Your favorite streamer EditionPrevious Thread: >>>>8786829>LOCAL UIreForge: https://github.com/Panchovix/stable-diffusion-webui-reForge | Stable commit: 6964cebComfy: https://github.com/comfyanonymous/ComfyUI | https://comfyanonymous.github.io/ComfyUI_examples>NOOBAI-XLhttps://civitai.com/models/833294/noobai-xl-nai-xlhttps://huggingface.co/Laxhar>NOOBAI SHITMIXES102d custom: https://civitai.com/models/1201815?modelVersionId=1491533291h: https://civitai.com/models/1301670/291h>RESOURCES4chanX Catbox/NAI prompt userscript: https://rentry.org/hdgcbEasy booru tag export userscript: https://github.com/Takenoko3333/Danbooru-Tags-Sort-ExporterTags: https://danbooru.donmai.us/wiki_pages/tag_groups | https://danbooru.donmai.us/related_tag | https://tagexplorer.github.io/#/Inpaint: https://files.catbox.moe/fbzsxb.jpg | https://huggingface.co/Wenaka/NoobAI_XL_Inpainting_ControlNet_FullUpscalers: https://openmodeldb.infoBooru: https://aibooru.onlineControlNet: https://rentry.org/dummycontrolnet (OLD) | https://civitai.com/models/136070 (look at the green links in the description)>TRAININGGuide: https://rentry.org/yahbgtr (WIP)Anon's scripts: https://mega.nz/folder/VxYFhAYb#FQZn8iz_SxWV3x1BBaJGbwTrainers: https://github.com/derrian-distro/LoRA_Easy_Training_Scripts | https://github.com/bmaltais/kohya_ss | https://github.com/Nerogar/OneTrainerOP Template/Logo: https://rentry.org/hggop/edit | https://files.catbox.moe/om5a99.png
should i sub to nai until we get something new?
>>8795001based op
>>8795004Yes! Its only 25 dollars after all
F for Kanata graduation
>watching vtranny shitgrim...
And that's for now, I have this time new things to post so there is that
what's good local model rn?
>>8795017base noob vpred with loras102d custom291hr3mixall shitmixes are more or less the same only difference is the slightly style bias but any of those are good enough for everything
>>8795017WAI the rest are all outdated
>>8795017on nai's server
>>8795020>>8795021shut the fuck up and go to your /hdg/
It's antisplitschizo
where is noodlekek?this thread was less hostile when he was in charge...
>>8795022she's not wrong though
>>8795014Why isn't it a queen of hearts? Karin is BWC only
>>8795026This is a nice opportunity to the fluxchads (lol) to flex on us and do that
>>8795027Qwen sucks at anime nsfw so the pasties look overlayed. Qwen edit animetune would be a dream come true
>>8795037That's a little better than I was expecting, solid base it just need a little of shading and the nipple outline I may do it myself later
>>8795038I'm sure you could get a better result but qwen is excruciatingly slow on neo and I can't get the Turbo lora to work, so I didn't feel like experimenting
>>8795044could just inpaint the pasties with noob in 20 seconds, instead of rerolling with the unreasonably big model
Thanks for the controlnet help
>>8795052wrong thread but you're welcome
>>8795050I do it for the tech
>>8795010Another one bites the dust.>>8795014That's a real nice Karin.
Still feel regret for all those shiny skin gens I did the other day
>>8795141dick-o-meters are nice. shame that all the loras sucks and it has to be inpainted every time
>>8795141Used condoms are disgusting.
>>8795146You should rather ask the question why there are 20 tally marks but only 4 condoms
>>8795141>4 condoms>4 loss runeskek!
Why never none of you notice the flood fill remains or the little inpaint halos here and there, that's what piss me off the most
>>8795161why is she so uninterested while looking at (me)
>>8795157Again. You are the only one begin too schizo about it.
>>8795163she can't believe he's charging money to let (you) watch
>>8795181I am a schizo yes but maybe chat should also put his shit togetherAnyway, we love vtumors here (I haven't really watch anything beyond YT clips)
I remember an anon saying that there is not enough POV material out there and I found this the other day>https://exhentai.org/g/3682722/63636d63eb/
>>8795187Nigga I look at gens on a 10 year old 1920*1080 monitor. They look fine to me. Really can't expect me to zooms across every image for 5 minutes to then nitpick about details.
do more graphically intensive 3d games have dlss/fsr? whats the availability of fsr? im picking up a new gpu and im thinking of swapping to amd but im just concerned with fsr not being available
>>8795245unreal 5 will look like shit no matter what
>>8795245>do more graphically intensive 3d games have dlss/fsr?yes almost every new AAA game has those
>>8795163You know what you were doing when you sat in the corner chair, anon.
>>8795157You got a point, Yoda, but as >>8795200 said, I only recently got a 4k oled so I now do see everything and make sure to try to low denoise blend fixes but prior to that I couldn't see nor cared enough.
>>8795245kek. Anon, no devs can code or optimize these days so EVERY game has those as a band-aid for performance.
>>8795262also im not getting paid to make this slop so why does anyone cares about seams or inpaint halos like wtf lol
mass sloppers get out
>>8795265Fair but I'd even go as far as 90% of sloppers don't even do that, anon. When some of the niggas here start posting rando pixivslop asking for artist mixes, a lot of that shit doesn't even have the courtesy of fixing fingers/toes/eyes.
>spend 10 minutes fixing gen>someone looks at it for 1.5 second and thinks "yeah thats alright"
>>8795253i meant for h games. ik every triple a game has at least one of the options but i wanna know if h games generally have fsr or dlss more often or if they have both or neither
Found this old HP server/workstation on facebook marketplace for around $400 bucks. The specs are as followsIntel Xeon E-2276G (6 cores 12 threads)64GB RAM DDR41TB SSD M21TB HDDNVIDIA QUADRO RTX 4000 8GB GDDR6Is this good enough for generating locally? I'm on the verge of getting it but couldn't find any benchmarks with this card. I like the 64 gb ram but the 8gb vram worries me. I've seen a lot of talk on zimage on other sites wonder how good it will be for hentai.I will not use this rig for gaming, just static image generation.
>>8795268Get popular on social media, then hundreds of thousands of people will look at your gen for 1.5 seconds. They will spend much more time looking at it in total than you spent fixing it.
>>8795342>Is this good enough for generating locally?Uhh not really, you'll need to slap a decent graphics card on it (lowest you can get for a ""decent"" experience it's a 3060)
>>8795342you already asked in the other thread.and the answer doesn't really change. RTX 4000 with only 8GB VRAM is, well, it's serviceable on SDXL but it's going to be shit for anything heavier. z-image requires basically 16GB VRAM.
>>8795351>Get popular on social mediaI have social anxiety
>>8795361imagine having adhd and consequentially rejection sensitive dysphoria, social anxiety and performance anxiety.certainly couldn't possibly be me.
You know chat, at this point I wouldn't even doubt most of you have a legit mental illness
>>8795365>most of you¿does that include (You)?
>>8795369Since I choose to keep interacting with you, I suppose I must have some form of mental illness myself as well yes
>>8795365I only have adhd and anxiety related things(anxiety as in aversion, not the break down and make a scene or hyperventilate type stuff), which stem entirely from being a kid with adhd and getting yelled at/excluded as a result. It's technically a mental illness but I consider having fucked dopamine receptors pretty mild and inconsequential outside of the introversion and how damaging it is to employment opportunities.
>>8795374yay~
Fuck fae folk in the forest.
>>8795365Is it really surprising in a 4chan thread dedicated to generating computer-hallucinated anime porn?
