Titjob EditionPrevious Thread: >>8115919>SETUPWebUI-reForge: https://github.com/Panchovix/stable-diffusion-webui-reForgeComfyUI: https://github.com/comfyanonymous/ComfyUIAutismMix SDXL: https://civitai.com/models/288584Loras: https://rentry.org/ponyxl_loras_n_stuff | https://rentry.org/ponyxl_lora_previewsTags: https://danbooru.donmai.us/wiki_pages/tag_groupsArtist hashes: https://lite.framacalc.org/4ttgzvd0rx-a6jfNovelAI (Paid): https://novelai.net/>RESOURCESTraining: https://civitai.com/models/257749 | https://rentry.org/59xed3 | https://civitai.com/articles/3105/essential-to-advanced-guide-to-training-a-lora | https://github.com/derrian-distro/LoRA_Easy_Training_Scripts | https://github.com/Nerogar/OneTrainer | https://github.com/picobyte/stable-diffusion-webui-wd14-tagger | https://github.com/toshiaki1729/stable-diffusion-webui-dataset-tag-editorUpscalers: https://openmodeldb.info/IOPaint (LamaCleaner): https://www.iopaint.com/installControlnet: https://civitai.com/models/136070/controlnetxl-cnxlBooru: https://aibooru.online4chanX Catbox/NAI prompt userscript: https://rentry.org/hdgcbOP Template/Logo: https://rentry.org/hdgtemplate2/raw | https://files.catbox.moe/xk99gy.png | https://files.catbox.moe/35bdt5.gif
That training section needs a newline
>>8119146rebake faggot, OP looks like dogshit
Did this place became /b/ or why all the .jpeg images?
>wrestling match where you can rape your opponenti'd watch it
>>8119164to save muh environment + retards can't afford dirt cheap drives
>redo MMD shaders>redo training lora>redo KK model>replace dataset>use partial masks to improve silhouette>no more style burn in, no more silhouette mistakes, no more shadowing mistakes, now recognizes the tassels attached to the collar>eyes are completely shit now because it thinks the highlight are her pupils because i included too many images trying to reproduce her facial expressionspiss
>>8119165
>>8119170cool ellen. missed opportunity with her shark teeth tho, make her grin
>>8119146>unironically linking only to a fucking shitmix and not the actual base modelfucking retard.
>>8119179>didn't bother reading more than two linesbased
>>8119157God I fucking love Jackoff the Ripper. Giga bush or catbox please?
>>8119179Do people even use the base model anymore?
>>8119180?there's only a link to autism, a shitmix.there is no link to the actual base model.>>8119182if you're not a retard you should be. Autism offers nothing beyond making lora's worse.
>>8119183>Training: https://civitai.com/models/257749
>>8119146Why didn't you put anytest and that other new controlnet model instead? Also thought you were going to put ponydiffusion under training.
>>8119164Because he's a faggot hiding his prompt.
>>8119170will you release this lora when it's done
>>8119185>>8119179You should train on autismmix anyway, what's the use for base pony in op?
>>8119185Oh never mind I see pony now.
>>8119185>Why didn't you put anytest and that other new controlnet model insteadbecause you suggested it right after it was bakedlol
>>8119184>completely unlabeled random link no one is going to know what it's even going to be forwhy is it even stuffed under training and not a regular model?Training shit should be, exclusively, about training shit.>>8119189even more retarded than the OP.
>>8119179>shitmix>autismWhy are ponyfags like this? How much does the ponyfucker pay you to shill? Oh, you do it for free...
>>8119194>>completely unlabeled random link no one is going to know what it's even going to be for>exactly the same as previous opanon again the links were up literally all daywhy not say anything before it was baked?
>>8119146>NovelAI (Paid): https://novelai.net/change link to https://rentry.org/hdg-nai-v3
Is there an artist that do schizo colors I can use at low weight to improve colors?
I don't think I've ever clicked on any of the OP links outside of the ones for loras.
>>8119176>make her grinno>>8119187yes. i was about to upload this version before i noticed that the eyes were fucked.
>>8119195...do you even know what a shitmix is?Autism has no extra training. It's not a finetune. It's something that some retard tossed a few lora's and the base model at various strengths into supermerge and hit a button for.>>8119196It was under the label MODELS"Training" means fuck all.>why not say anything before it was baked?The last thread is an actual garbagefire for the last few hundred posts that read like a /b/ raid. No one with an ounce of sanity would willingly actually try and read that shit.
Anyway I don't have the code to the rentry anyway, if you want to change it then yer gonna have to do it yerself ;)
>>8119202>I-I didn't read it because muh /b/lmao
>>8119195based retard
>>8119170oh wait i'm just retarded>went to add monochromatic backgrounds to my training images>forgot to turn off the setting to replace 0/0/0 with 128/128/128 that i was getting to get partial background loss on the loss masks>this replaced her pupils with grey in 50% of the training images while fucking up lots of other details
So I finally got some better results with a bake, that's exciting. It's not anything really major though.
>Fall for the base pony meme again>Have to use 100+ tokens just to get the style right>Model loses attention before i even finish describing my 1girlback to autism we go
>>8119170Can you remind me of how you did this again? I'm still trying to understand. This is a weak style lora + your character lora trained with only mmd and KK screencaps? And this was after you trained your own style lora with this character and merged it into the model? Can you post a catbox or is this inpainted?
>>8119230bro got baited by anon above... lolme too
>>8119230why would you need to do thatare you not using style loras
>>8119212>What values do you guys use for XL Loras?>>Network Rank16>>Network Alpha16>>OptimizerProdigy but I think I'll be switching back to AdamW8bit>>Learning Rate SchedulerCosine with retards>>Learning Rate1 for prodigy, 0.0003 base LR for Adam>Curious to try other settings but I've seen very contrasting info
>>8119234It's a bi-weekly occurrence. maybe i'm just easy to gaslight
>>8119238>Cosine with retardskek
>>8119236Haven't found a lora that looks good on base pony without score and/or artist tags yet
>>8119242>Haven't found a good lorafixed that for youif your lora needs score tags or a shitmix to work then it's garbage
>>8119238 meadam bake on topprodigy on bottomit's the character not style, just testing for compatibilitythe hair is more on model and the eyes, and general coherence are way better, even if adam has a bit greyer backgrounds
>>8119230Use turbomix instead
>if your lora needs score tags or a shitmix to work then it's garbageBut that's all of them
>>8119246Not much you can do when your base model is as shit as pony.
>>8119246it's 99% of them, yesespecially on places like civitai
>>8119246mine work fine without them :)
>>8119249>nogen
>>8119239I do this every now and then...
>>8119250>nogen reply to nogen
https://danbooru.donmai.us/posts/7578133what tag could be used to describe white part under skirt? "underskirt liner" didn't worked out
>>8119256>burden of proof-kun...
>>8119260>None of the posts in the chain posted a gen
>>8119232i'll post an explanation for how to do it later today probably>>8119250why would i lie>>8119259that's a petticoat
>>8119259Maybe the layered_skirt tag?
>>8119201catbox?
>>8119262Is there a way to run the first few step of a gen without lora?
>>8119262hah, this petty coat has nothing on my blazer
>>8119262>"petticoat lucky_star">10 hits, half of which are non-school uniform despite me having ~30 uniform ones in dataseti hate boorus holy shit, thank you for help>>8119265this one goes for gothic lolita dresses i feel like but thank you anyways
>>8119267this is using an unreleased lorahttps://files.catbox.moe/2g2zpc.png>>8119268not that i know of
>>8119268With comfy, yeah. Also with prompt editing on A11/forge, but you'll have to figure out the actual syntax. It was something like [:<lora:abc>:0.3] for "switch from nothing to lora abc after 30% of steps."
>>8119244How does adam8bit compare to adamw or have you never used it?
Requesting Mizuki Nakahara buttjob.
>>8119283i am using adamw8bitapparently regular one is supposed to be better but it's 2x as much vram and the results aren't much better
>>8119246For a more serious response, my answer is that your lora is supposed to work without score tags because there are no score tags in your dataset. When you prompt score tags, your lora should change since the score tags affect composition and colors. In my experience some loras can be unaffected by score tags. Finally if you add score tags to your dataset and prompt score tags, it should remove the effect of the score tags. Namely it should look the same with or without score tags.
>>8118866It's Japan so she'd probably end up with a boring office job, oh well
>>8119283I prefer AdamW vs AdamW8bit since it gave some weird results since it depends on bitsandbytes library, while AdamW isn't.Also maybe kohya scripts is bugged but I had same VRAM usage.
>>8119286We need more alcoholic hags.
>>8119289ols are based though
>>8119288Some styles work fine with the full score suite, I think it mostly depends on the style itself. If you're having trouble give it an activation tag, that solves the score conflict without retraining what those tags mean.
>>8119289Very nice. Boosted my morale today unironically. What is your prompt for the skirt (how did you avoid bleeding pencil?)
less hags more petite girls pleasehag overload is making my penis to rot away
>>8119289wife...
>>8119293>you're having trouble give it an activation tagHmm yeah this is worth a try.
Well /hdg/, how many saved (not autosaved) gens are you sitting on?
The adam setup is probably gonna need a few more epochs since for Rin I took a futher step epoch than prodigy as best The it/s are faster tho so it'll even out
will setting LR to 0 and unchecking "TE Learning Rate" make easy scripts not train TE?
>>8119300105 total since Nov 2021Most of them recent
>>8119294Just police uniform, pencil skirt. I think the bleed is too weak to do anything in a dark alley double handjob situation. Maybe if her hands were free or there was a table nearby.
>>8119306Excalibur...
>>8119302okay there's an option for that in network tab, sorry
Reforge guy, is you here?
>>8119314yes, any bug to report?
>>8119317No but would it be possible to add a XYZ option for loras?You know, like you can click the button and get all loras from a folder into it, like with Checkpoints or Samplers or whatever? Preferably with a setting to check which folder to use for easier lora epoch testing.Additional Networks used to have that, but SD didn't.
>>8119325>bought himself a spare cock in the grocery storebased
>>8119306NAI looks like THAT?
I've got a question if anyone has an answer: If I wanted to set up a twitter account to post shit, would I need to censor it to make it accessible to jp bros? I read somewhere that they have access to uncensored stuff, so long as it isn't actually produced/sold in Japan.
>>8119337who cares about them
>In my experience, the best epochs have loss rates ranging between 0.10 - 0.13, with values below 0.10 tending to be overtrained and above 0.13 as undertrained, however your results may vary.Does this mean I'm training for too long? Doesn't seem quite right.
>>8119340me. its why I asked.
>>8119343it doesn't mean anything at all
>>8119343it's bullshit
>>8119343But average loss varies from dataset to dataset. I saw another anon claiming a particular drop in loss (a particular negative slope on the curve) is where he had the best results but you still need to pick it after a certain epoch so it's irrelevant. Just make a grid and use your eyeballs.
>>8119319Oh I see, if you want to test some loras in x/y/z now you can do prompt S/R, so if you have to compare<lora:a:1>, <lora:b:2>, etcThen on prompt S/R you enter that long string, and on the original prompt, you use the first one (in this case, <lora:a:1> and it compares all the loras. I use it like this for now and it works fine.Now, gonna check if it's possible to do your request. Additional networks add the camp since like you select a lora from a list, i guess something similar exists from the lora menu that comes on the webui.
>>8119301Style also stabilized well before final epoch so maybe not.
>>8119354Yes that's what I do now, but it's definitely not convenient.IG the question is whether you can activate loras without putting them into prompts. Or maybe if a live injection "hack" can be made if not.
>>8119359Putting loras into prompts is a hack in the first place
>>8119363certainly confusing for newfags.
>>8119337Some jap messaged me on twitter and they didnt seem to give a fuck
>>8119361sir please share the vibe you transfer
>>8119363That makes it easier then.
>>8119353unless you're using loss masks you also can't really be sure what the loss is reporting. iirc something like 70% of any given 1girl, standing image is going to be miscellaneous background information that you don't care about.and the cool thing about loss masks is that the loss barely goes down as the AI learns the character anyways. it's just an almost entirely useless metric outside of knowing if training has completely died
>>8119300from various threads, 1944
>>8119300around 30k 12k nai4k ponybut I have sorta hoarder mentality when saving gens, I should throw out most of them, especially 1.5slop
>paizuri+ejaculation>cum is just a trail between the dick and the tongue 90% of the timehow hard is it to just make a flying rope uh
>>8119337Only if you live in Japan. If you post to Pixiv you have to censor to avoid the ban hammer.
>>8119372projectile cum
>>8119300the only gens that have escaped deletion are ones that i photoshoped before uploading since those go in a folder that has other things that i don't want to delete and i'm too lazy to sift through
>>8119374I obviously prompted that and it either does nothing or add some random projectile in the wrong palce
>>8119376not really obvious, plenty of anons don't know their booru tags. well unlucky.
>>8119375hot
>>8119377Yeah you are right indeed, might be my loras fucking with it tho but I would think something like that would work well most of the time
>>8119366I'm still toying with it, in no way an expert yet.https://files.catbox.moe/a1rzan.pnghttps://files.catbox.moe/jvdg9t.jpgThose two as low str positive vibehttps://files.catbox.moe/vnyvml.pngThis one that I made myself as a negative vibeDoes more harm than good with certain other artist combos, can be pretty fickle.
>>8119380>This one that I made myself as a negative vibewhy? is it same idea as negative embed or something?
>>8119380>negative vibewhat?????
>>8119380baysed thank you
>>8119383you can manually set negative strength to the "vibe" seems like furry naifags use it a lot
>use negpip and negative prompt to remove aaaaaaaaaaaaaaaaaaall the pony/furry/cartoon influencewhat a difference, might not even be necessarily good but holy fuck
>>8119390>nogen
>>8119392https://files.catbox.moe/b1oh0c.pnghttps://files.catbox.moe/lybsb2.png
>>8119395both looks like shit
>>8119395waifudiffusionbros, we are BACK
>>8119395i forgot to actually upscale this shit, anyway, the difference in general luminance is easy to see
>>8119390I noticed the same thing but it massively fucks up hands and anatomy as well? What weights/tags are you using? I experimented with (source_pony,source_furry:->1.0), (shiny skiny:-1.7) which worked well and (score4,5,6:->1)
>>8119401just read the metadata
>>8119401I thought there was no score 4 and 5
>>8119381>>8119383Yeah just pop it in and put the value to negative, most bad art scribbles work pretty well for it, toyed with 3d women and ugly bastards in negative to push certain artstyles more towards anime, works sometimes.
>>81194045 is the lowest
>even cascade's strongest shill b-sama is saying that the current partial epoch looks like shitit's jordachover
deja vu
>retraining a bunch of shit because of a 10% improvementAutismchads, we are so back?! At least I deleted 2/3 of my loras.
>>8119411let him cook. don't underestimate the power of dual 3050s
>>8119386>>8119408wait i don't get it... where do you choose negative vibe??
>>8119403based negative vibe poster
>>8119414it's a shame that he had to switch to the 3050s and now the bake will finish in 36 months but at least this lets him monitor the TE LR more closely!
>>8119386>>8119408ohhhhh writing -1 in the input box holy fuck I didn't know that was valid
>>8118256Similarly I find it amusing which copyright tags appear in the Jaccard index for anal versus vaginal.
>>8119421>anal>boomer>vaginal>zoomerTOTAL SENSE
>>8119337I assume that you're worried that Japanese people will be blocked from viewing your tweets at a domain level because of their censorship laws but I just tested looking at twitter from an Osaka VPN and I could still see uncensored porn. I think that law only applies for websites originally in Japan.
>>8119262>all these nice artists>shit still looks like ponyGrim. Even though their styles are absolutely there I would never be satisfied with a single one of these loras. Thank god I no longer use pony or I would go mad.
>>8119424based naichad
>>8119419We're getting even more partial epochs now and it's pretty amazing: 0.265.e8 has massively improved over 0.248.e8.
>>8119424>nogen
>>8119424based64 looked like based64 with loras toonai looks like nai with artist tags toomodels are ultimately recognizable after you train your eye for itit's just you hate the pony look
>>8119424NAI has the same problem. Their training data just wasn't curated enough and none of the artist tags look faithful, that's why style mixing is so popular. NAI has amazing stylistic range, recreating one particular style just isn't possible yet.
>>8119422It's probably gays too, FF and MHA at least
>>8119429>nai looks like nai with artist tags nai is the only one of those that i sometimes can't recognize at a glance if i don't already know that it's ai generated>>8119430it's true that nai's artist tags aren't accurate for the most part, but the point that anon is making is that even accurate artist loras for pony still look like pony
>>8119429Sure, but that's the point. Pony is 10% anime and 90% cartoon/mlp/furry, so if you _are_ satisfied with pony look while browsing /h/ of all places there's something wrong with you.
Sirs, a second nogen has hit the /hdg/
>>8119262Can you run this grid on base pony and not on autismmix?
>I'm team neggles because i lurk in the cascade discord>Would probably be team jordach if i lurked in the wd discordBakers are their own worst enemy
>>8119437lurk both so you can be team astralite instead
>>8119429That's just not true NAI gens have no particular look to them and even pony can't be identified as such if you go for extreme style choices.
Who would have thought 308 pics of some random pictures from danbooru I liked would be enough to sorta convert it to vpred.. I'd say it was surprisingly easy. Style-wise it's a total annihilation of pony, still far from NAI though. I'm aware some anons shit on Arti for a reason, but Neta 2.0 + vpred/ztsnr finetune unlocks it a little, dull slop filter no more, although I suspect to make the images actually darker you'd need much more images, now it looks too bright still if you prompt for darkness. Although it's probably me fucking up TE in this run. I can also see where I overcooked it with those THICC and smooth lines looking just like pony and some poses are fried as well, but the potential is there. Shame HLL-anon gave up on SDXL, I'd only wish you guys would to try and appreciate it for what it is before putting your (ces:1.3)es on display.Lowres gens, same seed, Neta+lora on the left, Neta on the right, random prompts:https://files.catbox.moe/qtg4gj.pnghttps://files.catbox.moe/p270ar.pnghttps://files.catbox.moe/1ckjew.pnghttps://files.catbox.moe/ue2gv0.pnghttps://files.catbox.moe/j7t7i7.pnghttps://files.catbox.moe/t9lqqv.pnghttps://files.catbox.moe/mq42v5.pnghttps://files.catbox.moe/9zpgo6.png
>>8119432>nai is the only one of those that i sometimes can't recognize at a glance if i don't already know that it's ai generatedskill issue
>>8119424>a line up of images that all use the exact same seed/prompt/negative prompt with a simple background look similarwhat did you expect>>8119424that is base pony
>>8119442Those gens do look nice, but until something with equal hands to autism comes along, no one will move.
>>8119439most nai gens are recognizableI mean if you go for very stylized shit like gyate yeah you can'tbut when it comes to rendering complex stuff you can usually see the "matte" style that it has.I do agree that nai tends to have less digital/plastic look compared to other models though
>>8119442Bro what's the point, you are trying to add good colors/style to a model that can't generate humans. Go finetune pony or something, even though that would probably be pointless too.
>>8119435>>8119446actually nevermind you're right, forge didn't change the sampler
>>8119375Can you do some Eunie?
>>8119424style loras just don't work well with pony for whatever reason.The few artist tags that aren't hashed work pretty well, it's only loras that have a very small effect on the "pony" look
>>8119451i have loras for most of the female cast
>>8119453lora issue
>>8119447It's not just hands with arti. Severely mangled hands are just the most visible indicator that model is not very good with anatomy. You need to actually test with something harder than 1girl standing to see that it's completely fucked like this with the whole body and not just hands.I know I linked these images plenty of times already, but a I think "the pose" is a very good 1girl test for any model:https://archiveofsins.com/data/h/image/1714/73/1714733003650765.jpghttps://archiveofsins.com/data/h/image/1714/73/1714736046923073.jpghttps://archiveofsins.com/data/h/image/1714/73/1714735643013605.jpgI know for sure non of the arti/neta can output anything coherent with this particular pose/angle.Reweik/autism can do it, but it looks boring because composition is infinitely simpler.Only NAI can do these as good.
>>8119437Every time they start talking about how they're gender special mentally handicapped snowflakes I want to an hero out of second hand embarrassment.
>>8119452I kneel. This negative vibe thing just opened my eyes to new possibilities
>>8119450What do you mean sampler? I am right that it's autism? Will you rerun it?
>>8119458Controlnet allows for all the interesting angles and poses that you want. If we had working loras/style tags and maybe ipadapter we could gen everything on local with enough effort.
>>8119453People intentionally handicap them
>>8119411i feel bad for b-sama...
It will never not be funny to see ponygooners criticize objectively better models lmao.>"b-b-b-but it can't generate my wife with a penis fucking me in my nipple hole so it's dead on arri-"shut up, all your gens literally look the same no matter the content
waifudiffusion is so bad
>>8119462Controlnet isn't magic. It will mangle shadows and clothes in these kinds of difficult poses and angles. You are forcing a model to work with something that is just out of it's depth.
>>8119413What was the 10% improvement so I can retrain mine too>>8119465If another model is so good where is your gen?
