Newfag edition, be nice n_n>NAI LEAK SPEEDRUN + SDG FAQ + HDG TROUBLESHOOTING FAQ(EMBED/MAGNET COLLECTION)https://rentry.org/nai-speedrun | https://rentry.org/sdg_FAQ | https://rentry.org/hdgfaq>GUIDESInpainting/Outpainting: https://rentry.org/drfar | https://rentry.org/inpainting-guide-SDUpscaling: https://rentry.org/sdupscaleAnimation: https://rentry.org/AnimAnonWildcards: https://rentry.org/NAIwildcardsCreate Embeddings / Hypernetworks: https://rentry.org/simplified-embed-training | https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Textual-Inversion | https://rentry.org/HNSpeedrun>NEWS AND UPDATES + SD GOLDMINE (RENTRY COLLECTION)https://rentry.org/sdupdates3 | https://rentry.org/sdgoldmine>LOCAL INSTALLNvidia GPU: https://rentry.org/voldy | https://github.com/AbdBarho/stable-diffusion-webui-dockerCPU: https://rentry.org/cputard | https://rentry.org/webui-cpuAMD GUIDE: Native: https://rentry.org/sd-nativeisekaitoo | Docker: https://rentry.org/sdamd | Onnx: https://rentry.org/ayymd-stable-diffustion-v1_4-guide>BROWSER SITEStxt2img: https://huggingface.co/spaces/stabilityai/stable-diffusionimg2img: https://huggingface.co/spaces/huggingface/diffuse-the-restinpainting: https://huggingface.co/spaces/fffiloni/stable-diffusion-inpaintingPaid Hentai: https://novelai.net/>AI BOORUhttps://aibooru.online/>CLOUD HOSTED INSTALLColab: https://rentry.org/sd-colab-automaticPaperspace: https://rentry.org/865dy>RESOURCESArtist styles: https://rentry.org/anime_and_titties | https://rentry.org/artists_sd-v1-4 | https://files.catbox.moe/9wgqx9.jpg (embed) (embed) (embed)Big index: https://pharmapsychotic.com/tools.htmlTags: https://danbooru.donmai.us/wiki_pages/tag_groupsTag effects: https://pastebin.com/GurXf9a4 (embed) (embed) (embed)First pass calculator: https://preyx.github.io/sd-scale-calc/Mix Recipes: https://rentry.org/hdgrecipes>Related>>>/aco/sdg>>>/g/sdg>>>/d/ddg>>>/e/sdg
the poor girl was a bit pent-up...
wish generators could do hands better
I was going for a simple concept and then decided it would be fun to have the algorithm invent covers for hypothetical manga.
Really I just threw tentacles and torn clothes together, and decided the results looked like manga covers anyway, so why not have AI generated titles and labels.The images before that feel a little empty in comparison.
>>10517167yeah, it'll do that sometimes
quad boobpress
>>10517163>>10517165
>futa reimu drawn by ratatatat74now we're fuckin getting somewhereit's as easy as>throw picture into img2img (taken from one of these very threads)>use original prompt + add "in the style of ratatatat74" and whatever else one could feasibly edit into the picture ( i added larger breasts here)>set lora and denoise to around 0.5
>>10517197wish the ones i have could do img2img, but the genetor i have only has image influence
>>10517205I will pray for your financial success this year, that you too may generate infinite coom
>>10517212Amazing.Mind sharing the prompt/catbox?
wish the belly was a bit bigger, generator does chubby/fat stuff somewhat well, but sometimes not big enough
tfw you just wanted some nice girls as food cannibal content, but instead you've created surprisingly good PETA poster. Damn I've got mixed feelings
>>10517240>accidentally type spit roast instead of spitroast
just a little arm jank in the middle one.
Looking into making a Lora and have some questions. Is 512 x 512 the typical image size NAI-adjacent models are trained on? Also, can I assume it is optimal to use uncensored images for Lora training?
>>10517189>>10517185>>10517181>>10517179>>10517176>>10517173>>10517137i can't tell since I'm stuck on my phone, but someone from another 4chan thread said the images are small, are they? because they seem full size to me
>>10517461yes to both, I believe. haven't created my own lora though.
>>10517215AOM2_hardhttps://files.catbox.moe/3wh30n.png
>>10517460can I get the prompt/settings for this one?
>>10517464Good to know. As a follow-up question for anyone in the know, I've seen that it's generally a good idea to include a text file with the same filename that helps the AI identify the content of the associated image. If training a Lora to work with NAI-based models, would it be a good idea to just insert appropriate tags into said text file?
>>10517494using prompt interpolation extension and blueberrymix model. Prompts are something like:beautiful woman, blue hair, tank top, jeans, standing, city streetand the interpolation prompt:beautiful demon woman, blue hair, succubus, lingerie, spikes, intricate glowing tattoos, glowing pentagrams, neon city streetI think this one was the same but with green hair. Above ones were the same with blonde too.
>>10517739Would be willing to do an interpolation one where in the left side there's a couple (a guy and a girl) in some fairly normal setting like a caffé or walk in the park and on the right side she's pegging him? (or the analogous with a futa)
>>10516956Anyone got a chastity lora or hyper network?
>>10517742do you understand what interpolation means anon?
>>10517742These models don't really support multiple people very well yet or allow for specific things like a woman pegging a man without major effort. My corruption stuff on the other hand basically works perfectly.
>>10517739>>10517730>>10517464>>10517460>>10517457>>10517455>>10517452I love your corruption generations.
AHHHHHHHHHHHHHH
>>10517746That makes sense, thank you anyway
>>10517751thanks dude. it's such a niche I think, and this is so easy to crank out content with. let me know what you want to see and I'll try it.
I think the dragon ate her friends.
this one is actually pretty cool.
having some difficult training ass up futa auto fellatio, since im working with somewhat limited amount of images i consider high quality enough,its sucks but i may have to resort to using gayshit like yaoi/bara to train the concept to my specifications.
>>10517752Very nice, love your stuff. Someone mentioned a goblin animation kind of along these lines, do you have it? I'd love to see.
>>10517752How did you do it without a shit ton of flickering ?