What's the best ui for tagging datasets?
novel ai
>>8795622>let me just take this anime girl to the most scenic and uncomfortable location before I rape her
>this anime girl goes to the most scenic and uncomfortable location before she takes a huge shit
I'm so close
bbc...
It's a good thing we have two threads, this way I get to read the same shitposts twice.
>>8795657You could just overcome your fomo and stay in /hdg/
>>8795353>>8795354Ah I see, I was under the impression that newest models were able to work with lower vram plus I didn't realize just how old the quadro 4000 is. Thanks for responding, I would've wasted my money otherwise.
>>8795157who
Oh nice, a Ame-chan thread. Just in time for a return.
>>8795806I can't remember the last frontal thighjob gen that didn't have the dick in some completely implausible position in relation to the dudes hips
>>8795809And isn't this one too low?
Feeling lazy, but might cook up a batch later.
From what I've gathered looking around, decent video generation is too impractical at the moment.Have you guys had any experience with it? What sites did you try?I made this one with an image and a simple prompt with pointers, but I'm wondering if I should use a more descriptive prompt instead.This is probably more related to the other thread, but that one's absolutely rotten to the core.
>>8795819It is
>>8795843>decent video generation is too impracticaldepends on what you deem decent. it's pretty easy to run wan locally, but all vid models are primarily trained on 3d stuff and it shows
>>8795853Easy to run is one thing, but it's apparently very hardware heavy. Not exactly worth it for the results.I generated a few more videos here and there using the same reference, but only that one and one more I didn't save were any good. The other one looked better, but it changed the artstyle too much.
>>8795661Nah they are trolling you. 8 vram is plenty enough if you don't mind waiting a couple minutes between gens.
>>8795843I have tried on local. At least on local, the damn thing is still pretty much in diapers.
>>8795861retard
>>8795884Ah yeah my bad, let's pretend it's still 2023. You surely need a 4090 at the very least broo
>>8795843After fiddling around some more on the few sites that have free trials, only this one seems any good. On the free plan anyway, but it's not like I'd blow money just to try this.It's unlucid ai, if anyone else feels like giving it a shot. Though it doesn't even give you enough credits to make one video after signing up; you have to grab the free ones, then wait a whole day to do it again. And that's for a single attempt at making a video.
>>8795843Local is still dogshit and no one wants to make progress on it I think, come back when say, I dunno China decides to help us. Because I doubt anyone outside of there is doing it.
did something happen to Kemono? I can't reach it anymore.
yes
>>8795622Why is she creaming like that?
>>8795622That hilichurl is jesus.
Something something a 2 and a 5
dollaridoos
>>8796026log in
>>8796026it's .cr now instead of .su in case you were using the old url like me
>>8796206I'm still using .party and it redirects to .crbut it is very slow today
sorry, we can't
I kneel /aco/girl
>>8796288
what caused this?
>>8796282I'm sure that Janny can afford a whore like Rin with his 4chan salary.
>>8796206>>8796261Thanks! it wasn't redirecting me from .su
nigger
Chat, it is gay to be concerned about the male anatomy while the girl is perfectly fine and cute?
>>8796332Depends on how glaring the issue is. weird looking muscles or proportions? nobody cares. quantumdisentangled leg? fix it
>>8796334The issue is not that bad since I didn't even notice at first but the upper part of the male (chest and shoulders) are a little too wide to the point it's concerning
I exclusively prompt morbidly obese ugly bastards so I don't have to fret over them looking weird.
>>8796341That's a nice hack, too bad I don't like ugly bastards
>>8796341this but bbc
So yeah, perspective and proportions look better on the first image right?
why does noodleick deserve his own general
because I am cute and sometimes funny
It's my general actually. I just asked noodlechan to make it
>>8796349Hello, highlight schizo.
>>8796350I don't particularly care for vocalfoids tho
>he doesn't like migungmi
>>8796346I cannot tellthe double mouth is really distracting
>>8796365double mouth?
samefag
>>8796366red upper teeth? I don't know
I want more vram 24gb is not enough
>>8796392I can see it but they both are base gens, all of that can be fixed with a little of inpaint, what I was looking for was the gen with the better perspective
>>8796396just buy an rtx 6000 pro
>>8796417I have to save up for eventual robot waifus
Rtx3060 12gb vs rtx5060 8gb Which one is better for local gen? Both are priced pretty much the same.Sorry for reposting, I don't really know what's the difference between generals.
>>87964558gb is pure suffering
>>87964553060 100% VRAM is king for generative AI
>>8796461Thanks man. Gonna throw that shit into an old shitbox I have laying around with 16gb ddr3 ram and a new psu. Gonna goon like there is no tomorrow.
>>8796455I think 5060: it has more compute and generating should be faster especially for videos.
>>8796480Till you run out of vram which you will with only 8gb. >>8796455Really you shouldn't get a GPU with less then 16gb.
>>8796562anything under 24gb is totally unusable>>8796455you can pick whatever but mark my words, you are going to regret buying either of those and you'd look at 3090 gods with awe and cry about being a gpulet
>>8796562> which you will with only 8gbIn what cases? During 4k upscale? Training loras for 8b+ models?
>>8796585If you want to use anything newer then sdxl based models.
>>8796589Like what?Also> what is offloading
>>8796590z-image is rumored to be getting a first party *booru finetune by the chinks themselves. and the distilled version has a 12.3GB unet. And forcing it to run at less with random vram hacks makes it look like absolute fucking ass.
Man, buying that 4070ti super with the wukong game bundle has been one of my greatest decisions so far
this but a nai sub
I'm still pissed openAI fucked up the 24gb 5070ti super release
>>8796607is this nai? box?
>>8796612It's noodleai.
>>8796617box then?
Anyone have the unedited op?
>>8796619yes, me
Anyone happen to have illustrious XL loras for oppai lolis?
>>8796695Nope. Didn't seem necessary.
>>8796696Any prompts that could help me? i try to use the tags but they don't stay loli most of the time when i use the large/huge/gigantic breasts tags
3080 good, bad or decent? I been wanting to try some slop myself.
>>8796711only vram matters, 10gb is decent
>>8796706Think you'd have better luck on /b/. Moderation's gotten a bit overzealous over the past year so most pedos left.Mot my thing, but I can make some blind suggestions. Maybe you can achieve something with prompt editing, add in a very biased artist for the first few steps, so that he only affects body shape and composition and not style or details. Think like [rustle, porforever: :0.2]. Also, the oppai_loli tag was not hidden from Danbooru, which means it's mostly used for older girls. Maybe avoid it and only use loli together with boob size.
>>8796734Alright, thanks for the advice.
>>8796591> And forcing it to run at less with random vram hacks makes it look like absolute fucking assDid you test it already?>>8796718> what is neural network
>>8796706prompt edit the tits and prompt an artist that does oppai lolii
>>8796711Pretty decent for image slop
noodlechan suki, tsukiatte kudasai
https://files.catbox.moe/s41mz3.pngfucking hate the new captcha
>>8796998I had to finally switch to using 4chanx via tampermonkey because of this bullshit because it spitting out "no captcha required" broke posting and the actual extension probably won't get updated since I don't think it's compatible with "new" chrome shit.I hate it.
>>8796998Why? It's faster.>>8797002Oh, well, I've been using violentmonkey the whole time.
>>8797006have you gotten the one that you have to actually scroll through and pick out the "odd one out" or whatever?it's pretty gay. The one that's just plain numbers without any dumb pattern shit is fine though.
>>8797007>have you gotten the one that you have to actually scroll through and pick out the "odd one out" or whatever?Yeah, I think it's the fastest one.