>>8119411you know it's bad when the yes men start saying no
>>8119468>What was the 10% improvement so I can retrain mine tooIs just adamw, I baked on prodigy beforebut the dev version of easyscripts also introduced a bunch of garbage that handicapped my bakes so beware
>>8119469multiple of his regulars called it jank and he said he "doesn't see the problem" and got defensive bruuh
>>8119465the only model better than pony is Nai and its debatable
>>8119465>b-b-b-but it can't generate my wife with a penis fucking me in my nipple holeI'm just asking for a girl lying on stomach with legs up while camera is focused on her face:)
>>8119465I gen my wife happily smiling at my pov while I consentually missionary her :)
>>8119467It might as well be magic in ponys case. Apart from upside down stuff it hasn't failed me once yet
>>8119461i did mean the modelthis is actually with pony
Stop fighting, guys.There's 2 different philosophies in this thread that always end up clashing with eachother and derailing the thread.>1: model just has to have good anatomy knowledge and prompt adherence, it doesn't matter how the end result looks>2: model needs to produce aesthetically pleasing images, prompt knowledge and anatomy take a backseatBoth philosophies are valid, you don't have to keep fighting over it.
>>8119465>shut up, all your gens literally look the same no matter the contentyou can just use a lora
>>8119472>the dev version of easyscripts also introduced a bunch of garbageHuh, I haven't touched the dev version so hopefully safe. Good luck on your rebake ventures
>>8119481That's what you call working without score tags?
>>8119482Nobody is fighting here, it's one guyThe NAIkek already got deleted and got revealed as a ponyfag multiple times
>>8119482It's one schizo who replies to himself when everyone else ignores him, not joking
>>8119481Bro, these completely lost "the pony look". Which is what's been said here 100 times: pony look = autismmix.Look at the smooth shading on 0202ase xl if you don't believe me.Nice anatomy though. That's the giant fucking tradeoff that makes base pony impossible to use.
What do you think B-sama's hair smells like...
>>8119484Going well so farI think Prodigy has a bit of that weird composition bake inBut it's maybe recency bias kek
>>8119481BTW Ishikei is hashed
>>8119492as what?
>>8119494I don't remember but it's labeled in the op listIt was multiple ones iirc
>>8119482It's just a single nogen calling every post dogshit, bullshit or similar. He's a regular. Ignore him.
>>8119494it was something like 8um
This one still has the base pony look despite no use of score tags and using a lora+ artist tag combo. Something about anime girl bodies is just burned into the model since backgrounds and clothes don't have the same problem
What do I say to get him to join?
If Resonance is bad because the TE and UNET aren't aligned, then why did they train them unaligned? *Shrug*
>>8119502kek
>>8119449This "lora" can work with pony too, sorta.
>>8119502Tell him it's a horsecock image sharing discord
>>8119503fake trip
>>8119505Left is cascade 1B output.
>>8119492the hashed version kind of sucks though
>>8119502He's already in it unless Jordach kicked him, check the member list
>>8119509i agree
>>81194948um or 7x90um
>>8119508Right is pony. Wait...
>>8119509Are there any hashed tags that work well?
>>8119514gjem vum baka
>>8119510Don't see him, found doctor head tho
>>8119514i like kpg, thon and psaz
>>8119514my favorite is aqemjm is really good too
>>8119502tell him jordach wants to give him some tagging advice
>>8119458To be honest I just think Arti/Neta is severely undertrained, especially Unet, and that's the reason it fails with anatomy. I just saw how well Unet trains on Arti, and it's kinda funny how much it improved.
>>8119525That's the exact conclusion that the llm community came to months ago: everything is undertrained and nobody wants to go with a long sensible lr training because it's expensive
>>8119526Weren't there papers written on how basically everything has 15 times too little data and was trained that many times too little too?
>>8119528veeeery nice
>>8119194Regular pony should be strictly used for training, not genning, it's dogshit and so are your gens.Non-schizos gen with AutismMix.>>8119191NTA but anytest has been out for almost 2 months now.You mean to tell me people don't know about it by now? Are you really that retarded?I can understand not putting in the new promax shit, since it wasn't talked about all that much and appeared in previous thread.But there's 0 excuse for not putting anytest in the OP.
>>8119534do it yourself then faggot
The Op should have a link to the csr-jk-jd style and nothing else.
>>8119534op herei agree with >>8119536>knew about it for months>didn't put it in the op because lazy faggottsk tsk
Anyone find dakimakura prompts seem to nosedive the quality of the gen? Always seems like I can get decent results right up until "full body, dakimakura" goes in, and then it looks like shit.
>>8119536>>8119540I'm not a baker and have no intention of stealing the retards job and then get shit upon for making something better.>>knew about it for months>>didn't put it in the op because lazy faggot>tsk tskI haven't even been here for almost 2 months anyways lol.
>>8119525funni https://files.catbox.moe/v5syrn.png>>8119526>and nobody wants to go with a long sensible lr training because it's expensiveNobody even wants to use more efficient architectures, what are you talking about.
>>8119542then shut the fuck up niggereasy pz
>>8119541not particularly
>>8119542>Im not thing, know nothing about thing but I will totally complain and insult person doing thing because I'm retardedbased
>>8119544>Nobody even wants to use more efficient architectures, what are you talking about.What better architectures do we have? Cascade is debatable and all the DiT ones have huge vram requirements that lock out large parts of the community
Hello, Astra, you're a good sport for lurking here despite all the negativity
>>8119550>the eternal hdgbro
>retrain shark girl lora with fixed dataset>eyes look fine now but the LR change i made completely mangled the handsi am going to start making a general training rentry just so i can write down why i switched LRs instead of forgetting and trying old ones again
Hello, Astra, you're a dumbass for lurking here, pruning artists, censoring concepts and still doing it despite everyone telling you not to.
>>8119551>all the DiT ones have huge vram requirements that lock out large parts of the communityThat's not really an arch issue, it's just the model creators pick large sizes. No one has applied MoE to imagegen yet, for example. Mamba vision who?
>>8119550Wtf are you even smoking mr schizo? Know nothing about what? Making threads?What made you project such a retarded idea?
>>8119558>Making a thread is not that hard, schizo!>still hasn't done a single thread
>>8119559>>still hasn't done a single threadAnd how would you know?>schizo calling others for schizokek
>>8119553the only baker that lurks here is jordach though
to me it's the ten months of not fixing the op and not wanting to contribute when someone links the rentry and when somebody does it you have 30 experts on ops
lol
>>8119562euge and kurumuz are behind you bud
>>8119561>H-how would you know!? I'm ackhually a Master thread baker! lmao
>>8119567Did you forget your meds today schizo?
>>8119562It's just you (neggles)Me (jordach)And astralite here
watch next thread be a spitebake that uses the old op format
>>8119566i doubt that euge's english is good enough to lurk here desu
>>8119570remember the war last time? kekhistory repeats itself
>>8119560cathag anon please address the allegations
>>8119575what are these allegations?Also, I believe people refer to the blue-haired catgirl anon as cathag, not me. so you might have the wrong man.
>>8119565based
>>8119422it’s west vs. east.
>>8119595FRG v DDR
>>8119479Based and whitepilled.
>>8119462 >>8119480>Controlnet allows for all the interesting angles and poses that you want>It might as well be magic in ponys caseIt's not. And it doesn't allow shit. First try to gen this kisaki. I already tried and it fucks up her neck area and top of her swimsuit in every way possible even at 1 strength. And 1 strength is just not flexible at all. How much will it fuck up if I use mannequin from some posing software instead of an already 2d gen? Assuming that I don't want to just copy other people's images 1 to 1.Then try to replace a girl. Sure I can change hair color to red, I can change swimsuit color to yellow, is that the extent? What if I want to generate yuzu who has voluminous long hair instead of double bun? (spoiler: it mangles hair completely and outputs a mess because prompt doesn't match what controlnet tries to force, and if you reduce controlnet strength anatomy goes to shit because model is retarded without controlnet) Or give her a frilled dress instead of a swimsuit? Should I just ask the nai guy to generate me new base image? In that case I feel like it's the nai guy doing all the work and not the controlnet.Show me I have a skill issue, please.
>>8119163>rebake faggot, OP looks like dogshitthis but we are stuck with it till the next thread
>>8119625The third aos link you posted? I'll try it tomorrow
>>8119555Do you remember which LR you changed? An anon claimed that settings was the reason why my lora hands would deviate too much from autism but I don't see how that's possible.
>>8119635TE at 0.00003 has pretty OK hands, TE at 0.000175 has very bad hands. but i'm not really sure how applicable that is to base pony since i'm training on a lora merge.
>>8119632I'm doing the first one. I think it's a more interesting angle.
>>8119638I just had low TE ruin my backgrounds so I'm baking pure base LR now
>>8119641yeah higher TE seems to look more greebly and interesting. the issue is that it will greeble fingers as well
who are best big tiddy subs
>>8119651you wouldn't be able to gen them with that inferior model
>>8119651the ones that aren't subs but compelled to fill the role
>>8119651Mei
>>8119620based
So do anons really prefer 1:1 dim alpha ratio? I've been doing anon's 128:1 ratio and it's been working well for me. Is this all placebo?
>>8119625that's a lot of words to say that you have a skill issue
>>8119667i am the anon who came up with the 128:1 ratio and i am also the anon who has been telling people to switch to 1:1. i was having a very hard time getting things like knee high boots to recreate a character's knee high boots at 128:1, but i switched to 1:1 and got it to work within a few attempts after doing a bit if napkin math to convert the LR.
>>8119670Very interesting. Do you recommend it for style loras too?
>>8119671i use 1:1 for both.IIRC it doesn't seem to really matter for style information
>>8119672Alright well my current bake is already 1 hour in so I guess I'll do it again overnight. Hopefully this means it will take less steps to learn everything now.
if I train a lora on a boy, tagging 1boy and stuff, and then on my prompt uses 1girl and long hair and stuff, will it manage to genderbend him? not trap, actually a girl
>>8119675yes
I hate how the pov tag force the pov to be lying down
>>8119675yes but the model doesn't really understand the concept of genderbending so how convincing the result would be as a girl version of the character is mostly up to luck
Everyone gets animal_ears now.
>>8119678kneeling works perfectly fine
>>8119681Are those the same catboys and now they're girls?
why does the xl lora im trying to use just not seem to work? Even when I copy directly what someone else has used it really just seems like its completely skipping over the loraI run the same prompt with and without it, and it makes no difference.I feel like I'm being a retard somehow but I'm just not sure
>>8119684No but maybe they are cousins. The lore is not fully established.
>>8119687>The lore is not fully established.I can't masturbate to this, unbelievable
>>8119685it could be a newer type of lora then your webui can run
>>8119683I dont see how thats related
>>8119685is there __ in your lora name?
>>8119690because pov doesnt force lying down?
>>8119691Yeah it's <lora:Keitaro_XL_V10:0.9>If that's the problem I assume the solution is renaming it without the _s, what causes the issue?>>8119689I'm using forge, fully up to date
>>8119693forge or reforge? forge doesn't have dora support
added new panels, it's kinda repost, but i dont care q_q
>>8119694I didn't even know there was a different version, I'll try installing and using that and see if it works.
bunny
>>8119696effort posting is always appreciated, thank you for sharing
>>8119694forge has dora support but not innately
>>8119701ty, it took some hours yeah q_q
>>8119702yes, you can pull a pr for it but i assumed it was more likely that this anon hadn't done that
>>8119706yeah obviously not, I was just being a pedantic cunt
>>8119698Would
>>8119696I like it. Hope to see more in the future.
16 channel vae please
So after reading the copier method rentry, isn't this how the night lora was trained? Probs 16-32 pictures then edited those same pics (preferably until Tagger recognizes them as "dark") to make them darker and trained for a short while. Done.
>>8119723you get a 16 channel vae but it can only do grey and yellow
>>8119723only 42 more months until resonance 3.6b!
>>8119727straight upgrade over pony
>>8119725Just ask the guy already
>>8119730He's a japanese retard who specifically doesn't want people to know what he did. Why are you this dumb?
>>8119731why are you assuming that he most have done something special anyways? it's just a lora
>>8119731Being uninformed is not being dumb you meany
>>8119733It's a lora that fits nearly perfectly over all others and fixes one of the fundamental flaws of the model. Much like with the copier method it would be absolutely mindblowing to see someone training on game CG but getting none of the style. Why wouldn't you want to have that technique? Look at all the different applications of the copier method. Come on anon.>>8119734He's not uninformed, he's ignorant. The difference is that an ignorant person is proud that he lacks knowledge and uses this flaw as grounds to start fights. It's the newfag that doesn't know anything about the conversation but feels the need to butt in rather than just keep scrolling.
>>8119739you sound retarded
Is there any way to change two variables at once on the xyz plot? I have an idea for something but that involves changing both the prompt SR and the Controlnet weights at the same time and I don't want to make x^2 images
where is the night lora? should be possible to pick it apart layer by layer to get more insight
>>8119698catbox?
>>8119744https://civitai.com/models/432511/low-light-for-pony
>>8119723there's actually one made for sdxl (new)BUUUUT it requires a full model retrain :)
><pony_outline_lbw_5x>What lora is this?
>>8119751Proof? Where?
>>8119742Why not put the other variable on the other axis and increase the batch size?
>>8119758https://x.com/ostrisai/status/1808213169749242160
>>8119763but that says that it isn't for sdxl and they have to train something to make it compatible, not the model
>>8119414[BREAKING NEWS]Cascade's mystery donor donates another $2k to the project. Jordach to soon begin training on newly purchased 2x GTX 1060s
>>8119773from what i understand there's 2 ways to use it1. train an adapter w/ lora that replaces the in & out block in the model. this way you can use it with existing models (he's in the midst of training some but idk if they'll be good enough to replace the standard SDXL vae via the adapter)2. fully train the vae into SDXL, which would require an entirely new model, thus more work, but the end result would theoretically be better.
>>8119778we're so jordachback
>>8119778cant wait for an epoch to be an actual improvement!!
>>81197782xAda6000 achieved lesgooo
>>8119782an epoch? no need, because he's training on a 1060 now he's decided to print 0.01 partial epochs every day0.69e8 is already a massive improvement over 0.68e8!
>>8119778>isreal jordachbros...
that finetune was only shilled on 4cuck and discord, and just by doing that they caught a big donor, imagine when they shill it on reddit lmaobut that would require the demo to be able to do good pics
has ControlNet++ been implementing into A1111 or reforge yet?
>>8119789i mean there's no guarantee that they'd get catch any whales on reddit even if they had a good prototype (which they don't have anyways)
>>8119791reforge guy here, you meant controlnet union? I added more support to it (I wonder because the file i tested is named controlnet++_union_sdxl.safetensors) on the dev_upstream branch, but I think maybe it works on main branch?A1111 has the extension updated I guess
bigger batch size or gradient?
>>8119796same thingone is more vram other is more time
>>8119760Because that will still create x^2 images, what I want is the prompt editing and controlnet weights to shift at the same time
>>8119442wait this is really cool can you please share your last working finetune config? i'm very interested in flipping models over to ztsnr vpred if it's workable!
>>8119792>7 toesChrist.
>>8119793they could already have a good enough model if they were baking with the 3.6b one, just look at what tree says about the current bake https://files.catbox.moe/mx58jd.jpgtalking about this genhttps://files.catbox.moe/x1sium.pngand i do believe it, the text encoder seems good enough but 1b effnet is so painfully undercooked its not even funny, at this point they will end up doing something that sai and warp team were supposed to do: to fully cook 1b. and i think not even the 11th epoch will be the point where 1b is fully baked
What is even the point of donating when he doesnt want to use anything more than a couple of consumer cards and treeman claims to have unlimited secret enterprise access? is the money even going towards training anything at this point?
>>8119805as the warp team already said the 1b model is shit and no amount of epochs will change this. i agree that the 3.6b model could have potential though but at this rate we won't know until months from now>>8119807>treeman claims to have unlimited secret enterprise accessapparently they don't really have access to the good cards yet>is the money even going towards training anything at this point?they're saving the money to buy hardware, so no
that Neta v2 model has it's advantages and disadvantages it's easily susceptible to artists and styles but my oc character loras I made for pony don't work like they did on pony and I have to include tags
>>8119808stupid little baby 1b
>>8119810pony loras only work on pony
>>8119808>no amount of epochs will change this.how are you so sure?
>>8119750mostly empty.OUT01 (Style) and OUT05 (Tint? Luminosity?)with transformer blocks 1/9https://files.catbox.moe/455o6o.jpgtest:https://www.mediafire.com/file/l57iog6cpfwppnn/lowlight_1721964122_1721964450_dim32_sv_fro_dp0.92_1721964790.safetensors/file
horsy reminder that pony took more than 20 epochs to deliver
>>8119818to explain, transformer block (_block0 -> _block9) is the effect of that transformer block dropped.OUT1,3,4,5 is with that block active.workflow: drop unused blocks, rescale dim32, save as fp8
>>8119821please explain like im retarded (i am)
>>8119816i'm exaggerating, of course. but i'm still doing my own tests on every epoch of the 1b model and i don't think it has significantly improved over the last few epochs, so it's hard for me to believe that it has that much room for improvement. i don't care about being wrong about this though, if it miraculously ends up being good at epoch 25 or something then i will happily abandon pony
>>8119819And it fried the tags so badly that it almost can't understand natural language anymore.
>>8119827but it's good at sex position !!!
>>8119807It goes towards them buying their own hardware, but they're supposed to have a bunch of Nvidia funding they need to blow threw.Tree has said to Jordach over and over they need to spend the grant money within y amount of time to apply for a higher amount and Jordach yaps on about not wanting to change anything or try anything. Autism ruining a golden opportunity. Tree needs to put his foot down and get Jordach to pull his head out of his ass if they want to get anywhere. I'd have a lot more faith in them if Jordach was locked in a room somewhere being filtered through Tree if he has to say anything, letting him speak unmuzzled is horrible PR. Tree going on about being a 5th soul native gender special with six heart attacks, pronouns for his trigger warnings he lists to medical staff and tranny suicide attempts doesn't inspire confidence either.
>>8119831>letting him speak unmuzzled is horrible PRthey're both pretty autistic but i agree tbhdesu
>>8119826>i don't think it has significantly improved over the last few epochs,Jordach turned the LR down way too early. He's shown he's aesthetically blind and his idea of good is when he cranks the cfg to 15 and outputs fried, eye burning garbage. In his mind the model is in refining stages winding down and doesn't need to learn more, only needs to clean up. 3.6B will not be able to succeed under these standards.
>>8119831i love when tree tries xer hardest to convince jordach to use their nvidia funding to switch to 8x a100 while jordach cries about not feeling comfortable unless he's using his 4090
>>8119832The difference to me is that Tree's genderspecial medical retardation will be dumb at worst to normies. Jordach on the other hand goes on tirades about how the real destiny of humanity is to fuck little boys in the butt. Tree seems open to feedback and ideas, Jordach digs his heels in, becomes defense and denies/ignores issues presented. Liability.
>>8119824it seems like the lora was trained on a fat bastard DIM, but the lora was mostly empty apart from specific blocks, and very sparse layers in those blockshow was it trained? still no idea, but it has revealed the function of OUT05 (blue tint/luminosity?) so thanks for sharing.if that's what that block does, then lyco (or at least block scaling) is beneficial to capture more dynamic range.in other words, to replicate: large DIM, lyco, more blocks (for example, full preset or pick blocks manually, including OUT05)optionally after training: cut blocks, resize down, and stuff to dump out empty layers and garbage.pic related
>>8119834This is one of the most frustrating things to see. I wish he'd lurk here more so it'd get through to his thick skull that he'd get a lot more support if he was being innovative with the tech and opportunities they've been offered instead of hiding in his autistic comfort zone. I used to root for Cascade here but him hiding in his echo chamber has only reinforced bad habits that will ruin things going forward.
>>8119839When even B-sama and the other shills are tip toeing around saying its bad, its bad.
forge dev, can we exclude the contents of models/Stable-diffusion from the repo or add to gitignore? Each time I symlink to my models folder it blocks updates.
>>8119839>I wish he'd lurk here morehe pulled out a SD3's "skill issue" card for this, even though he haven't posted any gen that he worked on (phoshop fuckery, upscaling etc.) to show how cascade is a tool that is better than NAI's "aesthetic slop generator" and pony
>>8119831dont talk about muzzle near a furry you might over excite them
>>8119838isnt that just putting a filter over the picture
>>8119795I found most of them yeah, weirdly when I select any control type and manually pick a preprocessor, I get different results to when I select the control type of the preprocessor I pick.
>>8119847seems like it yeah. you're not gonna get any kind of real luminosity breakthrough on SDXL regular or derivatives, only stuff like this.if you want real dynamic range you'll probably need to smash your model into cosXL or whatever, which never seemed to take off
>>8119849>smash your model into cosXLonly 60 more days until the wdV alpha!
>>8119844He keeps trying to weasel out of people wanting controlnet by saying it's a skill issue and they should use i2i. Doesn't even understand controlnet has its own use cases. Starting to pull a pony saying he wants to make things for THE ARTISTS instead. Amazing he isn't on the censorship bandwagon. >>8119846Whatever gets them to wear it.
What is the state of SD3? I read that the license yadada and that it's pozzed and censored can't even do girls or limbs correctly but I also heard that there's a more powerful model?
>>8119838I'm actually going to try the copier method later on and see if I can reproduce this. I figured he had trained a lora at 32dim and resized to 8 but we'll see.
