhhhhhh
i bet if the clip is completely destroyed it loads the clip model from your previously used checkpoint. same as how it used the previous checkpoint if you tried to load a lora as one.

>>45811906
Thank you.

https://litter.catbox.moe/qhutbr.png

>>45809828
>>45811906
How will [I] avoid being noticed as a Lamy? Can I bandage up her chest, put her hair in a bun, put a hood up to (mostly) cover her ears and try to wear baggy clothes so I can make it to a doctor to get my condition examined?

>>45811363
Got back from making food, tab says file finished uploading. Click it, site reloads, and it's just like before I uploaded anything there. Clicked my upload history, and it's not there either. If I go make some more food, will it fix itself? My x/y plot ran out of memory as well, I was fully expecting that one, however.

>>45803653

>>45804986
Look at them! They're so happy :3
>>45807226
Lulu NOOOOO

>>45812772
i don't remember hooters being like this
>>45813460
oh also it does this stupid shit where it verifies the files for like 10 minutes after you upload
>>45805192
read request more carefully. interrogated clip on the rosemi test image using NAI, then sent that to the prompts and ran the two checkpoints again.
43/75 tokens used.
interrogate clip on the rosemi standard image:
a woman sitting on a chair with a flower in her hair and a flower in her hair, wearing a red dress and green stockings, Eizan Kikukawa, official art, a detailed painting, rococo

https://files.catbox.moe/cxq6x7.png

>>45805825
I've trained a lora based on Kaela's designer, Yasuda Suzuhito. The dataset is made from somewhat equal parts Durarara screenshots, art from Suzuhito himself, and thumbnails/screenshots from Kaela's own streams.
https://mega.nz/folder/ITowiSZJ#qnum28NHpwfyCj8pu8Hq1w
Final may be a little overbaked at 1 so I included the penultimate epoch. Also included is a lora for Kaela's kimono outfit, which may need more work.
This https://litter.catbox.moe/t6j1zb.png
did not even include her name in the prompt, I believe, and this one has the artist lora and the kimono lora activated:
https://litter.catbox.moe/5rzqk0.png

>>45791467
Request for Lize Helesta in see-through dresses like in >>45784021

>>45806195
>>45807339
so since it breaks clip, supposedly, i just tried something
raw:a painting of a dog sitting on a red bridge over a pond of water with a red bridge in the background, Alice Prin, anime art, a storybook illustration, fantasy art
glaze:a painting of a person sitting on a red bridge over a pond with a dog on it's back, Alice Prin, anime art, a storybook illustration, fantasy art

same thing
base: a woman in a bikini holding a sword in the water with a fish in her hand and a fish in her other hand, Ai-Mitsu, official art, a detailed painting, sots art
glaze:a woman in a bikini holding a sword in the water with a fish on her head and a sword in her hand, Baiōken Eishun, official art, a detailed painting, rayonism

>>45806195
>>45807339
i just realized. they might actually be genuinely retarded. Remember how i said this might not work for lora because they didn't think about it? I think the way this works is that it makes it so you cannot train a model using CLIP interrogation. but nobody uses that shit, they interrogate with deepbooru or wd14.

Here's the deepbooru on >>45814386
base: 1girl, bangs, bare_shoulders, belt, bikini, blue_eyes, breasts, goggles, goggles_on_head, hair_ornament, katana, katsushika_hokusai_\(fate\), looking_at_viewer, medium_breasts, ocean, octopus, one_eye_closed, purple_hair, short_hair, smile, swimsuit, sword, tokitarou_\(fate\), tongue, tongue_out, water, waves, weapon, white_bikini

glaze: 1girl, bangs, bare_shoulders, belt, bikini, blue_eyes, breasts, closed_mouth, goggles, goggles_on_head, hair_ornament, jewelry, katana, katsushika_hokusai_\(fate\), looking_at_viewer, medium_breasts, octopus, one_eye_closed, purple_hair, short_hair, swimsuit, sword, tokitarou_\(fate\), tongue, tongue_out, water, weapon, white_bikini

>>45808927
Iwys!
>>45810786
It sure does look nice.