(got filtered on the requirement to press Next at first though kek)
https://files.catbox.moe/c91to2.png>>8797009idk, I had one earlier where the "difference" was extremely minimal. also having to actually scroll through all three images is gay. it's better than the fuckass recaptcha "click on all the <thing>" where half the time it doesn't actually register or it has such a vague notion of what <thing> even is that it isn't consistent at all in what it actually wants you to click and then makes you go through 5 more sets but it's still pretty gay.
>>8796998Que quiso decir con esto?
Sometimes, netorare is fine.
Mmm.. netorare mix lora with wai
Sometimes fried gens are fine.
no one says this
Your mom is not very awa
>>8797108nice tits, catbox?i feel like i recognize the style of those nipples, is it powzin?
>>8796998How to ‘friendzone’ in JapaneseIt can be rather awkward if you get confessed to by someone who you only see as a friend. You may feel like you are more suited to being friends rather than a couple. In this instance, you can say:Tomodachi toshite suki (友達として好き / ともだちとしてすき) which means ‘I like you as a friend’.You could also say tomodachi de imashou! (友達でいましょう / ともだちでいましょう) meaning ‘let’s just be friends!’.
>>8795843Alright, I collected enough credits for another try.I was more descriptive this time, and asked not to do that zoom-in.It worked, more or less. It's pretty much a loop of the original image's scenario. As far as making a video out of an anime style image, this might be the closest I've gotten.
>>8797369I don't want to be friends with chat
>>8797015very clean style, box?
>>8797419<https://files.catbox.moe/86w7ff.png>just testing a really old mix with really low effort lora trains.both loras output really bad hands and are kind of schizo though so not really worth uploading until I do more runs and try and fix that. artists are kokaki mumose and menma (enaic31)
>>8797411Why are you being such a tsundere?
Is there a version of OP's pic without the watermark?
>>8797435Because chat is a piece of shit, garbage, waste of oxygen, good for nothing>>8797443Yes
someone has any idea how it's possible to generate something like this? maybe im fucking stupid but I cant recreate this, I have no idea how it's possible
>>8797447"like this" in what way?
>>8797448basically everything , I wish I could see the prompt so I could learn from it It's a lora? some new model? idk the pose looks really good, the clothes, the hair, the hands, everything looks really good
>>8797449>basically everythingpose? prompt for itstyle? train a lorahands? don't gen on some fried shitmix
>>8797453How would you generate rudeus dad? it looks really good, he even has his ponytail
>>8797455use one of the shitty loras on civit which will probably ruin style or choose a more popular character to fuck your waifu for you.
>>8797423https://files.catbox.moe/vzjn7c.jpgVery nice, looking forward to when you drop the lora.
>>8797457You get the lora and then what? If you use that lora it will generate 2 dads fucking eachother is there a trick or a plugin like forge-couple?
>>8797461If you prompt for 1girl, roxy, 1boy rudeus dad or whatever the model will most likely get what you want and not generate gay porn.
>>8797461ntasome loras work fine with 1girl 1boy scenesor you can enable the lora when inpainting and then add his unique featuresor you can use controlnet together with regional lora schedulingetc
>>8797463>>8797462thanks
>>8797470>>8797447I tried a couple of times and this was the best result, the original image still looks 10 times better, fuck
>>8797475shading and face+dick shapes look like that pedo artist. possummachine or cromachina? I always mix them upmust be in a mix though
To whom it may concern: If you ever had issues making consistent outfits/characters, use wan or some other similar i2v model. Have the character perform several actions and poses. Take screenshots of said poses, img2img with whatever you like, clean the noise that the AI gives the image, and then train a lora on a fairly decent sized set. It has worked wonders for me. Much better than relying on dumb luck.
>>8797480Qwen image edit can give you a side view or back view or a character. Both it and wan are just guessing though, at how she's supposed to look from another angle. You move the dumb luck from genning to preparing your dataset.
>>8797455>How would you generate rudeus dad? it looks really good, he even has his ponytail>>8797447 This one is NAI I think, it just knows Paul Greyrat.Take this >>8797453 anon's prompt and try adding Paul's lora.Style is kakure eria + ?
>>8797502I tried with that adding the dads lora and got this >>8797447Im losing the battle anons, NAI looks really good haha
>>8797447localkekbros, how do we cope...
>>8797585 (me)eh well not quite, but still interesting enoughhttps://files.catbox.moe/oaof0t.png
Why the fuck Align Your Steps doesn't work anymore on reforge?
>>8797624I was about to work on some old gens but when I try to gen them again reforge straight up doesn't work and just gen mostly noise
Best resource for Image2Video?Grok has gone to shit
>>8797728wan videoonline services are all either already censored or figuring out how to censor, but you can still find ones that run the local models
is there any easy way to get rid of extra toes during generation on an illustrious model?tired of manually having to paint them out.there's an embedding for neghands that seems to work for fingers but i haven't been able to find anything for toes
>>8797736Manual upscale instead of hires fix. You roll the upscale multiple times and you'll get 5 toes eventually. Lazy way to not inpaint. Depends on speed of your gpu if it's fine or not.More advanced is rolling multiple upscales and layering the good parts from each in photoshop. That's pretty much my workflow for several years now.
>>8797736There's an...even harder way, you can figure out why you're getting so many extra toes in the first place. I find the issue to be very rare. Perhaps you're running the base res gen above 1 megapixel resolution, or you're using a badly baked lora. Or your prompt that's all "hurr durr shiny quality aesthetic" instead of what you actually want to see in the image. To be fair I don't know if that one causes any harm, the previous two do.
>>8797750What sort of denoise would you have to use for hiresfix to add more fingers? 0.5 with latent upscale like it's SD1.5 again?
>>8797760Depends on sampler. Euler A CFG++ changes very little even with 0.5 denoise, dpm++ 2m does huge localized changes even at 0.3 denoise while keeping composition pretty much the same. Also the higher cfg - the more the image changes.
>>8797736That sounds odd. Are you sure you aren't denoising too much? Over 0.3 is asking for the RNGods to fuck with you.
>>8797767i'm not using hires fix at all though just generating in 1600x1280
>>8797767Wife
>>8797775That is your issue then. 1 million pixels give or take, you're doing 2 million.
>>8797764Interesting. But then you would adjust denoise to match your sampler, no? Still want to end up below the finger-adding treshold.
>>8797793I love pushing it to where image actually changes though. Of course if you are working with trashy WAI-type styles where high denoise adds million wrinkles, folds and background clutter which just makes everything look AI-generated, ̶i̶n̶d̶i̶a̶n̶ and ugly, then high denoise isn't an option. I spend great deal of time searching for styles that don't do that.
>>8797796>Of course if you are working with trashy WAI-type styles where high denoise adds million wrinkles, folds and background clutterThis is just a matter of the model losing focus, and will happen with any style that isn't ultra minimalistic
>>8797799Minimalistic and also painterly, where small details are fuzzy. Well better fuzzy than 4ch VAE artifacted, but that's me...
>>8797799I wonder if it's the style itself causing those folds, or more the way the style is achieved. I have a lora that's extra bad at this. And seeing as WAI itself is mostly a collection of loras piled on top of a checkpoint merge.
>>8797831/e/... also have some mercy with dark mode users
turn your brightness down its not that bad nigga
>>8797973surely areola peek and panties are /e/ enough
>>8797978aaah nvm it's the other thread
>>8797625P-pancho.. onegai.. just one update.. *cough* tasukete onegai..
>>8797625for me it's how reforge CN implementation is somehow entirely different (better) from literally everything else
>figured out an issue with a dataset last night>it's fixable, but since it's a synthetic datasett it effectively means having to start over from scratch and redoing things in a really specific way>planned to do it today>except I woke up with an awful migraineI want to die
>>8798093There's always tomorrow anon. Hydrate, eat basics, put on an ebook or youtube and relax in bed. Hope you feel better soon.