>>8119855Everyone has rightfully dismissed and burried it. /g/ is hopiuming for pixart bigma now mostly
>>8119855they coped that the released model was a beta and rushed, hinting towards more models coming (a la 8b and 800m)
>>8119855they're supposedly releasing a less terrible 2b model in 2 weeks
The only thing that will get him to start actually doing shit besides wasting months on a garbage 1b prototype will be if neggles unironically delivers. and neggles will never actually deliver, so enjoy the rest of 2024 with nothing new. pony won
>>8119861I hate this timeline...
>>8119861euge will save local with a hunyuandit finetune, trust the plan
someone convince Astracuckie to join Jordach's server and shit on him so his spite baking can be refueled. If the bake continues this slowly I'm done with it.
>>8119866Yeah I wouldn't even care if outputs were this bad if it weren't that its baking so slowly we are not going to see a good epoch within this lifetime.
>>8119848I have no idea what I'm doing but using inpainting preprocessors with either model found here: https://huggingface.co/xinsir/controlnet-union-sdxl-1.0 doesn't seem to be working.
ugly bastard-san is very lucky that he's a nondescript male character that prevents the network from constantly generating selfcest
this just in: zeffy will be dropping cascade project because of sketchy donation amounts. zeffy saving that one guys wallet.
>>8119849Well I can just do that on my own with better accuracy through ps then
>>8119873if he can drop 10k on cascade that easily he's probably rich enough for that to be pocket money. and they can just find a different place to take donations from anyways
>>8119861HAH! poor little localkeks, having only just nothingness to witness while they all be left to rot into dust meanwhile oh, my dear nai is very... busy...finetuning...... an useless llm for the 70th time..... :(i -i mean they're just pretraining v4 right now! y-yeah!
>>8119872Type in his features. I always prompt my features with natural language like "boy has black hair and blue eyes."
>>8119871If on main, not sure if it's tested. Since it is like old forge on backend, maybe doesn't doesn't support it.if dev_upstream, man i'm not sure how to make inpainting preprocessors work as it should after comfy upstream changes. (or maybe I'm using a wrong model)
>>8119872selfcest is based though
>>8119882>describing male's appearancei'd rather "head out of frame" thanks
>>8119881I am more confident in them releasing a V4 than neggles and jordach delivering anything decent at this point
>>8119885>he doesn't gen himself fucking his waifuUnironically why are you even using AI if you don't want custom porn?
>>8119885use faceless man instead, works the same but it doesnt include the female
>>8119868remember when the main selling point of kekscade was that it was supposed to train lightning fast and shit out epochs daily?
>>8119888if you care that much about what the male looks like in your porn then i have news for you anon
https://files.catbox.moe/u7qksw.gifSpent an hour on this shit.
>>8119892Here we go
>>8119881nai would be fine if they did quarterly updates with new characters and addressed issues. but 25/month for them to sit around doing nothing is wild.
>>8119893I really hated Chifusa's power because her tits were the perfect size at the start of the show. Oh well.
>>8119888"male pov" is barebones enough for self-inserting
>>8119899I'm a pov chad but I also like to gen myself fucking her in front of the mirror.
>>8119898Liking regular large breasts in the transformation genre is like being a classical liberal in the Democrat party
>>8119902Sure but why couldn't she have started at A or B and have gotten to H cup by the end of the show is all my I'm saying. I mean part of her character arc was her returning the breasts she stole back to the girls because she was sadden by how devastated they were by their loss. You could have worked it into her character where she hated being a flatty and now felt bad for making other girls flatties.
For goat anon. I sleep now.
>>8119916did this a while ago
>>8119917Based and saved.
>>8119430>>8119430I don't think that's true. After playing around with NAI a bit you just need to put a couple of "{}" around the artist tag. For example, prompting for "mdf_an" looks really bad, but "{{{{mdf_an}}}}" looks exactly like mdf_an's style
>install reforge>loras not showing upamazing
>>8119932all your lora are gone bro, reforge was a bait all along
>>8119933FUARK.................
>>8119932Sorry man took all your loras.Did you get any error on the console? Seems to work fine here
>>8119936Doesn't appear to be any errors in thereI've also noted that the extension list refuses to load, and settings don't update even after pressing the button. I also cant reset the UI as well.I've put all of my installed extensions into a different place, to see if that was the case, yet the above problems still persist
>>8119951I have this happen sometimes. Refreshing the webui in your browser should fix it.
>>8119952>load it into edge>it worksfuck, i didn't think my browser would ACK like that, guess I just need to update it lol
>>8119803There's literally nothing interesting about the config exceptzero_terminal_snr = truev_parameterization = truemin_snr_gamma = 10ip_noise_gamma = 0.2ip_noise_gamma_random_strength = truenetwork_train_unet_only = truenetwork_args = ['preset=full', 'enable_conv=True', 'algo=full', 'dropout=0.1', 'train_norm=True']network_module = "lycoris.kohya"Ah yes, as far as I understand, kohya doesn't have vpred working in sdxl_train, at the minimum you should bring this https://github.com/kohya-ss/sd-scripts/blob/25f961bc779bc79aef440813e3e8e92244ac5739/train_network.py#L875 over to https://github.com/kohya-ss/sd-scripts/blob/25f961bc779bc79aef440813e3e8e92244ac5739/sdxl_train.py#L593, but it should work with sdxl_train_network out of the box.
>normalfag uses cascade's current epoch, makes shit gens>lolifag uses cascade's current epoch, makes a gen 10 times betterwhat is even happening why is it like this excuse me but how even, are lolifags some kind of aesthetic magicians
>>8119955>update browser>still malfunctions the sameguess it just doesn't want to work with waterfox, for whatever reason
>>8119560She's so awesome
>>8119959Lolifags tends to be more devoted and always tries to make gens to look better even when using limited resources, they usually find their own way of how to do it. It's kinda natural for them to look out for aestheticness. This rule applies only to any lolifag that is not a jeet namefag garbageposter that comes from /b/ though
>>8119959>>lolifag uses cascade's current epoch, makes a gen 10 times betterwhich gen?
>>8119959can you show which gens are you talking about
>>8119959surely you're not talking about the prillya gen, right?
>>8119968>>8119972https://litter.catbox.moe/m6gr2w.pngbest 1b gen ive ever seen atm and its still kinda bad and very simple but miles ahead of this one >>8119805
>>8119974lol
>>8119974jordach, maybe I treated you too harshly... it still looks like shit though.
>>8119974hmm i wonder why the upper body simple background gen looks cleaner than the cowboy shot outdoors oneand i wonder why the one who prompted real artists has better aesthetics that the one who only prompted for the synthetic dataset stylesmust be some aesthetic magician
>>8119974I remember the 1B gens all being shit as fuck specially the earliest epochs, this one is decent, kinda gives me NAI vibes too
>>8119981Are you implying some kind of... Skill issue?????
>>8119983if you mean naiv1 vibes then sure
>>8119985Hey I just said kinda, it's clearly not the "pony killer", details are just not there.
>finally sees a decent character lora that is NOT baked by nochekaiser>downloads it and check tags>"score_9":69, "source_anime":69many such cases
>>8119699I might die if we don't
>>8119696What's the clac sound in the third panel?
>>8120018button of her short popping
>>8119998>:69Wtf is this supposed to mean
>>8120022tag occurrence in dataset
>>8120020oh, rightthat makes more sense than I expected
Who is the best LOZ side piece. That is, who is the best girl in Zelda who is NOT a main character.
>>8120023Oh, that makes sense sorry half asleep. For some reason I thought you meant that was their recommended prompt
>>8120031gerudos point blank or whoever that white haired sheikah girl with the red eye tattoo is, not the one with glasses. unless that's some incarnation of Zelda so out of the running. I've only ever played the N64, GBA games and twilight princess
>>8120034>GBAGB*
>>8120031for me it's malon/Cremia.
>>8120034Paya
latent upscaling hatealso genderswaps are a pretty useful test for a lora's flexibilty, though i've mainly tested using male characters as females and not the other way around
>>8120034Gerudos huh, not any specific one just gerudos in general.It's probably for the best it didn't give her the OOT pointy nose.
>>8120031Din from OoS for sure
>>8120046same boring style, same boring pose, same offtopic post related to the characterstop being so pathetic
>>8120038box?
>>8119974Isn't that just about artists tag? I believe that's by pumpkinspicelatte (no metadata in the box).
>>8120052https://files.catbox.moe/gwo1xh.png
>>8120050>same offtopic post related to the characteri get the other two but this is stupid, do you want people here to talk about penis in vagina all day? training topics are even more off-topic on a hentai board
>>8120055How could training be off topic when it's directly related to the creation of the hentai. It's like complaining about drawfag /h/ generals.
>>8120050>nogen
>>8120058You can only discuss how the penis is inserted inside of vagina in hentai style, sorry. No tech talk, that goes to /g/. No character talk, that goes to /a/.
>>8120036A fine choice>>8120055Ignore nogen posts, their opinions don't matterAlso missionary is based
>>8120060>>8120061samefag-kun...
>>8120063nogen-kun...
>>8120063Yes?
uh oh melty...
>>8120065All me btw
https://files.catbox.moe/kqz6yw.pngno metadata on that so here's the lowres that doeshttps://files.catbox.moe/f9kss0.png
>>8119716>>8120003good luck q_q>>8119748>>8119698https://files.catbox.moe/m5ccpo.pnghere
>>8120075cute sack
Basido posting hours
Hag posting hours
>>8120061>Ignore nogen posts, their opinions don't matterIs that how you cope with being creatively bankrupt low iq mongrel?Why then just repost the same image 300 times and become the best contributor to this thread.
>>8119024>>8119024kittenbox if youre around?
https://files.catbox.moe/bt2q9p.pngara ara
nogen keeps having a melty every thread.
>>8120061She should have appeared in more games...
>>8119667Lower alpha has ironically been baking in more garbage into the loras in my tests.
>>8119796I liked the results of batch 1 gradient than batch 2 or batch 2 gradient
>>8120110Functionally, lowering alpha is the same as lowering LR. You can even calculate by how much exactly. So whatever you saw is just coincidence and/or confirmation bias.
>>8120114That's why it was ironicalBut it could make sense in the context of <too little training to understand concept looks similar to it frying it to appear all the time>
>>8120112so batch 1 gradient 1? you can change gradient size
>>8120100Based eunuch
>>8120119Batch 1 with gradient checkpointing
>>8120123what about Gradient Accumulation? do you keep it at 1?
>>8120124It's unchecked
VRAMlet question. Grad accum perfectly emulates batch size, right? Batch size 1 grad accum 4 is exactly the same as batch size 4 with no grad accum?
>>8120120poor 1boy.. I should have given him a small penis...
>>8119974this one is still far behind the same guy's own 3.6b gens:https://files.catbox.moe/ykpz9c.pngjordach's 3.6b prototype trained on a fraction of the full dataset with slopped together settings is still absolutely destroying his 'carefully crafted' LR adjusted 1b junk trained for double the epochs. and he wants to use this 1b to try and shill around on reddit... why he doesnt just spend ~$500 of the donations and finish this shit and get on with the good model is beyond me. the more funding pours in, the slower this project progresses
>>8120132Don't forget he's gimping the TE by bolting it on the junk 1b when it is supposed to be trained together with the 3.6b (he did the same for proto I believe). You will NEVER see cascade's full potential because of his retarded decisions.
What are your top picks for underrated style LORAs?
would anyone care for a dump of the same 15 loras i retrained like 5 times here alread?
>>8120137Just dump it and say that those are the same 15 loras you retrained 5 times already. If somebody cares, they'll download them...
>>8120138should be ready lateri should retry some of the failed old concepts too ahh
Was meant to be Takizawa, but she doesn't have a single "looking back" image so the lora won't work.>>8119292I guess you're right.
>>8120134kanzaki muyu https://kemono.su/fantia/user/164shidenchan https://kemono.su/fanbox/user/134033
i also now realized i have an updated rizdraws dataset
>>8120132>>8120133could he not just train the TE with the 3.6 and then freeze it for "Reso V2"? how can you even tell it's training properly on the 1b when it looks so bad regardless?
>>8120150wat? 1b is pretty solid already
>>8120150It's too slow or some other cope. In reality it would be better to see epoch 3 of TE+3.6b training by now than 500 epochs of this retarded gimped training run.
>>8120142That shidenchan guy draws really well.
>>8120155yeah, he's wasted on drawing children
>>8120156early today, aren't we
>>8120156ah the jap principle
Fat hag meltie!!
>>8120157Hey, it was just a friendly jab. That guy takes it a bit too far.
>>8120160I'd like to melt into the titties of a fat hag...
>childrenMore than 90% of shidenchan drawings depict perfectly fertile females.Much unlike hags.
>>8120160fat is not oldchildren can be fat too
>>8120163Yeah? Show the fat folds and cellulite then.
>>8120142shiden does nice stuff but their style is boring as fuck.you like the stuff from them due to the concept of their work and because their style is anatomically correct and isn't offending to the eyes.but that doesn't change that their style, by itself, is really boring.
>>8120100based
>>8120166If only the tits filled up the whole screen, that would be NOT boring!!!
>>8120166I've realized that goes for most artists desu>went from 35 artist loras to a single one i kept becuase i didn't care about any others
I like boring styles, modern moe is the pinnacle of anime
>>8120166yes, you are right posted those two because they are decent at drawing small breasts and anatomy that isn't "generic anime barbie doll"
>>8120170I keep a bunch of styles only for inpainting specific things, cum ropes, nipples, eyes, etc.
>>8120175There's also the generic tianliang toddler. Just by means of not being any of the two this artist is a head above most.
>>8120177I realized I liked most of them because of the characters they drew not the style
>>8120178>zhedongrent free
What happened with Rustle anyway? Used to be really popular when pony released, but I haven't seen a single gen in months now.
>>8120181>fengshui tingye lan-nimfe
>>8120181go post more toddlers on 8 chan and wonder why that place is deserted
>>8120169that would be the exact opposite, though. that'd be significantly more boring because you'd lose half the CONCEPT of his stuff and he's pretty much a purely conceptual artist.>>8120170Yeah. It's really important to keep in mind that, ideally, we're effectively skinning a lot of the same shit with different artstyles - so the goal should really be toward using one that isn't boring.Granted, you can also get bodytype differences between artists but if it's on the extreme side of things there's a good chance they're not really a 'boring' style. >>8120175Yeah. It's nothing against the artist. Like I said, they're conceptually very good - just stylistically boring. Pretty high tier thinbooks tho.
>>8120182rustlwho?
>>8120182>What happened with Rustle anyway? he was exposed for tracing>Used to be really popular when pony released, but I haven't seen a single gen in months now.his fans are probably using nai
/hdg/ rule As a thread discussion grows longer, the probability of a comparison involving tianliang duohe fangdongye approaches 1
>>8120189That would explain the accurate anatomy huh
>>8120186I baked non boring styles before but eh I really am too disconnected from social media to care about any of these people enough to want to gen with themI have two or three artists I'd consider liking
>>8120190That's not how his name is spelled, bro
>>8120193just get fang-dong right, that's the coolest bit the rest doesn't matter
mao tse-tung is my favorite artist
chubby brats erotic....
>fatdong (fatdonguser)
falungong zedong
>>8120186>they're conceptually very good - just stylistically boringI'm open to artist recommendations from your great eye. Similar body types, stylistically not boring.
cunnyuoh bratty tailand
I should point out that we have an ironic hagposter running around, otherwise 35-ish, Komi's mom, sugoi dekai mom, the teachers from dumbbell, boss takizawa, masane, poké-milfs, etc are the max.
>>8120211is the ironic hagposter in the room with us right now?
>>8120211bring back the 2006 studio style milfs and fight back against the newanime-youngmomist revisionism
>>8120214one gen at a time
eu hours are fucking awful
>>8120212nta but hagposting usually starts after somebody posts catboxxed loli
>>8120207shiden doesn't really have a super special specific bodytype profile to match against. Again, good artist but very conceptual. "Boring" in this case can be read as "generic." About the only thing unique going on is the breast weight profile which is going to be a bit of a bitch to fully emulate.just go look through random artists and try and find some that you actually like the look/style of separate from the content. it's not that difficult.
>>8120219thissomeone post black penis
>>8120219Isn't it too early for EU? 2PM everyone's still at work.
>>8120222obsessed closeted blacked cuckold
>>8120224i'm a spiritual individual lyrical
>>8120223eu has plenty of neets, remote workers and do-nothing office "workers"
>>8120220Based detective
>>8120223I'm still off from work until next week
Someone needs to screenshot all the crit for cascade bake here and post in on Jordachs trooncord so he wakes up and at least speeds up this shit bake. I'm not making a trooncord but you cowards are already on it.
>>8120221I can separate the style from content, AI can't. Go take a possumachine lora and see how every single girl gets incredibly short. It's unusable.It's not "a super special specific bodytype profile", it's realistic teen proportions.
weird how local models are free yet nobody wants to post images with them. maybe they're just shit?
>>8120231I should have also mentioned how the fingers become small and stubby. You can't escape that no matter how many negs you put.
well we had our fun talk time it's sadly time to shit the sheets again
>>8120233>>8120236You're too obvious
>>8120233This gen was made with MiqoMix, based on AutismMix Confetti, a local model.
>>8120237Anon... did you really not get what the second post meant?
>>8120236what's today melty topic?
>>8120114The ratio of the dim/alpha multiplies the size of the LORA's weights as they're being shown to the trainer and then converts them back down to their real size or something. IIRC the higher weight values require a higher LR to change, but it's not 1:1 with adjusting the LR.
>>8120239true, the only active local posters are ones that just do it to troll the general with ugly shit:)
>>8120243same as last seventy five days desulocal this local that nai this nai that omg it's so badit already started so i'm just gonna hibernate and focus on bakingget my 3ds out
>>8120236Better to spend the weekend outside anyways. See you fags next week.
Why is the ControlNet model list empty (even after clicking the refresh button next to it)? Am I supposed to install the models separately for whatever reason?
>>8120249more like last 200 daysthis has been going on uninterrupted since v3 release
>>8120252>Am I supposed to install the models separately for whatever reason?Did your model and lora list come pre-installed?
>>8120114apparently alpha is a gate for how much is learned, like a threshold. So a higher alpha would bring in more features, and more noise?I always thought alpha was just a scaling value, because that's how it looks in inference code
>>8120249you are not up to date anonlast week or two it's everyday meltdown about loli shit
>it's the "epochs are too similar to differentiate" timeaie
>>8120249move to treechan the schizo can't access it.
>>8120258That comes in waves. Once it gets too out of hand jannies will start hair-trigger banning for loli again like they did two-three weeks ago. Then they ease up, child posters come back and the cycle repeats.
>>8120257>I always thought alpha was just a scaling value, because that's how it looks in inference codeIt is
>>8120258This but it's fat hag meltdowns. Sisters just can't handle their undesirability...
>>8120263speaking about the devil
>>8120265due to how float rounding works, a scaled float will end up above or below some threshold that gets truncated off, effectively acting as a gateon one float, that's negligible. across however many thousand, it's almost like a transistor
>>8120256https://github.com/Mikubill/sd-webui-controlnet/wiki/Model-downloadThere's no depth model compatible with SDXL? That's the only one I really wanted...
>>8120276anon no electronics, it triggers my ptsd
>>8120231then don't use possumachine?or at least, be smarter about it and do a gen with a generic style first for whatever bodytype/etc. you want and then use different lora's while upscaling to 'overwrite' it. We did actually get a controlnet tile that actually works really well semi-recently and that lets you crank denoise values up to .8~ while letting you actually retain composition. Style transfer will be a bit limited but you can do two passes to really enforce style if you really need to.
treechan:>Israel and India rangebanned by default >loli only threads allowed >mods will listen to users >schizo spammer can't access it because he's rangebanned 0 downsides other than posting being less, but if we go there more often that won't be a problem either
>>8120245Ugly?
what LR schedulers are you using and why? cosine looks fine but i wonder if im missing something
>>8120283cosine with retardsit just works
>>8120277there is but I lost the linkjust google
>>8120277The link for the good one is in the op. SDXL depth is on illyavs GitHub just google it
>>8120280>or at least, be smarter about it and do a gen with a generic style first for whatever bodytype/etc. you want and then use different lora's while upscaling to 'overwrite' it. We did actually get a controlnet tile that actually works really well semi-recently and that lets you crank denoise values up to .8~ while letting you actually retain composition. Style transfer will be a bit limited but you can do two passes to really enforce style if you really need to.I cope like this all the time and it's never the same as just using a good artist that does what I like.
another rebake in progress
>>8120282don't listen to haters mistress you are beautiful i love you
>>8120289boobs seem a little small for the character
>>8120282i knell before the true queen of /hdg/
>>8120293it's so so over
this general is less of a forum and more of a chatroom and no it's not about the low amount of images posted>is my post low quality?>is my post off topic?>is my post completely fucking meaningless?if more people asked themselves these questions before pressing submit we'd have a better general with less shitposting plus bad posters would stick out like a sore thumbyou shitpost ironically, you normalize worthless irrelevant one liners, you cultivate this kind of enviroment and that inevitably attracts schizos which is only natural since they feel at home here
>>8120300ironic
>>8120300Too bad you didn't follow your own advice when you made that post
>>8120239>>8120282mornin reposter fweind
Are we back?