>>10517751I've tried some, but i can't do them like that, i made one set of a knight being beaten and corrupted by slimes, but i would have to put them one at a time, and even then supposedly the images will be small even though they look big to me
>>10517760Can you try a girl turning into a robot, or cyborg, or mecha girl?
>>10517816iI's not that similar, but by the same artist.
>>10517760Race change?
>>10517461LoRA training can take any image size, but try to keep the aspect ratio reasonableI try to keep it down to 2:1 at most>>10517533The text files usually have all the tags that would describe what you don't want the AI to learnBut this tagging is mostly done by an AI tagger, and you just prune the tags you want it to learnThis should have enough information, and is also what I followed to learn about ithttps://rentry.org/lora_train#training-lorasPS: Thank you once again, veiny anon
is there a model that can do bdsm stuff like ballgags properly yet?its the only thing keeping me from difuuuuusinggggg
>>10517890>>10513809 from the last thread had a ballgag LoRA
>>10517786Expand your search to Sankaku Booru, Pixiv and Nijie and you might find moreBut even if you give up and get males, proper tagging may teach the AI to do it with futa tooI'd say to go with at least 50 pics, but I think 100 for something decent
Waiting for chastity embed anon~
kinda want to see aomeone try to touch up some of my generations
>>10517786if you can use image influence look of something like you want and have it set to weak influence, you might get the pose you want, but it's not guaranteed
Is any current model good/viable at making lactation stuff?
>>10517932haven't used them myself but there are two LoRAs for ithttps://gitgud.io/gayshit/makesomefuckingporn#lora-list
I know fat Hex Maniac with a penis is as common as Nitrogen is in the atmosphere, but I put in little effort to make it. Love how easy prompting gets using LoRAs.
>>10517760can you try the fantasyraces lora?the races included are I believenight elforcgoblindraenei
i've been trying to look at some images made using novelAI on pixiv and using the png info thing on webui to extract the prompt so I can make some adjustments, but for some reason just repeating the prompt as is gets quite a different result than the original image and i'm not sure why (been trying with pic related)
>>10517212heh
>>10517997>check post date>from OCTOBERshit dude that's basically ancient historycheck this option, don't forget to apply
>>10517730which one is the prompt interpolation extension? i can't seem to find it.
>>10518096They're probably using this script:https://github.com/EugeoSynthesisThirtyTwo/prompt-interpolation-script-for-sd-webuiI think this extension is better but only outputs individual frames and mp4/gif/webm, and no grids.https://github.com/Kahsolt/stable-diffusion-webui-prompt-travelI used the second one to make >>10518095 and a separate script to put it together.
Hyper ass : Best assSaw a face-sitting model on civit, there is something to try
>>10518095First one of those that I like, well done
finally a good fucking thread
have a good day anons
>>10517847here you go>>10518096>>10518100correct, I'm using eugeosynthesisthirtytwo's interpolation. I just prefer the grid output. What script are you using to assemble it for you?>>10518126this is hot.
>>10517860here you go.>>10517967not sure where this lora is, but I'll try some fantasy race stuff without loras next.
>>10517967orc ended up being sort of goblin.
>>10517967night elf is passable
it has no idea what a draeni is. Still hot though.
>>10518216Nicely done, anon!
>>10518174Catbox, sir?
>>10518174>>10518177>>10518179Catbox, sir?
>>10517752God tier, do you have anywhere else you post this stuff? Twitter?
>>10518095can you do another one like this but like just do the stages in between the first and second image, perhaps also having her be surprised and aroused to be growing a cock?
>>10518216not that guy, but it's part of the fantasy race lora in herehttps://mega.nz/folder/ZPkFhbpS#kioBVYEuUsGbJMv1qm883Q
>>10518319ah, thanks. Taking the Incase lora from there because I love his work. Posting some random from the grid folder.
>>10517760I've been working on some of these, but yours always come out so nice. I'd love to see your take on prude demon girl -> slutty gyaru angel girl. I'm having some trouble getting the starting position less nsfw, picrel
>>10518368Mixing those LoRAs can get some pretty nice results.
>>10518216>>10518208more race / robot change?
>>10517848Thanks m8>>10518095Fucking YES
>>10518371have you tried removing nsfw from the initial prompt, or adding sfw
everyone should get the catbox browser script from /hdg/https://gist.github.com/catboxanon/ca46eb79ce55e3216aecab49d5c7a3fbit's only configured for /h/ out of the box but you can easily expand it to every board>>10517939what's the prompt/model for this?
>>10518423Just a heads up on how to do thisCopy line #4 and change it to /d/// @match https://boards.4chan.org/d/thread/*Add stuff to line #16, maybe like this const TITLE_CHECK = ['/hdg/', '>NAI LEAK SPEEDRUN', 'diffusion', '/ddg/'];Then go into Settings and disable auto update for the script
>>10518050tried that and also adjusting the weighted parts of the prompt to have the proper weight in WebUI that they would in NovelAI, but still end up with it looking quite different
>>10516956
>>10517888Appreciate pointing me in the right direction. Some of the instruction seems to be Windows-specific while I am running this on a Debian-based Linux environment and some of the commands may need tweaking to work properly here. Does anyone have experience with this?
Does anyone have new chatGPT workarounds for writing smut. My old prompts no longer work since the Jan 9th update. The first one passes, but all the subsequent ones have chat reply with something along the lines of "As a responsible AI, I am not able to generate explicit or unethical..."
>>10517919Chastity embed when?
>>10518371Oh that's fun. Played around with it a bit - it likes to make demon girls pretty slutty already.
ah fuck, I can't believe I've done this
>>10518553more>>10518401might make some more of this later too.
Very new to this.guide says to download animefull-final-pruned and animevae.pt but doesn't provide a location to download.where is a safe download?Please help I am a brainlet with coomer fog
>>10518564I think the angel side is drawing heavily from a specific image in its training - look how similar this is to the last one.
>>10518556nice, I almost forgot about those sea angels
>>10518577me too, don't know why they popped up in my head after all these years. a semi-interesting discovery - despite being trained entirely on sea angel images prompting without their tags gives a pretty decent clean general style in NAI.
