>>29491 unironically why didn't you make her pregnant?

>>29492 Not to nitpick but is rouwei linked just in case it becomes good?

>>29494 call it something between hopium and "well, it's *technically* an option"

>>29495 It's definitely plausible for the next version to be good I think. It just has to be less underbaked.

anyone have an idea what artist mix this guy's using? https://x.com/dakini_tencho I'm guessing a decent amount of wagashi

>>29491 Not sure it you want to include noob_v_29_checkpoint-e0_s4000 but I still have the original link though it isn't actively seeded The cyberfix and itercomp merges of noob vpred v1 may still be worth including https://huggingface.co/Doctor-Shotgun/NoobAI-XL-Merges/tree/main Also https://huggingface.co/Toc/toc/tree/main/models is the only place I can find something close to NoobAI-XL-Vpred-v1.0+v29b-v2 [88fc3afcef] >>29497 Looks more like mikoscrub

>>29498 >mikoscrub nice, I'll try that out

>>29498 I found 1.0+v29 here: https://huggingface.co/autismanon/sdxl_modeldump/blob/main/Vpred-v1.0%2Bv29b-v2.safetensors and added that along with the Dr. S itercomp/cyberfix HF. >>29497 Also, I'm fairly confident one of the the artists in this IS wagashi, given the textured eyelashes. This is the classic wagashi + healthyman with both at 1.0, for reference.

s~h~u~t~ u~p~ m~o~m~o~u~r-- ah, wrong tab!

._.

kinda feeling the makings of an andy warhol meme

>>29503 Good gen premise.

pngs are back :DD >>29505 it's nice when the model isn't shitposting.

>>29508 It's consistent on my end, one of us might be doing something retarded.

>>29509 it was a lora.

This body type does things to me...

>>29512 cute wife

>>29508 >>29510 JOVI! :DDDDD

https://danbooru.donmai.us/posts/9045835?q=qusouzu_mogari >>29514 jovi (‿ˠ‿)

so this is the power of finding a sweet spot

>>29516 artists/catbox? that's some nice skin and face

>>29517 nta but I see some parororo and hotvr

>>29517 yeah sure https://files.catbox.moe/o2ckts.png it just tries to write one of the artist names sometimes with body writing tags https://files.catbox.moe/e7qp5y.png it's also more mature female biased, just trying to build some kind of dataset with it

4um really did dieded. i thought it was a shitpost 0_0

>>29498 >the original link Can i have it please? my dumbass forgot to download the torrents after all this time

>>29521 https://files.catbox.moe/p9sr25.txt

>>29519 thanks bwo

Shameful repost with bonus image

>>29525 Now do this

if anyone ever tells you the base model is better because [incomprehensible gibberish about fringe styles] you are well within your rights to blow their head smooth off

>>29527 i mean, i don't think there is an objective winner here. all noob vpred models (not counting completely slopped ones) are pretty similar with their own quirks

I'm gonna repost this here because I remember one anon asked me to share this lora but i just couldn't do it at that time because it was a shitty one, I hope he's still around here. It was trained on noob vpred 1.0 Also, I'm sharing this one. She's Monaca, an OC from the artist ShiShigaJ5. Here's his pixiv: https://www.pixiv.net/en/users/13302316. He used to post on /v/, and I think he still does from time to time. Her OC was partially responsible for turning me into a ped-- I mean, lolicon. I really, really love her, to the point that I downloaded koikatsu and her card just to take a bunch of screenshots to make this lora. Yes, it was trained only on koikatsu images, so it might have some issues. Besides that, I remember making a lora with this dataset about 8 months ago, but it was tagged really bad and it used to generate random stuff out of nowhere in the image. So I retagged it way better this time, with a confidence of 0.1, and spent a lot of time curating the tags. (v13 and v14 are loras trained with the old tags and v15 with the new tags). Without further ado here is the lora: https://files.catbox.moe/4xmkhi.safetensors and the trigger words: monacaN, short hair, [light brown hair:brown hair:0.4], blue eyes, short twintails, twin drills, hair ribbon, pink ribbon, thick eyebrows, Well that's all for today, thanks for reading my blogpost. Until next time, anons.

these wildcards were long so i'm sharing.

>>29527 >base model is better because [incomprehensible gibberish about fringe styles] i haven't seen anybody saying that base model is outright "better", just that it's more flexible which is correct. I use 102d and while it holds back some styles (painterly and sketchier ones), 95% of the time I don't care that much and stick to 102d now because it's much easier to use and I like its style.

>>29503 boxes?

>>29529 Thanks, great versatility for a lora trained only on koikatsu screenshots

>>29534 Catbox? That's some quality PVC

>>29535 https://files.catbox.moe/nds919.webp https://files.catbox.moe/tz1745.webp https://files.catbox.moe/lnzosz.webp https://files.catbox.moe/xp3zc8.webp PVC model: >>>>29450

>>29536 Shit, why didn't I just ctrl+f the last thread, I knew it looked familiar. Thanks tho.

I will test to see if I can post it.

>>29542 delicious flat chest

>>29543 very erotic

>>29526 Gave it a lazy try and I'm impressed my shitty prompt works, cheekjob isnt even a tag

>>29545 UOOOOOOOOOOOOOOHHHHHHHHHHHHHH

>>29542 very nice.