>>8798094I'm still powering through it, it just kind of sucks. I say "woke up with" but it didn't really come into full swing until an hour or so after getting up, at which point I already took stimulants so I'm fucked regardless.(un)luckily migraines/headaches in general is just a random garbage thing I've dealt with a lot since I was a kid. And my migraines, while they are, in fact, migraines, aren't the super horrendous debilitating "have to sit in a corner under a blanket to not want to commit suicide" type of migraine. Instead it's just a sharp pain gouging into the back of my eye on the right side of my head when I shift focus or move around too much and that taking advil does absolutely nothing for. having something to actually focus on (sort of) helps, it just dulls the thinking a bit.I'm just bitching because it sucks but the process I'm doing is vaguely relevant, of which;>>8798093to expand on the issue and what I found, which honestly should be super obvious in hindsight, is related to iterative datasets. process:train lora on single image for style - works - cool!expand dataset with images output by that style, with intention to make it more flexible and to add/modify the style a bit - works - cool!expand dataset with images output by the v2 lora - sort of works? but loses too much due to, effectively, style averaging issues - not cool!solution: you have to keep using the v1 lora to increase the dataset, anything from a v2+ output will be bad for training.basically, it's just the telephone game. You're not going to really be able to replicate a style from synthetic datasets no matter what you do or how hard you try. The only reasonable solution is to minimize the loss/changes as much as possible by not using synthetic data from any iteration other than the first.of course, the actual solution I'm taking is even more dumb and extreme but still falls along the same lines, but I've got a different sort of goal and that's a story for later.
>>8798099literature calls it "model collapse", you keep feeding a model its own outputs it will compound any AI issues it already has, even ones that aren't initially visiblev1 is already bad, just maybe not enough to really matter
>>8798109Yes. Well. Sort of. Model collapse is usually referred to on a completely different sort of scale and has different kind of connotations and (generally) different results. In this case I'm doing a lot of process to mitigate a lot of the issues that would arise from model collapse so random errors aren't really the concern at all. It's just compounded style deviation.here's a comparison grid to show what I'm talking abouthttps://files.catbox.moe/kvmr4l.pngtop is v1, trained on a single imagemiddle is v2, trained on the original image+4 outputs from v1bottom is v3, trained on 6 output from v2ignoring the obvious color palette shift that's occurring(since it's mostly intentional), it's also gradually getting more detailed and linework is getting a bit more fine. Some people might actually prefer this? But it's not the direction I want it to go.and you're right, it's technically sort of a form of/occurring because of model collapse, it's just that this isn't a context it would commonly be considered in.>v1 is already badalso correct. No matter what you do there will be biases introduced by the models default output. Point is though is that if you're doing synthetic additions, you should stick with v1 as much as possible to minimize this issue.
>>8798131Give me your .json I need to bake
>>8798141for those I'm just doing the stupid 1dim 512 step bullshit configs that are in the stupid experimental shit directoryhttps://app.mediafire.com/folder/dfjm9587kflegjust keep dataset to 1-6~ images and it'll at least get you something. for larger datasets use the FULL 1dim config(might want to increase step count on it to 2800~ though) or the glora configs.
>>8798000
Requesting assistance.I'm really into 'netorase', specifically the type of stuff where the girl jerks you off while showing you a sex vid on a TV or with a phone in her hand.I see some AI gen'd stuff like this I'm posting - how do you think it was achieved? I have a NovelAI subscription but have no idea where to start.
>>8798163Just copy the tags of any image like that>https://danbooru.donmai.us/posts/5843259?q=holding_phone+netorare+
>>8798163>prompt hand job holding phone whatever bullshit>make another image of girl getting fucked>use photoshop to put that image onto the phonewaowsuch difficult
>>8798163you're disgusting
>>8798163>I have a NovelAI subscriptionJust... write
https://github.com/Panchovix/stable-diffusion-webui-reForge/pull/453What is pancho cooking
I'm fucking retarded as shit and I've been trying to mess around fusing artists together with no luck. Floox is likely dead or somewhere I don't fucking know. Anyone got the art mix for this shit? Blushyspicy and ???
>>8798220That's my dawg. That's MY dawg.
>>8798259box?
>>8798152hmm...but whence cummies?
>>8798390Cathay-rahtThe Cathay-raht are a bipedal furstock of Khajiit. Khajiit of this furstock were nicknamed as "jaguar-men" by the Imperials, though many do not actually resemble jaguars. Khajiit of this furstock naturally resemble Cathay, albeit being larger. They are said to be more agile than a werewolf, and make great warriors. They possess phallic barbs.During the Five Year War, the Cathay-raht were deployed as powerful front-line infantry against the forces of Valenwood.
No
Can pixel art hentai be made? How innovative
>>8798407It looks really bad most of the time
>>8798260https://files.catbox.moe/0yu9oq.png
>>8797758Good taste.
Do you guys have any pointers for genning FFM POV stuff that has some variability but looks decent? I don't care about keeping character traits distinct because it's fun to see how they get randomly arranged, but if I try to include any description of acts at all I wind up with nightmare contortions like one girl holding up the other girl's leg and it twisting in the wrong direction.
>>8798562Any example?
V5 when
>>8798562Should post examples or metadata, you describe an issue that I do not run into. It's a little random like this without regional prompter, but mostly seems to work fine.Guess I can give general advice for variability, keep your prompt short and leave things unspecified, and use a model closer to the base checkpoint.
>>8798569I already scrapped most of my shit for the day because none of it was of any interest even as an example, prompts included. I just threw a couple more together and it's annoying because if I'm too barebones I wind up with stupid shit like this.
>>8798589Oh god, I didn't specify any of this and now I can't help but contemplate the mechanics of it.
>>8798597You missed the part about not using fried shitmixes
>>8798598All that's using is illustrious so who knows. This was the entire prompt I was testing with:2girls, threesome, vaginal, sex, penetration, sexual coaching, holding another's ankle, fingering, french kiss, 1boy, pov
>>8795620https://github.com/starik222/BooruDatasetTagManagerLast time I tagged a dataset I used the Kohya Gui, then cleaned it up with this this applet.
>>8798602WAI illustrious =/= illustrious.
So I found an AI account on Pixiv creating images of characters from a gacha game I'm playing that I'm pretty sure lacks enough fanarts, and long story short I found out about NAI's character reference optionWhat is this magick? Is there an equivalent function available for local models?
:^)
Heh
>>8797728Grok could animate sex? All I've seen of those was ecchi.>>8797732How does that compare to, say >>8797383 ? Any examples?
kek im notice
>>8798800>Is there an equivalent function available for local models?Nah, loras is still the best we have for that.
>>8798935>IPAdapter
>>8798938>memeGo back, Cathag.
So Kurumuz won after all? Or is China gonna save us still?
kurumuz lost a year ago
Nope, NAI is striving and every local general is dead or dying, even non-anime ones.
>pic related: NAI at its peak
>>8798970Not surprising, banana caters to the normies / global audience and has ways to use it for free rate limited
>>8798983forgot pic
wawaweewa...
From /ldg/ thought you anons might find this useful>>>/g/107611121 >>>/g/107614373 >>>/g/107614498
Buy an ad.
>>8798935So is it a proprietary algorithm that NovelAI somehow developed on their own or bought, or is it based on some paper someone uploaded on arxiv?
>>8799022with how bad it it is it's probably a vision model that tries to tag out traits and possibly high denoise i2i's the reference.
