Previous Thread: >>8904372>LOCAL UIForge Neo (ANIMA SUPPORT): https://github.com/Haoming02/sd-webui-forge-classic/tree/neoreForge: https://github.com/Panchovix/stable-diffusion-webui-reForge | Stable commit: 6964cebComfy: https://github.com/Comfy-Org/ComfyUI | https://comfyanonymous.github.io/ComfyUI_examplesSD.Next: https://github.com/vladmandic/sdnext>ANIMAhttps://huggingface.co/circlestone-labs/Animahttps://thetacursed.github.io/Anima-Style-Explorer>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/zy2xlr.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_SxWV3x1BBaJGbwsd-scripts trainers: https://github.com/67372a/LoRA_Easy_Training_Scripts (check out refresh branch) | https://github.com/derrian-distro/LoRA_Easy_Training_Scripts | https://github.com/bmaltais/kohya_ssOther trainers: https://github.com/Nerogar/OneTrainer | https://github.com/tdrussell/diffusion-pipeOP Template/Logo: https://rentry.org/hggop/edit | https://files.catbox.moe/om5a99.png
I haven't been around in months, has SD hentai generation progressed anywhere?
post here if you support mexican hate
>>8912792Novelai has progressed a lot
nais :DD
Aesthetic V1.1 soon, friends
>>8912809its already out nigger
>>8912809>he is unawarehttps://huggingface.co/circlestone-labs/Anima/blob/main/split_files/diffusion_models/anima-aesthetic-v1.1.safetensors
>bibirugsell keeps releasing nothingburger after nothingburgerIs it over localbros?
is slop an aesthetic?
mhm
>>8912813wow 1.1 is just worse in every way possible?
>>8912813seems worse than 1.0b, which isn't exactly better than just base.
>if it doesn't look exactly like wai sdxl it's garbagethe whole AI anime "community" is legit beyond saving. your brains are friedaes1.1 is actually good btw, easily the best "high quality" anima finetune. whether you prefer that to base is a matter of opinion but it does very well at what it sets out to do. better than animayume, wai, cat tower at least
Anima base is so good that it hasn't had any good finetunes yet
why can't you morons be serious for once
Good morning sarsヽ(^o^)丿
>>8912843I liked aes1 but the noise, frequent finger issues, and the fact it came apart at any bit of upscaling made it unusable. Aes1b seemed like a sidegrade of base. Does some things better. Does some things worse.Looking forward to try 1.1 later.
>>>/g/109266871come here tdrusell is talking
>>8912970Why did saar get mad about this pic specifically? kek
i feel like tdrussell has no idea what he's doing. preview 1 was the best model, it all went downhill into unironic ai slop and boring rigidness.
>>8912943>moving to another base modelCan't wait for the krea2 finetune that won't be able to gen sex and will take 15 minutes for a single base resolution picture
>>8912862sir this is a brahmin only thread refrain from posting brown pakis with white Indian women
>>8913156ruined
god anima sucks for porn
>>8913184are you sure this isn't a skull issue?
>>8913247I'm sure it has to do with using a lobotomized TE and refusing to train on e621 (but deviant art is OK because... IT JUST IS OK???)
Anima can't generate a decent pussy to save its life.
>>8913271it can't make juicy horse cock and balls either
If anima could make some good genitals it would be the only model you ever need.Right now the only way to generate somewhat passable porn with it is to generate the base with anima and then use an illustrious based model to inpaint the naughty bits.
>>8913156improved.
>>8913271vagoos are ugly anyway who cares
>>8913275nooob-basedor to use a simpler anime style
can't wait to turn all of your gens into BLACKED for my patreon with the new anima edit model
https://old.reddit.com/r/comics/comments/1uw1jok/sequel_of_healing_and_big_poles/
>/hgg/ thread>no ad in the opfinally a place to post
>>8913491define ad
>>8913501it makes me extremely mad
>>8913503but not as mad as it makes me when it gets removed
>>8913419>implying i don't do that to all bibicii posts already
Made my first Anima lora and testing it now. Some initial thoughts- Anima Base seems to be extremely volatile and the results are often just wacky and low quality, jumps between not really showing the style at all to Flanderizing it- Base results are just so all over the place that it's hard to even see how different lora weights are working- Differences between epochs is really substantial, seems like you pretty much have to download and run a bunch yourself to really know what's going to happen, while Illu epochs usually follow a smoother curve of development where you can more easily zero in on the Goldilocks zone- Trains rapidly and heavily, I was literally looking at like epoch 4 as one of the viable options- STILL fucks up hands from time to time
>>8913623Recommended LR was one third of default iirc, it trains very fast. I had no trouble getting styles to stick though, much easier than on noob.