Lora mixing is pog
>>10518584It was just one artist that did them wasn't it? I can't remember his name.
>>10518602https://www.patreon.com/Soutaro_SasizumeA lot of weird shit there
>>10518599I always found that using multiple loras turns it into manmade horrors beyond comprehension, what kind of settings are you using?
>Does she bite?>No, but she can hurt you in other ways
>>10518565iirc it's inside the torrent leak. Basically when you add the torrent to your client (i.e, the program you use to download torrents, qbittorrent for example or utorrent if you're a retard), you just take those two files instead of downloading everything.
>>10518565wait, nvm. you're actually right, i can't find the leak on rentry.It's okay, you can use the anything model instead.
>>10518565>https://rentry.org/hdgfaqThird time is the charm. you can find the leak ont his page.
>>10518208>What script are you using to assemble it for you?a simple python script I asked ChatGPT to write. I generate more inbetweens than I need and pick the best ones, and run the script which just collates all the pictures in a folder into a row.>>10518306for something like a dong, it can be hard to get the diffuser to add it gradually instead of immediately all at once, unfortunately. Even breast change sizes are often very sudden, and sometimes the composition changes radically. Also I suck at prompting expressions. I like your idea though, and I'll try to cook something like this up>>10518565nai magnet link is in second rentry
>>10518657lol it's AI all the way down. You're right about the 'all or nothing' bit too - that's super evident in some of my sequences.>>10518401for you
is amputee allowed in this thread? i've made "asanagi art & amputee concept" thingy but idk if its okay to post the result in here.
>>10518674
>>10518674awesome. more if it;s fun for you, love it.
>>10518681they're basically all coming out perfect, which is nice. This has a lot less variation than my usual corruption prompts, which is probably why.
Anyway here is some result - livestock concepts
>>10518681whoops spilled some demon on it
>>10518696last for now
>>10518679>>10518690>>10518696hecking awesome. I hope you can get the big lips right, or light palms right some day, but so far it's really great.
it got more spicer than i expected after long training
Anyone getting AttributeError: 'dict' object has no attribute 'calls' with pickle scanner?I am using python from venv folder too.
>>10518208Thanks!I realize you are stopping for now, but I'd love to see do some more robots when you get a chance. But again, I like all of your stuff so I'm happy with what ever you generate.
wish my generator did half as good as some of these
For those getting good results with LoRA, what learning rate, step count, batch size, and gradient accumulation are you using?
>>10518747you gotta get a pc, anon>>10518801it's a little outdated, but check out the readme for futaveinyv3 https://mega.nz/folder/0FlxWDJb#hHTSgK3jyivYk65CJV6PaAThe model turned out okay but it's a little overfit, so I wouldn't copy paste anything from there without scrutiny (also it's set up for colab anyway).You probably don't need to bother with multiple folders. I'd start with a higher learning rate than what I had (5-e5 to 1-e4), aim for a minimum 1000-2000 steps. Batch size depends on your VRAM and dataset size, but 3 is a safe bet. Apparently you should not use gradient accumulation unless you can only do a batch size of 1.There's many many variables and I don't know them all, and they're all highly dependent on what you're attempting to train for, and the size and quality of your dataset. Have a look at >>>/h/hdg and their OP rentrys.
>>10518801Check the lora_train rentry, it might be in this thread but it's part of the /h/ thread opening post. The default settings are good enough. Batches are whatever you can handle, and gradient steps might make output worse. The most important thing for LoRAs is having a good dataset (clearly shows the subject you want to train on/ no useless noise like patreon tags or text boxes/ actual backgrounds) with fitting tags. Any number of images can work, you want quality over quantity.
>>10518747if youre using a futa model from civitai it may have trigger words
>>10518698lol this sounds racistfuck the android tf stuff is pretty hot too
look at these fucking cannons lmao
>>10518871She's gotta be able to fire lasers from SOMEwhere.
>>10518877>lost the tummy windowperhaps science has gone too far...
>>10518880you can always have shady legacy modules on craigslist
I'm following the training guide for LoRA, but I've gotten myself stuck and I think something fucky is going on with running the necessary scripts. Pastebin for context: https://pastebin.com/zp0nkpdn
not as good as I'd have liked but I feel like I gotta upload somethingbeing a gpulet is suffering
>>10518869>the android tf stuff is pretty hot tooI know right?
>>10518864I'm mostly forced to use an app since I'm stuck on my phone, i can't use my PC for now, i was using magicAI sonce it had limitless generations, but that suddenly changed, so I'm using imagine and unidream, getting femboys was also hard, but i recently found out using "small male" can get what i want for theost part, also can't get memebers the size i want or orgasms sadly
>>10517746> that first pic> those eyesSHE KNOWS
>>10518126Fucking kek, the last two fucking got me.The dick comes in and she's like "can you wait one fucking second?"
CHASTITY CAGE EMBED WHEN
>>10518500Even before the Jan update those didn't work for meWhat does is telling the AI that it is helping me write a book, then stating the setting has a different set of ethicsBut I do have to avoid explicit terms, so it's nowhere near as good as legends of those workarounds>>10518692>>10518913Those are great>>10518864Is that the monster cock model?>>10518899Try step #4 againhttps://rentry.org/lora_train#how-to-use-this
when pc is too expensive for ai overlords: just use humans - it's cheap.amputee dataset is too hard to collect, I'm pretty sure i have to generate one with img2img to get a good result i guess.still dunno if i can upload spicy one
>>10517240Sometimes cannibal content is gay.
>>10519135>Is that the monster cock model?yeah
>>10518676>>10518692>>10518699>>10518701>>10518887this is based
>>10517752now we're talking
>>10518500Use KoboldAI https://github.com/KoboldAI/KoboldAI-Client
New to all of this. Finally got setup.any model recommendations?
>>10519214AbyssOrangeMix2 and EerieOrangeMixhttps://huggingface.co/WarriorMama777/OrangeMixs/tree/main/ModelsAnything v3https://civitai.com/models/66/anything-v3Anything v4.5 (not really from the v3 guy)https://huggingface.co/andite/anything-v4.0/blob/main/anything-v4.5.ckpt
>>10519259Sorry another noob questionAre safetensor files really safe? Is there a way to scan them?