>>8799026>how bad it it isis it really? from the few examples i saw it looked a lot better than other zero-shot garbage
>>8799029It is.
>>8799013Ok, how does it do lewds?
>>8799033how exactly? stiff? too much style/composition bleed? faithfulness seemed ok for characters that you don't want to gen on the regular.
>>8799026I don't know, the few examples of that gacha game images I've seen, it seems like it was able to replicate the small details like the costumes and artstyles and stuff quite well in all sorts of different situations for it to be just picking out tags
>>8799037Those are generic elements that can be picked up by tags.
how is ramtorch for training? nothingburger or actual savings on sdxl?
>>8799067there is a grown up version of that one?
>>8799069more grown up than a 1000 year old oni hag?
>>8799067#NotMyLoli
>>8799085https://files.catbox.moe/kpw4mi.png
virus
>>8799085Shuten isn't loli and prompting her as such is a disservice to her character
age is just a numberthe only thing that's important is if they can get knocked up
I want to post here but I use local and dont want to start a fight
https://files.catbox.moe/4m9p1g.png
>>8799069Why would anyone want that?
>>8799036it's nai which means it's awful, okay? now shut the fuck up shill
>>8799086What model/lora is that? Looks impressive
24.. hours... since... last... hag...posted... BEEEEEEEEP-----
I wish I was black
i wish i was brahmin
i Wish I was in /hdg/
thank god i'm white
yt boy winter
>>8799369what's the point of being white when everyone around you isn't? you can't kindle the flame of civilization alone
>>8799372ever been to a zoo?
>>8798970holy FUCK?! one of the biggest corps in the world beats a small company??? no way!!???!?!
i want to do this chara but the prompt gave the game version how hard is to make a lora for noobai ? or i could ''''convert a lora""" to noob ai ?
>>8799381There's a bunch of illustrious loras for her on civit that will work with noob out of the box, but you can also prompt for carrera, viper \(series\), horns, forehead jewel, lipstick, red eyes, dark-skinned female, earrings, and that will give you a pretty accurate result.
>>8799384>Illustrous lora work on noob ?
>>8799385you'll get shit eps colors but apart from that, yes
>>8799381it's really not very hardif you have bad hardware make and tag a dataset and i'll bake it for you
>>8799391i have 4060TI 16go ryzen 5800 and 48 GB ram it enough ? kinda wanna learn how to do too >>8799384 >>8799386 it worked Thanks
>>8799392more than enough
>>8799386even that isn't always true
>>8799394there are blog post about how to train or the links is good enough ?
>>8799386prompt issue
>>8799397Having to modify your prompt just to accommodate a lora is already a negative. But some are overfit enough on brightly lit scenes that you can't push against it enough.
>>8799361In the troll split of a general that is itself barely usable. You expect too much.
>>8799400don't use shit loras, that's obvious
>>8799396Everything is out of date and shitDownload and install gitgit clone https://github.com/67372a/LoRA_Easy_Training_Scripts/commits/refresh/ run install.bat, follow the instructionsget a dataset tagger, tag your datasetrun run.bat when you want to train, ask someone here for a toml file (training settings)train on the base model not on a shitmix, ask questions if it fails
>>8799403>ask someone here for a tomlkek
>>8799396tl;drmake your dataset: good images only, high enough resolution for the buckets you are using (1024x or better), add closeups if necessary, avoid too many similar images to overfittag everything: start with an autotagger, smilingwolf's wd03 tagger is still best imo, there is a tagging extension for webui but there are other programs too; use a dataset editor to remove extraneous tags and consolidate and/or concatenate things (e.g. pick one of skirt, blue skirt, pleated skirt; or, change to blue pleated skirt); you usually want to associateuse sd-scripts or a UI such as easyscripts (they are all based on this repo anyway); look at other people's training settings and test, test, test, test everything, and remember the scientific method, change one variable at a timeif you want to do what i do:https://github.com/kohya-ss/sd-scriptshttps://github.com/starik222/BooruDatasetTagManager/sample scriptsstyle: https://files.catbox.moe/m0k8xo.ps1concept: https://files.catbox.moe/ac50ji.ps1
>>8799405*you usually want to associate a lora with a tag and put it at the start of every image's tags
i honestly don't know where you get a tagger besides the extension sorry
>>8799407taggui comes with the v3 taggers and there's some way to install them in waifuset as well
>>8799404Or search the archive for the dozens that were already shared this year. It's all so tiresome.
just found out noodlekek is leaving forever
>>8799313It's 2 loras both trained at 1 dim 12 conv with 2800 steps (aka dogshit). One of them was done by another anon in here, and the other I just trained myself.https://www.mediafire.com/folder/h6jb6h7q9f5d9en,73hchuiggo8g3ey/sharedHere's both, but bare in mind they are kinda shitty and might mess stuff up (especially anatomy)
Can this low dim meme die already? almost worse than captionless and no negs
>>8799527I refuse to dedicate my vramlet GPU to training in excess of 2 hours, therefore I 1 dim.
>>8799527It's not a big issue for style loras. You get the rough contours of the style, instead of the exact way he draws nipple holes or individual pubic hairs.
>>8799529You live!
>>8799527>almost worse than captionless and no negsit actually works, though.I agree it probably shouldn't be used for everything but it's low investment, really easy to share and it works better than the vast majority of other random configs you'll find strewn about the internet.
From all of my intense and difficult work training loras I have deduced that I don't know what the fuck I am doing.>>8799013I tried NewBie out and apparently the native implementation is worse then the old implementation so probably best to wait a month or so for everything to get worked out
>>8799594it's important to recognize that some artists/aspects of artists just aren't really very trainable at current(due to VAE, just plain object permanence issues and a few other things). You generally need to set goals and curate the dataset towards them.Also don't test loras with captions directly from the dataset, you're not really testing all that much that way. You need to use neutral/constructed prompts that are intentionally removed to see how it actually gets applied(leave concept stuff in if that was your goal of training but change the context as much as possible).
>>8799619>it's important to recognize that some artists/aspects of artists just aren't really very trainable at current(due to VAE, just plain object permanence issues and a few other things). You generally need to set goals and curate the dataset towards them.This is some 8gb vram advice
>>8799620have fun trying to train mixed halftone that doesn't look like complete ass.
>>8799619The top caption is from the dataset but the bottom one is not, I tend to test that way because the dataset caption tests how fried it is and the non dataset caption tests how well it generalizes. It just seems that training artists in general is more difficult compared to training characters. Of course it could just be that I'm tougher on styles then characters but then I look at how the nose is all wrong on the white-haired ones...
>>8799622tbdesu I'm not entirely sure what you're aiming for training on the artist, but from my experience looking through what's up on *booru, dataset'ing them would be absolute fucking ass.they've got a lot of floating body with floating background shots, a lot of gachashit tier detail bobble spam slop compositions, extremely schizo eyes, etc. it's just a lot of stuff that's usually pretty hard for SDXL to properly digest.also what's the difference between full and full-all? is the later just a larger dataset? if so that makes sense since I would think a smaller dataset is probably the way to go with this(and it seems? like the first is sort of better).
>>8799622It's either bad generalization or you're bleeding style on other tags.
I was running some tests making a new lora. It does look like the artist's style, but something feels off and I cannot figure out what it is.
>>8799646which artist?
>>8799647it's a troll
>>8799647Piutahayo. I think the model may be fiddling with the artist's "flatter" style.https://x.com/Piutahayo/status/1993443082281664643
>>8799649Have you tried not using a shitmerge or quality tags/negs? Seeing as with them you're really testing a mix of your style and theirs.