>>8913623did you try nai's vibe transfer? it's really good at strengths 0.6 and below
what caused decline
>>8913680you forgot to attach an image with your post
my nai gens are too good for this place
https://x.com/tnwdprorok/status/2076940141558456502
>>8913623From what you are describing you used too much steps, trained with the images at 1024 and to make things worse you generated the samples with the base model.-Don't get over 1000 steps, no exceptions-Train using 512 instead 1024 (seems counter intuitive but at 1024 it will learn too much small details like textures while ignore general aspects of characters and artists)-Generate your samples with wai illus or some other mix that has a bias, the base model is way too much of a wild cardIf you don't want to use a mix, generate your samples with the turbo version since it also has an style bias that will keep things someway coherent.
>>8913731well, at least you can't give any worse advice
>>8913623IDK bro every problem you describe points to LR being too high.
>>8909730>Summery Of the Situation>To sum up my situation, CivitAI took down some of my favorite lora's because something something wrong think. I got pissed and decided to leave the site and take down my lora. The lora that were taken down were: Cool Rance, Disgaea RPG and Himukai Yuuji style.>After sleeping on it I decided not to take down my lora and continue to provide content. This is solely because I want people who still use this site to continue to enjoy the lora I make. I still very much despise this site for the decision they made, but if they want me off this platform, then they're going to have to straight up delete my account, otherwise I'm not fucking leaving.>(Fuck Them!)BASED! FISHBROS EATING GOOD FOR CIVYEARS ON END!
>>8913847fish loras are the epitome of slop, if he leaves the hobby it's good news
oh nyoooo~Fishbros.. I thought we had won 1 (one).
why all shitposters moved to /hdg/
>post beautiful NovelAI images>shitposterThat's why! FUCK YOU!
>>8913856His anime style LoRAs are blue biased, fried slop???
>>8913859You'll never deplatform the fish, civjeet. We swim against the current to prepare us for situations like this!
>>8913623>-Train using 512 instead 1024 (seems counter intuitive but at 1024 it will learn too much small details like textures while ignore general aspects of characters and artists)you can just adjust the sigmoid scale if this is happening to you
>>8914048how do I do that on civitai trainer??
>>8914049idk, I guess you don't since it's a piece of crap designed to pump out slop with minimal compute cost
>>8914048whats a sigmoid scale?
>>8914052how the sigmoid function for sigmoid timestep sampling curves
>>8914062what dat meeeeaaaaannn?
>>8914066set big number model learn big thing about image. set small number model learn small thing
>>8914069Bullshit. Higher sigmoid scale = less focus on middle timesteps and focus more on early and late timesteps. What tf is "learn big thing" and "learn small thing"
>>8914082this is why people should just grab a toml from the thread, not fuck around with settings nobody ever needed to change
>>8914082this is why people should just pay 25 dolalrs to novelai, not fuck around with old shitware nobody should ever need to use in 2026
>4 pictures posted in 2 days>of which maybe 1 is not a shitpostgreat thread
there's like 4 posts in total within the last 2 days. this place is dead dead
>>8914095>>8912711
>>8914095>>8912022
>>8913847>>8913869>I will simp for whatever Discord or Civit troon in hopes he joins this totally dead general!In your dreams bisexual faggot
i hate anime
>>8914104You'll be swimming with the guppies soon enough, little fish.
guppies are pakis of fishes (inbred)
Hi, I’m a weakass bitch that hasn’t really wanted to get more than the tip of his toes wet when it comes to ai so I signed up to SmutGPT. It’s been kind of fun. Would you all recommend a local program on your machine, and is it faster than web-based shit? I’m kind of a normie in that I don’t know coding or any other shit, I just used to draw a whole lot.
lol
>>8914521You want speed? Do you have a 5090? Or $25 per month?
>>8914626why is her cleft of venus longer than that weird city they wanted to build in saudi arabia?
>>8914626why the nigger?
>>8914932intimidated?
>>8915020finally, henti that I can consume as a proud white man without feeling humiliated by Jewish interracial messaging.
>>8915175>>8915175Hmmm, I wanted to ask you guys *scratches the back of my neck* could you have sex? I mean, um, I don't know how to say this, but I need to masturbate
>>8912789>2026>people still using god awful python for their back end
>>8915287feel free to rewrite the entirety of pytorch in rust or whatever is popular among troons these days
>>8915318why wouldn't you want a marginal speedup at the cost of 10 times slower dev speed and most new features never making it to you?
>no chloe lora for anima and thus no way to do her summer outifts. local pedos are pretty disappointing ngl
>>8915287>>8915318>>8915321you know libtorch is written in c++ right? right?