>>29545 nice

>>29529 thanks, loved the lora. Could you please share the artist mix or catbox?

>>29529 Speaking of which, NTA but I made one based on the more realistic version of Monaca (from Honey Select 2) if anyone wants it I can drop a link as well. >>29529

>>29552 >if anyone wants it I can drop a link as well. Yes please.

>>29553 This one is for Illust V0.1 though, (I'll train one on noob vpred 1.0 once I figure out how to rain on Vpred) the trigger tag is m0naca should be same hair and eye color, spoilered the preview image due to it not being very anime. (hope its okay) https://files.catbox.moe/i8h214.safetensors

Anyone who switched to lossy webp/avif, what quality settings are you using and have you ever encountered issues with inpainting/upscaling?

>>29540 nice

hort

>>29547 thanks >>29548 uoh

Has anyone ever made a quick rundown(tm) of pros/cons of comfy vs automatic-type ui? i never tried comfy but people always said it gives you "more control" but idk what that really means in practice

>>29562 Comfy pros/cons: + Hyper specific workflow creation and automation + Latest features + Developer friendly - Awful UI/UX compared to matured software (Blender, Houdini, etc.) - Unreadable raw metadata format (unless using something like https://pastebin.com/AJ6avY5f that) Invert for Auto and the other derivatives.

Lazy slopping

>>29491 haven't been around since nai vpred 1.0 dropped. any new models worth using since then?

>>29566 no major developments unless you count naiv4. we've gotten a lot of mixes of noob vpred 1.0 that have made it much more pleasant to use. some of the popular ones are linked in the OP rentry but otherwise you can try dumpster diving civitai.

yeah I still prefer v3's aesthetics

>>29563 I see, thanks. I think I'll just stick to reforge for now since I'm not generating too much lately

>>29569 imo there's no reason to use comfy if you aren't developing bleeding edge extensions or making highly automated workflows for some sort of slop mass-manufacturing

>>29567 i see thanks for the update. it's ashame that civitai is so full of trash that it's barely worth browsing but we all knew it would turn out like this

Innate artist mixing feels different on all illu/noob versions compared to v3 and pony. I don't get why. In my experience both v3 and pony seemed more capable of producing unique or interesting looking blends. Or maybe I just haven't used illu/noob enough.

>>29572 >I don't get why. I mean, they're all different models with different datasets, training strategies, and loss scaling. Artist mixing on Noob v-pred derivatives feels fine, it's just very sensitive and what artists are stronger than others can feel completely detached from their style or number of works. You can always try scheduled prompting, too: https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#prompt-editing I haven't done it much myself but sometimes I'll tweak artists slightly for the upscale step or adetailer.

this is probably the only AI place I know that is not extremely cringe in some way, even if it's slow thanks for keeping it up anons

>>29574 i just wish it were a tad bit more active, i'd post more but I dont want to spam the thread either

>>29574 I'm amazed that this place is still around. >>29575 I'd post more if I had the time to gen as much as I used to.

>>29574 Yeah weird how schizos took over almost every general on quad chan now, well here you know its so slow but comfy at least

>>29574 This place has almost no schizos, mostly because it's dead. The moment anyone talks shit about a model someone else likes, or posts a gen most people around here don't like, i.e., anything that's not 1girl, pov or kodomo only, they're most likely to start schizo fighting just like on half-chan. Don't think this place is superior to anywhere else. It's just another place.

as shit as v4 is for style consistency, at least my wife still comes out cute

>>29579 very cute

>>29578 >The moment anyone talks shit about a model someone else likes, or posts a gen most people around here don't like, i.e., anything that's not 1girl, pov or kodomo only, they're most likely to start schizo fighting just like on half-chan thread doesn't have to be a total hugbox. some arguments are fine unless they go overboard

>>29578 look, it's very easy: i've spent years talking to myself and turning my brain into a gay minefield full of stupid-ass rules that you simply need to not break or i'll shit myself in public

>>29567 >we're still stagnant fuck, back to my eternal slumber

>>29584 If anything major comes up I can implement it and have repo permissions to merge it on my own. Most stuff can be facilitated by extensions anyways though. You're already using hella old code by opting to use something like reForge since the Comfy backend is rarely synced with upstream at all, so I don't see merit to view this as a reason to abandon ship (yet).

>>29585 >If anything major comes up I can implement it and have repo permissions to merge it on my own. Thanks for the hard work. >You're already using hella old code by opting to use something like reForge since the Comfy backend is rarely synced with upstream at all, so I don't see merit to view this as a reason to abandon ship (yet). Fair, I just don't think comfy is comfy at all. Not even sure how you do stuff like inpainting with it.

>>29584 sad but yeah I don't see any reason to move from reforge for a while if you're only interested in SDXL-based models

I feel like I want to learn how to draw, I think. Any way to incorporate AI in learning?

>>29588 Gens can make for decent pose references and general design ideas. I'd be careful about using them to learn anatomy though.

someone hacked half-chan lmao

it was me

https://www.soyjak.st/soy/thread/10615723.html GET IN NOW

>quad chad hacked How in the fuck

who the fuck signs up to be a janny with their .edu address

I can't believe it's unironically over for 4chink. Anyway, my new home.