>>10519265safetensor files are apparently safer, but I'm not knowledgeable about model safety and can't guarantee anythingI haven't got pickled yet at leastYou can check for pickle on ckpt and such with https://github.com/zxix/stable-diffusion-pickle-scanner
>>10518500Check the /h/ thread about it>>10519211Can't recommend it, as it's WAY stupider than ChatGPT from my experience
>>10518692prompt and settings?
The thing was for a possession fetish but I guess it can be used for 100 other different fetishes
Hi, researchanon from /g/ here:I just released a new massive in-depth tutorial on advanced merging.>pic related, collateral damage from my researchIt's still WIP, but should still contain some stuff that you probably didn't know:https://rentry.org/BlockMergeExplainedAny feedback is appreciated!
>failed attempt 7 to get slugs in the right places
>>10518825i have a PC, i just don't have enough room to actually use it right now, that and I'm worried it will die when i can use it agian since it's been off with no updates since late 2018
>>10518100>I think this extension is better but only outputs individual frames and mp4/gif/webm, and no grids.>https://github.com/Kahsolt/stable-diffusion-webui-prompt-travelI've been playing around with this since it looks promising but I'm not sure why the inidividual frames and the final video look blurry as shit, single gens I make without the script and all other things equal look fine.
I don't think snake anatomy works this way
>>10519430the webms it generates are indeed shitty quality unfortunately, but the individual png files work perfectly on my end. I couldn't work out how to modify the extension to increase webm quality or be bothered to make an issue on the github. the readme does mention some tools (including ffmpeg) which i've downloaded and haven't tried yet.mostly I just prefer the extension for the extra options, including being able to have multiple "keyframes", and I'd like to experiment more with the travel mode and frame genesis options.
Sometimes the gacha pays off
In the midst of correctly installing SD-Scripts on Linux and the guide is pointing me to a Windows-specific build of Xformers. Anyone have a workaround?
>>10519458> If you use a Pascal, Turing, Ampere, Lovelace or Hopper card with Python 3.10, you shouldn't need to build manually. Just launch the repo with --xformers. A compatible wheel will be installed.That being said, i can't guarantee this will work. I do not use xformers so i didn't tried to install it this way.i would try to set up everything without it, generate an image to see if it work, and then relaunch the ui with the argument.
>>10519453>00261-*I know that feeling. Because I'm on an old 970 it takes at best 30 seconds to get an image, so each time it's like opening a lootbox.
Sirs I need the male chastity embed or I might die
>>10518697You know when you discover you have a fetish?
how do people create that "transitioning" matrix or whatever? i wanna try making a pic that shows a gal in one of those victorian poofy dresses with a smug expression, only to have the second panel be her all embarrassed with a reveal that it's just a fat ass
>>10519487read the thread before asking questions m8see >>10518100
>>10519580Cute!
Not /d/ enough in here Have some dick-sucking creatures
>>10519338excellent guide to something I'd been wondering about. I'd go insane trying to put it to use on my current hardware, though. Only thing I'd add is a note to clarify exactly what base_alpha is.
>>10519580Impossible!I call photoshop until a catbox is provided
>>10519463Attempted to run the LoRA training script, but it seems like xformers is needed. Tried going about installing it the way you suggested, but the terminal spat out an error code at me as seen in pastebin. https://pastebin.com/wA0kM7sAJust from what I can tell, I may be missing a submodule called CUTLASS. Any idea where to run the command it mentions on line 43?
been tryin to get a TFTG image done of a female kight turning into a male dragon, but no luck, i did end up with this though, dragon taur, anyway, any tags you guys can think of i can use? i even have the genderwap tag but that didn't help
one of my oldest gens, and my favorite mostly because of the nipples, haven't been able to get nipplesz like them sense even with the "heart shape nipple" tag
>>10519803what about "heart areola" or even just "heart tattoo on breasts"
>>10519785ngl, i'm running the stuff on windows, and i'm not using xformers. i know \g\ SD threads have a tendency to fill up very quickly but you will get better advice on them. Still, out of curiosity, what kind of Lora are you trying to train?
>>10519580Is that a kasumi toyama lora?
>>10519811use xformers it will change your lifeI thought because I had a 3090 I didn't need it, but god damn it will change your life.
>>10519832You bet!https://mega.nz/folder/XMgjVDja#hSYAT_EsNSokys7oAIKm2A
>>10519681 https://files.catbox.moe/8ot3e0.pngThe image was cherry-picked, but this LoRA does a pretty good job of capturing shemales at various angles when you experiment with some of the negative prompts, such as "testicles" or "disembodied penis".https://mega.nz/folder/XMgjVDja#hSYAT_EsNSokys7oAIKm2A
Holy shit, I did not expect this to work.
>>10519967Trigger word is cyclops. https://mega.nz/folder/47Yj3ZIS#klaoBwVZI_u5DbjmCjkqRQ
posting only because I was amused that the text on the headstones came out
man I wish upscalers didn't eat hands for breakfast
>>10519307I can't share the setting because the tuned model is so unstable due to my coomer brain.I'll share fully replicable image when it's done. read the image if you want to see the prompt
also here is some hires image that almost melted 3090
>>10520028>almost melted 3090>only 550kb
>>10517752Any like this with penis ass tits expansion? Hot!
>the only decent looking dicks I roll aren't attached to anythingSon of a bitch.
Baqua LORA soon.
>>10520203>>10520206>>10520210i have a hard time getting fat like that, even when i use the tags fat/chubby or round belly, what do you use?
>>10520236I use this model:https://civitai.com/models/3748And I trained it on baqua's style, but it gives fantastic fats on its own, with bellies similar to that. Try belly grab, fat, deep skin, kipteitei with it for great squeeze.
how does this lora shit work anyway? is there a guide?
>>10520282In fact, there is, although it is as short as it is simplehttps://rentry.org/lora_train#for-using
>>10520271sadly I'm not on pc so i probably can't use cicitai unless anyone can say otherwise, I'm stuck using imagine or unidream, i got this, but wanted the belly bigger, it was part of a pokemon trainer weight gain group i did
>>10520299Good news, you can. There are a number of colab or paperspace based ways to use SD.