>anime model>no show-accurate anime lorasWhy is anima like this bros
only mentally stunted individuals would want to replicate show styles
Succubus
>>8916849she stepped in somet snot with her 4-toed foot
>Final_after_extra_detailerwhat causes this
>>8916859it's cui
>>8916849succubae are supposed to be attractive
Interesting. Anima can reliably make mirror images.
>>8917159it can't generate in a style that's worth a damn though.
>>8916849How'd you get the tattoo to apply so well?
>>8917875Gave a toe for it. Law of equivalent exchange.
>>8917875NTA, but you can simply paste a decal yourself and then either img2img or inpaint to make it fit.
>>8918102Shouldn't have grown 'em this big if she isn't enjoying the experience. Bitches never think before acting...
>>8918102purrfect position to launch a big poopie!
>>8912831welcome to local generations, where all hope is futile.
i like 1.1
How do you effectively differentiate features for different characters in Anima? I'm trying the 'to the left: ....; to the right:....;' thing but it's not working very well.
>>8918380If they are tagged characters, usually the hair color is enough.If it's OCs, though shit
>>8917715nice
>>8918331It's pretty good but get rid of those score tags in pos/neg for 1.1. Even master & best aren't recommended as TouchDown Russ baked them in. You're overslopping.
can any1 recommend some good loras for anima turbo? and how to make the model not generate absolutely at random even when i gave the characters name? im pretty new to this.
>>8918380You mean like this? Just prompt like this, and it should make it easy for the AI to know who you are talking about., masterpiece, high quality, uncensored, mmf threesome scene, indoors scene on a bed, 2girls and 1boy, all sitting on the edge of a bed. On the left is miyuki iijima, she has green hair and blue eyes, a yellow ribbon and a ponytail, she looks surprised as she is looking at the other too, she is completely naked with medium sized breasts, she is facing right, and is leaning towards the right, performing a handjob, she is looking at the boy.On the right is minamizato ai, she has verylong brown hair, she is kissing the boy in the midde, she has small breasts and has her eyes closed, a small heart is floating on top of her head, she is facing left.In the middle is yamaguchi tsuyoshi, he has short brown hair, is fairly toned and is sitting between the two girls, he is kissing the girl on the right, he is facing rightm he has his hands on his own knees.>>8918616An anon posted a toml recently. It should still be around, either here on on the other thread.
>>8918649I've played around with your prompt as a base - it seems the point of failure is when you try to make more extreme differences between characters. For example, trying to make one character short and flat and another character short with gigantic breasts - you need to use emphasis to force the sizes and then it leaks those individual emphasised tags throughout the characters.
>>8918669use attention couple for that
>>8918669If all else fails, you can also try using controlnet to img2img. Image related was made using img2img after some judicious use of Krita.
>>8918713why would you need controlnet for a composition that needs maybe 5 tags with anima?
>>8918714So the Ai knows which of the characters is to be how tall and how large the tits should be, in case prompting is not enough. Thus my "if all else fails".My assumption is that anon is prompting a character off-model, thus his woes.
What's better between generating at native 15362 and at 10242 + hires fix, with anima?
>>8918893anima is not very good at img2img so just rawdog 1536^1536
>>8918893Native is better because it gives your gens actual 'room' to generate better images than just lower res and upscaled
How do you guys deal with inpainting halo/desaturation/color banding issue?
>>8918893You can do as >>8918894 >>8918896 suggested or use that cn lite or whatever it's called. It's not perfect but it mogs hires, i2i diffuser/upscale, etc.
>>8919064it's not a problem since 2 years ago bwo
>>8919064Do not use some outdated UI/cumrag node for your inpainting
>>8919179I'm using Forge Neo
>>8918893>>8918894>>8918896>>8919081Conversely I'm doing SD1.5 base resolution with hiresfix 2x
>>8919181oh it's the giga retard from /g/
>>8919181Lol, ashamed your identity got exposed? Don't worry, I will keep quiet, so you can start over.
>>8919183Shh, let him start over!
>>8919183>>8919184>>8919185can I get a qrd
>>8919186Certified beautiful anon, hyped for idol master, cowboy shot, inpainting halos, and girls in bikinis with thick thighs and tummy aches. He is a being of light, if he wandered in here, it's because he saw something in your thread, let him be.
>>8919186>>>/g/109359654>>>/g/109359705https://desuarchive.org/_/search/boards/g.desu.meta/filename/idolmaster/page/1/
>>8912789>2026>Inpainting is still fuckedMight've been my skill issue or that ComfyUI code is shit or that Gradio is shit>>8914626This is the kind of shit you'd see on /trash/>>8915175It just needed the visible black bars on both sides, and upscale artifact>>8919187>not even a single mention of anime coloring, shark teeth, and heavy abuse of official artstyle because I have default Hassaku slop style fatigue>>8918718Is that Roll?>>8917297The artifact and black bars on the sides are what make it soulful>>8918102Fix the number tattoo on her should anon
>>8919181>>8919190Umm henti?