>>29595 it'll be back online soon enough will probably have significantly less jannies though kek

qrd?

>>29597 4chan was hacked, janny info leaked, some of the code was leaked, website is shut down by admins

>>29598 NTA but wow, they actually leaked source code, the absolute madlads

>>29579 >painterly, impressionism, [wamudraws], [opossummachine], fokkahim7, healthyman, tianliang duohe fangdongye, rustle, {{{asanagi}}}, [[[zankuro]]], lam(ramdayo), mika pikazo, aesthetic, digital illustration All this yet the output is 90% tianliang style NAIv4 style mixing doesn't feel right

I can't believe this happened

>>29600 Yeah its a mixed bag, you can take the same mix into a different prompt and itll change completely. The newer emphasis tagging works a lot better but, its still not the best.

>>29592 I'm so happy this happened. Fuck them, fuck the site and fuck their garbage moderation. The site deserves to burn in the fires of hell.

>>29596 >source code leaks >probably the captcha code too >it'll be back soon Don't hold your breath.

>>29604 they also don't likely don't have anyone to fix this mess seems like they were running a freebsd install that was EOL 9 years ago the soyfag had root access this will take a while to recover from

>>29601 It's over...

God i'm glad the board is unlisted or I'd have to actually do work for once if /hdg/ flooded here

>>29607 /hdg/ is pretty much dead anyway, its only created and bumped by shitposters thinking they're trolling other people when in reality they're the only occupants; they wouldn't go out of their way to find alternatives otherwise they would have been here or other ai generals long ago on 4chan

>>29608 yeah /hdg/ had turned into pure crossposter shitposting cesspool ages ago, those early comfy NAI 1.5 leak days are long gone

Right now, I guess we are split up between here and treechan.

anyone getting some strange behavior with tag autocomplete on reforge? if i have any letter enclosed in parentheses with a specified weight, typing that letter elsewhere in the prompt yields no results for autocomplete. for instance, if i have (v:0.5) in my prompt, all of a sudden the listings for v (i.e. very long hair, violet hair, virtual youtuber) disappears from any autocomplete search results

>>29610 people still use treechan?

>>29612 Just today.

>>29610 there is another one tho. meta.4chan.gay

>>29614 if you're into 3dpd I guess

We don't need the cum splotches here, right?

>>29616 nah there's plenty /e/ type stuff here

>>29616 Don't post ilegal shit Don't be a dickhead Absolutely no realism or anything that could look like a real living person Other than those rules, go nuts.

good to see that you guys are still around no gen because either comfy or my gpu is shitting itself >>29608 yeah i haven't been to the other hdg in over a year far as i know its a ghost town entirely inhabited by bots and not much better >>29610 did a quick check and this place is vastly better

>>29605 Their 3-6 hour estimate for down time is optimistic. >>29610 There's an /h/ here too. It's mostly dead though.

>>29618 Maybe moving here isn't so bad.

>>29619 restarting my computer fixed it hands are still fucked but thats normal

rip 4chinks

nice to see this place a little more active with halfchan nuked >>29620 catboxes? nice style

analtan teased V4 VT with this, it looks way more flexible than V3 VT but you can't convince me that they're not gonna treat it as a bandaid fix for artist mixing

>>29625 It's released actually, they just haven't announced it for some reason.

>>29625 >>29626 Yeah, it's out. Now you can share "vibes" and other shit. https://docs.novelai.net/image/vibetransfer.html

i missed you guys

Ah right I forgor about this place, haven't been here since the pony days

For Civitai Lora previews are you "allowed" to use inpainting, adetailer, etc or just raw gen? Weird that you can't use their generator before publishing the lora if you baked locally.

>>29584 its over

>>29630 civitai generator sucks so much dude just use something like tensor art or run your own colab

>>29632 I'm just trying to post a lora for upvotes, but it needs sample images

>>29633 Can't you add images without gen data?

>>29625 trying it and it seems pointless

>>29626 >>29627 >supposedly major hyped feature is out >no announcement why are they like it i'll test it out later with some v3 mixes and other stuff

>>29636 like this* thanks autocorrect

How come people are still using nai here?

>>29638 Why not?

>>29638 I use it in tandem with local.

Haven't really kept up with anything since Noob's final release. Have there been any actual finetune improvements since, or is it just mixes? I vaguely recall some anon posting comparisons between noob and a further finetune he was doing but I'm not sure.

>>29641 some minor ones but thats it

>"modules will never return" >turns vibe into modules for image gen turk what the fuck are you doing

>>29634 I meant to ask if they should be reproducible

>>29644 reproducibility doesn't matter you don't even need to have gen data in there

>finally feel like prompting gura >she announces her graduation w-whoops...

>>29636 >i'll test it out later with some v3 mixes and other stuff Well, I like that it extracts more of the style this time around (but it's still not enough) and that setting the reference strength to 1 doesn't turn it into a glorified img2img. With enough wrangling, the new emphasis syntax and (multi-)vibe you could probably get in the same zip code as your old "complex" V3 mixes but calling this a bandaid solution would be delusional at best. I really hope V5 will be what they said V4 would be.

>>29648 Also, it sucks massive ass that you have to pay to encode images for VT even if it's just 2 whole anals but at least it might encourage people to share the encoded images.