>>8120251Lel, I know it's cumrag ui but anyways, box?
>>8120304lol at the showerhead height being scaled down shes actually crouching just to fit under it oh, ai
>>8120307WE'RE BACK
Year 2047: AI now projects out of your mind into your VR glasses augmented environment >Resonance 1B is still baking on Jordach's 3090
>>8120280 >>8120288I should also add that you don't need anything advanced like controlnet tile for this because you don't need to do it at upscale stage. You can just i2i the pic at the same base resolution and that will allow you to use any denoise value you want, THEN you upscale.Even easier, you don't need to i2i, you can just specify [style X:style Y:0.4] and you'll get bodytype and composition of style X while the actual style will be style Y. The problems arise when you also want composition from style Y and only bodytype from style X.
>>8120249it's all the failbakers fault. we were supposed to have better by now but we don't. we were promised the world and didn't even get a patch of dirt.
>>8120317The funnier thing will be people complaining that it doesn't do it for them anymore like two weeks after that releases
>>8119916Neat!
>>8120337Cute 20-year-old /aco/ woman. Dick is ugly as fuck though. Too bad it's at the center of this composition.
>>8120281I tried posting and it 404ed the images.
>>8120345The dick is great. Plump and drooly.
AAAAAAAAAAAAAAAAAAA I CAME 2 TIMES
>>8120337fuck that guy, keep it up
>>8120374proof?
Post the last pic you cummed to
>>8120378>>8120375
>>8120354It's kinda weird though. Dicks don't look like that outside of deranged cartoons.https://files.catbox.moe/jwsdms.png
>>8120375AAAAAAAAAAAAAAHHHHHHHHHH AHHHHHHHHHHHHHHHHH MISTRESS AAAAAAAAHHHHHHHHH AHHHHHHHHHH IM COOMING AHHHH
fengdongh...fangdoh..fangdounde..fuck it
>>8120383fangduooohh cnyuzheng
For every slight against pony, astra takes away one artist tag
>>8120387I have a good one!What's worse? Toddler rape or the noises todcons make?
>>8120383antifreez3
>>8120380There's a demonic goat girl in the image.
>>8120403She's aesthetically pleasant while that dick isn't though. Why is the focus on it anyway? That's just gay, /h/ is supposed to be female focus.
so I think I understand that I can just use git to turn my A1111 install of the webui into ReForge.Should I first backup&move things like the loras,models and other thing I know I changed/added?
>>8120405sorry anon, /h/ is all about tasty dicksgo to /u/ if you are fag
>>8120409Ah, got it, thanks.
>>8120416no, he just realised that local will never be better at genning fate lolis than NAI, so he gave up
>>8120416i'm ready to bet my virginity that he doesn't give a slightest fuck
>>8120416>vramlets sending death threats due to dropping forge support>artistlets sending death threats due to making Paints-UNDO
>>8120418As a big fan of fate lolis, can't blame him.
>>8120424cascade would*If TE was trained with 3.6b. It will be retarded now though.
>>8120424>majority of good quality cunnyfag material it's not on gelbooru
tianliang duohe fandongye
>>8120432how somebody can be a troon, look everyday in a mirror and don't hate lolis
>>8120433>he knows
Bras have not been invented by this ancient tribe yet.
>>8120433ominous aura
>>8120444Cute style.
>>8120444finally, black representation
>>8120444Hello BobsBlazed, crazy good gen as always!>SSARR
>>8120345I tried to cover up the penis a little bit, sorry about that.
>>8120424it won't even be as good as arti
ting tang walla walla bing bang
>>8120449>>8120444>>8120375Outstanding gens. Thanks for contributing.
>>8120454u i u a a
sneedliang feedohe seedongye
fedongbased basedonghe tianbasedang
I'm very shy to post. What if anonies don't like me?
>>8120467I won't like you. With that settled, post away.
>>8120244Can confirm it's definitely not 1:1. Still trying to figure out the right LR after such a huge increase and I'm starting to think that although it makes sense in theory there won't be much practical benefit.
antisneed3
>>8120472i use 0.00003 TE or 0.0001 TE and 0.0003 or 0.0006 UNET, using different TE and UNET values depending on what i'm training
>>8120469OK here we go
>>8120475forgot pic
>>8120475>>8120476owari
>>8120474Yep I'm baking 0.0006 unet now because 0.0003 unet was close but not quite. I had been doing 0.003 unet at 0.25 alpha but I think after all this testing it's only gonna create the same lora rather than a better one.
where can i even find his art
>>8120484whoops, i just misspelled the name i couldnt find it
>>8120311How is it AI's fault? He typed shower head in prompt but limited the size of the image so it had to put it somewhere.
>>8120310I don't use comfy but I can already tell you it's the 3d figurine model not anything you're using.
>>8120277>>8120286>The link for the good one is in the opWhat did he mean by this? This thread literally began with people complaining that anytest and promax weren't in the OP.
>>8120480if you're doing a style lora then it should mostly be the same afaik
>this threadhttps://www.youtube.com/watch?v=ncRi8bZL1q4
>>8120497Machine learning everyone.... I guess I have to wait and see with the character afterward.
>>8120498lmao
>>8120300Checked. How do we go about improving this general?
latex fox
>>8120504post more cunny
what does a xerson have against fictional children?
she was only 30 you sick fuck! #fightfor45
ok so if i wanted to only train the IN08, OUT00 and OUT01 blocks with standard lora is this the right configdown_lr_weight=0,0,0,0,0,0,0,0,1,0,0,0up_lr_weight=1,1,0,0,0,0,0,0,0,0,0,0
>>8120508estrogen
Are we getting raided again?
>>8119818>>8119821What extension are you using?Also why transformer blocks? I thought cross-attention part was touched the most with loras
>>8120513define 'we'
>>8120510yes but afaik for some reason you'll still train other blocks unless you use lycoris. kohya just sort of ignores certain blocks with sdxl
>>8120517i see kek thanks
>>8120524should've just said you are american
damn that was hella fastcrazy how a single ironic nogen post can make jannies seethe massively but your daily off-topic dalle spam is completely fine
Poast lolis please and thanks :3
reforge guy here, seems Illya will update stock forge?https://github.com/lllyasviel/stable-diffusion-webui-forgenoticed because some commits difference, pretty nice. hope he updates controlnet, code is painful hard for me IoI>>8120407More than backup, probably I would suggest doing a clone in a new folder and then moving your a1111 files there
shitposter tried to be so funny that he will be raidedkek
>>8120538illya already said that he would update forge to test experimental stuff
>>8120538he said in support drop post that forge repo will be used for gradio experiments
>>8120539People seem to forget that this subtext only applies to /b/. Saying you are a pedo rapist on any other board gets you a visit from the FBI (they want in on the action).
>>8120536Based request.
>>8120544You should try to actually make the girl attractive: 1- make the breasts smaller2- remove the disgusting piercings3- remove the goth clothing4- remove the black lipstick5- cut the hair so we can see her eyes
>>8120546>>8120547only for experimental stuff? :( hope he takes a look at controlnet
>>8120539i just hope they don't shoot him
>>8120548good to know i'm a turdworlder, you wish there were something similar to fbi here lmfao
>>8120556there already is, it's called the cia
>>8120561Based request. (x2)
>>8120561Based request
>>8120551can't tell if you are trolling but I've done it before but, now shes just a generic cat girl and not unique.
>>8120567500% cuter
>>8120567She looks better now. If you want to make her unique give her unique hairstyle and clothing. Like drills and frilly medieval clothes. And some cute bows.
>>8120567nta, but holy shit, day and night
>>8120556The ICPO exists.
>>8120567Now lolify her please- I mean errmm just make her a flat and short adult ofc
>>8120567Also, I'll edit number 5 to be optional. Anime girls can be cute even if the hair covers their eyes. So if you want to keep that it can look good but the rest of the hairstyle still needs to change a bit.
>>8120567Wow, she looks more /h/. Good job.
>>8120567modern cat girls are the result of decades of gooner research and testing. this is just convergent evolution
>>8120567Great. Now the style needs less plastic.jkjk
>>8119640>>8119632>>8119625Masked union cn gave some okay results https://files.catbox.moe/5xmi2e.pngI think think the anatomy is a little borked because i have loli, child and petite neged and the afro obscuring everything around her head doesn't help either. This is what I got without a cn tho, so i don't think cn is really necessary here in the first place https://files.catbox.moe/sdxzbt.png
that's a lot of (You)'s specially for a ded thred
>>8120590Youises?
(You)sisters, did we get caught?
>>8120583poor guy lost his head
>>8120584this but unironically
>>8120597Seems the TE took "head out of frame" too literally. Rolf!
>>8120598it wasn't but i tried the "haha i'm totally not insulting you" strat
>>8120589You forgot bare shoulders though which is the part pony struggles the most with in this position. And with original swimsuit I never got the straps right behind kisaki's hair, it just doesn't seem to have enough object permanence to ever get it right. Otherwise yeah, I guess the girl is different enough from original image.
>>8120602i'm not surprised that the guy with a skill issue was the object permanence schizo
>>8120610I got better results with this controlnet. I'm just not satisfied with it.
>>8120602Wait a sec, I'll try a tomboy wearing a spaghetti top
retard here, what is promax? an iphone?asking seriously
>>8120616yep it got really good with latest coreml optimizations, 13 pro max is pretty much on par with 4060
>>8120602https://files.catbox.moe/o8rukp.png This one is alright I guess. Promax seems to ignore my masks for some reason so all the clothes stick to her/have the swimsuit shape. Have to try with anytest later
>>8120468Fishnet blush!
>>8120627The strap is all kinds of fucked between the hair though and that top is broken on the other side. It's not how NAI would look even at base res. It's much more detailed and consistent which controlnet cannot solve.
>>8120639uoh
>>8120627Also now that you exposed more of her body than in the original gen, you can see that the muscle outlines on her back are split by that top.Object permanence?
>>8120639Good shit
>>8120639cute butt
>>8120563what art tags are you using? I'm sensing at least part saltyxodium and sususuigi, maybe
Why doesn't OneTrainer respect the `Include Config` setting? It's worked like once and then every other time it's not worked. Really hard to keep track of all the parameters/ epoch I chose without the the information on the lora itself.
>>8120642>>8120648Issues that can be fixed with inpaint. I never claimed that it could beat Nai with a simple base res gen, I said that most poses were possible with enough effort.
>playing with some oppai loli prompts>body type comes out slightly resembling miqo spammer cat's>burn it with fire and start from scratch
>>8120570>>8120573>>8120577>>8120579>>8120580>>8120582>>8120583>>8120584tabbed off to do something for work and came back to this.I didn't realize you anons all felt this way.... I am still partial to the hair-covering eyes as her unique factor though. I may do more less goth stuff though.
>>8120664Don't compromise on your OCs design due to samefagging shitposters. Do what you enjoy.
>>8120563nice background and nice colors, can you share a box?
>>8120664I liked her more before, she even inspired one of my oc (tried something strange with this one)
>>8120662Except if I want any sort of intricate design with complex shapes, cutouts, straps, and not a plain simple shirt, I wouldn't really be able to fix anything much with inpaint, it would just spawn two errors while fixing one (if fixing anything at all). At this point it becomes a waste of time and energy, which the "magical" tool that you claim controlnet to be isn't really alleviating. The model should be able to draw complex stuff in difficult poses without hand-holding, only then will it be real magic (i call that naiv3).
>>8120672sounds like you don't know how to inpaint
>>8120671Glad miqoanon is inspiring such amazing, beautiful OCs! We need more of those at /hdg/.
>>8120674Sounds like you don't know how to troll.
>>8120672can't you train an intricate shape lora?
>>8120664goth curvy black lips stuff reeks of "goth gf" meme that shifts overall gen's tone to /aco/ but my post is invalid because>nogen
>the goth aesthetic is no longer hentai just becauseWhat am I even reading
>>8120683gothic lolita is fine and /h/, "goth gf" shit isn't
>tfw no gothic waifu to colonize crimea with>tfw no gothic gf to bludgeon people with
>goth>noses>lips>large breasts>penisesAlright so far this is what I have for my list of "/aco/ traits" and should not be allowed in these threads. What else am I missing?
>>8120683It just looks like shit and she is extremely overused in this general so I guess anon is just making up excuses to shit on miqo posting.
>>8120677i don't think there's any model that will help you when you have this much of a skill issue, sorry bro
>>8120689styles that show highlights and aren't perfectly flat and using a 32 bit color pallette
>>8119916>>8119918lora?
>>8120689pony gens
>>8120684>gothic lolitaYeah I haven't genned this in ages, it is time.
>>8120664not aco>>8120671aco junk
>>8120696suigintou if you're on NAI
>>8120665>>8120671I don't mind some criticism, I like being able to customize my character and other things, it gives me more things to prompt if I change her look for certain stuff too. I don't mind it. Nice work with the elf OC.>>8120684this good enough for you?
>anime doesn't have goth girlfri-ACK!
>>8120698local but thanks
>>8120699based>>8120701basado
>>8120702>short black hair>choker>gothI meeeeeeeeeeean
>>8120702persona 5 is westoid-pandering garbage
>>8120702thats punk not goth
>>8120701Do you actually just GAN upscale the whole pic and then inpaint the parts?
>>8120698>genning nsfw with Rozen Maiden girlsnyo......
>>8120701His foot bruh
>>8120718He was born that way, anon...Not cool.
>>8120718that's a youkai, he'll be fine
>>8120693>32 bit color palletteThat'd be hyperrealism-tier. Absolute limit for /hdg/ is 4 bit. I see more than 16 distinct colors in your image, I hit report for off-topic. Simple as.
>>8120701based af
>>8120661One trainer is shit, use lazy scripts
>>8120692Sure bro, I have a skill issue, or people who don't see the difference in level of detail between nai and pony that no amount of inpainting and controlnet will fix have an eyesight issue, one of the two.>>8120683>the goth aesthetic is no longer hentai just because<- goth aesthetic>>8120239 <- just some blob of fat
>>8120722Everyone should be using that pallettelize extension to limit them to 16 and I am completely unironical
>>8120723>>8120705Samefag
>>8120725based based based
>>8120725>black clothes and a stud>gothpeople please
>>8120725Good shit, mate. I'd give her the ole rubba dub, if you catch my drift.
>>8120717nah i'd only "suigintou, (feet, feet focus, feet worship:1.5)"
>>8120725Just upscale it if you want more detail, It's free on loca l;)
>ask for gothic >get scatwas it bait all along?
>>8120735There is no amount of kneeling I could do to congratulate your basedness
>>8120738you hit the secret hash anonie
>>8120741I don't have ces anywhere in my prompt, is there another one?
>imagen defending the 25$ stabilislop tune by some roaches just because you can't afford a 60$ midjourney subscription
>>8120748I don't think midjourney is going to generate nude uncensored JCs for me, so... nai it is.
>>8120748>subscriptionThat's what poor people do when they can't afford a GPU.https://en.wikipedia.org/wiki/Boots_theory
>>8120751>discworldb.a.c.k to reddit you go
>>8120751Doesn't apply here desu. In case of local you're buying a pair of shitty leaky /aco/ boots for 2k dollars instead of elite ones for 25.
>>8120758Keep coping and buying your Abibos ExtraMax every month, poor.
yet another meltie...why do i even come here
>>8120763You should come in your waifu instead
>>8120768her lora is still baking though...
>It's another NAI vs local episodeSkip
>No Nai posts>No discussion about nai genning>Naiggers still feel the need to come here every dayNot even jordachs discord has so much mental illness
>>8120774You ever considered that someone who posts local images still holds nai supreme?:)
>>8120774You ever consider someone who has a girlfriend still holds BBC supreme?
>>8120774there is a nai post though
>>8120774You ever have a dream that you that you
7 months later and localjeets are still using pony? kurumuz was right, local bakers are pathetic
>ComfyUI is surprisingly much easier to use for phoneposting since it doesn't lag af
>>8120785>Now I need you to login to the NAI and click the "try for free" button>Now give me that code for the 30 day free trial so I can confirm BUT do not redeem yet>DO NOT REDEEM>DO NOOOOOOT REDEEEEEEEEEEEEEEM>NOOOOOOOOOOOOOOOO>*he redeemed*
>>8120789>NOOOOOOO*electronic intererence*OOOOOO
>>8120785Fried style designed to make nai look bad!!DO NOT be deceived!!!
>>81207851.5 my beloved
>>8120795BASED BASED BASED
>>8120795tits too small and give her the piercings back
>>8120795bro wtf where is her lipstick
>>8120774I gen both local and nai and just come to shitpost in this thread as both sides.
>uses config that supposed to bake loras at ~3000 steps>barely becomes stable at ~4500i don't know what to believe anymore
god i cant wait for v4 to watch the freejeets shit themselves just like when v3 dropped. even had proto-astralite seething over artist tags. local hasn't recovered since, and probably never will
>>8120806You can throw all your pony baking knowledge out , once we get rid of it, regardless
>>8120664it seems to me like it's mostly the black lips that make her look more western, especially combined with the style you're using. majority of danbooru images tagged with black lips are either western style or realistic, that tag seems to even effect the style on naihttps://catbox.moe/c/mta78cwell I don't care either way, do whatever you like, I was just curious and wanted to see how she'd look in other styles and without lipstick
>>8120806>baking based on stepsliterally ngmi because that's not how that works
>>8120813then why people keep aiming at them instead of epoches? im running 317 steps per epoch
>>8120813it is though
>>8120806If you're using my config it's 3000 steps because I always use 100 pictures. Just set it to 30 epochs instead.
>>8120818bcuz they're le stupidebut yeah i posted a graph before my character bakes range from 700-1400 steps to be stable, it's better to use epochslike now i baked 50-160 image datasets and they all stabilize around the 11-13 epoch despite obviously not being same steps
>>8120810never gonna happenif you want a picture of the future, imagine a pony hoof stamping on local's face - forever!
>>8120826Wholesome.
At this point I'm convinced that jordach is trolling and just reuploading the third epoch over and over again while the 4090s are mining etherium
>>8120834you must be blind anonthe partial epoch jordach just uploaded is clearing up so fast it's scary!
>>8120799>>8120800back to normal operating procedures>>8120812I didn't realize how much of an effect just the lips would have. I still like the look of lipstick though. Nice little album of gens too.
>b-sama said that she doesn't even want to test the partial epochs anymoreit's cascover
>>8120837unbelievably based. this is peak body type
>>8120837Obviously the lesson here is that you shouldn't be avatarfagging or trying to please anonymous. You should only be posting whatever you like.
>>8120785I actually like this style, artist(s)?I feel like Nai must've scraped Weibo or something, they have a lot more of that nice Chinese artists painterly style
>>8120837Can you do the same image but with abs
>>8120834it's actually so good but you just don't notice it. to better visualize it, imagine if he released a partial every 1%. surely you wouldn't notice the difference between e7_01 and e7_02. that's the case with it now. it's actually so much better than pony and nai but you just don't realize it.
>>8120818what step count you want depends on what you are training. more complex or diverse concepts will need to be trained for additional steps. the actual dataset size does probably play a role to an extent. 10 epochs of 10 images will probably learn faster than 1 epoch of 100 images. but at the same time 10 epochs of 100 images and 1000 images don't have a huge degree of difference, which is why step count is a much more valuable metric.
I think I came 6 times today already
>>8119146Ok so, if I have a good graphics card should I be using the normal automatic1111 or the reforge one?
>>8120845>artist(s)?I don't think he has them, bro. It has a pixiv name, probably just a repost from there and not his gen.
>>8120853yes
>>8120840Jordach pumps them out every .1% of progress none of them are bothering. He needs to use the good hardware and release full epochs.
>>8120853no
>>8120837good culture
>>8120855Well shit.
>>8120857do they even have access to the better GPUs?
>>8120857He needs to closely monitor the learning rate. It can't make too much progress between one piece episodes or it'll fry. the current dual 2080TI setup is ideal.
>>8120853use reforge
Which is faster OneTrainer or Kohya? Or is it just down to preference?
>>8120846but of course>>8120861very nice
>>8120865I'll never understand his logic. if they were using the better hardware even if it fries he can just rebake that epoch in the same amount of time. its so stupid.
>>8120868i believe they're about the same speed but onetrainer has more options
>>8120868one trainer is meme'd ass with a token limit and forced background settings you can't access, use lazy scripts if you want a good UI for kohya
>>8120657>>8120666https://files.catbox.moe/wjlpxn.png
>>8120870Don't quote me on this, but 3.6b is supposed to have a larger boost from being trained on paid hardware, which is why money is saved for that.
>>8120840here i did a comparison of his imageshttps://files.catbox.moe/agkxrh.pngtop set is from july 11, bottom one is from today. is it getting better? worse? staying the same? i honestly can't tell.
>>8120872onetrainer has less options, maybe you are confused by how they label settings.
Really pisses me off even though I went through the trouble of having cropped images of Rio's halo and it still doesn't want to train it properly.
>>8120657>>8120666Are you pleased to learn that it's nai?:3 PAY UP!