>>8919191Is this the bare minimum for /hgg/?
behold>https://aibooru.online/posts?tags=gary1145149315&z=1
>>8919193Oh he does full body?
>>8919192>>8919190>>8919181It's fine, post whatever you want, maybe like there was some sex in the scene, maybe a used dildo or cum on the thigh or something like that if you are a refugee.
>>8919190>Hassaku slop style fatigueCan you show an example with your own gens? One with "Hassaku slop fatigue" on, and one with your personal "non slop style"? Genuinely curious how you're defining "Hassaku slop." And especially where your slop threshold actually sits.
>splitting by posting in /hgg/ againRetards...
Oh are we posting on /hdg2/ this week? Alright. Didn't get the discord ping.
>>8919199Here's something I made using WAI-AnimaIt was described by other anon as "simultaneously insanely sloppy but clean". Not sure if it looks "Hassaku"-y to either you or me thoughI can post an example of "Hassaku slop fatigue" but that requires me to steal some other fag's slop
>>8919190Does hating on Hassaku give you a free pass on your own slop?
>>8919207Fuck go back SEAfag
>>8919207Didn't realize you needed another anon to hold your hand and show you an example before you could feel "fatigued" by Hassaku slop yourself.
>>8918896I've heard more steps should be used for native high res, but I'm not sure how many
>8919207Guys I hate Hassaku too, can I belong here? I'm not a slopper, I swear! :(
>and one with your personal "non slop style"My personal "non slop style" has always been imitation of official artist's artstyle from that particular franchiseThis one in particular is an imitation of Annin Doufu's artstyle, made using some Illustrious/NoobAI shitmix>>8919209Nah, it's overrated in my opinionRemember AOM3 and Misstoon from the SD1.5 era?
>>8919207d-dios mio.. la creatura mas mojada..
>>8914521What GPU do you have?>>8919197Is anime girl holding a dildo (no sex) also okay?
sirs how do I make sex on pixai? thank you
I want to talk about NAI v5, how long do I have to wait between when the anons decides which /h*g/ to post in?
>>8919224Bake your own thread
>>8919220Yes you can also add a cum jar or she is holding a bottle of cum with a dude blurry in the background. You can go creative in all sort of avant grade hentai
>>8919217>>8919207>>8919192This are very nice, can you post more?
We should all bake individual generals to discuss what each of us wants to.
>>8919231Bake whatever you want, I will post NAI in all of them.
>>8919232But I don't want to discuss that in MY general, anon.
>>8919228I'd post more but I'm afraid that my slop doesn't meet any criteria
Freelance thread engagement poster here, taking posts commissions.
>>8919238Come on, just a bit of coom, don't be shy. We all start out the same.
>>8919238Come on anon, you can add cum to that girl!, It's like adding mayonnaise!
>>8919246You thought of it as mayonnaise? I always pictured cum more like white glue, going off how sticky it gets.
>>8919225This place is dead already, if anything the Nai discussion would be a good thing
>>8919249>stickyWhat
>>8919252My cum is sticky. Isn't yours?
>>8919254Is cum supposed to be sticky?
This the real general
if only these lolcows could see what NovelAI can actually do
>>8919255The magic of cum is that it can be thought of from different angles, as something sticky, as something liquid, but also as something in a semi solid state. Each artist is responsible for giving cum their best interpretation.
>>8919257Yep
>>8919256>>8919260more AIDS infected chimpzilian Miku anus, add some pozzed nigger cum in her asshole please
>>8919262
>>8919256>>8919260inb4 long ass sausage shit
henti?
>>8919264I'm not the one spamming negress shitholes
Noodle won
we moved to /hdg/
>>8919267>sees Brazilian Miku>Mindbroken that he thinks that's a niggerMy condolences retard-kun
>>8919275tsmt
>>8919274mmmmh save some of that pozzed BBC juice for us gurl
>>8919190It is indeed DLN-002 Mk II "Hyper Roll" model.
>this shit-skinned character dressing like a nigger prostitute is not a nigger because... she just isn't ok???lmao
why is blud crashing out?
>>8919283why do you love nigger assholes so much
>>8919283he's crashing out because nai broke him
Fags wanted me to post more so here I am>why are the dildos rainbow coloredIt's a reference to this AI slop meme from 2022https://danbooru.donmai.us/posts?tags=gaming_penis_ikebana_%28meme%29+&z=5
>>8919196Occasionally yes
can some anons post their image gen times? i want a comparison between different gpus
>>89192923080 12 GB, 1 minute 5 seconds for a 1536x1536 30-step image using Euler a and Simple.
on my 1080ti my NovelAI gen takes 3 seconds
on my voodoo 3 the NovelAI page can't load
>>8919292RTX 5050 8GB Laptop, 61.9 seconds for a 1024x1536 30-step image using Euler and Normal (hiresfix upscaled from 512x768)Forge Neo
>hiresfix upscaled from 512x768what
cunny?