Had to use img2img , but turned out quite well.Trying to train my own LoRA atm, but when using the trained output, it doesn't have any effect on the generated image, so probably still messed up somewhere.
>>10520339can you link to them? i normally end up finding "wumbo" which doesn't do the stuff i want
>>10520350Every time I try to use img2img I get either blurry results because the denoising is too high, or useless results because the denoising is too low. I'm really new at all this though.
>>10520353set it to 50-55% that could help
>>10520353You should make an inpainting model>on the checkpoint merger tab: put the normal SD 1.5 model in A, your model of choice in B and the SD 1.5 inpainting model in C>choose "Add difference", set to 1 and start (you need to have "inpainting" in the result's name for it to work correctly but the latest build does this automatically)You should also know about Inpainting conditioning mask strengthhttps://youtu.be/VdDhGjkKD1A
>>10519811I actually managed to find a resource to get around the error I encountered and have xformers installed now. kohya's script still refuses to recognize it though, so I'm not sure how to get around that.As for your question, I have a thing for multicock futa and want to see if the AI can be trained to generate that better.
>>10520352There's a link to colab in the OP
>>10520374do i just down load it? it's also asking for a sign in(I'm on incognito mode)
>>10520384i should mention i am not a smart person
>>10520385It's all online, you need to be signed in to a google account. Just make a fake one.
>>10520353Denoising shouldn't affect the definition of the result unless you are doing single digit steps or something as dumbMaybe it's your CFG setting what's wrong? It's better to keep between 6 and 11. Sometimes I accidentally give it focus while wheel-scrolling down and drop it low enough to get garbage results for a while until I realize what's going on.
>>10520271It merges pretty well with the orangemix at 0.3
>>10520369Hope you'll find a way to make it work.pic related was totally accidental but i kept it for some reason, It's not much buy maybe you can add it to your future dataset.
>>10519915>kasumity anon this is really good
>>10520449maybe you have it in the prompt that I can't see from the filename but you'll have better luck with cumming if you have "ejaculation" and "projectile cum"
Where can i find archives of hypernetworks? im new to this and all i know is that hypernetworks are like mini models to layer on stuff
>>10520528https://rentry.org/hdgfaq | https://rentry.org/hdglorarepo | https://gitgud.io/gayshit/makesomefuckingpornYou should also look into LoRAs.
>>10520413>Just make a fake one.Not that anon, but it wants a phone number.
Do you want some threesome ?
>>10520692Huh, it didn't ask for one when I made my fake account. I guess I was lucky.
>>10516956Why the fuck is there not a zip file with an exe in it that I can click to do this? Why does every fucking project expect me to compile the source code. I'm fucking lazy, why do you think I want to AI generate art????
>>10520730Because the code compiles differently depending on your system. Just use a google collab if you want braindead mode.
All right, model. Calm down.
>>10520057Never done ass before; here's a test.
>>10520870BRAAAAP
>>10517752that's fucking incredible
>>10520736this is ai generated? holy fuck what model?
>>10517903what models do you use to get such good futa?
>>10518095which ai do you use? which models, is it a custom model?
>>10519184>>10519135where do i pick up the monster cock model?
>>10520918Get the Catbox script mentioned here >10518423https://files.catbox.moe/52trq8.pngIt's a combination of LoRAs, but the futa part is from the veiny and futa loras, I believehttps://mega.nz/folder/Obp0nIoR#5IzAbrHX3wkVYA7vo6sfNw>>10520928https://civitai.com/models/4165/monstercock
>>10520353new inpaint kinda sucks balls desu, just get a separated old commit just for inpainting its a good way to get access to old useful features as wellalso too much hyperfuta on these threads, post more tiny cute peckers
>>10520987whats the best sampler for making good futa? or hentai in general?
>>10520727Noice.
I'm not sure, but this should be some new farm irrigation method
help I'm retarded, what does asking for a catbox of a pic actually do
>>10521082on generation, full generation settings are written into exif. 4chan strips exif on upload. a catbox of a pic lets you drop it into the pnginfo tab and copy settings exactly, so that you can play around with it yourself.
>>10521087aw shit thanks for answering a retard
>>10520736was thinking of a kawahagitei model
>>10520736is this a lora?
>>10521016samplers are pure preference, it's all about the model and getting lucky with the seedthe way I go about it is>roll gacha seeds with euler/euler a because they're the fastest>once I get a nice composition, redo it in one of the dpm ++ karras samplers
>>10520369>>10520449Okay, progress report. The solution to getting the script to recognize xformers was to clone the git repo into the sd-scripts folder itself. Unfortunately, a new problem has arisen that's likely going to be a more major hurdle. Essentially, the kohya scripts are written specifically with a Windows environment in mind and although Bitsandbytes was originally written for Linux, the scripts seem to use the Windows-compiled binaries seen in pic related. I'm not entirely sure how to go about this, but I theorize that it would be possible to modify the kohya scripts to use the original libraries as they were compiled for Linux. As always, insight is welcomed. If a solution is found, this could benefit everyone in the long term.
>>10520916AOM2_nsfwhttps://files.catbox.moe/6nfjrh.png>>10521101>>10521107It's a lora I'm trying to train, will post when finished. Tag pruning this artists' concept is a shitshow (too many things at once)
>>10519785>>10521198Oh boy. This was me yesterday. It took an evening to get it all organized and set up right, including figuring out auto-captioning with WD1.4 for the 2,500+ images I wanted to se to make a LoRA.Only to find my computer couldn't handle even 1 epoch.
>>10521263thank you for not leaving me hanging anon, he is my favorite h artist. only artist iv ever paid for
>>10521268Same, I find it very difficult to find others that can do cum inflation like him.
>>10521268his pokemon arts are godsend. if anon makes this ill try reskinning a pokemon game in his artwork. myabe chuck in a few loss conditions
How do you get the faces to not look shit?
>>10521263what tags did you use to get cum inflation?