>>8799652fucking kill yourself already noob schizo
>git pull>Already up to date>cry
>>8799729That's a good thing. Do you want a Bethesda moment?
/edg/ has managed to discover inpainting and using artists tags in the span of a single thread, meanwhile the so called thread for ML experts can't even bake a single decent lora after 34.
>>8799779kanikananika was decent.eencya was decent.There's an unspread pussy cameltoe shape on panties.
>>8799779cute gen, box?
>>8799779>/edg/ has managed to discover inpainting and using artists tags in the span of a single thread,kek it's truegood for them
>>8799779when will /hgg/ discover videogens?
>>8799833hopefully never
>>8799834why not?
>>8799833I posted 2 mp4 in this thread, but the general consensus seems to be that the technology is not quite there yet.
>>8799779>meanwhile the so called thread for ML experts can't even bake a single decent lora after 34.we can, we just don't share them
>AssertionError: fused_backward_pass does not work with gradient_accumulation_steps > 1>AssertionError: fused_backward_pass does not work with gradient_accumulation_steps > 1>AssertionError: fused_backward_pass does not work with gradient_accumulation_steps > 1Fuck you, now I'll have to manually resize my entire dataset
>>8799820It was a Prometheus kind of deal, thankfully there are no eagles left on /h/.Well, there's not much of anything left.
>>8799922The liver of a pixai user probably tastes awful anyway.
>>8799925I wonder, how much would a pixai user be able to tell you about inpainting.
>>8799909that has been the "general consensus" for a while now
>>8799909it's fine for realism and 3D, but the movement looks unnatural in anime styles
>>8799943I think the 2nd one looks alright, especially since breasts pressing against glass would seem tricky.
>>8799779I hope you are talking about the NewBie thing
>>8799984I'm not downloading a whole ass cumfy fork for a single lumina finetune, sorry.
>>8799992I read on the real thread that it has native support on comfy now
>>8799994forge support or bust. actually bust anyway because it's undertrained dogshit
>>8799997but muh 16ch vae
>>8800000I still don't get if everyone is training their models wrong or if it's just overhyped, because both local 16ch VAE models and NAI look like garbage full of melted details anyway.
>>8800021googles new nanobano has "melted" details even when generating at 4kits just an AI thing
>>8800021nai looks bad due to its styling, the details are good though.
>>8800021>>880002216ch doesn't absolve you from upscaling/downscaling, you'd need pixelspace for that
>>8800030>nai looks bad due to its stylingeven if you keep saying this it doesnt make it true lmao
>>8800036? and what is the truth
>>8800036Good for you if you like it
best sampler for controlnet upscale?
>>8800038the styling is fine maybe use artist tags
>>8800043there are some gens that look fine, but majority of 4.5 gens are unadulterated trash probably below civitai tier. fighting your model to not look like shit doesn't sound like an experience you should pay for
>>8800047>but majority of 4.5 gens are [shit i tell myself]proof? they look fine to me
https://danbooru.donmai.us/posts/10472161?q=user%3Araredragonhelp find artist mix please
>>8800054isn't it just kakure_eria again
>>8800054Love how he took the the time tagging every conceivable thing about the nog, but the girl barely got any
>>8800054>ai-generated pics are not permitted on danbooru>but the tag still has 13k approved picsI don't get it
>>8800059browse down to page 5 and the tag is emptythey're probably hidden by a separate system, like loli
>>880005913k is all the pictures uploaded, not all the ones approved. There is only 1 active (not deleted, not still in queue) image tagged ai-generated.>>8800062You can use "ai-generated is:deleted" to see all the other images
>>8800054M-my oshi would never...
I really need some artist tags recs for noob. even that typical mature women shiny skin look you see everywhere I have an hard time reproducing it. wtf do people use anyway? Style loras sucks, they reduce pose variety and range in my testing, but maybe I'm wrong.
>>8800095>wtf do people use anyway?artist tags mixed with sometimes style loras at low weight
>>8800095>wtf do people use anyway?full finetunes
>>8800096>>8800097Damn, finetune styles could be anything.But I managed to extract metadata on some pics of the style I wanted after all
I really need to go back to improve my workflow and quality instead of just prooompting and gooning.
>>8799779Hey, don't bully /g/. That's not nice.
What's wrong with shiny skin?
we love WAI here
>>8800054mochirong
https://files.catbox.moe/kzptdz.png
honestly bad
>>8800180based
No artists req? Just post your favorite or go to tag, come on.
>>8800225>https://danbooru.donmai.us/artists?commit=Search&search%5Border%5D=post_countJust search your own, everyone here have way too different tastesAny artist with a higher count than 150 posts (((should))) work fine on base noobvpred or it's shitmixes
>>8800232>way too different tastesYou're right about that lmaothanks for the link I didn't think the post count mattered
>>8800042I stick to what I base gen with: DPM++ SDE Comfy
I wish finetune schizo was still here
WHOMST
>>8800525Guy that made this rentry https://rentry.org/lora-is-not-a-finetune#migrating-from-easyscripts-to-sd-scripts
16ch vae model when
>>8800021i've been wondering for a bit if it's an issue with dataset pre-processing. i had an issue with my own bucketing script where just using opencv inter_area was causing a pretty massive hit in quality before switching to a weird super resolution > decimation process, and just that was enough to get SDXL's VAE to start picking up on more fine details.
>>8800576NetaLumina has been out for a while. Sure, it's underbaked and kinda shit, but it uses the FLUX VAE.
>>8800724is that desirable?
>>8800724box please
>>8800822No, as the FLUX VAE is universally considered dogshit.
>>8800960Don't like the shine on that one. why are you experimenting with that style so much anyway?
Should I give latent upscales a try? I was never satisfied with regular upscaling method, it aads fingers and toes and shit when the base Noob gen is often just right. I have to ask cuz you guys are posting pretty high res pictures that still looks good.
>>8801068what do you mean by "regular upscaling method"?
>>8801068You can try it, but it has some hefty impact on style. regular upscaling to 1.5 and controlnet from there is probably the best universal solution. But try latent/deepshrink, it might work for your style.
>>8801074I mean the 2nd pass "spacial" upscaling at x1.5 res>>8801083I'll give it a try but I didn't try with controlnet yet. I heard Noob vpred doesn't have controlnet released still.
>>8801090>Noob vpred doesn't have controlnetIllu controlnets work just fine, at least the union ones.
>>8801092thanks, I'll check it out, I know very little about controlnet and didn't know you could mix models like that.
>>8801093>noobaiXLControlnet_epsTile>CN-anytest4_illustrious2_B>CN-anytest_v4-marged>diffusion_pytorch_model_promaxI've found any of those to work with tile upscaling, if that's what you're on about.
>>8801100>>CN-anytest4_illustrious2_B>>CN-anytest_v4-margedwhy not A lol
>>8801090>>8801092>>8801100why do you retards want to use controlnet and inpaint with vpred so much, there's literally no point in that since you don't set denoise to 1.0 thus ztsnr is not required, you can just use eps directly
???????????????????
stop samefagging
>>8801110>just switch the model for every inpaint and upscale
>>8801110I don't even have eps downloaded lol
>>8801113>vramlet a1111 legacy troon cope
>>8801101I never tried A to be honest. Some anon was talking about B the other day so I tried it for myself and it was fine. I still prefer the noobai model overall, but there's a few styles, in particular 3D/Realistic, that promax and anytest handles better.
>>8801110Why would I ever want to use eps in the first place.