>>8120879before: slight pose varietyafter: same overfit poseit's fried bub
>>8120878They aren't using the money on training they're using it on buying their own hardware. They got a deal and funding from Nvidia to use their GPUs for micro cents per hour, they're being stupid for no reason other than jordachs autism. I wish the guy donating thousands would buck up and say he wants his money spent towards speeding it up, as they originally promised. Maybe he is cucked and really does want to buy Jordach new 3090s though.
>>8120886it still have a few thousand anlas but I'm waiting for a new model to use them
>>8120886>>8116183
>>8120888There should be no pose variety on the same seed with only the artist tag changing
>>8120879Shouldve added the screenshots of him saying what an improvement each partial was to this, holy
>>8120861catbox? that is some delicious plump
>>8120890As someone who donated $15 I implore you to issue Jordach an emergency directive to spend all of my donation on renting an H100 for 10 minutes to finish the 1b bake and move on. It's a sacrifice I'm willing to make
>He says clip is goodIt's never been more cascover
>>8120897They went mask off and said they don't give a shit about those of us who donated because most of the donations are from that one whale. Quick to alienate their supporters.
>>8120880i heard onetrainer had a couple of things before kohya did like multiple captions and better embedding training because it doesn't use sd-scripts as a backendi just use sd-scripts straight from powershell because ive had bad experiences with the various guis
>>8120901>spitting on paypigsWTF I love Jordach now?
>>8120865dual gtx 560 Ti rig when
>>8120901I got a new job and was planning to donate part of each paycheck to resonance before they started saying that. Glad I saved it, it doesn't look like they would have used it for the bake. Feels bad I gave my last hunk of savings before that and they turned around to say it was worthless. Lost all potential respect in that moment.
>>8120906He'll finish e11 on intel integrated graphics
>>8120911nice fanfiction
>>8120914Might read that way but it really happened. I was dumb and full of hopium to have given them my money when I had none, but I did it.
>>8120917bawhahaha retard
You were told from the beginning, lmao. Simping a furry is a sign of mental illness just like being a furry is.
I have faith in punished jordach. once he catches all his donors genning with waifu diffusion V he will go nuclear and spent ever last drop of donations trying to catch up.
>>8120921I'm glad I donated to astra. Already got the e-mail with the 5 hashes i wanted.
Congress just passed the Jordach Appropriations Bill...
>>8120923>he will go nuclearholy, I'm now afraid
>>8120900I think he only refers to the one that Cascade has, which I agree, that one chunky 1B TE is miles better compared to the retarded duo of TE's that SDXL hasBut there's no actual way that such an incredibly awful and retarded technology architecture such as CLIP itself in general could be any "good"
I bought a Stability membership instead. No matter what team you're on, we can all agree this wouldn't be possible without the amazing folks at StabilityAI. Thank you Emad for your hard work in bringing us Stable Diffusion.
>>8120921>>8120922Yeah, I was retarded. I won't deny that. Its still really scummy of them to spit in everyones faces after begging here for so long.
Don't worry bros, another Cascade partial just dropped and this one is definitely better OMG
>>8120900>>8120930I don't mean to say CLIP is good, but generally speaking, what are the disadvantages of it?
>>8120852Update: I came 7th time
>>8120923wdV isn't real anon i'm sorry
>903/118christ
>>8120937>newfag spottedSeems you haven't used pony or any other SD finetune or base model enough time in order to experience the retardation of CLIP.
>>8120941cope imagetard
>not even on pace for 3000 repliesnogensisters, our response?
This is very based https://nhentai.net/g/345451/
>>8120947Nope, I just wanted to make a comprehensive list.I'd start with the length limitation of 75 characters.
>>8120941Grim
>>8120950antiloli schizo got buried today so you must try harder
bleak I'm going into a coma
The worst part about jordachnigger is that he once again reinforced the opinion that any AI crowdsource project is a scam.
>>8120939its real and releasing within 54 days from now. neggles will actually deliver
>>8120968The open model initiativ isn't a scam. They even formed an anti-scam committee
>>8120954>characters*tokens
>>8120954Token bleed. Often, a token or a series of tokens that are being used in a word along with other tokens may trigger what the token itself means. It's why "hands out of frame" doesn't work consistently, so it's better to put "hands" in negs along with it.
>>8120874isn't it the same 225 token limit as sd-scripts?
>>8120968You didn't learn this after the unstable diffusion scam? What's sad is that there's clearly people who want to give money to a project, but every baker always ends up being a troon or some other shit kek
>>8120954after using a certain number of tokens the model starts to lose focus, unrelated to token limit
>>8120970i would love for that to be true but seeing neggles restart the bake every couple of days because of some new bug in the training code doesn't inspire confidence
>>8120980does he work in a coal mine?
>>8120983>>8120970Neggles has been promising a model for months, Xir will never deliver and just wants validation from the trooncord.
>>8120954Due to CLIP basically working by "blocks" and because it starts to read tokens from left to right on each "block", the first tokens it starts to read on the left will always have more relevance or priority compared to the ones on the end of the prompt, this causes some tokens at the end of the prompt to be ignored about the output image sometimes
Surprise sex is best sex.
>>8120993Is that what Sexual Assaulty Attracted persons call it?
>>8121014Did you mean to say "rapists"?
>>8121015that's bigoted
https://files.catbox.moe/vgschn.jpgbatch 1 feels better but one-piece swimsuit is still gacha, x and white bows either swap their places or don't gen at all, would probably be fixed with extra epoches or unique tag (lame)all other outfits looks passablecould maybe pull off miko through pony alone but red-white hair bows would be aids to gen
Taking a few NAI requests. Give me your best.
christ only one rebake left and a test of the bigger dataset
>>8120849>but at the same time 10 epochs of 100 images and 1000 images don't have a huge degree of difference, which is why step count is a much more valuable metric.Wait I don't understand. Doesn't this still mean epochs is more important? What I've actually noticed is that if I had a character with 80 pics finish at 30 epochs, the same character with 200 pics would need 35 or 40 epochs to get all the details. Rather than get bogged down in step calculations it would be better to just use epochs.
>>8121026Houshou Marine from the popular vtuber group "hololive" kissing Oozora Subaru from the popular vtuber group "hololive"(kissing on the butt :>)
>>8121029epochs don't matter
>he doesn't bake 1 epoch with 300 repeatsngmi
>>8121026nsfw, 1girl, 1boy, dino (dinoartforame), houkisei, 2023, {{best quality}}, {{amazing quality}}, {{very aesthetic}}, black hair, long hair, loli, petite, bent over, sex, hands on another's arms, {{motion lines, speed lines}}, arms behind back, overalls, from behind, pov, pussy juice, bottomless, restroom, best quality, amazing quality, very aesthetic, absurdres
>He doesn't bake 300 epochs with 1 repeatgmi
>>8121026Chloe von Einzbern tummy
>asura>(asurauser)
>>8120989It's not really an issue with the CLIP itself, it's just a flaw of a workaround around the flaw of CLIP.>>8120981Fair, but I don't really like the way you worded it. More like, CLIP gradually starts losing confidence with every individual concept as you add more of them to the prompt.>>8120975Fair.
>Cutesexyrobuttes>(CSR)>(cuteSR)>\(cutesexyuser\)
My mind is blown, is there really no tag for finger sex gesture?
>>8121048penetration gesture
>>8121026megumin, [[[[artist:tottotonero, artist:om (nk2007)]]]], 1girl, solo, lying, on back, {{legs together}}, anal object insertion, mage staff, sweat, moaning, shiny skin, fat mons, witch hat, panties around one leg, black panties, indoors, on bed, uncensored, nsfw, best quality, amazing quality, [[[[very aesthetic]]]]undesirable content:worst quality, bad quality, [[[[very displeasing, sketch]]]], sex toy
>>8121026{{{{{{{{{{{{{{{{{{{{{{{{{{{{{{{{{{{{{{{{{{{{{{{{{{{{{dark skinned male, kfc manager}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}
>>8121031kek that just doesn't work.>>8121036>>8121039>>8121052It's all in there : https://litter.catbox.moe/yhhpuh.rarGrab it quickly.
>>8121028What is this?
>>8121063Bro this isn't a MAGE STAFF in her butt.............
>>8121066I just needed 1 image on topic for the board. Forgive me.
You know what? I just realized NTR is gay.
>>8121063based Kuro tummy providerthanks
>>8121070Masturbation with toys is on-topic though.
>>8121063Amazing object permanence
>>8121029>Doesn't this still mean epochs is more important?no
>>8121074The brown thing on the right? true.
>>8121071NTR is the straight man's gay porn
>>8121048default tag doesn't work too well but this LoRA works goodhttps://civitai.com/models/483844/sex-gesture-comission
Tried foot focus on resonance 75.e8...
>>8121063localkeks will never see this level of object permanence in their lives
>>8121085Essentially, yes. It's as gay as BBC fetishists.
>>8121087Tree bros, I don't feel so good...
>jordach is insulting b-sama nowit's over
>Sorting my Ai folder>29k images>15k are loliWhoops, I never understood addiction until this AI shit happened.
>>8121074>>8121088I'm a newfag and I can't tell if these posts are ironic or not
This month in local news:>ponyslop mix #18382 got released on civitai!>resonance bakers are having a mental breakdown live on discord!>WDV is still MIASubscribe for more news in august!
>>8121097we had some guy posting yesterday about how the biggest advantage nai has over pony is that it understands object permanence
>>8121092I wish them the best but I'm not dealing with that clown behavior
>>8120774There are plenty of NAI posts, but they don't have NAI filenames.
>I see, and are these "NAI posts" with us in the room right now?
I do both local and NAI. I only care about AI making progress.
They must've gotten so blurry that they started melting into the yellow background here
>>8121028box?
>>8121097don't take anything seriously herealso when you post NAI gens, it's likely that shitposters will crawl from their holes
>He doesn't know how to inpaintWhy bake a model when you can't even use it?
>>8121115it's more that you have to jump through hoops to inpaint with cascade because of the way the model works
>>8121117What do you mean? Inpainting set up for it is exactly the same as any other model, except you add an additional ksampler.
>>8121117Doesn't excuse that awful mask
Um...gens?
>>8121119you have to play with the compression to get a latent that stage c can work with properly
When I post NAI gens with pony names no one cares.When I post pony gens with NAI names no one cares.No one can even tell the difference between NAI gens and pony gens if it's not obvious from the file name.
>>8121121i'm bakerino
>>8121123Inpainting only uses compression 32 or its a blurry mess.
>>8121125>986/131Oh i'm sure you post tons of these!
cascade's compression nonsense will be its downfall. even with an ideal bake nobody wants to wrangle that shit.
Holy moly Jordach is turning on his biggest shills. You almost feel bad for B-chan. Almost.
2k bros, we're on track
>>8121130catbox screenshots from trooncord, it sounds so retarded that im curious
>>8121133this. someone give screens i want to laugh.
>>8121121whats even the point anymore
>>8121063>>8121121Bro there's something wrong with your nai settings, it shouldn't look like naiv1...
>>8121137shut the fuck up.I need you to just shut the fuck up.I literally just got home from work, just want to share some cool pics with my bros and you're here ruining the vibe.Just shut up.
>Astra lurks and I don't want to give him the satisfaction of seeing pony gens>Nai sucks>Genning with arti/animagine is too much work and gacha
>>8121133>>8121134Bro that's doing live feed from cord is exaggerating for shitposting's sake. He's just farming the post count. There's nothing interesting there.
>>8121140That convo of tree and that other guy comparing mental illness histories was interesting
>nai continues to top the pixiv charts>local falls deeper into irrelevance as jeets beg furries for partial epochs like a cargo cultglad I joined the winning team
>>8121136I'm still having fun. Not sure what's happening in here though. Did local really become that bad?
>>8121144gen, sirs?
>>8121146you don't deserve one you pathetic catbox begging brownoid
>>8121140fuck, i got baited already had mental image of the autistic furry spergout
>>8121148prease sirsu i needo da nai quarity
>>8121145Nah, the only people left posting here are just schizos who hate everything.
>Watching seasonal animu>background chars where clearly genned with ponyDamn, local won
>>8121140Screenshot catbox so we can see its nothing? I don't want to make a troon account to gauge autistic meltdown status.
>>8121152fuck i was actually watching nickelodeon my bad
>>8121151It's not schizophrenia, it's burnt out brain. Image gen is worse than heavy drugs. Everyone is salty and mad because they don't have joy in life anymore!
I checked the cascade discord and it's just jordach posting more 1girl anthro grids that look no different from previous epochs kek
Is NAI down again?God, I hate how often I literally can't use this service I pay $25 a month for.I should just play world of warcraft instead at this point. It's only $15 a month and I don't feel like the servers are down 24/7
>>8121155>were already complaining about lack of new thing a month after a revolutionary breakthroughmakes sense literal druggieshttps://www.youtube.com/watch?v=qrxv0JNVtgY
>>8121138I'm so sorry but can you adetailer at least please??
>>8121155Actually true, I'm just happy this hasn't happened to me. But I guess it makes sense because I gen in moderation and never did bother doing it for like 12 hours straight for weeks.
>>8121157It's not, I've never had issues.
>>8121151I wanted to try out local again so I came back here after a year. Far fucking cry from the original leak months. I swear the few gens posted look worse than back then as well.
>>8121164thissssback then we had cute 1girl brown maido standing and 1girl brown maido standing and 1girl brown maido standing
>>8121164I keep coming back here expecting some new things but for fucks sake we're still stuck on Pony.
>>8121166Brown meido poster was the soul of /hdg/.
Are we back, /hdg/bros?
>>8121167I'm genning on cascade 1b and it looks amazing damn this cooks its already 10x better than pony
Alexa, google "Hedonic treadmill".
Does it look like she's fingering her pussy or her urethra now? I tried to adjust the tags so her fingers would be so high up but I'm not sure.Is it my Model?
>>8121171>tfw 3 years ago it was either learn how to draw or commission some art school dropout out the assYeah im happy with what I got rn
>>8121156>>8121140jordach got mad I pointed out a possible error to try to help, then told me to fuck off for using the hardware he kept telling me to use because I guess its off limits when hes in a mood. I left the discord b/c I'm not dealing with being shit on for trying to be helpful. it really wasn't much of a spectacle/not worth a screenshot, but thats all that happened. I still hope the project succeeds, but truth is they never asked for my input, help or anything, so I'm going to put my time and efforts into something else. thats it - b-
This never would've happened if mods didn't ban the cunny posting.
>>8121170I'm just baffled at the decision making of these model bakers. They want NAI but better. They know somewhat what NAI is using, and yet we're continuously deviating from that.
>>8121169BAAAAAAAAAASD
>>8121164Probably bait but local is far beyond what it was a year ago. People able to gen coherent gangbangs/etc/etc is crazy.
>>8121130>can't even develop the Gradio frontend stuff without people wondering why it's not workingMy 3090ti runs a stable instance for a reason. The 3090 (non ti card) runs the indev version.Funnily enough I was working on a *minimum* masked area strength slider so I don't have to paint the whole damn image, as well as other minor UX tweaks that affect the 17 open tabs in Firefox/Chromium derived browser to be reloaded/reset.>>8121177I'm trying to get code out of the door and this frustrates me while I'm working. I've also conveniently finished working on changes for now. I'm sorry when I'm in the middle of something I tend to get aggressive, you forget I have a very strong tism that does not cede anything for social norms. I'd already mentioned I was working on something before you even mentioned it.
>>8121177Listen here little baby. You're gonna get a lot of hurtful and degrading comments, but that ain't what I'm about. Let me just say, you are perfect the way you are. You hear me sugar? PERFECT. Don't ever change. You deserve anything and everything you want. Stay safe for me, baby girl.>mfw thinking about you hurting
>>8121177I'm sorry you were the victim of domestic violence, b-chan
>>8121182Didn't loras already do that shit like, a year ago?
>>8121172it can be spotty on miqomix sometimes, I think it is heavily impacted by the position your character is in. Try "on back" and see if it makes it any better.You can always try a lora for fingering too.
>>8121177Don't lie, b-sama has always been considerate about jordach's outbursts and would never do something as vile and disgusting as LEAVING THE DISCORD for such a petty little thing!!
wait is he actually falling for it???
>>8121182Which again points to hedonic treadmill>ohhh that model does 30 times more things than the previous one an in hd???>GRRR IT DOESN'T KNOW X even THOUGH THE EARLIER ONE DIDN'T EITHER>*black penis* *black penis* *black penis* for 5 months now
>>8121187Not even close to what we can do now and anyone saying otherwise is retarded.https://civitai.com/posts/1892686These are all done with autismmix/pony for example, no concept/character loras needed.
>>8121177>>8121184nice larp session
>>8121196lmao people forget there were 0 black cocks in the pre XL days because any attempts to gen peanus and 1girl looked awful
>>8121184pretty sure that person is just pretending to be her
>b- joined astras discordGreatest ntr plotline of the century
b-sama.... NOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOO
>>8121201at least i know people are really loving local nowbecause people didn't even bother filling out a full thread in three days in 1.5 twilight days kek
>>8121199Why shota porn of all things...?
>>8121200https://github.com/Jordach/CAPGUI/commit/ba7b1efbb611bcd377c32f21fda9669c6eaaae2fI wasn't fucking kidding about the software maintenance period. Really was cooking.>people genuinely forgetting that i have supreme autism that prevents social cues working on me
>>8121207When is that example comfy workflow coming? the one in the rentry is mid and copying values from capgui ones doesn't work
Is there a way to have NAI automatically gen images one after another in the background? Without having to click every time I mean, and for free ofc.
>>8120847>it's actually so much better than pony and nai but you just don't realize it.that's right. the outputs are much better, but they're not finished end products, because that's something only lazy fuckers would enjoy. that's also why they're much better than pony and nai
>>8121212Get the API to run it in A111/comfy/forge.
contributing for da boyshttps://files.catbox.moe/gqqehm.pnghttps://files.catbox.moe/s5rwgy.pnghttps://files.catbox.moe/qlfrrt.pnghttps://files.catbox.moe/jhpl3j.png
>>8121212There are extensions for webUI. But also if you abuse that you will get banned from infinite genning and will be left with 10k anals you got for your $25-tier, so...
>>8121206Just goes to show how powerful local is now
>>8121207sasuga Jordach-sama kek but i'm pretty sure that you responded to a shitposter earlier
>>8121216the boys in blue, eh?
>>8121213Dear god I hope this is bait
>>8121219kill yourself
>>8121177If this is the real B-sama prove it with your trip!
>>8121224That's not an argument sir
>>8121216Did you just remove both thick thighs AND huge ass from positive prompt?!??I thought that day would never come... There are tears on my eyes.
I nominate bluvoll as b-'s successor
>>8121211I can print some examples out for later - but right now I'm busy with other stuff.>>8121220You can tell I struggle with social stuff. I can't help the laser focus to getting shit done, it's my curse, I am best left cooking with others to handle the more social side of things for me. I literally test stuff many times over before pushing those changes to GitHub.
>>8121223it says gullible in the ceiling, anon
>>8121215>>8121218I see thanks. I'm too paranoid so I won't.
>>8121232You should still use the API though, that's not prohibited. At least it will allow you to use adetailer and stuff.
>>8121227Yeah I made a mistake, never gonna happen again!
>>8121228bluvoll nudes WHEN?
>>8121212https://github.com/Metachs/sdwebui-nai-api/tree/main/scripts Can do it with the standard generate forever I believe. But as others have mentioned, NAI can detect this and will eventually suspend or ban you if you do it a lot.
>>8121230jordach-sama, when are you going to start using the nvidia credits?
>>8121235It's much better though. When they are plump and short it just makes them more like toddlers instead of making them older. Thin is definitely more hebe.
>>8121157>error 500keks, our response??Boy was I glad when I finally left NAI.
>>8121230It's probably for the best if you don't use these threads for now. Only schizos and shitposters talk about cascade here these days.
>>8121239Actually you have a point
>>8121238Literally stuck on waiting AWS to unfuck their retarded motherfucking asses and hand over A100s that aren't spot (killable by other customers) instances. Sadly, they're yet to release that restriction. Kinda stuck on that front. The two 4090s are whirring away in the mean time and *is actively* improving and becoming less muddy.>>8121242My lack of social skills gives no fucks.
>>8121235Don't listen to >>8121239 they were so deliciously thicc before, please keep posting with those on positives
>>8121247DAMN BASED
>>8121247What does that in the corner read? tingling fatdong shenhe?
>>8121247Can we get something new unironically? I don't even hate miqoposter but I don't want what few pics we have to be 50 different versions of the same thing. You could at least change the hair colors ffs.
>>8121247Now she's perfect In your face, miqokek anon
>>8121235Don't listen to both of them! Gen with and without for variety!
>>8121253He said he'd eventually grow tired of the character back in march...
>>8121255ok that's fair
>>8121248You're right as well. Will try to have them older but not too plump and not too thin
>>8121234>>8121237I'll try soon thanks. I'm kinda dumb so I don't get what adetailer does that NAI ui doesn't do though.
>>8121064NAI
>>8121255Yep
>>8121260>older>not too plumpmy god pls nononono go back go back don't redeem
Pregnant Tomoe Gozen pouring sake onto her huge saggy tits.
>b-chan really left jordach's discordit's so joever
>>8121269Why did she leave?
>>8121266Don't worry bro
no one wants to see blurry nai slop especially when they can only do ONE style because prompting on that garbage model is a huge mess
>>8121269This is what you get for an outbursts during queen's pms! Don't fuck with the queen when she's on a red day...