>>8919318You know the rules anon, no cunny on /h/
>>8919288Based, we love dicks and NAI here, ty!
>>8919292RTX 4070. 32 secs for an image in this resolution.
put the cunny in the box. NOW
needs more hairpins
>>8919385what does the N stand for?
>>8919391nigger cock love
>>8919391netorase
>>8919391nai
I'm still using IL/Noob since I like using danbooru data and taggings. So should I still wait for better Anima or Krea2?
>>8919407you can use pure tags with anima too
>>8919407krea2 w loras CLEARS anima and it's outdated slop technology
>>8919407just try out the models yourself
>>8919409>krea2 w loraswhat's with this weird krea trolling?
why are shitposters shitposting ahh question
what files do i have to change again top uncap the max res on forge?
>>8919409*its
>>8919190Rika-chama!?
fuck we died again? we are noodle's apprentices
>>8919648classic saturday evening phenomenon. everyone is out partying
>>8919649what losers
while you were partying I studied the pyTotch documentation
>>8919648I'm busy doing things I can't talk about or post.
>>8919656STOP RIGHT THERE CRIMINAL SCUM
>>8919656look at this guy with his fancy V5 testing NDA
>>8919358Catbox? I could never figure out how to do multi-character gens like this.
rip noodlefriend
>>8919934This guide is all you need.https://dskjal.com/deeplearning/circlestone-labs-anima#%E5%88%A9%E7%82%B9
>>8920014isn't that written in the same shit language that the guy who invented LLLite uses?
>>8920014>The Agency for Cultural Affairs' view on the LoRa art style>Artistic style and style are not protected under copyright law (Copyright Seminar, 2023, Copyright Division, Agency for Cultural Affairs, p. 7) .>Creating LoRa itself does not constitute copyright infringement, but LoRa files and images created using LoRa may constitute copyright infringement. Infringement of AI-generated content is determined by similarity and reliance, just as with regular copyright infringement. (Copyright Seminar, Agency for Cultural Affairs, Copyright Division, p. 16) .>Question: 6-18. A LoRA file bearing my name has been created and distributed. It appears that several illustrations I created are being used without permission as additional training data for an existing generative AI, and using this LoRA file allows for the generation of images similar to my illustrations. I would like this LoRA file to stop being created and distributed. How should I proceed?>Answer: [...omitted...] When using a small collection of works by a specific creator as training data for additional learning, the reproduction of these works may be considered to be for the purpose of having the AI generate all or part of the creative expression contained in the collection. In such cases, the reproduction of these works without the permission of the copyright holder and not falling under any other limitation of rights provisions may constitute copyright infringement.>Therefore, copyright holders may consider taking measures against those who reproduce their works for such additional learning purposes, such as requesting the removal of copies created as learning data (injunction against infringing acts) (Article 112 of the Copyright Act, etc.), claiming damages (Article 709 of the Civil Code), or filing a complaint seeking criminal penalties for copyright infringement (Article 119 of the Copyright Act, etc.). kek, japs are cooked
who cares about the wapanese
ai is a ploy by big joo to artifically replace the japaknees
>>8920114How the hell is this supposed to be enforced?
>>8920115>>8920116Even right-winger japs are against AI too, but mostly for copyright reason>>8920114No wonder why so many Japanese artists hate them
gaynal goes to /y/
anime fucking sucks
>>8920464Pixiv sorted by new ahhhhhh gen
>>8920433>>8920464Beautiful character. What's her name?
>>8918439Ram getting RAMMED
>>8919192looks pretty great to me. idc what anyone thinks
>>8920586Haruhi Fujioka from Ouran High School Host Club
>macrocunny
>>8920780VIEW THE SMELL
>>8920780Catbox my nigga
>>8920780lovely anime with a tasteful hint of realism and sexy shading... can we have the cat box??
>>8920780her age???
LLLite models for Anima really suck. With controlnet union for SDXL I could do wonders.
>>8920464love how she put lipstick lines on her lower lips
>>8920929I don#t think any original code from kohya has ever been good
I don't suppose you guys have a guide for absolute retards with no talent on how to draw facial expressions? I find that I am reaching the limit of what I can prompt, and need to use inpainting to get exactly what I want.Please and thank you.