>>8800861https://litter.catbox.moe/udqxlcoc68um0bzr.png
waow this thread is very bad
https://www.reddit.com/r/StableDiffusion/comments/1pucj4g/spectral_vae_detailer_new_way_to_squeeze_out_more/ new vae snakeoil for comfytroons
>>8801445looks like it fucks up contrast and finer details, it just looks worse
>>8801445>gives everyone cataractswow
>>8801459Another win for milfchads
>>8801445I don't dislike the results, but I can't help but feel that bro is trying to counteract problems caused by his shitty checkpoint or settings, rather than the VAE. He says it himself, the tool never leaves the latent space, which means he fixes the place that is the responsibility of the checkpoint, not the VAE.
>>8801445>3dpdwho cares?
hello?
>>8801756no, we're dead
>>8801756this is patrick
good morning
Why did hgg lose all steam after #10?
>Why [it was real in my mind]
the fall of /hgg/ will be studied
>>8802323good
>>8802323I wonder why she's so popular (relatively). Design is ass and the only thing I can remember about her character is that she's annoying and likes the mc. Does she get some arc later in the story that focuses on her?
>>8802333she's an absolute midget. and a dumbass. this combination can often make characters popular.and the series itself is actually really popular and she's the poster girl. also the only other primary girl was made to be more on the homely side and introduced to be unlikable.
https://files.catbox.moe/l29uqo.png
>>8802348Nice, thanks for sharing
>>8802323Woah, did you install ExpressiveH or something?I'm kidding, cute faces, very moe!
https://civitai.com/models/2167995?modelVersionId=2532077v0.2 of Chenkin Noob is outis it good yet
>>8802665>yetMaybe once they ditch EPS and start over with something that at least has potential to get good. So never, forget that civit page exists.
>>8802665>every single preview image is grey
>>8802665the year tags are a really nice addition and they work well, however if you're a lora abuser and don't care about the updated dataset there's not much point in moving on from v-pred 1.0 or its shitmixes yet
>>8802665 how the fuck is eps baking still a thing going into 2026? it’s caveman tech at this point and genuinely embarrassing that it’s still being used
infant bobblehead bros.. we won't take the bait.
>>8802684Russian and chinks are born with a black and white color filter, they don't notice things being grey and dull
>>8802323Hot, high quality, no /d/egeneracy. Thank you, Anon.
Long torso or not? how would (You) fix it?>https://x.com/mipu29336421/status/1951915945754337628
I hate when people refer me to some AI general thread on how to create AI and I come here and there's 50 different links and every tool has multiple options with no clear way to know the advantages and disadvantages of either without sinking an incredible amount of time into figuring it out yourself. I just wanna make porn on my 5090 man
>>8802806just send me your 5090 and i'll set it up for you, bro
>>8802806Just grab something at random and stick with it.That's how I got into Comfy when I started generating boobs couple of years ago.
Noob here, may I ask for advices?I've installed ComfyUI, tried to replicate an artstyle I liked on Civitai, downloaded the same checkpoint and loras but the results are completly different, any suggestions on what to do?Here's what I'm trying to do: https://civitai.com/images/85140312 while pic related is what I get.
>>8802862anatomy machine :B:roke
>>8802859The difference here is likely the low cfg, which is bad practice but gives a unique look.
>>8802859Well for one, your image seems to be in different dimensions than the one posted behind that link.Also, make sure all those loras are at strength listed in your link and all of your ksampler settings are also set like in original image.
Im using forge to edit some previously generated lewd AI images from a site. Im trying to use inpainting to remove the panties on certain pictures, but it just wont do it. Am I prompting wrong? Is there some built in censorship or something? Im pleased with the changes the inpainting is making but.. remove the fucking panties already please
>>8802908Forge is the user interface, not the AI model. You may well be using a censored model, no way to tell without more info.Otherwise, read this. It's way easier to remove them if you first edit the image in an external tool (or inpaint sketch) and change their color into the color of her skin.
>>8802908Of course I post the ONE image that actually worked as intended. This image is usually what I get whenever I try to remove underwear/bras.>>8802909Im using nova3DCGXL_illlustriousV50 and I can generate hentai images no problem, its just the inpainting thats giving me trouble. It adds the anatomy but very rarely removes the panties.Ive never used inpaint sketch.. Ill have to give it a try. I appreciate the guidance
>>8802909Thats actually fantastic. Inpaint sketch gives me consistent results! Very much appreciated anon
>>8802859>Resources used>WAI-ANI-PONYXL>InCase LoRa/LoCon | Style | SDXL Pony>ExpressiveH (Hentai LoRa Style) エロアニメ>Cum slider Pony/IllustriousXL>EasyNegative>Pony PDXL Negative EmbeddingsLet me guess.. you NEED more?
>>8802909That's a great pic, got one for Comfy?
>>8802942Functionally, everything works the same. Just a matter of matching those settings to the comfy equivalent, which is MaskDetailer from ImpactPack. >>>/e/3034204 except for masked padding, not sure about that one yet.
>>8802884Changing the cfg didn't seem to meaningfully change the results. As far as I could tell. >>8802885I've fiddled with parameters, and it would seem the problem come from the sampler? ComfyUI doesn't seem to have the one used by Civitai (DPM++ 2M Karras), but dpm_adaptative seem to produce somewhat closer results? >>8802940Half the time I had no idea whatever I was doing and would just copy the parameters of cool pictures then just slap more shit on top of them. It even worked, sometimes.
>>8802979>ComfyUI doesn't seem to have DPM++ 2M KarrasIn Comfy that's:Sampler dpmpp_2m (the "pp" stands for plus plus)Scheduler: Karras>Changing the cfg didn't seem to meaningfully change the resultsextremely low CFG (like his value of 1) dulls colors and makes outlines less sharpgoing too high does the opposite, thought how much is "too high" depends on the checkpoint
>update comfy>breaks>regret it>reinstall everything>the prompt encoded removes commas after each genInfuriating.
>>8803001nai doesn't have that problem
Why are newcuties using cumragUI again?
Why are you samefagging again?
>Lets try doing something simple "hug_from_behind, kissing_neck". Surely, this will be easy.>Necks are unpredictable eldritch beastsIt was, in fact, not easy.
>>8803001what the fuck is this shitty bug, it's doing it on every text encoder
>>8803041now shit is completely unusable, bravo.
average day of a comfyui user
Speaking of which, is Comfyanonymous, well, an anon from these parts?
I think I fixed it kek. It's fucking fast now too.. but I have no idea what happened. Hopefully it doesn't do it again.Are you guy using other uis?
>>8803058Nah, I'm using Comfyui too, it's comf, and I like seeing the process going step by step. Building up a workflow is the most fun part part as far as I'm concerned.
>>8803038Necks are a bane for some reason, anything involving them like kissing, neck biting, hickies is a bad time.
Why are you still samefagging?
so bored of local and naishould i try illustrious api?
>>8803055He is my uncle
good evening schizos. any suggestions for style loras of lesser known artists? tired of browsing booru. no promises though
I cannot parse what he wrote.
X-Ray Physics engine WIP
>>8803130morningif I had a style worth baking I'd have done it myself 14 months ago
What is the best model to generate hentai with easy diffusion?
novelai 4.5
>>8802994It's not perfect but it mostly work, thanks!
>>8803058so it wasn't a bug but a feature. It was autocomplete plus, it handle commas and blank spaces automatically, that's pretty cool, if they didn't use the worst possible setting as the default all of a sudden.
>>8800696>>8803149There are lots of /h/ of that one but not of her sister with longer hair. Why?
I'm still using R3mix as my main anime gen model, can anyone who was also using it and moved to something else recommend me your current checkpoint? I want to try something else, but none of the highly rated mixes on Civit look good to me.
>>8803281>but none of the highly rated mixes on Civit look good to me.>not liking WAI, NTR-Mix, Prefect, Yiffy, Plant MilkSounds like a personal bad taste issue, anon. I'd gladly gen with any of those, any day of the week.