>>29649 that's the main reason why I dont wanna bother trying to tard wrangle it to attempt the v3 style

>>29650 I get that but at the same time it's just two anals. *IF* gathering a bunch of single artist/mix pics, tossing them into VT and sharing the encoded files proves to be a reliable way to actually replicate styles/mixes we could have our own "have LoRAs at home" meme.

>>29651 Sure its two anals, per vibe image But, how many two anals are you gonna burn through trying to find the right vibe images+emphasis mix to get the look?

>>29647 Just gen it

>>29641 All the idiocy surrounding it aside, I do think there's something in Illustrious 1.1. I'm playing around adding it to various previous mixes of older Noob or whatever. Dunno if it's a placebo or some shit, but it does feel like it adds more control over the prompt. Like making natural language more accessible or something.

Are you excited for Illu 2.0?

>>29656 if its better

>>28887 >https://mega.nz/folder/ctl1FYoK#BYlpywutnH4psbEcxsbLLg >Is the anon who owns this mega still here? I've seen some gens made with guweiz lora of yours but I don't think you ever posted the lora itself. Would be very nice if you shared it, thank you lmao is the guy who posted this still around? added it to the folder

>>29656 no, I think only latest vpred versions could really compete with noob checkpoints, but it's unreal to reach at this rate of their goal completion

>>29659 nta but how does illustrious 1.0 stack up to noobvpred 1.0 anyways? wasn't here for when it dropped but it doesn't seem like there's been much excitement over it

>>29660 it was pretty much just disappointment when it got dropped as on site gen exclusive on civ and others, because of their way to try make any money out of it, while the model is not really great at all, it's like, if there was no noob at all, yeah, it could be an upgrade over 0.1, but every other their release will be just inferior to a little fucked up noob vpred just because of the technology behind it and I personally think it's not worth it

>>29660 1.0 and 1.1 are only barely improvements

>>29656 Local has been on SDXL longer than 1.5 now. Just surpassed it last week. Pathetic how he's charging money for this outdated shit.

>>29656 People are actually still donating?

>>29664 maybe >>29664 >>29664 i'm suspecting that angels company is artificially raising the amount a little every day to make it seem like there's still support for the project. The individual funding goals for the models only combine to 1.2m stardust.

>>29656 >53 million memepoints for that """experimental""" shitbake lol

>>29652 Like three fiddy (you don't need to pay to change the emphasis, you need to pay to change the information extracted value and that's best left at the default for copying the style)

>>29656 reminder: illustrious is a kohaku fine-tune

>>29668 Remind me what a Kohaku is must've been a nothingburger if I remember artiwaifu, tofu, 4th tail etc but not this.

Didn't think i'd be back in this place (much less because 4cuck kill), whats up niggers, how has the last year and half been?

>>29670 Howdy, things have been pretty gud.

whats the new local meta? whats the new training meta? I want to get back into working on some unfinished projects Was working part time with an AI company but my gig finished and im back to neet maxing until something comes up

>>29672 illu finetunes/shitmixes same as before just with some new args and random shit to try. personally i've been liking lanzcos resizing for bucket since the sharpen effect helps the ai learn curves instead of blur, and network rank stabilization is just overall better for anything you're doing (tested both of these empirically to my satisfaction)

I took a deep dive on the illustriousXL models since finally some shitbag got enough good boy points to finetune Danbooru ontop of XL "correctly" (waifudiffusionXL lmao), but I saw that the base they used is a fucking shitmix on top of another as well. Wanted to see if I could probably tackle a clean finetune of XL with some tricks I learned from my gig.

>>29669 https://huggingface.co/KBlueLeaf/kohaku-xl-beta5 one of kohaku's MANY alpha failbakes

>>29675 Yea, reading all of this shit pissed me the fuck off, its amazing the models are even an improvement to the peaks of 1.5

ooooh goooood, we have our very own "aibro" now huh. maybe the cucktainment site dying isn't a good thing after all.

>>29677 I used to post my stuff here before, I know how to behave.

>>29678 shut up momoura

who?

>>29618 >absolutely no realism Where do I put those? Does Civitai take them, long as they're gachaslut cosplays and not real celebs? I only know of Deviantart, but that place is awful both on a technical level and culture-wise. pic unrelated

>>29679 >おらなんご >龍分枚 what did she mean by this

Btw are there any news with Pony version whatever? I thought they were about to release it this spring?

>>29684 astralite and his cronies have released some preview pics but it's shaping up to be a bigger disaster than cascade. completely fucked hands, sepia rerun, the only thing it seems to do coherenly is ponies (lol). the fact v6 was "usable" was nothing short of a miracle.

>Wake up >Not forced to read same old shitposting memes or dumb repetitive political ideology preaching yet again >life improves a bit Wow amazing how that works >>29682 Civitai will take them as long as theyre not IRL celebs or figures sucking dog penis or some shit, if you do wanna post shit like that, then im sure theres an /s/ copy here or another 4chan clone that might have a thread active I dunno.

>>29685 what exactly are these idiots doing wrong?