>>8121221HOLY SHIT
Is "b-chan" pronounced as>be minus chanor>beee chan?
>>8121244Are the A100s faster than your 2x4090's? Asking because they do not handle bf16 as good as a 4090 or an ada6kWould a rig of A100s be faster than that at least?
I have an radeon r9 255 laying around that i could donate to the project.
>>8121221You lucky bastard. Good job
>>8121028WDV looking good so far
>>8121277In my head its>beee chan
>>8121269Literally told someone before to hold their horses while I unfucked my dev instance last month. I've said many times over that my 3090 is not exactly safe to work on, as any development time cuts into my own personal generation time and I don't even want that to occur, so I try and get it done ASAP.>>8121278A100s are "slower" at bf16 because they work in fp32 then just downcast the weights at the end. However, more VRAM allows me to turn off gradient checkpointing and go 50% faster. They're about the same if the Ada/Hopper line are using gradient checkpointing. If Ada/Hopper cards are not using gradient checkpointing, then those cards run twice as fast as an A100.
>>8121223on one hand it is, on the other hand those are just things jordach said, so
>>8121288I see, so I imagine if you had the 40k donation milestone you would go for the ada6k only, am I right?
>>8121207>muh social cuesi agree even more with anons from earlier that tree should be the sole spokesperson for your group, you're a walking PR disaster
>>8121297Tree gets filtered by that captcha
>>8121297tree already posts the public updates though. do you want him to have a spokesperson to talk on 4chan or what?
>>8121228No thank you, I'm fine in my little corner training underbaked shit, if the resonance guys end up with something good, that's nice, if they do not, then its business as usual.
>>8121288how confident are you that the 3.6b effnet will be able to lock into the TE that was trained with the 1b model? it's hard to tell how good the current TE is because the 1b effnet is kind of jank
>>8121261NAI ui doesn't let you img2img inpaint. Only infill. Adetailer just automates the process of img2img inpainting faces and stuff. It makes things faster, but doesn't do anything you couldn't do manually.>>8121274You discredit yourself with every sentence you utter when it's obvious to anyone with eyes that you're a delusional bad faith troll. You couldn't be a more effective advertisement for NAI if you tried which is why you are the single person I think is most likely to be a shill. The alternative, that you are in here 24/7 stirring shit FOR FREE is simply too ghastly to comprehend. That such a worm calling itself a human could even exist...
I nominate Astralite. Not only did he actually deliver a model that revived local, he has the hardware/money to deliver yet another one.
>>8121307I thought that your pony finetune idea was pretty neat, you definitely did a better job with that than the 4th tail guy. Any plans to do that with more artists?
>>8121311Nominate for what?
>>8121310the secret third option is that it's a spambot
>>8121297>Could turn off discord and "focus" on other things>Could probably resolve his biggest shills PMS if he gave a sincere apology>Blames his raging autism instead>Working on his issues if he knows they're that bad never occurs Autists need an LLM that autocorrects their autism for every message they send. This is the only true solution. >>8121307>Even bluvoll wants nothing to do with ResonanceCascadebros, is it over?
What if...somebody just leaks V3? Wouldn't that be funny?
>>8121296There would be an outright permanent installation of a 4x A6000 (Ada) machine in the DC.>>8121297I'm a get shit done person that ignores social cues. Somethings will get caught in the crossfire because I'm trying to solve an active issue. Obviously, I'm going to feel like shit afterwards, and that I want to make things right, simply because that's the correct thing to do.>>8121308The Epsilon prototype reached decent coherence within about 25% of the first epoch, which to say is very fast, given that the official paper states a 16x speed up, I measured a 20x speed up on the first partial of Epsilon. Make of that as you will.>>8121311Mf got kicked to the curb by all investors.>>8121319I would but I have no avenues of such a thing currently.>>8121319Giving up is quitter talk.
>>8121320>>>wdv trooncord
>>8121288are you aware of how shit and underbaked the (base) cascade 1b version is? because at this rate youll end up doing 5 or 7 more epochs more compare to what it would take with the 3.6b, i mean you literally will finish to bake what the warp team didnt do in the first place cuz they already said 3.6b was their goal
What if...somebody just included artist tags in v69? Wouldn't that be funny?
>>8121310oof looks like i struck a nervethey dont call you nai sissies for nothing
>>8121325>>8121324>>8121323>>8121320>>8121319>>8121316>>8121315>>8121313>>8121311>nogen
>>8121323you wont defeat asstroonlitehe has market dominance
Not even kurumuz and astra could have a nai vs pony discussion as good as this thread
uh oh nai melty
>>8121320At this rate we're better off waiting for v4
>on pace for 2500/300based desu
>>8121336UH OH I WILL FUCKING BEHEAD YOUYOU BETTER SHUT THE FUCK UP
>>8121313Eventually yes, testing more tokens as I end up hitting latent knowledge for other stuff and that takes a lot of time, I'll try my hand at it again, or perhaps release a LyCORIS/DoRA like HLL, dunno, I want to get better results as this one has some bleeding from StardustPapi, Spicyfuse, Tousen and Damda into base gen, putting their tokens into negatives helps, but 200USD+ a pop for a checkpoint with 80k images is hard for a 3rd worlder like me
>>8121330That should be this generals new name. /ngg/.
SIR NAI GROID REDEEM MELTED JPEG SEPIA YELLOW ARTIFACTS CASCADE SHILLS B- TREE BLURRY ARTI FRIED NOGEN TROON ASTRALITE NEGGLES wdV 2 WEEKS
shitters are uppity, time to exit
>>8121052>>8121063Always fun when old prompts pop up out of nowhere
>naikeks consistently the most invariably jealous and vitriolic people around>same thing as was in /aids/must be something in the water
damnI came 8th time
>>8121347time to call the jannies to archive this crap you mean
>>8121325I've stated this many times: the text encoder is the most important piece of the puzzle which the large 3.6B model depends on critically, having a reasonably competent 1B model in the meantime is also worth a lot. Especially since I keep finding training free optimisations that can be applied as part of the generation workflow itself and thusly automated.
>>8121349They don't pop up out of nowhere, I copied it from your gen obviously!
fellow nai-sissies just got word from the higher ups at novel ai they said to tone down the seethe in our posts so it looks more believable
>>8121323What this guy >>8121308 is saying might have valid point in another way: are you sure TE will be able to properly learn ALL the concepts trained together with 1b if, for example, top-down bottom-up currently looks like this: https://files.catbox.moe/2llekt.pngCould the difficult concepts just straight up not work later on 3.6b because TE never reached the point where it could learn them because effnet is shit?At least they don't work on the proto which was trained in the same way.
>>8121353aaaah so this whole time the 1b bake was just your guinea pig to give you time while you develop side project and experiment with other things, okayi trust the plan
Hentai? Anime girls?
>>8121350/aids/ was always /naig/ doenobody was using local when i left
...most of us are here in the /h/ threads because of some floox porn we came across aren't we?
>>8121366none this is whatsapp chat group now
B-sisters, where do we go now? Did B-sama really join Astracuckies discord?
>>8121363Basically from what I can tell, the text encoder has completely bottomed out on it's learning data and can no longer improve anymore. It's reached the total informational capacity of the dataset. Shuffling tags can help but that'd only cause it to lose focus on the real things that matter more.The prototype had a literal wet farts worth of data (~440k images) and I'm surprised that even works as it is.>>8121364Pretty much, fuck around and find out information from the smaller model that also applies to the bigger brother model, while also removing bugs and other issues at the same time.
>>8121372B- just sent a $300 donation to astralite
>1151/142
>>8121372B- just sent nudes to astralite
>>8121375potentially 2300 posts this time lesgoo
>>8121373What I'm saying is could it have bottomed out and stopped improving because 1b effnet is holding it back?
>>8121376it should have been me
>>8121370I'm here to find good examples of local gens to copy and modify. Can't find any...
You know what, I'm gonna be honest, NAIv3 is simply the best model available for hentai right now. Seethe all you like, but everyone with eyes can see it. If it was leaked, every single gen in this thread would be using it or a mix based on it. SMEA alone is a complete game changer and only recently has local gotten even a slightly accurate implementation. To be even more honest, someone who can't afford $25 a month is probably a dalit living in the slums on Delhi desperately hunt and peck typing on a barely functioning windows xp laptop that someone in a real country threw out 20 years ago, in between bouts of taking liquid curry shits (conveniently in the street right outside) and waiting for the opportunity to rape a dog or a crippled girl with 4 limbs. They just need to rangeban the whole fucking Indian subcontinent already (from the entire internet), and we can just pretend they don't exist while they die of corpse aids from bathing in the open sewer of a river they hilariously call holy while dumping the dozens of corpses of those hit by trains everyday (if only there were some way to know when they were coming!) in it.
Good news bros, trump just announced that he will release NAI3 if elected.
B- is helping Neggles transition
>>8121379then go and send them
>>8121380>>8120304
>>8121378No, because 1B still has space to overfit and it's nowhere near that yet, even with the brutal sledgehammering of high LR initially. It's still got space to go, and more or less needing to settle down. I know for sure it's getting close as I slowly lower the text model's LR.
>>8121381you're so brave for saying things everybody knows already
>>8121381Yes, yes, let the hate flow through you
>>8121381why is your gen so blurry?
>>8121381Bro? Hello? Maybe use that powerful model to it's fullest extent then?..If all I saw were your gens I would think that NAI is the stiffest most boring model on earth.
>>8121382Inshallah ibnTrump al-Khalifa
>>8121381I would agree with everything in this post except for that I opened the thumbnail and saw the image attached had two belly buttons. now I can only agree with the second half
>>8121381I have 35k anlas and haven't touched NAI for months my good sir. It's not a question of being poor, NAI is ass to use compared to local with a 4090 speed wise.
>go on sadpanda>another 500 AI galleries from new Patreon e-beggers who are literally all using the exact same fucking truck-kun BLACKED promptJesus fucking Christ.
>treech has india and israel rangebanned>Only local postsReally gets the old noggin joggin
>>8121396It's an extra fold to add more texture when you pet it.
>>8121373Y'know, something I appreciate from you is that you take your time to go here an explain things of this project specially at these times when we're suffering a nogen troll syndrome, something that cowards like astrakek or neggles would never do, they honestly prefer to stay with their cocksucking shillers while not hearing a thing from the unpopular opinion, if we have a technical opinion about whatever, you know we're not sugarcoat itDo never let these current troll retards to scare your ass
>>8121381they seethe, they gnash, but they can't argue with us.
>>8121402it has a higher percentage of nai posts than here though
>go on 4chan>another 3 images whole thread from new /hdg/ nogenners who are literally all using the exact same fucking BAIT postAlmighty Allah.
>>8121406Astra is lurking all your feedbacl
>>8121406Real, astralite got ran off of this general after being exposed.
>>8121399>sadpanda will never have the option to filter AI galleries Its over...
>>8121399its nai
>>8121410Bro we are speaking with jordach-sama, go stand in a corner and wait for adults to finish!
>>8121416>internet is completely flooded with ai slop artists were right, unironically.
>>8121411Hope he knows damn well how much I hate both xir and his garbage fried joke of a model :)
>>8121407this image is really disgusting for some reason
>>8121410Just embrace that nowadays this general's only purpose is shitposting and maybe occasional model gossip and have fun
>>8121418>go stand in a corner and wait for adults to finish!what ur girlfriend says to you every day
>>8121399>half of the sadpanda frontpage is blacked griftersai was a mistake
Bros... does this mean I have to stop shipping Jordach-sama and B-sama? Do I gen it as hatesex now? With Meanoldtree watching, of course.
>>8121421yes, the sight of a vagina is repulsive to the homosexual.
All this arguing, yet none of you are genning anything to prove your points. Come on you fucks, pick an interesting prompt and compete on who can gen it the best.
>>8121428Share tree lora, kudasai
>>8121406Like my lack of social norms and cues cares about that. I'm open with what I'm thinking about proceeding with because someone might have a more realistically achievable idea than mine.>>8121419It's only correct in that most websites are for profit at low cost. If it was being made for free/passion it wouldn#t be slop.I just feel like I fucked up massively earlier and feel like shit for it. I know enough not to go chasing, just know to let things chill.
>>81214321girl, standing, smile
>>8121432Why?
never interrupt your enemy when he is making a mistake
>>8121416what? just filter the other:"ai generated$" tag
>>8121432We aren't arguing, there's a civil discussion going on, don't spread misinformation!
>>8121432NAi could never
>>8121436Kills two birds with one stone. It lets people shitpost at each other while actually posting hentai in the /h/ general.
>>8121445you first retard*epic*
>>8121432>nogen
I'm going to make a brave statement. I still enjoy genning, and would be fine if we only had pony for another year. You all need to relax and go play some games, not gen 24/7. You've fried your brains.
NAI is amazing
>>8121445this place is unsalvageable who the fuck cares about hentai
>>8121451Holy shit, look at that object permanence
>>8121453>most people post /aco/ or /e/>yet we refuse to move to /e/weird
>>8121450and you've fried the eyes of your poor character
>>8121457So true, I like to zoom in on the eyes when i'm cooming.
>>8121457Bro? That eye is perfect.
>>8121458same but unironically
>>8121457>best eye of the thread>friedgrim
>>8121434>I just feel like I fucked up massively earlier and feel like shit for it. I know enough not to go chasing, just know to let things chill.Bro let it go shes PMSing, roasties are unstable
>>8121450Box please
>>8121460>>8121462i suggest checking your eyes irl
>>8121466if you're implying you're an anime girl irl thenanime tits or gtfo
>>8121465Have you never seen an opossumachine lora?>>8121466It's a perfect eye, you can stop trolling.
>>8121465https://github.com/ashen-sensored/sd_webui_stealth_pnginfoCatbox should have metadata as wellhttps://files.catbox.moe/gflbee.png
>>8120958Mind sharing a box?
>>8121463...You know you can have female friends that don't involve sex right? Are you that shallow of a person? Has hentai manga fried your brain that much my guy, then again this is /h/.
https://www.reddit.com/r/NovelAI_Hentai/comments/1e16oiu/the_benefactor_bountiful_care_part_2/lool
>>8121474You using heart of apple or something? Why are your prompts not in the alpha layer?
>>8121477>You know you can have female friends that don't involve sex rightlol cuck
>>8121476https://files.catbox.moe/h8rnwt.png
>>8121484Sex is fucking boring and this comes from someone who had an absolutely crazy ADHD GF 12 years ago. No sex is ever going to match that energy again.
>>8121487did she go to the other school kek
>>8121485Thanks.
>>8121487Just please make the comfyflows instead of responding to every stale bate here
>>8121475It's weird to find pics that already have metadata using the extension, thanks anyways
>>8121502thanks for stopping by!
Uh oh, guys... I think i have some object spermanence in my pants...
>>8121505kek
>>8121502i wonder if you will get jannied for this post like other shitposter earlier kek
>>8121513let's try
>>8121432>smug, looking at viewer, from below,squatting GO
kek
>>8121516holy kekone thread two retards who gaslighted themself
>>8121522She got worse bone spur than trump
>>8119286https://litter.catbox.moe/vbcbaz.png
>>8121450thank god this was deletedtext posts only please.......
>>8121527us law no lessanons will be visited by feds
>>8121527>>8121537shut the fuck up ESL samefag jesus christ
>>8121527>>8121537>>8121540samefag
>>8121529I like it! Thank you!
What's the model used?
>>8121551animagine
>>8121551object permanence smegma + dyn
>>8121551based64
Ok so, I'm on the latest commit of reforge dev branch (Or so I think) and my adetailer is doing weird things, the face comes out very blurry what could it be?
>>8121551waifudiffusion 1.5
>>8121558uhh... certain v3 disease???
>>8121558Is your adetailer resolution set properly? Also, what sampler?
>>8121551To give you a serious answer: it's very obviously NAI (but with no inpainting and at too low a resolution for the composition).
Bigger resolution: https://files.catbox.moe/w21cj0.png
>>8121564jesus what happened to her eyes
>>8121564Anon, the 1.5 eyes?
given their massive volume of human-trained feedback, civitai should publish an autotagger. it's even in their business interest to improve the caption quality of training sets
>>8121558Another exmaple>>8121561>Is your adetailer resolution set properly?yes, I have it on 1024 for both width and height>Also, what sampler?Euler SMEA Dy with SGM Uniform as schedule type
>>8121571If you just switched to Smea, then it's solved because it really struggles with denoising.
>>8121569They have an autotagger but its built into their shitty website and is not superior by any means to the existing ones
I'll make a tagger based on e621 tag human-tagged hentai images. Trust me.
>>8121562Don't wanna pay for nai but if I want to copy the style to at least 70% what models can I use
>>8121579there is none unfortunately
>>8121575they use wd 1.4 for their on-site lora trainerwhat i mean is they probably have the most humans voting on tags and providing metadata in realtime, that's opportunity for training
>>8121579ipadapter?
>>8121573Hmmm, so I just needed to gen more with it? that's super weird but seems like it doing it fine now
>>8121584what no lolit has inherent bluriness in it when denoising, you just got lucky igimho set your adetailer model to euler a and be safe
>>8121392Another hot take: missionary vaginal is the best overall sex position. If you could only pick one type of sex for the rest of your life, you'd be a fool to choose anything else. Oral is great too especially as foreplay, and sure you can change-up positions for vaginal now and then, but missionary is comfy and puts you and your partner face to face. As for anal, anal is just inferior vaginal that gets shit on your dick and makes the room smell like ass. I can't understand why anyone but faggots or Indians would fixate on it exclusively. I'm 99% sure that most anal fetishists on the internet have never actually had anal (or any other type) of sex in real life. It just doesn't feel as good for either party and it takes a lot of fucking lube.
>>8121579>>8121583That sucks but tnx
>>8121571Lmfao the schizo naihater was right, SMEA is Indeed blurry, or it's just that the one available for local just sucks so much
>>8121573ive had fine experience with SMEA for inpaintingfor some reason it's only full screen img2img where it consistently suffers for meive come full circle back to regular euler a however now that i gen with RAUnet
>>8121591nobody who posts in this thread has had sex my guy
perky breasts lora sexy
I haven't checked here in a while, is there anything better than confetti for general use yet
>>8121591Eh, prone bone would be my pick. There's nothing quite like the feeling of putting your entire weight onto her ass and back and pinning her down. Doggystyle is a close second.
>>8121601lole
>>8121589>>8121593But it literally werks now, I didn't change anything, you could see it on the boxes I have been posting
>>8121604wellenjoy i guess
Nico Robin is perfect for SD renders lol
>>8121606So that's the hourglass figure I've been hearing so much about
>>8121579>not a troll after all>people actually keep coming here to ask how to gen the same style over and over
>>8121591I mean a lot of people says paizuri is mid irl but damn it looks great
>>8121591it's a good thing you can't smell images huh?if every hentai artist drew nothing but the kind of sex acts they'd like to engage in irl then we'd have next to no hentai
>style dataset twice as large as the original one>last time i baked it the result was worse than the ogmoment of truth here
>>8121604There's probably a huge bug going on, related either to how ReForge interacts with adetailer or with the SMEA scheduler. ReForge anon, check please.
>>8121601moreconfetti is a straight upgrade
If the reforge guy comes here I also have a question
>>8121620cool, thanks
>>8121616my experience with comfyui was, it's fine for inpaint, bad for full field img2img (like hiresfix)but it's not among my preferred choices for inpaint regardlessand i also ultimately abandoned it for base gens too
>>8121394Why the fuck was this deleted? "Girls with small breasts = loli" in their mind???Is Hiro so afraid of losing his sponsors jannies have to behave like twitter roasties losing their mind over anything that doesn't look like a /co/ pineup with oversized boobs???
>>8121620>>8121628I must be retarded because i can't find anything, I checked like the past 10 threads and google/civit ain't showing me shit
>>8121639https://civitai.com/models/184813?modelVersionId=638095Do some test with it yourself before deciding if you want to swap or not
>>8121637>thread so fucking filled with text messages i can barely archive search itoh thatmaybe janny is based and knew to delete nai kek
>>8121591>>8121602That'd be a hard choice, but I'd pick cowgirl probably. If she's straight up it's hot and puts a lot of weight onto your body, and if she comes close to hug you get to kiss her as well. Plus it's versatile, you can let her do everything, or you can be the one doing everything if she collapses onto you and you thrust her hard from behind + pressing down with your hands on her ass.At least it was by a certain margin my favorite position back when I had a partner, even though any form of doggy is probably what gets me hardest.
>first xyz of epochs for style with my complicated character>every epoch but one is perfect and has no weird glitchesthat's gonna take long
>unironic futashit now
>>8121111https://files.catbox.moe/p0dunh.png
only ironic futashit here
>>8121663>https://files.catbox.moe/p0dunh.pngthanks!
>>8121662>:(
>>8121651NTA but I would choose suspended congress (not the reverse one), it's also very versatile, you're face to face with your partner, can easily kiss or hug her, can grab her legs or butt and she can leglock me too, the downside is that it requires a great amount of stamina and strenght but you can support yourself using a wall or even sitting too, still I wouldn't mind at all because I go to gym and do calisthenics, and also my wife is small so it's a win-win.