>>8921082>>>/ic/
>>8921082ms paint using emojis for reference
Hey Im trying to generate Squirrel Girl and I realize the dataset is trained on stuff too early for her to be in the database for itIs there another good LoRA thats trained on newer stuff to help generate newer girls?
Is the RTX 5070 12GB a good choice for local imagen? Exclusively for anime/2D works obviously.From reading around people say it's 12GB is too constraining, that even a 5060 Ti 16GB is a better choice, or that you should anyway go with a 5070 Ti 16GB. The issue is that both of these 16GB cards are horribly overpriced now, and only the 5070 can be found at a sensible price nowadays. I don't know how long that'll last so I've been considering getting one before it's too late, as Nvidia is rumored to increase GDDR6/7 price by up to 40%.
(its*)Also, I was wondering how well artist mixing worked in Anima?From using https://thetacursed.github.io/Anima-Style-Explorer/ I got from the OP, I see that my favorite artists seem to work fairly well. Can I hope to mix them, say up to 4 of them, like artist mixing worked in some of the NAI models?
>>8921923>5070 12GBif you're solely using it for genning with noob/anima, i doubt you'll have much problems (maybe speed if you're upscaling)but if you wanna mess around with training or newish frontier models as and when they come out (like krea) then you probably should settle for higher vramwhy not get a 4070ti su instead bwo?>>8921930anima's style mixing isn't as smooth as earlier nai versions ('cause no more clip token bleed), its mixing is akin to nai4. you can mix them through prompt scheduling or use something like github.com/An1X3R/Anima-Artist-Mixer to do some cross-attn mixing (but its not perfect and constrained by the number of artists)if you wanted style mixing similar to sdxl at all costs, you'd have to go the lora mixing approach and train (or get) a lora for each of the styles you wanna play with
>>8921941>[...] its mixing is akin to nai4.NAIDV4 was peak aesthetic imo, and while the mixing was inferior to V3, I don't think it was handicapping per se (especially compared to how horrendous it got in V4.5). If it's on par then that does not sound too bad.LoRA training is something I'll have to be doing sooner or later, as even though NAI gets better, I need more control over the style and I cannot stay dependant on whatever ways the devs decide to stir their newer models. V4.5 already killed too much in my use case and I can already tell it's not going back in V5.>why not get a 4070ti su instead bwo?It's pretty expensive used here, so I'd rather just spend more to get something newer+new.I'll think about it. I appreciate the replies and the resources, thank you friend.
Is ~ 45-50 seconds for [1 image 1024x1024 + x1.5 upscale] is a good speed for geforce 4070 super with 12gb of vram? Trying to figure out if I can somehow speed up the process... I'm using forge with some illustrious-based model merge.
>>8921850If you mean finding a good lora in civitpajeet, no. You are better off making one yourself.
>>8921974It is, but I will give you a mildly unrelated advice first: Do not upscale until you get a good output. Then do the upscaling.
>>8922039The point is that if I'm not using upscale - art looks undone...I mean only after upscaler I can seen what I really like, cuz I apply it with 0.35-0.4 and it change genned image a bit and add those bits of style that I missed on the first generation of the original image. Can I skip it and make image looks the same without upscaler?
>>8922045You want "undone" but close first. After you get the appropriate image, then use img2img and inpainting to further refine your output.
>>8921947>V4.5 already killed too much in my use case and I can already tell it's not going back in V5.I'm curious what your use case is. The V5 samples I've seen seem quite diverse in style and much less flat than V4.5.
>>8921947>I don't think it was handicapping per se (especially compared to how horrendous it got in V4.5).This doesn't make sense. v4.0's mixing was essentially nonexistent, while 4.5 can at least mix on a satisfactory level - which isn't bad for a post-clip model.
>>8922108My use case are key niche artists that simply stopped working in V4.5. I'm not saying V4.5 is worse than V4 obviously.I'm still looking forward to thoroughly test V5 in any case.>>8922111Nah, it mixes worse than V4, in my case at least. Try multiple views prompts for example, you'll see each instance being assigned an artist, instead of all instances having all artists mixed in.
>>8922115>My use case are key niche artists that simply stopped working in V4.5.Yeah, but which ones? If you say this kind of shit, you can't expect people not to get curious.
>>8912789How is krea2 , I'm hesitant to move away from illustrious and so far krea2 looks like a hassle to get around censorship although it's able to generate some very high quality tentacle monsters Miles above anything illustrious ever produced out of the box
>>8922336Wrong thread anons, I'll fuck off to ldg, did a mistake here
Is she alright?
>have like 500 GB of illustrious stuffdid all the cool kids already move on to Anima?
>>8922442illu is still ok if you like simple 1girl'ing / pov stuff
>>8922442Anima still suffer from outdated dataset though (not to mention it doesn't have furshit dataset)
Why does the node manager not look like the screenshots on github and why is the search function not giving me any of the stuff I'm looking for?In what specific ways am I being an idiot here?