>>10521291>It's a lora I'm trying to train, will post when finished. Tag pruning this artists' concept is a shitshow (too many things at once)
>>10521263>>10521274very interested anons
>>10521288use the waifu diffusion VAE, and check high-res fix (makes small details better in general)
>>10521291cum, inflation, overflow, stomach bulge, navel, belly, cum pool
Hyperbreasts + Huge Nipples LoRAs are a scarily good combination.
>>10521288Practice.Include some negative modifiers like:>(ugly eyes, deformed iris, deformed pupils, fused lips and teeth:1.2)What this anon said also helps:>>10521310The high res fix adds a lot of processing and time, though, so most people save it for after they've found an image they like and want to improve it.The general flow for most newbies is to generate a batch of images with a quick generation method while fine-tuning parameters. When you see something you like, feed that image in and try to refine it until it looks great, THEN use the high res fix to upscale it.
>>10521344>inconvenient breasts:1.2does that even do anything? Is it an activator?
>>10521344w2c huge nipple lora
>>10521347No idea. I stole that idea from another anon's prompt.>>10521349At the very end of this list https://gitgud.io/gayshit/makesomefuckingporn#lora-list
>>10521350fantastic resource, thanks anon
>>10521350>https://gitgud.io/gayshit/makesomefuckingporn#lora-list>concepts:shortstacks
>>10521366anon's almonds are active, everyone
>>10521344Previously, it took me a shitton of inpainting to get good gigantic tittes. Now, I just press three buttons to get better results. I love technology.
>>10518701please for the love of god share the LORA i need human cattle!!!!
I saw something about a tentacle HN on 5ch, anyone happen to know where that could be?
>>10521381i used to photoshop them bigger then sending them into img 2 img
God img2img too powerful
>>10521366There's nothing more amazing than checking out a mega link for a lora and finding a completely different lora that you didn't know existed
Quick question for confirmation: what are the current nvidia graphics driver and CUDA dependencies for running sd-scripts?
Why is it that pic related keeps happening at the last step? It is usually fine during the webui previews until before the last step.
>>10521611try lowering cfg
>>10521611It looks like what happens when an image is cancelled before completing the minimum steps.
>>10521614Tested it with 11 and 7, no luck. Will try lower next.
>>10520698Gimme that catbox actually.
messing around with 7thv3holoabyss, surprised at how well it natively understands futa
Damn can we get a Rebis LORA? Specifically for the very dark but saturated skin tones he does? I feel like his choco-girls are unmatched and any attempts at dark skin in the past were very weak.
>>10521656Honestly I'm trying to figure out more robust ways to increase bulge sizes for use with normal outfits. Right now with the model I'm using if I ever get a bulge it's definitely pussy lips, or a package the size of a god damned golf ball.
>>10521661I suspect a Lora's gonna need to be made, because I've never had great luck with bulges on anything except for NAI's furry model, and even then it simply can't do them erect or too large.
>>10521662I'd gladly use a bulge one. I'd make it myself if I didn't have my hands full with finishing up another LoRA
>>10521661>>10521662Yeah I've been in the same boat, like anon said someone's probably gonna have to make a lora for it to actually get good, every model I've tried has failed at itIt's on my list of loras I should make but I've been slowly chipping away at a giant insect lora, manual tagging is a bitch and trying to train different species isn't working out so well but it's probably a skill issue because It's my first one and there's so many different settings to tweak, if it ever gets good enough I plan on posting it here
>>10521669It doesn't help that many many artists don't draw tenting erect-bulges, and if they do, sometimes it's drawn what I'd consider 'incorrectly', if only we were able to train models off of real pics and somehow generate that concept as drawn images.Seriously, check rule34 and e621, even kinky places like that have a small selection of properly erect-tenting bulges. Too many are drawn 'sideways' / tugged up to the side.
>>10521267II believe I'm having to relent here as well. Figuring out how to get LoRAs to work on Linux is starting to become an extraordinarily deep rabbit hole, delving into concepts that would take a great deal of time to even begin to understand.
any news from kawahagitei anon?
>>10521686still fixing some issues, the x-ray is a little detached from the stomach sometimes
need more super tiny flaccid peckers
>>10518589damn. need more chastity for SURE!
>coom>start creating a tanned elf farm living tradwife>hours pass>start making her hotter>give her a dick>repeat
>>10521692does it always xray with the lora?
>>10521703living the dream
>>105216567thv3holoabyss?>>10521694I should try training a LoRA for that again...>>10521703Same experience with ChatGPT
>>10521708you should. the tiny flaccid pecker reigns supreme. needs to be tinier though
>>10521708https://huggingface.co/HentaiDiffusionGeneral/7thHoloAbyss/resolve/main/7thv3HoloAbyss.safetensorssomeone posted it on /h/ a few threads back and I'm giving it a spin
>>10521704no, but the higher the weight you use for the lora, it is more likely, along with inflation size too.
>burned through all my google colab resources making the dumbest shit>didn't even think there was a limitI'm so fucking stupid. I don't need 1500 images of some ice witch squatting on ice dildos surrounded by cock shaped ice sculptures. I'm not spending money. Any idea how long it takes to get more poverty rolls?
>>10521717brilliant, once released ill start my poke project
>>10521611Tested several cfg levels and sampling methods and facial restorartion. It seems it has more to do with the random seed, if I get a god seed I can change the other parameters freely, if not there is no fixing it.What is really getting my attention however is how only the last pass is being affected. I found out I can right click the preview right before the last pass and save it as index.png, those are fine.
>>10516956>>BROWSER SITES>txt2img: >https://huggingface.co/spaces/stabilityai/stable-diffusionI tried this and it gives an error if you try to create /d/ content
>>10520870Tried to follow in your footsteps with the prompt-travel, but I'm lost. Any tips on how to achieve such pleasant pulses?
>>10521892I think I'm getting closer but still lightyears behind. What I'm doing:Prompt travel with successive genesis at 0.65 denoising.Prompt looks something like this:>base prompt>base prompt (transformativeA transformativeB)>base prompt>base prompt (transformativeA transformativeB:1.1)>base prompt>base prompt ((transformativeA:1.1) transformativeB:1.1)>base prompt>base prompt ((transformativeA:1.1) (transformativeB:1.1):1.1)>base promptAnd travel steps something like >10,2,10,2,10,2,10,2The problem here seems to be that the "base prompt" only prompts (to pulse back) are too strong and quickly coalesce to a consistent form despite the few travel steps and relatively low denoising.This time I managed to fix my kdenlive to make an actual video.