>>45791467
>Kiwawa cliprel
https://youtu.be/0I7w8TzG45U

>>45813846
killer,
also when did loras become so tiny?

>>45816250
Man she's built like a goblin here...

>>45813846
i fucking kneel.
can't wait for more.

Good morning!
I was browsing civitai and found something I had wanted but didn't think would ever prop up, a new VAE.
https://civitai.com/models/22354/clearvae
Somewhere in beween NAI and the other one, more vibrant than NAI but seems less prone to color-deeprfy

>>45792283

>>45791467
Perhaps a postapocalyptic Moom doing some maintenance/cleaning of her firearm of choice?

>>45816678
Thats pretty thic. Can I request a shortstack Marine?

>>45816820
Several of the checkpoints have "custom" vaes. A lot seem the same as sd vae though so ymmv

How's hll4p1 for those trying it?

>>45798628
Kronii's out this list now

>>45816665
For a month or so now many lorafags have been experimenting with settings, including optimizers and previously-believed-to-be-recommended settings. The loras I make nowadays use the lion optimizer and are done at 768 res, 32 DIM, 16 alpha; before that I was using 64/64 with lion.

>>45817730
Messy. It has enough style difference from NAI that add difference doesn't work right. A new one is being made with certain artists removed.

>>45817772

>>45817830
you said this minutes before i was about to make the first namie glaze lora

>>45817901
uncensored please?

How do i limit max memory? Having this issue where if im watching youtube or doing something else stable diff freezes or errors out until i close that tab. Though sometimes it will be perfectly fine. Feels like if i started rendering with the video on its fine but if i change vids then it errors out.

>>45791467
paizuri with request anchor Scarle, who is visibly enjoying it

>>45817830
what kind of learning rate have you been settling on? I'm still new and have been using 1e-4 with 4 repeats

>>45817730
I'm having fun experimenting with mixes, it does have a lot more style than hll31 which makes it less of a 'slap onto anything to have it know vtubers' model and more of its own thing

>>45817730
thanks for doing Rein-o

>>45816678

>>45817449
Thicker than a bowl of cottage cheese

>>45818203
https://files.catbox.moe/jbkhlr.png

>>45818638
Squeezable belly...

>>45817830
>>45818332 (me)
nevermind i figured it out

>>45816820
Neat, definitely seems like it might be a nice sweet spot

>>45818852
>Forgot the image

>>45818638
>>45818829
Damn she looks good like that. Thanks anon.

Lamia rabbit

>>45818332
For a while now I havent changed my learning rates of 1e-4 unet and 2e-5 text_encoder, though that text encoder rate might be pretty low. Besides those, I have 6 epochs, batch size 2 (my max with resolution 768, if your batch size is higher you would need to scale up accordingly), and 3 cosine restarts. I haven't changed these in quite a while and the only changes I make are to my repeats depending on how large the dataset is.
Based on my experience with character and style loras, I start with 8 repeats on datasets with ~60 images, scaling inversely with dataset size; 100 being 5 or 6, 150 with 4, 200 with 2.

Concept loras elude me, I'm currently working on a futa lora and it seems to be woefully underbaked unless it's trained for twice+ as many steps as any of my other loras

>>45805377
Looks like it's working pretty well! using it on top of a pvc merged model seems to bring the realistic pvc look back out! My only issue is that I have trouble to get ina come out well of the hll models...

>>45819057
oh you got a couple more images than i do for this lora. lmao.
Trying to see if glazing fucks this namie lora but we only have 60 images because it takes over an hour to glaze them and about 1/5 images fails to glaze and has to be redone.

thanks for the repeat information. i saw this one used 6 but the hint of 8 will be useful.

>>45818962
I dunno, I still think klf8 looks really good. Love the vibrant colors that no other vae seems to provide. And I haven't noticed any 'deep frying' with it or whatever

>>45819191 (me)
dum
i ment i saw it had 4 repeats and 6 epochs

>>45818829

>>45819035
Np, I like seeing girls whose bodymass is concentrated in their tits, thighs, and ass.

test

>>45819191
I don't know why, but that bit is wrong, the dataset was 196 images for that one

>>45819387
It might have something to do with the bucketing and flip_aug

>>45819387
oh, must be counting the 4 repeats. 196*4=784

>>45818421
Which mix would you recommend for getting stuff similar to CreamyPesto412nuhard?