>>29687 >trainwreck arch they're only using to avoid paying for stability license >mixed dataset of anime/furry/rule34/pony/irl >captions are solely generated by an outdated ghetto rigged 7B VLM >project lead is a fucking moron, didn't understand the point of network dropout >"training cluster" is like 8 A100s in his garage >style clusters

>>29688 >mixed dataset oh no >autogenned captions with 7B model oh god >network dropout wait he didn't bother to set it? >style clusters huh? Overall yea this sounds fucked. And these are the Pony fuckers right? What about the deeper bullshit on the Illustrious fuckers?

>>29688 >network dropout NTA, but afaik it was not a thing with a finetune scripts? Only loras. Not trusting in that v7 shitshow anyway tho

>>29683 these gobbledegook signs are the only time i find myself grateful i dont know moon

>>29689 same thing he did before with hashes but with regular tags instead of random generated bullshit; the intention is to decouple words contained inside artist names from the tag for their style, which is kinda sorta a good idea on paper but practically it means you need to have a fucking spreadsheet open to do prompting instead of just typing the artist name for example, there's an artist who goes by "m k" who's impossible to prompt on any model without a lora because it's too hard for the model to lock on to just two letters like that

>>29692 i mean "v" works perfectly fine on most models. same with the expression tags like ":o"

>>29693 yeah but those have thousands of images as opposed to a few hundred anyway style clusters are fucking retarded

>>29692 Yea I can see how that shit would be absolutely hard to deal with on a full model unless you decide to just retag the name all together with one of his other handles as a fallback.

>>29692 ideally artist names would be inserted as unique tokens in the tokenizer like wdV (rip) planned on doing which would have sidestepped the issue entirely

>>29692 It's just to avoid mentioning artists by same same as in v6, legal and/or moral reasons. These prompt issues should be solved by the new TE, and if they're not then having hundreds of tags like style_cluster_187 would break things much worse.

>>29697 *by name

>>29694 yeah but it's certainly not impossible for a model to lock onto two letters as you said. the tag name just makes it more difficult for a model. pedantic, i know.

>>29697 sorry i had on my best intentions hat and you are almost certainly completely correct

>>29692 >impossible to prompt on any model without a lora There's this, but I guess it's cheating.

>>29689 >huh? I'm pretty sure he somehow clustered "similar" artists and all images from the same clusters get the same style cluster tag. From the 'cord samples, all the clusters look like (ai-generated:2) though.

>>29702 now thats fucking stupid

>>29701 thats vaguely accurate maybe it was just model issue idk man

sharkussy

So I guess this is the loli thread and the one on treechan is the regular

>>29706 this was the /hdg/ refuge from 2 years ago due to tranny jannies and some /g/ schizo, has nothing to do with with the 4cuck outage

>>29707 I will take that as a yes

>>29706 treechan has lolis too.

>>29706 you can post whatever you want on both threads as long as it is within the rules of the respective sites

>>29664 >>29710 nah, better to contain pedos here

looks like the shitposters remembered about this place

I guess webps are finally back

>>29713 >webps i will never understand why people use this accursed format

>>29714 It saves storage space while still being lossless and works the same as png everywhere I expect it to so where's the problem

>>29624 Sure. All the lora's are on civit. https://files.catbox.moe/7bxfcw.jpeg https://files.catbox.moe/84qoeo.jpeg https://files.catbox.moe/u4fwfc.jpeg https://files.catbox.moe/dwqyib.jpeg

>>29715 >so where's the problem photoshop doesn't want to open them

>>29716 Man, we have to box here too? Treech doesn't prune metadata.

>>29718 yeah but all the schizos went to treechan

>>29717 weird, even shitty gimp opens webps

>>29718 For me, jpegs are just slightly easier to work with.

https://8chan.moe/slop/ kek

>>29722 there are too many threads now

New addition to my slop wildcard

>>29725 what, "toddler?"

>>29686 ><lora:wetbody concept:0.8> That's a Pony lora, is your model >ntrmix35NoobAI1.1mix2 a mix between noob and pony? https://civitai.com/models/600231/wet-body-concept-pony

>>29727 Nah, some Pony LORAs work on Illustrious fine its not all of them though I think it depends how they're trained, that model is just an earlier version of NTR mix I liked blocked merged over a later version of Noob (which I did because I liked the aesthetics but wanted to obtain the later Noob's knowledge) Theres no Pony whatsoever in that mix.

Why is it that when I use the catbox userscript I can see the metadata, but when I use an extension to view the metadata, nothing show up?

>>29728 It doesn't seem to work on my mix but thanks

>>29729 If by extension you mean "infinite image browsing", it doesn't support stealth png metadata The latest reforge has native support for that in PNG Info tab

>>29726 Now you guys are here maybe it's chance to do this again How low do you consider "acceptable"?

>>29732 not a fan of 5, 6 but not gonna sperg out if I see them posted *nta

>>29732 you know where the line is, do i need to post the chart again

>>29732 I'm fine with 5 tbqh and maybe even with 6. I blame kereno for it, his lolis (toddlers) are really cute, small and hot. But I don't usually gen them, I'm more into girls from 1-4

>>29732 feral lemma proportions don't weird me out too much because the legs are still long enough to mentally file them as gremlins rather than true toddlercon. it's when you dip into realistic proportions that it starts to get iffy.

>>29732 I'm more disturbed by how wrinkly and shiny the skin looks in 6 than the size. Other than that, all of them can get it.