>>10519959Why is your regularisation dataset so similar to the training dataset? I'm new to this, but I thought the regularisation images should look something like the model you train on can easily reproduce and furthermore most likely produced by the model itself.
>>10521930Tried a different workflow. Prompt looks now something like this:>base prompt>base prompt (transformativeA transformativeB transformativeC ... transformativeN)>base prompt (transformativeA transformative B:0.5)Travel steps just 20,5.After generate, if output is garbage, delete output folder, otherwise send last image to img2img.If outputs converge, increase intensity of middle and last prompt, or add one transform to first prompt.Repeat until satisfied or outputs become too deep fried from the intensity.https://files.catbox.moe/b6qx35.webm
Can I ask a request?A futa schoolgirl teaching educationally how to use a condom
>>10521892>>10521930he mentioned a few threads ago - but he's doing LOTS of work after generating images to seam them together. It's not a straight output like yours are. That being said, the other prompt interpolation plugin makes much better gifs.
so I love the incase lora.it also feels a lot like I'm stealing from him now though.
mommy milkers
>>10521800theres https://dezgo.com/ which does for the most part but i don't think it's as advanced as somevof the generators being used, that or i don't use the right tags
Can you mix a dreambooth model with a non-dreambooth model? I want to be able to mix that Analog one on Huggingface with one of the anime ones.I want Aesthetic™ softcore pictures of my husbandos with hipster film camera colors while being able to take advantage of Danbooru tags to make it actually look like the characters I'm aiming for.
>>10516956Holy mother of cock
>check /h/ thread>futaba lorafinally... futa futaba...
the whole inpaint sketch > copy to inpaint > mask and generate has really sped up img2img for me, one of the best additions in my opinion
>>10522275I've got a folder of transparent dick&ball combos in various sizes and angles for quick copy/pasting onto images that have everything else perfect except the one thing I can't tard wrangle the AI to give me.
been trying for hours to get the anything v3 vae working, the new safetensors one and i just can't do it bros. every single time i try to load it with a1111 it fails. so mad at this point.
Almost done with my worm LoRA
>>10522447how exactly do you make your own lora?
Does anyone have any experience creating cock comparison pics for "futa is always bigger" kind of stuff?
Calling my first LoRA finished, it's not perfect but it's good enough. I don't even know how I'd finetune it more, I only found like 13 good images out of the 6 boorus I looked through, guess I'd gen pics and touch them up and throw them into the training dataset or somethingAnyways, Cocoon LoRA.https://mega.nz/file/ekhVkKDY#h5A_dfK3wbZXpOyGdMQMXnDjZUq8E1hoV1gsdUFPnOMhttps://files.catbox.moe/dlocx7.pngKeywords are;cocoon, silkgag, silkhead, silk, spider web, suspension, boundSeems to work best on AoM, 7th layer and NAI, has a tendency to turn gags into masks but putting gag, mask in the negatives helps
>>10522083>>10522037>>10522052>Breast expansionPlease do more. And maybe do one where the girl is topless at the end.
Is this https://mega.nz/folder/s8UXSJoZ#2Beh1O4aroLaRbjx2YuAPg/folder/cksUVKzDand thishttps://mega.nz/folder/ptk1wIbY#LOi45PfA1dqBiL5QUQW4LwThe same thing? Don't want to download duplicates.
Any info or tutorial on 3D compositions like:>https://www.pixiv.net/en/artworks/104848953
>>10522620yeah look at the datasets, they're identicalthere are 3 futa loras around and they're all pretty goodthat one, loranari and veiny_penisv3
>>10522541so its just a silk lora?
>>10522701yeah basically?to be honest it's just a consolation for not being able to get a giant spider x female lora to work, I spent days fucking around with settings and tagging and every one turned into body horror with a 10% hit rate at best, so I figured I'd at least get something out of all the work
>>10522722what were you using to make the lora?
>mfw I finally figured out how to get my prompts right
>>10522735aom2? i cant get prompts right, please guide me senpai
>>10522733https://rentry.org/lora_trainused the popup
>>10522739I was making a bunch of stuff that looks like this.masterpiece, best quality, detailed, absurdres, highres, (((Professional))), (((Solo))) Naked Smiling Elf ((Wearing a Large Bulging Penis on Crotch)), (Dark Brown) Skin, Slender Waist, Wide Hips, ((Abs)), Gigantic Breasts, (Long) White Hair, Blue Eyes, Blue Glasses, (((See-Through))) Formal White [Dress], Library SettingNegative prompt: lowres, bad anatomy, bad hands, text, error, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, username, blurry, 3d, nose, Window, Blue Pink Green Black Dress, Heterochromia, 1boy, 2boys, 3boys, 4boys, Disembodied, FloatingSteps: 16, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 3087729382, Size: 512x512, Model hash: 0fc198c490, Model: AbyssOrangeMix2_hard, Batch size: 4, Batch pos: 2Also using the OrangeMixVAE. The trick is to keep everything under 75 tokens, consolidate as much space as you can by grouping similar concepts together. Use as few commas as you can because they eat up tokens. I was having the hardest time avoiding multiple dicks and balls too until I later added something to that effect into the negative prompt and added solo to the positive prompt. "Wearing a Large Bulging Penis on Crotch" definitely tricks AOM2 into making futa without any LoRas or training. The bulging word is your workhorse because it will really help avoiding small dicks. Another important point is you'll want some contrasting colors in your prompts because if too many colors are similar you'll end up with a lot of stuff melding into eachother. I was having a hell of a time making an Egyptian themed dark skinned futa in the desert having a dick because it kept on disappearing into the Abs prompt. I changed the setting to inside of a building and added some blue smoke for contrast and the problem went away. Stuff like that.