>>45819463
oh yeah I forgot it treats repeats like that

>>45819469
the mix i'd reccomend, if you want to get like really similar, like super duper close to creamypesto412nuhard, is creamypesto412nuhard

>>45819297
I'm not even asking for ripped legwear, but thank you ai.

>>45818751
that's a nice tummy

>>45819555
La goblina...

Heart gacha

>>45818540
nta but no wonder she looks familiar.

neat

>>45818638
>>45818829
catboxes please?

>>45819555
could you catbox that too onegai, i didn't finish scrolling

>>45820103

>>45819503
>>45807339
cooking now

>>45817830
i googled this error code. Did you get it too or do something before you ran lion the first time?

>>45807226
>>45813662
>A... anon, I'm feeling hot now, let's stop for a while and do it
https://files.catbox.moe/i819he.png

>>45820827
Whenever python tells you something like that, it means you need to install a package to be able to use that module. In this case, you need to install lion_pytorch. Make sure you have the python venv activated before you install with pip install lion_pytorch, same would go for other optimizers (look them up if you want to experiment)
I'm going to bed while my sarav2 lora bakes

>>45820992
honestly, i didn't try that because i expected lion to be too unique to this lora training program for pip to find it. clearly i don't understand exactly how pip works

>>45821127
pip basically pulls from very live repositories, and if something isn't found automatically it may still be installed with pip install git+<githuburl>

>>45821170
>pip install git+<githuburl>
i was expecting to have to do that for lion

this shit is a cunthair away from throwing a memory error lmao

>>45821309
I'm surprised you can run within 12gb, my gens always take between 16 and 23

>>45821405
???? are you using xformers?? i can literally do 512*762 hires 2x (maybe even more i haven't tested) with batch size 2

>>45821549
>No module 'xformers'. Proceeding without it.
Well I do get this error

take a plastic rosemi made by my 12gb gpu in like 13 seconds
>>45821648
i forgor what i did to install xformers. you added it to your user bat as --xformers right?

>>45821435
Welcome to the church, Sister

Which one of you faggots is reposting in /vp/

>>45803011

>>45821699
what settings are you using? I'm using >>45805377 's lora but I can't get a checkpoint to both look plastic enough and have ina come out as not a literal eldritch abomination
And no I'm not using that argument

>>45821845
I know this could be anyone, but it looks really similar to this anon's model >>45335266

>>45821887
>Gets biblically accurate ina
>"Why is it not working?"
That's using the prompts off my standard rosemi test prompts (you've likely seen a few of them posted here) which I don't have on hand with basic creamsicle and the lora at weight one. It is the card image for that lora.
I don't know if you're still running that ina lora but if it's on an hll checkpoint you don't need it or can reduce weight a lot.

So many gens I dont understand how people are supposed to look at this angle anymore.

>>45822162
This exact image >>45338978

>>45821845
Posting cunny to the catalogue of a blueboard? Brave mf.

>>45822265
Could you catbox that please?

>>45822289
I highly doubt it's a generator that's breaching containment, it has to be an autist beggar or lurker
Gonna catch /vtai/ flak for being a retard and posting AI chuubas if anybody recognizes it like I did =`_'=
>>45822372
They have shota/cunny containment threads, but Holos are off topic regardless

>>45822450
Even with a containment thread, that's holo cunny straight to catalogue (not in thread) advertising a discord server

Kek

>>45822500
Dubs of truth
It's retardation no matter how you look at it

>>45817772

>>45822171
That's what I get with the lora at 1 using hll4-p1. Ina comes out pretty well but as soon as I try to merge it in something else, I need to do like (ninomae ina'nis:1.4) for her to even start appearing and even then it's completely messed up. Also the plastic effect is nowhere near your rosemi

>>45822265
Gens screw up more often than not when generating landscape compositions

>>45822616
oh hey, I've been wondering for a bit.

why do some posters only show censored versions without posting the catbox version?

just curious. not necessarily requesting the catbox.