>>29739 the skin looks absolutely freaky in all 6 images

>>29740 Yeah it looks glossy and shiny like the character got oiled up but I just dismissed it as AI things. It's when it gets all wrinkly in 6 where I get truly disturbed.

congrats, you all failed the normgroid test spectacularly! move back to 4cucks when it's back online, yeah? yeah.

>>29725 incredibly based

>>29742 geeeg

>>29745 Yeah this is some good shit, nice.

people training loras on wai is the new anylora-tier retardation

>>29747 jeets gonna jeet

body is no longer in use

>>29749 Hot

i tried boomerprompting for the first time and i genuinely feel dirty

why is this the only thread that doesnt auto refresh for me

>>29753 Are you using .moe? Switch to .se

>>29754 same in either case, just this thread which is weird

why are noob vpred models unable to differentiate "male on top" and "female on bottom" tags? it's almost impossible to gen small male on top of large female

>>29756 Didn't realize there's a difference between those two either. Also, do you mean some other models are better at it?

>>29745 Now I want big-tittied hag x fat, old batard for completion.

>>29722 Don't forget /slop/ on >>>/trash/

oopsies

Slopping

>>29742 On the contrary I hope anyone disturbed by characters on my chart >>29732 stay here Hugbox is the worst thing a loli enjoyer may come into contact Hugbox makes people bold, bold people get caught

I'm just a simple female form enjoyer.

>>29765 >Hugbox is the worst thing a loli enjoyer may come into contact >Hugbox makes people bold, bold people get caught Yeah ATF is the weirdest honeypot I lurk, the people who got caught there clearly got too apathetic.

V4.5...

>>29769 >they actually fucking fixed traditional media and its related (medium) tags better fucking late than never, this is unironically one of my biggest gripes with v4

>>29771 now if only they'd fix artist mixing

>>29772 funny you mention that, i was watching the channel at the right time

>>29773 hmm interesting, we'll see if they actually do it

>8chan full of naiggers and pedos >treechan dead >4chan DEAD grim

>>29775 post a gen

>>29775 time to move hdg over to 7chan

>>29775 You can make an alt-/hdg/ on >>>/slop/. They have /edg/ with highlights and request anchor there!

>>29775 Hopefully 4cuck will be back up soon enough and you can go back to shitting up hdg again. Just spend those two weeks on sharty.

>>29775 This is not strictly a loli thread, anon. Post hags if you want

>>29769 Interesting. >>29771 I'd be more interested to see if they fix however they fucked the VAE. Prompt for something like "glitter" and compare V4 to local. It's so grainy compared to local.

>>29780 pretty much this, you can post whatever you want just don't be a dick and start bitching about pedo this and pedo that or you'll be the first guy to be banned in quite a while

told you that while extremely cathartic the death of the containment site is a bad thing in the long run

>>29783 >[extremely obvious thing] >told youuuu woaw

>single low effort baitpost >triggers hour long pedo meltdown How is this place different from 4chan /hdg/ exactly?

>heh, i am... le threadshitter :^) wow i've never seen this bit before

>>29786 4cucks brought 4cuck with them.

Clowns gonna clown no matter which circus huh

How does one develop a superiority complex from using a dead altchan?

>>29789 Hopefully our janny here is better at cleaning up their messes.

Anyone want a piece?

stealing is a fun exercise

/hdg/bros (REAL /hdg/, not the 8chan spinoff) is this our new home?

>>29778 >They have /edg/ with highlights and request anchor there AHAHAHAHAHAHA

>>29793 tail eyes and mouth are fucked on the ai one but the legs look better

>>>51 kill thyselves

If any of guys miss 4chanX >https://greasyfork.org/en/scripts/533067-fullchan-x

>>29800 Forgor to mention is still a WIP by some anon on /b/ >https://8chan.se/b/res/474864.html#474864

Any news since 4tard was killed?

where are my pillars...

Dead thread..... But everything else is pedo and furry shit. I hate it here.

Illu 2.0 is out if anyone cares I guess https://tensor.art/models/840365567832646761

>>29805 Finally, you gave us your approval, cathag anon. Welcome, even though you've probably been lurking or even posting here for a while. Now let's see how long it takes for brittle and oekakifag to post here

holy shit, i might as well start posting toddlers here to get rid of fucking tumor that is cathagfaggot

Dead thread..... But everything else is pedo and furry kino. I love it here.

>>29810 He's got a point though, somehow all the hagfags disappeared.

>>29811 hagfags were a psyop, altchans are not raided by feds yet

>>29812 Maybe just out for Easter, or chasing Civit buzz.

thoughts on r3mix vs 102d?

>>29815 i prefer 102d, r3mix has more ai shine. chromayume is what you want for better artist replication but it fucks up anatomy and details a bit more often (not current chromayume thats ass, the one from a few weeks ago is what you want)

>>29815 102d is too stiff and overfit. i like r3 more, and if your artist mix begins to feel too sloppy unironically use https://civitai.com/models/1452858/slopslayer-lora-shiny-horrible-average-ai-style-to-put-in-negatives at negative weights

>>29816 chromayume v1 feels almost indistinguishable from vpred 1.0

>>29818 okay but if you plot them against each other it looks somewhat better

>>29809 this but unironically scorched earth might be the only way to get rid of these tourists

Let's just get along...