>>8922457Neither does illustrious.
>>8922339/ldg/ is basically a bunch of discord avatarfags exchanging their conversations between discord and 4chan to give it bumps.
>add the stabilizer lora at -0.3>get more soulvful gensit does kind of fuck up the colors though
>>8922941liked your old style better. this one feels mismatched between the scratchiness of the linework for outlines and smoothness of the coloration
Noob question, what's the current best checkpoint and inpaint setting/prompt for removing mosaic censoring? Is there one clever enough yet to auto detect without masking like deepcreampy/lada?
>>8923098Hello newfren we moved to /hdg/
>>8923098>Is there one clever enough yet to auto detect without maskingonly an edit checkpoint could do that, but we don't have one for anime yet (anima edit lora probably sucks?) your best bet is using some yolo model to recognize censored areas and automatically mask them. as for the model the currently best one is anima
The current state of hentai generation is extremely disappointing. It looks very AI and full of hallucinations. I expected moreI guess im stuck with generating non explicit erotic stuff with nano banana
>>8923133Nano banana does not look less AI than hentai models though. Unless you judge hentai models by civit sloppers output.
>>8923135>hentai models by civit sloppers output.Where can I check decent outputs then?? because these threads also are full of slop
>>8923136Here you go bro, cream of the crop of local AI https://mega.nz/folder/unJDCRRT#aJDT22UA2yRd_U9RCAIyWg
>>8923146thanks for the dolphin porn
>>8923146old links don't update with new images even when it's the same folder? interesting
>>8923154I wanted to link some more megas but they were all dead x_x
>>8923155oh noes, what a shame :(
>>8923136Also this anon's outputs look very authentic to me, and he indeed uses nano banana to edit in the accessories in a detailed and clean fashion, but the base images are a hentai model (idk if SDXL or Anima though but he started with this style before Anima):>>>/g/108673245 >>>/g/108673253
Bruh, I thought this /g/ post was some try-hard twitterfag, but this was noodlekek too.Look at the quality of the composition of this ART! Like a real anime poster.What happened after the Anima switch......That noodlekek......He is dead x﹏x
>>8923174why are you clown like this, seriously
Now he will post more thumbnails to make a message how he hates /h/ and then more vomit-skinned 1girls to prove that he still got it even with Anima.
>>8923175My sweet lil' Mexican bro, if you consistently uploaded shit like this to an sfw X account, you'd be raking tens of thousands of likes per image. Wouldn't get banned too like our mr. faget.Plus let's be honest, anonnoodles nsfw?>>>/e/3121460 these one-line cleft of venuses are as sexy as a fucking tile wall.Clothes are actually preferable.
I like noodlegens. one of the few good avatarposters (rip ramfag)
Who is Noodlekek? I asked this a few weeks ago, but I still can’t figure it out today. I can’t spot a noodle gen. What is the pattern? Long limbs and Chinese anime like faces?
Newfags should lurk moar.
Why? I’m on /hgg/, a general on decline, you have to welcome me with a red carpet. I could be the next noodlekek.
>>8923253You certainly like to use the on preposition where it goes and where it doesn't just as much. What remains is to gen good images. Pretty sure you will fail this part spectacularly.
>>8923258But isn’t Noodlkek ESL? Wouldn’t that be a good sign for me?
We'll talk more when you post epic gens. If I comment something negative about them it probably means they were good enough.Get lost until then.
>>8923268>>8923278How can I spot a noodlegen?
>>8923174lol at that alakazam that totally isn't hypno in disguise
>>8923158>but the base images are a hentai model (idk if SDXL or Anima though but he started with this style before Anima):this guy's using NAI lol
>>8923308Whoever both of them are, they 100% have nefarious plans for her unsuspecting cleft of venus. Evil, evil creatures.
I'm doin sometin. But also new but also an artfag who understands the power of sloppa. What is the best image2image model in the OP? iyo. Again put yourself in my shoes as an artfag you can draw well and will be using your own artwork as a base to gen off of.
>>8923618>But also new but also an artfagproof?
>>8923622answer my question damn it
>>8923618whatever krita diffusion supports
>>8923624If you want to use img2img strictly you can try anima with a high base resolutionjust be aware that it will mostly retain the composition but not the style
>>8923631>>8923632so it is probably better to go with the krita w/stable and flux stated on the github my goal is to gen animations very quickly as ethically as possible
Favorite AI creators OC's?