>>8121591>I only like pov missionary vaginal!!!!!!!Imminent penetration?Guy with cock out and girl is crawling away in fear? (since you love violence)Arched back, breast grab (both own breasts and by the guy, maybe together), foot focus (yes, during pov missionary), motion lines, motion blur, speed lines, trembling.Girls having any indication of enjoyment during sex (upper teeth only, blush, moaning, closed eyes, embarrassed, looking away, pick any)? All I see from you is either smug or in pain.Girl holding her favorite plush toy while fucked.Guy leaning in towards the girl. Girl leaning in towards the guy.Push her head into the pool and drown her while fucking, idk, do something new already.Any style experiments at all??
the hottest sex position is the casual flash during daily chores
reforge guy, should i update reforge from forge upstream? seems Illya updated to upstream A1111, but also Gradio 4.Or maybe in different branch?
>>8121675I only do stretching exercises so I'm not really fit, I'm not confident I could do this well other than when sitting. I do like it when sitting though yeah.
>>8121682Can't say I fancy a gradio change that probably breaks everythingBut anon, is it possible to put the <lora:blabla> at the end instead of the begining when selecting loras?
wtf https://github.com/Eugeoter/waifuset
Bigger resolution: https://files.catbox.moe/i0e1hb.png
do you think leaving tags in like "collarbone" or "teeth" help or hinder a character LoRA? not all images will contain said things so in theory you are teaching the model what the character looks like with them visible
>>8121682Not worth it, will break so much.
>>8121615>still shittierthat's almost funny>>8121705I don't think captions matter THAT much.Especially considering a lot of people train on almost nonexistent TE LRSame with tag order
>>8121704nice!
>>8121705theyre bad
>>8121682Gradio 4 has a mem leak dont upgrade to it
>>8121677>imminent penetration etc.I do and have done every single one of those things except drowning and foot focus, but you always come home to what you love.You also probably don't realize it's me when I'm doing other styles and content because they are other fucking styles and content.
>>8121677And another thing, you're spending all this time bitching at me for same content instead of people who are literally posting the same literally who donut steel oc in every fucking thread for months.
>>8121677And again, since I'm not a tripfag, you just don't realize when I'm posting other shit that I do like with local.
>>8121713To be fair, a lot of anons bitch and those avatarfags. You're just getting it now because that anon isn't there and you're slightly more visible. Don't take it so personally.
And I don't even try posting anything slightly adventurous because even if it's eastern style if it's not totally flat kawaii ugu anime style of anime schoolgirls with huge tits someone will maliciously report it and if the wrong jannie sees it you eat a 3-day which is annoying. Fucking got 3 days for megumin and nagatoro while there were active megumin and nagatoro threads on the board. Can't even be a yordlefucker without someone trying to get on your ass for furry or some shit these days, as if yordles were any more furry than miqotes or whatever other animal eared characters get posted all the time. It's just a blue skinned goblin girl with animal ears.
>>8119146>>8121696HOLI FUCKING MACARONI THIS IS THE BEST DATASET MANAGEMENT TOOL HANDS DOWN, THANK YOU EUGE
>>8121721Sir, this is a wendy's. Stop giving attention to the nogenner, he does this with every regular poster.
>>8121721Again anon, like with everything else on this entire site, jannies are stupid and shit at their jobs. Don't take it so personally. It's not that you did something bad but that someone doesn't like you for whatever retarded reason and decided to report. It happens to everyone on every board.
>>8121707>I don't think captions matter THAT muchgood morning nochekaiser sir
>>8121705it depends on how prominent they are, but in my experience autotaggers tend to add them more than necessary and i prune them most of the time. consider also being more specific and using tags like "clenched teeth" or "upper teeth only" instead
>>8121711>You also probably don't realize it's me when I'm doing other styles and content I do actually. After all, you're the only NAI poster here who always uses disgusting westoid artists in your gens, so it's easy to tell by how aesthetically repulsive they are.
>>8121732>nogen>thinking anyone cares about his aesthetic opinions when he's too cowardly to post something he thinks is goodPut up or shut up
>>8121737? I was only explaining to that anon why his gens are so easy to spot.
>>8121742If your aesthetic taste is so superior, demonstrate it. Post a gen.You can't, can you? Like, I literally think you come to this thread about how to gen AI hentai and yet you yourself do not know how to. Or, you're on that Delhi laptop and literally can't without exploding its pentium 2. Tell you what, you don't even have to post one of YOUR gens. Just find some gen you think is good and share it. It's pointless to have any sort of argument on the relative merits of different styles when one side doesn't even have a style to put forth.
>>8121063>litter dead NOOOOO, the only fucking good thing in this shithole of a thread and it's deadFUCK
>>8121696AAAAAAAAAAAAAA>deoverlapping (gloves, black gloves -> black gloves)>wd taggers>automatic aesthetic scorer>artist and character detection>does not lag>history of changes, undo & redo>searching>automatic conversion between prompting styles>many moreFUCKFUCKFUCK IM CUMMMMING
>>8121745rest in peace little blurry nai slop...
just saw this cutie some minutes ago and then suddenly this idea came to my mind almost instantly, couldn't help it (kani warning)https://files.catbox.moe/r552go.pngi will take good care of her
>>8121750Who's this? Recette?
>8121748thank god we still got the same bold lines aco ponyslopfuck off naimeltoid
>>8121753kinomoto sakura, from an old anime
>Haven't sloppa'd anything since February>Had to reformat my hdd that had my ai shit on it since>Don't remember how to re-install SD and Forge anymoreIt's over, bros...
>>8121745
>>8121753>zoomers don't know sakurait's bleak
nai sissies wont embrace the blur they still consider it an insultthey just need to admit the model blurs all the pixels together then they will no longer need to soil their diapers every 15 minutes
Nochekaiser get paid 25$ per loras
>>8121759>Don't remember how to re-install SD and Forge anymoreYou also lost your ability to read?
rajeesh, you're so fucking pathetic
>>8121759depending on how much money you make per hour, it might be more cost effective to just get nai. its better than local anyway and learning local can take upwards of an hour plus so if you make more than 25 bucks an hour, its something to consider
>>8121685Gonna check how that works, but does lora matter where it is placed on the prompt? since it is a hijack into the model.>>8121706>>8121710Did it on another branch, and yes fuck it breaks a lot of thingshttps://github.com/Panchovix/stable-diffusion-webui-reForge/tree/main-forge-upstream
>>8121769Yes
>>8121771what a terrible argumentso if it takes you 4 hours to bake a lora and you make 25$ an hour you should pay somebody 100$ to bake the lora for you?
>>8121778you dont need loras with nai though
>>8121779yes you do
>>8121779yeah if nai doesn't know your character you're shit out of luckwe'll be baking dozens of loras per day for nai3 if it ever leaks
Is it normal that hands often suck with NAI? I haven't tried higher res yet since I'm still learning.
>>8121781huh? you can just vibe transfer any character you want, takes two seconds
I just realized something. The reason there hasn't been a western local anime/hentai model made is because westerns don't like anime the way asians do. Westerns seem to prefer 2.5d/3dpd or /co/. Even in these threads dedicated to hentai, some people still can't resist posting /aco/ or 3dpd.Our best local model for hentai is actually a fucking model about sex with cartoon horses. The only model that's western and anime/hentai focus is stuck behind a paywall. Why is the west like this? Why can't the west just release a proper anime model like animagine but not be a retard about it and include extensive /h/ content?
>>8121784i like how you immediately contradicted your own point and kept writing anywaysthe reason we don't have a good local anime model is that all the competent weebs got hired by nai already
>>8121784is the /aco/ in the room with us right now
>>8121713>>8121716I have that flandre saved from 1.5 months ago, it's a good gen. And your local gens are so old I can't even find them but I have them too. I'm talking about what you've been posting for a while now. Can you point to a single gen like that recently?..This >>8121732 isn't me, but I agree that personal ami is trash. And you are making me hate asanagi, how you yourself aren't sick yet is a mistery, but to each their own.What I don't understand is I have several hundred nai gens saved and they are all very different and nice each in their own way, it's a field of infinite possibilities, yet your gens dating back from february are just so fucking similar with few exceptions.Nice posts after these though. Hopefully you can keep THAT up.
>>8121784>Why can't the west just release a proper anime model like animagine but not be a retard about it and include extensive /h/ content?That would be ArtiWaifu/Neta2
>>8121784please quote some the posts from this thread that are aco/3d, it doesnt have to be all of them
>>8121790>euge>west
>>8121791NTA but>>8121744>>8121733>>8121672>>8121606>>8121601>>8121542>>8121522>>8121451>>8121282>>8121221>>8121191
>>8121799you werent supposed to include the nai ones AIEEEEE
hell yeah my image didnt have a (you) but now it does
>>8121805yeah because your image SUCKS
How good of a model would we have if Google decided to make a nsfw anime model?
>>8121121ignore him anon, love your yukari gens keep at it
>>8121810>corporation>nsfwHow did Meta's model turn out again?
>>8121810it would make naiv3 look like a joke but why even consider an impossibility like that
>>8121784i was going to answer you, but then i'll get another 3 day timeout, so figure it out yourself
>>8121812oh right, lmfaocurrent one has like -1 IQ, you ask for literally zelda and does a link insteadjeets must be loving it
i wish that kohya hires fix didn't randomly output black images on reforge so frequently
Complete new user but before starting this, I was wondering if there's anyway to find out the sort of model settings and such from pixiv posts or is it most likely the ppl who make it onto the top rankings have custom models? (How hard would it be to try to replicate or have similar if starting with no experience?)
>>8121822with nai? 3 clickswith pony? impossible
>>8121812Im obviously talking about an alternate universe where corporation dont care>>8121813I dont know, just thought about it randomly
>>8121826then why not think about an alternative universe where you can just have your own holodeck in your house
>>8121825I got in the top ranking a few time with pony but its hard since jap really like the nai looks
>>8121829What about an universe where I would send the best model to this universe on my computer, still waiting alternate me, hurry the fuck up
>>8121830yeah? link your top gen, coward
>>8121825Thanks, I guess I should start learning nai.
>>8121799only a couple of these look "aco" to me. the raita ones are fucking troll garbage but the artist is definitely not western
you guys are mixing aco and ponyslop together.aco ***can*** look good.pony can't.(that was a shitpost, but partly. autismmix base ugly style it wants to apply to every image is the worst offender. i've come back to base pony with tons of schizo negs)
>>8121850I think I agree about autismmix, I feel like im fighting the grey/removal of color of autism more than im fighting pony style
>>8121834Why would I lie
I have 16 models archived basically none of them can gen nipples with their default style.
>>8121799>Raita is /aco/God I shouldn't have come back to this thread.
>>8121855too scared for it to be judged by superior taste of /hdg/ dwellers?
>>8121860Yes, also I only like 3 of those and I have no idea how the other 2 are in here and I dont think you can get ranked with porn anyways
>>8121799So /aco/ is just "pictures I don't like" rather than anything concrete?
>>8121862seems to be anything with skin shine
>>8121859Right, you should have stayed in /aco/>>8121862Let's be honest, if there were less pictures that look like that, the thread would be much less painful to look at
>>8121862welcome to /hdg/
>>8121862no, /aco/ is autismmix with full string of score_9... and style loras on weight 0.2 if there are any.https://files.catbox.moe/b9ynsw.png
>>8121816How much clothing bleed do you get when prompting alternate outfits?
>>8121871>[[best quality]], [[amazing quality]], [[very aesthetic]],
>>8121874yeah? i just stole the prompt from this thread, fuck if i careplus 'best quality' actually does something with 4th tail extract.
Another noob question but are the chinese ai pixiv users using NAI or something else? The "good" ones on the pixiv rankings.
>>8121879they're using bot farms for views/likes.
i love how freetards Stockholm syndromed themselves into liking garbage western toonslop solely because that's al they're only relevant model is capable of. they actually have to bend over for astralite's horsecock because otherwise they'd still be using 1.5. a really sad state for localjeets.
Why is there so much shitposting in a porn general? Shouldn't you all be genning after hacking off so theoretically you should all just be mellowed out? Why does AI porn generation make everybody so bent out of shape? This is all so weird.
>>8121872about this much. if i get more specific with tokens it will usually go away to an extent but certain tokens like jacket are basically unusable because there's to much overlap with coats
>>8121883autism, why do you think the most popular local model used here is called like that
If a man should defeat an elf in honorable combat, is he not entitled* to keep that elf in his basement/sex dungeon?*This is not legal advice.
>>8121883it's the downfall of general threads and why anything on /vg/ immediately goes to shit.active threads attract secondary's and people who are entirely uninvolved/uninterested in the actual topic. It's just zoomers trying to treat an imageboard like a discord server.It also attracts some seriously fucked up schizophrenic behavior by certain individuals for no reason other than their personally seeking any type of attention wherever they can find it.
>>8121879mostly nai, but some use pony with sekrit loras they only share among chinks
>>8121880I was mostly curious if ppl had an idea of what model they're using. Some of the top 10/20 images look pretty good outside the constant use of rape/blacked.
>>8121891Thanks, I guess I'll have to pay to make those styles.
>>8121891The chink secrecy is pretty hilarious. They treat their prompts like they were secret spells.
>>8121897well they barely understand english so it may aswell be a magic spell for them
>>8121901that's part of why nips literally call prompts spells (呪文)
>>8119146>can't promptFeels bad bros
How does local handle "upside-down"?
>>8121908it doesn't
>wdV removed every image tagged as loli from the datasetlol, lmaoDOA if it even comes out
>>8121913yeah, zero reason why it should be mentioned at all. carry on
how does nai handle "not blurry"?
>>8121917it doesn't
>grey background in negs>removes autism's grey filter
how does naiblurmeltie's pants handle pounds of feces?
>>8121919yellow background is the real tech
>bad in negs>autismmix makes images better than nai ever could
>>8121922Yeah that does something too, as does blue background or using all 3 together
>>8121917>>8121918>1 minute 2 seconds apart exactlyTalking to yourself schizo?
>>8121909Pony can do upside down prompts but if you get too fancy it'll start deforming a shitload. Here's a real quick gen.
>>8121924>>bad in negs>but not putting (((good))) in main prompt to unlock local's true power
>>8121922>people remembered my yellow background negby the way do not put white background in neg, it make things more yellow somehow
following blora, is there any way I can extract certain blocks from loras, and use them only?
>>8121930I think the complaint was that it doesn't do the hair hanging down properly.
>>8121943it doesn't do anything good related to the head, actually. it's the worst model ive ever seen that handles "upside-down" so fucking bad. in fact i haven't seen a model as trash as this one.
Lazy elves, always just hanging around...
>>8121937that's what blora is, but also it doesn't really work as far as i can tell
>>8121913what a fucking piece of garbagethat dogshit will be doa and i'm very happy about itbut sadly it only leaves us with jordach as the only option left, which is not very good at all we're fuckedmy god just leak v3 already
>>8121958why did you reply to a bait
How do you sharpen an image without upscaling it in Automatic1111? I've had some pretty poor results when upscaling too, even when using Ultimate SD Upscaler
>>8121961ironic.
>>8121961how is that a bait?
>>8121962Have you tried hires fix or multidiffusion at low denoise?
>>8121961It's a serious concern about model's capabilities of generating something else than shine skin black lips curvy /aco/ stuff. We already have pony that can gen in, and we don't need another model that does it.
>>8121937https://files.catbox.moe/f8jn42.pyadapt as required
>>8121961>Removing loli is bait
testo testo
>>8121975wider
>accidentally baked at 768x768>no real difference in quality
>>8121972Thx, gonna try it>>8121954Oh yeah it's just that I want to remove all IN blocks from lora on civitai I use for characters
>>8121981i guess that might actually be ok since in blocks do nothing
image post
>>8121984they dont?
>>8121961i'm the person who posted that, but how is it bait? if you don't believe me you can check the 4+4chan thread where a WD member talks about it. or do you mean that you're okay with them not using hundreds of thousands of high quality data because of bullshit moral concerns?
without b-, what does cascade really have left? to be honest i liked their little gang of misfits. tree the voice of reason, the fat tiger guy who generates baraslop and just says 'its cookin' every epoch, manni the actual retard, bluvoll desperately trying to slop together lora settings to save the model, and finally b- the loyal cheerleader who stuck by jordach's side... until now.one by one they're all drifting away, now only echoes remain in these empty halls. murmurs of what once was, and what could've been
>>8121989IN08 does do something, sometimes.
>>8121989for characters there's effectively no difference at all
>>8121993you say "for characters", but what about for styles?doesn't it help removing style bleed from character loras?
>>8121991It is in moments like these that we are reminded of the impermanence of all dharmas, and of the truth that the prosperous must decline.
>>8121995not a huge difference for styles. maybe for concepts but i'm not sure
Need help with a prompt. Want to get her hanging straight down by her legs (legs straight). Can't seem to get it, it really likes to have her bent over. One of those things where it's not obvious what you can even prompt to get it (legs together... but the legs are together in this picture, just not in the way I want).
>>8121995IN08 (layout?) controls the shape of things to some extent. For example narrow eyes. Useful depending on whether specific shapes are important. Gen with and without that block in an XYZ plot across a few varied gens and check for your specific use case.
>>8121984what about the out blocks, is only training the 00 and 01 blocks a meme then
>>8121990Why is it always moralfagging and grandstanding? first it was a schism because pony doesn't want to train on artist tags and is removing even more for v7, and now the two alternative projects are still self-sabotaged crap. jordach's "it all goes in" approach for not pre-pruning is good, but his spitebaking attitude and inability to see the slop leads the model into the dumpster. neggles has better technicals and resources with the ability to notice when it looks like shit, but is incapable of actually doing the needful and properly baking a prototype for even 5 epochs.
>>8121990Can't believe you did made me go there just to checkSo in the end it's just danbooru gold pruned, something we all already knewStill, the model has a high chance of being bad at doing nsfw loli tho
>>8121999Nice trips. For starters you have knees yet ass. How is that supposed to be if it's from behind? Do knee pits. Also be mindful of the resolution. Try full body instead so it isn't trying to squish her.
>>8122006that does seem to work to a degree but there's not really a distinction for what goes in what block wrt content/style. i tried training characters and styles and concepts with b-lora and each time there was enough conceptual and style info spread between both blocks that splitting them produced pretty bad results
>>8122007You forgot to mention how neggles has no taste and no sense of aesthetic since xer already said xer goal is to achieve a model that basically looks like aom2
>>8121999>gen 1girl, standing>turn 360 degrees>put upside down image into controlnetThere. Infinite reproducibility.
>>8122010Yeah knees were in there when I was experimenting with that. I've already tried the things instead and tried full body. The problem is, all of these things are also true of the bent over version. Her full body is in when it squishes in... The obvious thing is to put all the bent over stuff I can in negs, but the tags for what I want and what I don't want don't precisely exist.
>>8122015Try flying or floating
Getting closer, putting top-down bottom-up in negatives helps avoid that specific pose.
What the fuck is civitai doing? Surely they're paying out the ass for the amount of slop thats being spammed on the site. On-site lora training, no real limit to images/etc users can upload, on-site generations.. how is any of this sustainable? Also, I fucking hate that they added support for on site lora training. It makes it so difficult to search for loras when so many of them are pure slop. There needs to be some form of quality control, but I guess the site will be dead within the next few months anyway at this rate.
>>8122008>gold content was removed (= anything that can potentially have "loli" tag)>but anything that didn't have the loli tag was kept in (= anything WITHOUT a "loli" tag is in dataset)>the model will definetely know what a loli is at launch (???)if there's only danbooru dataset = there is NO loli content in model at all given those facts, and the only possible way to prompt it is to use artist styles or cunny characters that weren't tagged as cunny (if they won't be hashed!)
>>8122008>>8122019i was typing the same thing>was able to get the tag in>yet all images containing the tag are pruned?you can't even see the tag without gold, just like gelbooru you need to manually enable it. so how is the tag still there if the data isn't?
So what's the good hentai model that's second to NAI3 then
>>8122018Look at the bottom right when you see a lora, if it says "Created on Civitai" then you can ignore it 99% of the time. I wish there was a way to filter loras trained on civitai but I don't think there is
>>8122022>hentaiany ponymix with shitton of style loras, dropped score tags, "detailed nose, milf" and other /aco/ shit in negatives
>>8122022EF+HFF
>>8122019Neggles is a middle aged man with AGP larping as a cat girl in discord. They will never actually deliver anything that's better than pony.
>>8121637I reported it for loli and it got deleted for that. I just love exposing the absurdity of moderation here and will keep doing that.Enjoy your fat hag cuck thread.
>>8122024Yeah but you cant filter it like you said., so painful.
Anyone knows the original artist of this style
>>8122027people abandoned this far too quick, it was a lot of fun. it had a red bias but i liked it for the time i used it
>jordach after an epoch: "wow the model is improving">just looks slightly less melted but still underbaked>some random on his dicksword asks why it still looks melted>he answers "it's starting to focus, slowly">continues baking>every anon here: wow what an idiot he doesnt recognizes slop when he sees it, and he thinks this crap will kill pony lol lmaoi mean, if all of those statements were true, he would've just stopped baking at e5 and say "here you go guys, the finished pony killer, model is perfect and doesn't needs to be trained further"there is a reason why he hasn't stopped baking when you look at the sample of epoch 1 of your lora bake do you actually say "this is pure slop, it will never improve no matter what" and stop the bake because it looks underbaked, and never try to bake it again?