>>29823 you can't get along with shitposters, anon how haven't you learned this yet after a year and a half of /hdg/ being an unusable cesspool?

>>29824 Just ban them when they actually shitpost. We're not at the mercy of a retarded /h/ janny here.

>>29825 pretty sure this board's mod only deletes spam and illegal/3dpd shit and said he doesn't involve himself in anything else

>>29826 I would have assumed rule 1 covers a bit more.

>>29824 I can't tell which one was the shitposter in that argument.

>>29820 >>29822 based

>>29828 do your reps

>>29828 if you try to invade established peaceful board and try to enforce your beliefs, i count that as shitposting

>>29822 based >>29820 fake

>slowly finding out the styles I perfected for /hdg/ appeal are not liked anywhere else it really was an echo chamber in the end

>>29820 >>29822 very based, love me sum yuri

>>29833 >posting to appeal to a crowd of anons anonie...

>>29835 yeah...

>>29832 idk the doujin authors disagree

like seriously if you put madoka and homu together and add the spanking tag madoka's the top over 95% of the time

>>29833 >he tried to appeal to highlightschizo no one to blame but yourself

>>29831 Outside of a situation you made up in your head to feel the pleasure of being a victim, when has that happened in irl?

Madoka being the top should be pretty obvious.

I'm in the mood to test some new shitmixes, any recommendations? I have tried a few that has been mentioned in the threads already.

Not about imagegen but it's hilarious to see zoomio-normogroids mald harder and harder about AI in *every* field. I hope the creatures who were trying to poison the imagegen scene don't tell these ones that it's all pointless, the psychosis is too fucking funny.

>>29842 not really a shitmix but, have you tried illust 2.0? >>29807

>>29844 Not yet, but I guess I should.

>>29844 nta if i can't dl it it doesn't exist

>>29843 > I hope the creatures who were trying to poison the imagegen scene don't tell these ones that it's all pointless, the psychosis is too fucking funny. They're still coping and believing it works themselves.

>>29846 here it is https://hf-mirror.com/OnomaAIResearch/Illustrious-XL-v2.0/blob/main/Illustrious-XL-v2.0.safetensors

>>29848 Thanks

>>29847 I can't wait until they start overlaying trannyfesto-tier shit over the image at 50-75% opacity

>>29848 strange it's not linked on civitai

>>29850 https://danbooru.donmai.us/posts/8895522

>>29852 wow

>>29843 I'm not going to bother watching this shit but if they're talking about doing what I think they're talking about doing(which is "poisoning" it with super high inaudible frequency ranges) this shit purely does not work for music. You can just cut off certain frequency ranges from training and wow it's cured. Worst case whoever is poisoning it uses a mosquito tone or whatever which is technically audible(but won't be heard by "most" people) but even that can be clipped out and they'll just piss off zoomer and gen alpha consoooomers. This sort of thing does technically work against speech training where certain shortcuts are taken to dramatically drop compute requirement but for that shit the people doing the training are going to be using bespoke datasets so what does it even matter.

>>29851 It was posted literally few posted above >>29807 idk tho. It does something but feels way too wild. It feels like it's listening to a natural language better but is so unstable. I'll wait for some better shitmix

>>29855 that was tensor art and it didnt have a dl link

>>29852 I meant anti-ai stuff don't train etc etc but lmao that's fine too >>29854 >You can just cut off certain frequency ranges from training and wow it's cured. Not sure how they're doing it, I didn't watch the video because I judge clickbait exclusively by its cover but that's exactly what I was thinking, if that's how they do it then even the laziest negative IQ EQing job will completely nullify it. I already have to do that shit to fix terrible hi-hat/tambouring hissing all the time.

>>29857 tambourine* I can't type

>>29857 i only dabble in music prod but theres so much shitty mastering out there it makes me fuckin angy

>>29859 Can you hear this hell? It gives me a headache real fuckin' fast. https://files.catbox.moe/al3ewh.flac

>>29861 instant headache in the first 3 seconds thanks internet

>>29862 The real torture starts at 0:13

>>29843 ofc, it's the same retard who was claiming that there is no female producers because of anime boobs VST

>>29864 To be honest most "female" producers I've heard about in the last decade turned out to be trannies.

>>29861 tbh I listen to a lot of shit similar-ish to this all the time so it's not really all that bad. The retard was trying way, way too hard at the start, especially with the fader and they probably needed to dampen the bass when it comes in since the track focuses so hard on the highs but I've heard a lot worse. Not that it's good, just that I've heard a lot worse.

any good way to combat overfried loras? (same exact composition, bad prompt following) https://civitai.com/models/974110?modelVersionId=1615193

>>29867 remake the loras yourself

>>29868 uuh, no

>>29869 then suffer

After years of obsessing over a single post in one of /b/s degen threads, I think I finally figured out what model and artist Loras the genner used. Just need to tune the tags, and I can finally lewd my sleep paralysis demon. I'm so happy. I know it's slop, but it's my slop. The lewder tags are quite strong, the dress rarely even stays, I probably need to up the weight. When I moved some tags, the arcade tag turned into pixel art. Also, if someone knows good tags for ghosts, especially the nightmare fuel eyes on the second pic (or really anything to improve my slop), let me know.