>>45822706
>Hll4p1
Yeah use anything else. Raw hll mixes are filled with condensed chuuba essence and must be watered down to function. For now I'd use existing hll3 mixes because hll4 doesn't use add difference correctly.

If you must use hll4, creamsicle is aom2nsfw, hll, and NAI using add difference at weight 1 (add 100% of the difference between hll and NAI to aom)

>>45820835
If it means you'll stop driving drunk... then fine.

>>45822409
https://files.catbox.moe/6t36vh.png
>>45822821
Thats why I used a depth map. Made sure the feet and general pose were good most of the time.

>>45822894
>If you must use hll4
not really I just want ina in plastic and I only have one functioning braincell...

>>45822840
For me, its just depends if anyone is actually interested in it to begin with

>>45822894
>hll4 doesn't use add difference correctly
What?

It took awhile but I'm back until I get tired again.

>>45823050
Hll4p1 has so much style difference from NAI that it doesn't just add chuuba information when add difference is used, so it's significantly altering the styles of all the mixes we're more used to.

>>45819831
Marine probably isn't flexible enough to do this but whatever.
>>45820122
>>45820163
Here you go. Have fun with it.
https://litter.catbox.moe/gim7mo.png
https://litter.catbox.moe/ovmzf3.png
https://litter.catbox.moe/hwbvv4.png

>>45823353
thanks, i will

.... Suisei?

I saw this on catalog, is there any source and similar work?

for latent couple, can we generate a division of portion size not in the form of 1/n, ie 2/3?

>>45823032
i see, is this a "i don't want to waste catbox server space?" or a "eh, that's too much work" mentality for you? again, just curious. no need to answer.

>>45812108
but then what is it doing for people who use it without having previously loaded a checkpoint?
>>45813717
huh? I might try it myself later, I compared 7th_v2_C with/without fixing CLIP and saw a difference

>>45814174
>>45814386
KEK. so it's a CPU burn in tool that does fucking nothing?

>>45815676
FFS I think I've only ever used CLIP Interrogate on AI generated images, and I still usually click the Deepbooru button instead

>>45816820
Neat. Didn't realize VAE could be merged. Sounds like it's a merge between NAI VAE and WD VAE.
>Caution.
>NaNs errors (black images) may occur due to NAI's VAE.
>
>You can add --no-half-vae in the startup argument of 1111 or --disable-nan-check if you do not mind black images.
>
>About Variants
>Two types of VAEs have been created.
>
>ClearVAE_1 is a variant with less redness and more contrast than ClearVAE.

>>45821845
I don't think that's any of us. Same fucker spammed their Discord link in /vt/ using images of /vtai/ cunny a few days ago. https://archive.palanq.win/vt/search/text/https%3A%2F%2Fdiscord.gg%2Fcu3Pbj9h/
https://archive.palanq.win/vt/search/image/zwjMDs4reYWiCxcxRVHbyw/
again with a different discord link in a few generals and in the catalog the first time, because I remember seeing it in my general: https://archive.palanq.win/vt/search/text/https%3A%2F%2Fdiscord.gg%2FMeX5uCWM/

>>45822450
>>45822500
>>45822568
It's either a retard, or a fucking honeypot.

>>45823353
Time to try a different style now. Hope you guys like chocolate.

welp, I give up.
>>45805377
thanks for your work but I'm too retarded to get anything as good as you or as good as >>45821699
These are all the gens I tried in the last 8 hours https://litter.catbox.moe/nydwja.zip using only the pvc lora and using hll and hll derived models to get ina. I just can't get both the plastic look and ina at the same time.
I'm also knee deep in dependency hell for making my own ina lora and I can't get any scripts going. I'm just too retarded for this stuff

>>45824111
Too much work up until someone asks for it

I got tired before I prompted anything more.
Goodnight /vtai/

Hitfowl today

Bread?