>>8122017The final piece of the puzzle was to neg "ass."But victory came at a price...For now I am cursed with assless elves.
>>8122036who are you quoting?
>>8122028>never actually deliver anything that's better than pony.i'll ditch pony if they'll manage to deliver SFW weeb aesthetic with decent prompt adherence, style variety and character knowledge>>8122036bakers usually dedupe, reduce/remove JPEG artifacts and properly tag their dataset before bakeryjordach didn't do that
>>8122036how about we agree to just stop bringing it up until there's something worth looking atit's pretty clear it's not worth anything at this point
>>8122036anon, let's not talk about cascade right nowwe're all mourning b-sama's passing here...
>>8122039list of quotes would be too large lmfao
>>8122040Dataset work is what carries pony despite all the other bullshit that happened. Not curating the dataset is genuinely retarded and just makes baking a waste of time, which some of us said months ago. But it is what it is, you cannot be unburdened by what has been.
>>8122038inpaint the ass
>>8122040>bakers usually dedupe, reduce/remove JPEG artifacts and properly tag their dataset before bakeryastrakike did tagged the dataset but decided to transform the entire dataset to jpeg 2 fucking times before starting to bake, ah yes, very good, amazing >>8122041>waaah neggles will never deliver anything not even a prototype>jordach: delivers both a prototype and early epochs of the actual bake>waaahh its shit???????????????
>>8122047who are you quoting?
>>8122047>did taggedesl-kun...
>people laughing at astracookie not knowing how to bake a lora>meanwhile jordach doesn't even curate his shitDid you forget the main ML rule? Garbage in - garbage out.
>>8122049Poor anonie can't answer properly but only to use ad hominems now, you did well this time esl bro!
>>8122051Didn't he also decide to add tens of thousands of synthetic images that are quite literally nai v1 slop?
if pony is so bad why is it the best local option after all this timecheckmate
>>8122055yes, and some niji slop too
>>8122055Yeah, same as the animagine tard.
>>8122051And yet it has no compression or tagging issues - it may be one rule, but not the norm if there is other content that fits better. Companies like Google will just in-place analyse everything about your ad profile even if it's pure noise.>>8122055Less than a thousand with *manual curation effort*. Took me several days to pipeline this and setup manual review. And frankly, you weren't calling the NAIv1 merge-fest slop a year ago so don't even start.>>8122057The only thing it learned was the aesthetic lmfao. It discarded everything else learning wise as there is better data internally from Gelbooru for that.
>>8122046Trying
>>8122059>And frankly, you weren't calling the NAIv1 merge-fest slop a year ago so don't even start.I was
>>8122059Good morning sir
>>8122061Someone with taste around here, based.>>8122062KinoBut no in all seriousness, finetunes over a specific image count fall into "big data" category of analytics where /everything/ helps build the predictive model. It's why and how LLMs work, curation isn't at all needed over a certain data size where it becomes a few droplets of piss into a fucking ocean. Not a real concern.
>>8122066i can't think of any other big finetune that doesn't use quality tags at all
>>8122066that makes sense if you were making a model from scratch, but you are finetuning and people are going to want to use it as is
>>8122069It still applies in general with regards to predictive services, what's a handful (like 10) fucked up images going to do to an average weighting of 6.13 million images? Not a lot, if not, nothing. It's why my concerns are with the models themselves reaching a plateau where improvements aren't exactly easily achievable without further brute force. I get the concerns, but things that work for LoRAs do not apply to finetunes of a certain size, and outright base model training.
>Someone with taste around here, based.hearing this from a furry doing nothing but piss dog pinups in 10 most popular /aco/ styles to """ensure""" that his shit burns right is fucking hilarious
>>8122066Stop arguing with these idiots they don't even have a 4090 to try to make a finetune with more than 10k pics and yet they think of themselves they know so much about finetuning a model right after baking 10 LoRAs with 200 pics at most Just go to sleep and keep baking, king
One more epoch bros
>>8122060What once was lost, is now found.
>>8122072b-sama's position is open now, i'm just saying...
>>8122070yeah i didnt mean it in the sense that the synthetic would particularly fuck the model, but i dont think the millions of low quality images on boorus will particularly improve the quality
>>8122071rudebeing a furry doesnt necessarily means you dont have eyes to recognize something that looks good>doing nothing but piss dog pinups in 10 most popular /aco/ styles to """ensure""" that his shit burns right tell me youve never done and tested one single model finetune without saying you didnt:
>>8122077do you still use the tag "by novelai" when testing? i think that sets a bad precedent
>>8122075Astrakike is accepting donations, I'm just saying...
>>8122078you're replying to a schizo troll who is ironically pretending to defend cascade, not to jordachand no, people have mostly stopped using the novelai tag since a while ago
>>8122075Just ask luckypedoposter to both post trooncord screenshots and cheer Jordach for good bakery.>>8122077>being a furry doesnt necessarily means you dont have eyes to recognize something that looks goodThen I wonder why good chunk of your kind is still stuck with SD1.5 and (You) were baking fluff tunes while dismissing SDXL even after pony release.
>>8122078im not jordachbut testing a tag of a synthetic dataset to make sure its behaving correctly is important specially since that one dataset is dozens of times bigger than any artist inside the cascades dataset but if u don't want 1.5sloppa-looking gens just dont prompt for it or put it on negs its very simple>>8122084retard>>8122087>Then I wonder why good chunk of your kind is still stuck with SD1.5 and (You) were baking fluff tunes while dismissing SDXL even after pony release.im not jordach and not a furry either
Why do people in this general get so invested in this stupid shit? Sit down and wait for the next model that comes out of nowhere.
>>8122091that is my very point thanks for using your braincells this timego post some gens or go to sleep or go gaming whatever
>>8122089post your best cascade gen if you're seriously defending it and not just trolling
1526 / 200based based based
>>8122095who told you i'm defending cascade i just got fed up of you all troll idiots who talks big shit about any model in general (includes wdv and any other currently-in-the-oven model too) but the most theyve done is bake muh lora of 10 pics then think to themselves "im a finetune expert" and actually believe itall this drama gets me on the nerve and i cant keep pretending it doesnt
Rape.
>>8122100Who has claimed to be a finetune expert? You don't have to be an expert to use your eyes.
>>8122100nobody here has ever claimed to be a finetune expert and if it was necessary to be one to talk about how currently baking models are progressing then basically nobody could do thatbut you're either baiting or mentally retarded so i'll stop replying
>>8122097not gonna to reach 3k but somewhere near 2.5k that's for sure!
the absolute desperation of localtards, so invested in failbake jeetware they just can't look away. NAI doesn't even know who jordach, euge, etc even are. you're all so starved for content that you actually care about the day-to-day progression of a model training in some guy's basement like it's some professional project.
>post gen>"LOL THIS SUCKS FUCK YOU">....>"WHY IS NO ONE POSTING GENS ANYMORE?!?!?"
>8122110What a lame bait. Shame.
>>8122087People/furries are stuck on it because merge culture is arguably a very real cancer to those who don't know anything (see CivitAI) and just copy and paste what they see, which is also fairly true for some poeple in here. They're not interested in the finer culture, or partaking in the parental culture, just tourists faffing about with shit they don't exactly know in great detail.Or in other words, getting normies or a wider community to adopt a new standard is nigh-fucking-impossible and requires some extra benefits thrown their way (that does not include technical ones), because otherwise they just return to what little they actually know instead of trying something new, again, this equally applies to /h/ as it does for furries.>>8122091>just wait for the next basedslop model that has training requirements that make it impossible even on 3090/4090s>>8122100It's a right pain in the ass and a massive time commitment to make usable base models for things that are niche relative to the original model. I have a theory that most of the issues with CLIP based models are due to CLIP using 2018 era datasets (ie, it's shit) causing most of the issues in SD, SDXL and Cascade.>>8122104>>8122105The flipside to this is that I'm learning how it's working things out much like you are. If I can't understand wtf is going on then neither will you. That said, most problems can be fixed with hi-res fix and a bit of [soft] inpainting, I have zero idea why people want to be spoonfed so much. Future runs will benefit from this run's understanding of things.
was it this bad during the sweltering hot summer of 1.5? i feel like people are craving harder now than they were back then. is it NAI? is it dall-e? maybe there was a blissful ignorance in the 1.5 days, when people didn't realize how good things could actually be. now they only want more
>>8122115The hedonic treadmill, also known as hedonic adaptation, is the observed tendency of humans to quickly return to a relatively stable level of happiness despite major positive or negative events or life changes.
>>8122115We literally have models capable of NAI v3 level output, they just need more wrangling to get good results so no one uses them.Instead people just stick to their one-and-done ponyslop generators.The reason NAI is popular is because it's easy to use and gives good results.The reason pony is popular is because it's easy to use and gives mediocre results.The reason any newer model isn't popular is because they're hard to use but give good results.
>>8122118hello Arti shill
>>8122111This thread is literally this 2011 meme. About. Every. Single. Thing. Posted.
>>8122120Arti/neta are good, but not for NSFW. I'm talking about animeta and cartograffer v4
>>8122116except it's not, nothing is returning to a 'stable level' as evident by the post-to-image ratio. every day it's the same arguments about nai vs pony, and then the inevitable 'when will jordach save us' which then leads to the inevitable 'cascade and wdv and dead, local is done for' and the cycle repeats. it's quite evident that this general no longer cares for actually generating images with current models, be in pony or nai. and its not like everyone fled from the spam, because the alternatives are all dead too.
>>8122111Trips of truth.
>>8122115It's like playing an mmo for 10+ hours a day for months, eventually you'll burn out and start hating everything but cant bring yourself to stop.
>>8122125
>>8122124>animeta and cartograffer v4link to these? i cant find anything
>>8122130Here you gohttps://civitai.com/models/282061
ok who wants to donate to me to save local? i'll bake on my 3070 an if i get enough donations i'll upgrade to a single 3090.
>>8122045>you cannot be unburdened by what has been.kek this is really making the rounds eh?
>>8122135Kamala will save local
>>8122114>I have a theory that most of the issues with CLIP based models are due to CLIP using 2018 era datasets (ie, it's shit) causing most of the issues in SD, SDXL and Cascade.Wouldn't surprise me tbdesu
>>8122137>/hdg/ on one of the screens bas- i mean unbased
>>8122127No the reason why I hated the MMO I was playing is because the game is actual shit. It's the I was the one trying to make the game fun, not the game being fun and pulling me in. This happens usually because the game has something that really interested me like aesthetics or character design or combat etc., but in the end I gave up because I no longer wanted to polish the turd. It's not that burnout made me stop enjoying something good, but that it was never good and the honeymoon phase has ended.
>>8122012I went through neggles' posts and yeah he posts literal 1girl standing with exploded background all the time. This shit is so boring.
>I can't get hard unless there's atleast 1guy in the image
>>8122138i hate clip so much i just want to be able to do something like (girl: black shirt) (boy: green shirt) but you can only do that sort of thing with a separate clunky addon
Why must SD be so shit at handling more than ONE subject in the prompt?. Dall-e architecture when?
>>8122145Dall-e isn't that perfect at it either. It's just a fundamental flaw of prompt-based inference and the lack of segmentation and token relationships
>>8122145SDXL-based models aren't that bad at it though?
new wd tagger ""just"" droppedhttps://huggingface.co/SmilingWolf/wd-vit-large-tagger-v3
>>8119863i really wish someone would actually try a hunyuan finetune. it's way better at anime as a base model than anything else available right now.
I've been on a vacation for a month, have I missed anything?
>>8122152you missed everythingbut at the end of it allyou missed nothing
>>8122152You missed a whole three seasons of trooncord drama, we're working on a cast change now. If we get lucky next thread will not have a single image post unless its a screenshot of the new season.
>>8121990>>8122007>>8122008>>8122020>>8122019this is literally fake, he even fabricated the screenshot there in order to le bait you guys
>>8122027This but unironically
>>8122156your pic is just pr from his sideI don't actually think that neggles is pruning loli because xe's a moralfagging agp troon, it really more likely that zhe's afraid of some legal csam bullshit, xe said xhermselves
>>8120977No, for some archaic reason onetrainer has a 75 token limit
>>8122155>>8122153Damn, sounds lame and gay.
>>8122156so they're keeping it in, but actually not keeping it in, but keeping it in...so they didnt bother with gold images because theyre nono and garbage but decided to keep everything the autotagger detects because somehow thats not nono and garbage? but theyre renaming these 'child' tags despite there not being any in the dataset to begin with? and half the 200k images are trash and wouldnt pass the aesthetic check, yet somehow 12m images (more than the entirety of gelbooru) all pass and make it in?maybe this model just doesnt exist to begin with as nothing adds up. whatever, i wont pester too hard because we'd be lucky to get anything at this rate. just tell him to shart out an epoch to play with so we can go back to genning.
>>8122167>prunning shota/loliAnd another one bites the dust....
>>8119458What style is this?
spoken banana
tenth rebake of these, but fewer of them
>>8122185Wait till she finds out it's a plantain.
artistrizdraws https://files.catbox.moe/5lub3p.safetensorssuoiresnu https://files.catbox.moe/ysmvoi.safetensorscharacteramane misa (dn) https://files.catbox.moe/1e24kx.safetensorskaguya (kuroinu) https://files.catbox.moe/7bvg07.safetensorsl (dn) https://files.catbox.moe/wgppkt.safetensorsliru (wgwy) https://files.catbox.moe/npy7r2.safetensorsolga (kuroinu) https://files.catbox.moe/wphlen.safetensorsprincess shokora (warioland) https://files.catbox.moe/mxze95.safetensorsreiko (zone) https://files.catbox.moe/czcn8f.safetensorstezuka rin (katawa shoujo) https://files.catbox.moe/ng9ecv.safetensorszone-tan (zone) https://files.catbox.moe/ydqylo.safetensorsconceptcunnilingus gesture https://files.catbox.moe/jds6j8.safetensorsperky breasts https://files.catbox.moe/q0xvom.safetensors
>>8121753wtf
>still posting futacock
>>8121782Adding hand pose tags often helps. Enhance can also help, generally things that make them bigger.
>>8121593>Lmfao the schizo naihater was right, SMEA is Indeed blurryFor nai lowres gens, for like 90% of artist mixes it really is, honestly. A lot of NAI gens are blurry and washed out as fuck because people feel compelled to use smea for fucking 1girl posts. I think doing an img2img pass after the smea gen helps though, but most nai genners I've seen in other places don't care to do that.
>>8122208>but most nai genners I've seen in other places don't care to do that.because upscaling costs anals, they're reserving them for proper gens
>>8122177aaaa, found it
>>8122212I know, you can do an img2img pass without upscaling. You can also tile upscale without anal cost.And you don't even have to use smea, it's not always good, really.
smea should be a slider not a sampler
>>8121999Make the resolution really tall and thin. Either prompt in a way that doesn't show her ass or doesn't show her face. Showing both isn't possible unless you do from side or something.
>>8122008>keeps shota postsWhat a faggot.
>>8122220Oh?
>>8122118>We literally have models capable of NAI v3 level output, they just need more wrangling to get good results so no one uses them.>Instead people just stick to their one-and-done ponyslop generators.That's one thing some people don't understand. You can get the same results with local as with NAI, you just have to spend a lot more effort, inpaint, apply color correction, stack loras. On NAI you type in your basic prompt, tweak it over 10-20 gens and you're done, unless you're a perfectionist who needs to fix every small detail.
christ almighty
really why bother with anything past sd 1.4? you can just inpaint it all anyway, it's all just pixels in the end
it might sound crazy but i would be more inclined to post my local gens if there was a local only threadi could pay for nai but i have more fun with local and i seeing nai mentioned every 5 posts is boringnot gonna happen so w/e
>>8122235You're fucking delusional if a separate thread would stop all the shitposting at this point. I'm not even sure that the 3 nai genners that still bother to come here even engage in this shit.
>>8122235Just become a furfag and you'll get a full local thread with 1:2 image-post ratio
>>8122237This. Furry threads are always very cozy, no paypigs, no trolls, no kink shaming, just anons posting hot stuff and discussing tech.
I haven't touched Stable Diffusion in over an year, do I just update the automatic UI and download some new checkpoint and loras or is there something else I should be moving to and just nuking my old stuff?
>>8122242Go back to sleep, come back in another year
>>8122240>immediate meltie>furfag image in pocketlol
it only just hit me that b- is actually gone... does wdv have a fun cast of characters? it seems kind of boring and closed off in comparison
>>8122244Who are thou quoting?
As an outsider looking in, the amount of autism this thread has is abnormally higher than normal.
>>8122251yer mother hung herself
>>8122235Seems like it's the reason why there is /lmg/ for local LLMs and /aicg/ for SAAS stuff.
>>8122252This one is worse than usual
>>8122252blud pretending he doesn't post here every secondat least change ur writing style
>>8122255We had a couple NAI only threads, but they melted down because half of the NAI posters wanted blacked and the other half wanted loli
>>8122255There is a local-only diffusion thread on /g/, you're free to post there, including training shit (off-topic on /h/).
>>8121887>>8121908>>8121931>>8122074Hot!
>>8122257>blud
>>8122262no cap on a stack foo
>>8122258Thing is, all of them were created by you with explicit purpose to shitpost and troll, just like now.
>>8122265>Acknowledging who 99% of NAI users are>TrollingYeah sure
>>8122256no it isn'tit's always garbage
>>8122235This is a fully local thread broThe "NAI" shitposters were exposed as mindbroken localbros twenty times nowThere are two genuine NAIposters here
>>8122268I honestly feel nai wouldn't be brought up as much if there were 2 threads
>>8122272Of course, you bring it up every time, shitting up the thread with melt and blur shitposting with the obvious fucking purpose to make a split thread.Not happening bro ;)
>>8122235Just make images better than NAI? It's very simple, are you afraid of healthy competition or something?
>>8122272nai wouldn't be brought up as much if this general wasn't full of broken oldfags that were forced to gen 1girl standing for 1.5 yearsit's the same thing that happened with /aids/again who the fuck would post in a nai thread when there are two nai posters here
>>8122272i'll be sure to bring it up every time just for you
Damn, even US hours were pure seethe this time around
>>8122208some gens are blurry and washed out because of a bad combination of artist, smea, rescale and sampler, higher rescale with smea often makes it look really washed out>compelled to use smea for fucking 1girl postsdoesn't matter, like I said it's style/other settings>I think doing an img2img pass after the smea gen helpsyeah it helps making an image blurry and if you do that on a mix that plays really well with smea it'll completely destroy the detailtheir img2img is just weird, img2img on local sdxl and 1.5 ends up sharpening the image and adding detail to it if you're upscaling, even with relatively low denoise, nai softens the image and makes it somewhat blurry and requires much higher strength for detail, adding extra noise helps but that can also depend on the artists and other settings
>>8122281>even US hourslol
I'm genning some images that I'm almost certain will piss somebody off.
>>8122282>nai softens the image and makes it somewhat blurry and requires much higher strength for detailThat is also true for local, just not to this extent. Higher denoise = sharper details. Also the noisier the upscaler the better, you don't want a "good-looking" upscaler that makes it buttery smooth before doing an i2i pass, and you certainly don't want to skip an upscaler for the same reason.
>>8122284Daring today, are we?
>>8122292I might even add a penis into the image depending on how I'm feeling.
>>8122289https://files.catbox.moe/whnohv.pnghttps://files.catbox.moe/o26s26.pngFirst one is with siax, second without upscaling.No real difference in sharpness or detail at .4 denoise
>>8121887>>8121713i fked up, box?
https://files.catbox.moe/vmxqsj.png
>>8122305based
>>8122196small_penissisters.. is this a personal attack?
>>8122294You can absolutely see the difference even on your pic.0.4 denoise dpm sde++AnimeSharp: https://files.catbox.moe/0olpo6.pnglanczos (no upscaler): https://files.catbox.moe/1cw7y6.png
>AI CAN'T DRAW HANDS!!>https://gelbooru.com/index.php?page=post&s=view&id=9378199
>>8122313There is a difference, but not in objective sharpness or level of detail. Upscalers change the colors that end up in the denoised image, that's all you're seeing.
>prompt very small penis>it's the average real sizeRealbros...?
>>8122317>that>averageanon..
>>8122318>The average penis is between 5.1 and 5.5 inches>even above averageWhy do hentaichads mog us so bad...
>>8122319They're so used to arm length cocks and loli proportions, their sense of scale is absolutely fried.
>>8122319ugly_man can't keep getting away with it
>>8122319Huge cocks are just more aesthetically pleasing to look at
>>8122325Fruity
speaking of juicy cocksi was thinking of making that forearm sizing lora but 99% of these images are blacked 3d renders for some reason
>>8122316Bruh..... If you can't see the blurry lines without upscaler (hair, hello?) I don't know what to tell you. Look at the guys' thumbs, it's most visible there? I have no idea how you don't see it in every single detail in that pic though. Are you on some ancient cheap display?
is there any way to avoid anime screencap style bleed when training character loras? there's not really fanart I can use
>>8122330uhhhhh, try setting some dropouts when training
>>8122330train style lora via screencapsmerge with base ponytrain character lora on mergeit sounds good on paper but i haven't seen anyone pulling this off
>>8122330also that one differential lora thingie but i haven't done it