>>29871 make a dataset for it and ill bake u a lora

>>29864 >no female producers because of anime boobs VST wat

>>29871 For eyes you could probably very easily train one of those "addif" Loras by training on a before/after pair of images.

>>29866 You can't hear it then, be grateful. It's something in the high 12-13Ks and it sounds like absolute hell.

>>29875 i do av for my church and i keep having to stop these people whove gotten their eardrums blasted out their whole life from jacking it up so high it clips

>>29872 That would be very generous of you. I gen on a potato so I have no idea what you would need for training. How many pics would you need and how detailed do the tags need to be?

>>29873 there was quite some outrage about nani by wxaudio and he made video where he was saying shit like "this is why there is no females in the music industry" or something like this

>>29878 100+ is optimal (don't think this should be too hard tbh), if you want to go the extra mile do closeup crops, same number as the full res images. for concepts, the hard part of tags is testing which ones already fit the best so you can take them over. i think you probably already have a good grasp on that so again this shouldnt be too hard. (i figure it's something like horror (theme), ghost, sharp teeth, black sclera, red pupils, something like that)

>>29879 I see, thanks. That's a cute plugin.

>>29879 >the phrase "giant anime boobs" is actually in the one sentence pitch you know, maybe this timeline isn't all that bad

>>29867 Reduce the lora weight for the base gen, down to almost nothing if you have to. Then bring it back to 1.0 for the upscale step. Since by then almost everything is locked in and you just change the style. Add a CN, anytest or xinsir union if you need to raise denoise higher.

>>29880 Sorry if you have to hold my hand for this, but it's my first dataset. >100+ is optimal (don't think this should be too hard tbh) I'll try, but it might take some time. Should I exclusively collect pics with the eyes I want, or should I throw in some pictures missing one or two components I'm aiming for (black sclera, that black haze, black tears)? I should exclude character and artist tags, right? Does the picture's size matter? Specific resolution?

so uh, this is the place now huh

>>29879 While I can't imagine being the kind of cuck who gets offended on women's behalf about shit like this, I can't help but have my doubts about the quality of this VST. I've never cared about what the VSTs I use look like, only that things are laid out in a logical and easy to use manner. Sometimes the fanciest, most eyecandy UIs are the hardest to use.

>>29884 the most important thing is to make sure everything in the dataset accurately represents the look you want, i would advise not to include things that are only partially what you want since it can poison your dataset. it will also help overfitting if you include different camera angles. if you want to tag them as well just make sure to check the images manually to make sure that all the tags you want are where they belong. i'd suggest just adding em after autotagging; if you want to go the extra mile, add image composition / camera angle tags. ideally you want minimum res to be the native training res of 1024x1024, i usually just trawl gelbooru absurdres images.

>>29886 imo it's actually very good, control is low to the ground but it sounds good without need of tinkering with it, it has r2r version so you can try it yourself

>>29887 >i'd suggest just adding em after autotagging I just took all the tags from Gelbooru. Bad? I trust the booru tags more than programs/AI at identifying what's there, unless it's some ambiguous tag.

>>29889 nta but remove stuff from the "Meta" section like cropped, highres, spoilers, etc.

>>29889 yes, it's bad, they are terribly inconsistent, and inconsistency is the enemy with ai. autotagging is more accurate if you want you can tag your way and i'll compare them, it's no trouble

>>29885 my pillar...

>>29890 >remove stuff from the "Meta" section Would have removed anything except the general tags and highres, I'll toss that as well. >>29891 >autotagging is more accurate At least then I don't need a booru for the images. I already went through most of the relevant tags on the boorus to look for such pics, and I remember I found maybe a dozen that kinda had the look I was going for. Maybe I'll try Pixiv or any of the reverse image searchers.

>>29891 >yes, it's bad, they are terribly inconsistent, and inconsistency is the enemy with ai. this. when it comes to tagging there's no "it's better than nothing", wrong/inconsistent tags are more harmful than not having tags.

>>29894 And that is why autotaggers are bad. They give you more wrong/inconsistent tags. Human tagged images may be missing tags due to laziness, but they usually won't be blatantly wrong.

>>29658 Yeah I am, lol. Thanks, the nun is hot Spoiler cuz kinda aco

>>29896 the rules don't say anything about aco or /d/ stuff, or yuri, etc. just photorealism afaik

>>29897 Spoilering aco is considerate, I appreciate it.

>>29897 >Remember, furry and western art styles are not allowed here. 'Western' refers to cartoony art styles, realistic art styles or art styles used or often seen in the western part of the world and not where the artist or franchise is originally from.

I was lied to apparently >>29618 Still, where are you quoting these from? The rules only say: >3 : No furry or western cartoons

>>29900 >>1

>>29900 >>1

>>29895 >>29895 >They give you more wrong/inconsistent tags If autotagger refers to the algorithm behind the Danbooru tag button in Forge, then agreed. I'm always getting a lot of false positives. >they usually won't be blatantly wrong Depends on the tag. When I made my wildcard list, I found some tags that are ambiguous without the tag wikis, and so people mislabel pictures. For example, the "breasts_apart" wiki explicitly says it's not for small chests in the wiki, but there's a lot where they are incorrectly tagged that way.