>>10522744gotcha, i have been using this prompt for a minute simply changing the style and characters. guess its way too long since i keep getting fucked up renders.2girls, nico robin, nami, one piece, boob press, ((muscles)), steroids, (midriff), navel, abs, blush, ahegao, huge ass, wide hips, massive breasts, large nipples, beautiful lighting, masterpiece, sexually explicit, cum, (((bukkake))), nude, (microbikini), lactation, blue hair, subsurface scattering, weightlifting, workout, tattoos, erect sawaru
>>10522745I also have a basic style that I use to prefix all of my prompts.
wish i could get futanari more consistently with the generators i use, phone apps are nerfed compared to the online gens, but they're free for the most part
She will be goth soon alhamdulillah
>>10522744>3087729382ty, i used the same prompt and seed and am getting decent results, i will tweak it a lot.
just cum
>>10522822Looks like you're not using the VAE I'm using. Your colors don't look as vibrant.
>>10522953im most likely using an entirely different UI. im just using the basic bitch stable diffusion UI. i dont have any sliders and no access to loras atm. its version 2.4 so idk. i tried installing the newer one thats in ops post, but cant figure it out.
>>10522813my brother in corruption sequence
>>10522813>>10523130Can you generate these with a Prompt S/R grid that just replaces a word like "goth:0.0" with "goth:0.2", "goth:0.6", "goth:1.2", etc?I'm too new to know if you can use goth:(0.0-1.0)(+2.0) to generate 5 increasingly gothy images, but now I want to try.
>>10517197lol
>>10523169Yes, that's possible. I don't recommend doing 0.0 weights, it tends to fuck up the first image.I use the prompts interpolation script from github. Easily the best option imo. https://github.com/EugeoSynthesisThirtyTwo/prompt-interpolation-script-for-sd-webui
found a new site to use, not perfect but pretty good https://waifus.nemusona.com/
>>10523226Most people know about it, but I don't like it anymore because it switched back to AnythingV4. I prefer AbyssOrangeMix2.
>>10523205pointing out the hands is the N word of AI art
>>10523254Wish I could bruteforce reroll hands quickly but it would take way too long.
>>10523254Where do you think you are, nigger?
>>10523271It's probably easier to draw a basic outline of what you want the hand to look like first, THEN reroll, it worked wonders for >>10522275 the right girl's peace sign was more like a W, with the thumb being as long as a finger and regular img2img would try to keep the W, so I painted over the existing thumb and made a thin black line approximating how I wanted the new thumb to go
>>10523278You think I can draw?!?
>>10523281you think I can???BASIC outline nigga, if I can do it so can you!
>>10517131fixed eyes on photoshop
>>10523287I'm more worried about having a functional set of genitals then the eyes.
>>10523316>>10523287>>10517131If you can't handle a girl with a thick, high-capacity seminal vesicle and vas deferans you are a coward.
Ah yes. Thank you, ma'am. That is how I like my beer.
Anyone testing trying out this "new" PLMS sampler that popped in the latest webui? (a quick search sees it mentioned mid-2022)Its really fucking fast (like 300%) but without at least double the steps as usual it tends to leave faces undone or fucked up
>>10523388"new" is right, it's been there since the beginning. https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/3324f31e8463434796028c7bee6e695ef95c0681/screenshot.pngIt's one of the samplers with flaky support so some recent commit might've made it available for your system now. Either that or you never noticed it.
Gentle reminder that lolis are not allowed on any board other than /b/. Please don't get banned.
>>10523433oof you right.
>>10523485Goblins look like that where you're from?
>>10523485
how the FUCK do I make dezgo generate futas and/or traps reliablyI can get one decent one out of dozens of attemps, so I know it does do it, but it's just not fucking reliable at alldon't tell me to run local or use collabs, those aren't viable for me
>>10523487i get a variety of results from the goblin prompt.
>>10523490you cant run it locally?well unfortunately, you are kinda fucked.
>>10523490you're the only one here that uses those services. everyone else uses local or colab. you're trying to get blood out of a stone.
any love for blowjob faces?
>>10523562if you're not already using it there's a vacuum blowjob LoRA in the gitgud you might enjoy
>>10523567what's the activation phrase for that, by the way?I've tried vacuum blowjob and vacuumblowjob, but even at 1.0 it doesn't affect my outputs at all.
>>10523627suckface
>>10523630bless you anon
fresh from the porngineers of /h/((manga page)), (multi-panel sequence), ((speech bubble, text))
>>10523794mind sharing catbox?
>>10519959hey this seems to work well with other art styles. but these below don't seem to work well at all with not-anime models:>>10520987>https://mega.nz/folder/Obp0nIoR#5IzAbrHX3wkVYA7vo6sfNw
What's the best model to use at the moment? Been out for a while.
>>10523797https://files.catbox.moe/fygvu5.png
>>10523807abyss orangemix 2 is almost unanimously considered the best at the momentNSFW, hard, or R34-SF1.1 are the variants of choice depending on your needsthe current hot trend is to focus on LORAs rather than models
>>10523794i actually got the face in picture thing by complete accident once, i forget what generator i was using and what tag, i forget if i still have it or not I'll have to check
so whats the best futanari lora? I know thats subjective so I'll clarify that I like them full-package(pussy+cock+balls), so not thrilled by that shemale lora
Soon(tm)Testing it with the futa lora's atm.
img2img, on manga scans is dope
Think it's the hyperbreasts LoRA trning my loopback dark, or is this an effect of too-low denoising for too many frames?
How is this thread still alive? It's been on page 10 forever.
>>10524068Coming right up.
I know it's basic, but it I just never get tired of impressionist-inspired cowgirls in domestic scenes.
>>10524108I have a group of six characters I create in different scenarios. I typically keep ~50 images of them. It's helping me find out how to be more efficient with prompts because the more I learn the more complicated things I'll be able to reliably make.
>>10524117I know what you mean. I find myself kind of modulating my prompts - adding things to try and capture a certain effect, then removing the things I'm not really looking for anymore, stripping out tokens to be able to fit more in. It's perfect comfy.
>>10524059/d/'s always been one of the slower boards
>>10524068Catbox me on that one. One anon made a nicer futa lora that I've been combining with the veiny futa one at low str to some pretty decent results for a consistently meaty dong. It's just adjusting the image to not look like it's run through a blender.
>>10519974holy shit, model/prompt/lora?