>>8923678
>>8923624is this your art? very nice butt
Sir upscale from 128x128 to 1024x1024 work very good on i5-2500k
I shall use my new forbaden powers to create full hentai episode and my dream anime in under a year thank you saars >>8923689ye, I do like tush
>>8923795hope you have fun with AI but i do like the shading of your drawn art
>>8923797That's surprising, my weakest areas are shading/rendering with perspective coming in a close 3rd. My strongest skill is anatomy.
>>8922941>the stabilizer loralink me, famalam
is there a place like civitai except instead of being the finished loras people dump their datasets?feels a bit wasteful for each person who makes a lora of the same character on separate models, past, present, & future, all need to re-do the collecting, tagging, etc...
what did blud meme by this
rare good horse gen
sup peeps, have a wholesome couple
>>8913623I haven't trained anything in over a year since Illustrious. There anything specifc or new about anima or any new applications for it? I see some shit called Anima Trainflow but the github appears to be dead
>>8924912>>8924926cfg status?
>>8924967if you wanted to try your hand at training you can use this forkgithub.com/67372a/LoRA_Easy_Training_Scripts/tree/refreshi've only trained styles, but found anima to train fast (not fried) but is highly susceptible to dataset outliers - a couple of images of an artist's older works can often destabilize the learnt style representation. when training at 1024, i'd recommend you introduce crops of the smaller features you'd like the lora to learn. (unless you'd like to train at higher res)personally i'd recommend training with a trigger tag for styles, training up to a 1000 steps and saving every 25 steps. you can identify the lowest losses out of all the steps, observe where it clusters and test around the lowest 5-10 to pick out a decent step.overall its way less lenient than illu at being able to slap a dataset in and get something that averages between features but anima does produce some decent results if tuned well. anyway, have fun bwo
>>8924993>is highly susceptible to dataset outlierstry smooth l2 log or scaled quadratic
>>8924993>but found anima to train fast (not fried)I need like 1k steps min lol
>>8924993pretty much thisand don't train the TE
wow it seems (((they))) actually succeeded in killing hdg & hgg, lmao. this place is dead
>>8924993Only 1k steps?AdaFactor at 1e-4?
Hay gusy dumb question. I'm trying to generate catgirls but they end up with both cat and human ears. That's too many ears. Any obvious solutions that I'm too dumb to find myself?I tried adding big hairstyles that should cover the human ears, and that works sometimes, but it's not very reliable. Tried a couple different models, mostly illustrious shitmixes, and they all had the same issue. Trying out different kemonomimi artist styles didn't help either. Only thing that seems to work is prompting for miqo'te (FF14 catgirls), but their ears are badwrong and I don't like it.
>>8925137Have you tried using a model that isn't slopped dogshit fried on butiful realistic synthetic data from mid journey
>>8925137what models have you tried bwo?perhaps add extra_ears to the negatives if all else fail
>>8925124and /edg/, though the spam and shitposting patterns were differentalso /asdg/ and /ddg/ by proxy, /csdg/ got killed by mods, /udg/ was never alive to begin with,...
>>8925137first try on Anima, first try on noobAI, didn't even specify hairit's most likely >>8925274 as civit merges all involve some degree of realism bias
>>8925137"extra ears" is a proper tag with thousands of results, use it in negs.
>>8925346he's on illustrious shitmerges though, and illu is pretty bad at following negs
>>8925277Mostly a bunch of overcooked illustrious mixes straight from civit. PlantMilk, arthemy, waiNSFW, atomix, bismuth, novaAnime, that sort of shit. Basically my method of finding shit to try was looking at their preview images - if there were kitty titties present, I gave it a shot. I've stuck with illu because I like the booru tag type of prompting, and every flavor of the month reddit has been hyping up is some natural language prompt stuff I can't wrap my head around.>>8925338If the base model is the problem, what model should I migrate to? I don't want to write a whole novel in the prompt every damn time.
>>8925355There are no "mandatory natural language" models that can produce nsfw anime, everything runs off booru tags. Anima is the current model used by just about everyone itt, it can optionally use natural language if you want.
>>8925355the base model might not be your problem but it is a problem for getting advice, two generations past so only civit jeets still use it
>>8925359Can Noob really be considered a generation of its own when it's just an Illustrious finetune, which itself is just an SDXL finetune?
>>8925365the time we used them for is roughly equal between pony, illu and noob so I would say yesbut either way that's just semantics. point stands, nobody here uses illu anymore
>>8925369Eh, at the end of the day, all three are SDXLslop - 4-channel VAE, CLIP and 1 MP limit and all. Sure, going from Pony to Illu or from Illu to Noob certainly felt like a tangible improvement, but hardly a generational one - the aesthetics improved, the tag adherence improved, but besides those, it was all the same.
hello sirs how do i use krea2 for the xentai
>>8925449I say this.
jesus christ i have a weakness for violet
her age?