>>8803281The two in the OP (102d, 291h) and r3mix are about the same in terms of stability but each with their own style bias. There isn't anything better, just more sidegrades, the rest comes from loras, workflow optimizations and how much work you are willing to put into each gen.
Guess I'll keep coping on r3mix for now, thanks. Are we expecting great things from Zit Base or nah? The Turbo anime loras on Civit all look janky as hell.
>>8803301>great thingseven if the base model gets released and we somehow get compute for a finetune, you can bet your ass every retard troon and furry will come crawling out of the woodwork again to sabotage the bake with diaper scat and fnaf cosplay
>>8803308There's no need to be such a skeptic in this jolly holiday season.
>>8803308>sabotage the bake with diaper scat and fnaf cosplayand yet, it gets recommended here. Weird, isn't it?
>>8803363what's the alternative?
>>8803364lie downtry not to crycry (a lot)
>>8803364Just doing what NAI did.
>>8803371Add even more furryshit in the dataset and ditch the pretraining, so it has maximum impact? idk, I'd like an anime model for once
>>8803376I'd like something that works rather than pandering to the purity requirements of random retards on the Internet.
sup guise. who wants to get paid for proooooompting?i'll give monero worth USD20 to whoever uploads 100 galleries within 24 hours on exhentai. only 8 images per gallery. every image must be a variation. only one parody character per gallery. each gallery must follow a theme, like one gallery has a female wearing a red lingerie and getting fucked doggy style while giving blowjob to someone else. put the character's name in the title and the theme, e.g. "Asuka Evangelion Red Lingerie Doggy".
>>8803423For that price I could do maybe 10 gens for you, if you beg nicely.
>>8803423You got told on Exhentai for posting slop and now want us to be your personal army?
>>8803504>NOOOOOOOOOOOOOOO YOU CAN'T UPLOAD ON EXHENTAI BECAUSE YOU'LL HURT SOME RETARD'S FEELINGS :'(((((((((((((((
>>8803508t. retard who got his feelings hurt by the moderation on sadpanda and is now paying people to be his personal army
>>8803423>>8803508What are you trying to accomplish by doing this? I thought AI was already allowed on sadpanda.
hey anons, is there a modern guide I can follow to get my setup going? Like the one that ends with an asuka gen, so you knew your stuff was working properly. I upgraded my PC recently and I haven't prompted since SD1.5 so I'm quite behind.I just usually cannot get a clean image and I'm never able recreate the catboxed ones no matter what I do, kinda sucks.I gathered people use reforge and I got some popular checkpoints and tried some stuff already, this is the best so far, but I would really appreciate some pointers.
I always distrust this kind of spacing
is there custom node to easily schedule loras yet? >>8803423kys
i aint reading all that
onegai help...
>>8803587>>8803604looks ok, what's not satisfactory to you?
>>8803604I like her make up version
>>8803605that mix worked nice but yeah, generally my pictures are "fuzzy" even after upscaling and like I said, can't recreate images from people's catboxes even when using the same model, loras, and at least general settings.I suspect maybe my ui is outdated and doesn't work well with newer models or samplers or god knows what but idk what is current, it seems the opinion is split on forge?like pic related, I tried last month based off an anon catbox. His was sharp but mine's all... idk, smudged?
>>8803609could also be the emphasis and compatibility settings. try switching those around.
>>8803587maybe show us one of these catboxes and your try to recreate it
>>8803613I tried turning each emphasis and compatibility setting on and off but nothing seemed to work>>8803619sure, here is an example with a catbox I saved a while back: https://files.catbox.moe/ncwd4f.pngand when I tried it: https://files.catbox.moe/myuwe4.png
>>8803631my uneducated guess here is that they use adetailer but you don't
>>8803634For the whole body? I did use adetailer, but only for the face/head.
>>8803640dunno, that's the diff in the metadata though.I don't really see adetailer in yours, and the hires prompt is kinda questionable in both.
>>8803631you have different versions of whatever ui you're using, probably some back-end change causing the difference
>>8803643Speaking of which, what's the "meta" now? Reforge still?
>>8803647comfy spaghetti spilling
>>8803647forge neo
>>8803631Might be that 0.5 token merging shit you have enabled. Can't see any other difference that would cause this unless your specific UI version is busted.
>>8803289you're just a fucking slopper, if your post is genuine
Disingenuous post.
>>8803634why does everyone use this shitty lora
>>8803754Nyalia is the modern expressiveH except nowhere near as good but what is, chudai?
>>8797587Kinda late reply but what tags did you use for those sandals? Or does it come natively with the character?
>>8803758https://danbooru.donmai.us/posts?tags=gladiator_sandals+&z=5
>>8803754don't ask me, I don't use it
>>8803760Thanks, 4 years and I still come across tags I never knew that existed.
>>8802742box plz
>>8803754>>8803757samefag
nah that's you
me?
>>8803871A very rare OИE /h/ indeed.
i post but no one replies
>>8803901https://files.catbox.moe/98jlve.png
>>8803905No, fucking retard gger.
>>8803965what do you post anon?
>>8804050Don't worry about it.
>>8803965:(
>>8804039yes
Why are poketrainers idolized in ojisan nojukusha circles?https://files.catbox.moe/p6lrvy.png
>>8804118Nigga you gay
>>8804133no shit sherlock
>>8804135Faggotry is bad.
>>8804133It's not gay if it's pokeboys
>it's not gay if [fanfiction]
>>8804118I'm more surprised the dude has the energy for sex despite the metastasis of his melanoma.
>trap (formerly otoko_no_ko) is le gay!The shitposting here never ends. Take me back to pre #10.
>>8804144That's his natural camouflage
>>8803965so what, gonna cry? piss your pants maybe? maybe shit and cum?
>>8804139no shit sherlock, religious version
>>8804008thanks
>>8803965just samefag your own posts like the big boys do
>>8804503nice
>>8804568a box and a link to that kanikananika lora, my good sir.
>>8804793nice, I like styles that give cum its own outline
>>8804151There are lots of secular reasons to hate you disgusting disease speaders.
>>8804863it was upscale rng actually
How do I stop the tile CN completely changing certain hues? It completely ruins light blues and tends to make warm yellows and oranges shift to a flesh tone. I can usually fix it with paint.net but it's still annoying.
>>8804894You stop it by using hires fix instead
>>8804896I can fix wrong hues in one minute, but I can't fix my gen being blurry and melty all over without inpainting every single square inch of it.
>>8804793what style lora?
>>8804898it's just artist:sazaki ichiri artist:aomushi \(mushamusha\)
>>8804900im not sure what that's supposed to mean. Is that a tag?
>>8804617It's very inpainted, but the base gen ishttps://files.catbox.moe/y8ce85.pngThe lora is in the archives, you should find it easily by looking it up
>>8804908they are artists...
>>8805159what is her glove called tag wise?
>>8805185First day? Bridal_Gauntlets
>>8805187no, bad memory, thanks anon
>>8805184Yeah, but I mean, that kind of tag with a semicolor, I never saw it. Usually semicolon is used for strength of a tagWould you share the catbox?
>>8805260that's the format for noob based mixes since it was trained that way
>Z-Image is a new contender as the savior of localtf happened to Flux?
nothing ever happened
>>8805330the dataset for flux is absolute dogshit
>>8805340>>8805340>>8805340
https://files.catbox.moe/lwzyum.pngI hope this image is worthy of the standards of /hgg/.
>>8805639Good post anon, thanks for sharing. More gens with the space theme?
PAGE 11 BUMP
