Optimalhentai 39 Posted November 7, 2022 Share Posted November 7, 2022 @ihcah Nice, definitely a lot of good things to think about. I havent even thought of running an image through multiple seperately trained models or networks yet for some reason. Must give that a try I havent had much time at it as I just got stable diffusion working the other day but yes, on the current training I'm using anime abdl art images and using waifu diffusion for the initial generation I wanted to see what would happen if I showed a hypernetwork how diapers look in some anime contexts with a bit of variety and inevitably differing styles. So far I think its got an okay sense of them but maybe with other networks it could refine them even more. I'm completely new to this and I dont know if I'm able to do inpainting yet or if I need to download a module or something but it sounds super useful. It's only one network I've got so far (in varying stages) but I couldn't make the generations wet or darken much by editing the prompt and going through img2img. I'll probably try some training with real images before refining for anime,that sounds like a good idea. it'd provide the model with much more information and detail that it could use or discard. Thanks for sharing what you're doing, I feel like this has a lot of potential. Looking forward to working on it more when I get some time ihcah 1 Quote Link to comment
Optimalhentai 39 Posted November 13, 2022 Share Posted November 13, 2022 Fruits of some fiddling Diapheure and Varys 2 Quote Link to comment
ihcah 485 Posted November 29, 2022 Author Share Posted November 29, 2022 I worked for a while on a new one today when I got inspired thinking of an old favorite image.I'll show my process a little on this one, for fun. Here's where I started - it's a great shot from Animal House, where they go on a panty raid and he watches this beauty in the window: I started by doing some conversions on that photo using my model (the one mostly based on Always Discreet panties) and just had it paint over her panties to replace them, so it output things like this: Then I asked that to attempt some different illustrations, which it did: And ran those through an anime-trained model to get further in the stylistic direction: Then I moved those back through the higher quality model and upscaled as well, to get things like this: Then after many many iterations I combined my favorite versions together into a photo tool and used the layer blending to highlight a few different features, and finally end up with my final dry(ish): And final wet: And that's it, I have something I love now, that fulfills exactly the fantasy with which I started: "what if that famous panty-raid scene had the main college girl wearing a... different sort of panty?" Myralee, ral, tw82 and 3 others 4 2 Quote Link to comment
ihcah 485 Posted December 1, 2022 Author Share Posted December 1, 2022 A couple of hybrid ones I made recently, because I love designing product packages using AI art + photoshopping: I think there would be fewer complaints about AI art from artists if everyone would try creative things, instead of just making one prompt and posting a ton of instant results online. kofi06, rachelkirwan and MM2000 3 Quote Link to comment
Azazbel 14 Posted December 12, 2022 Share Posted December 12, 2022 (edited) Wow, I'm sad that there wasn't a lot of posts lately. I had great hopes! Anyway, I've put my hands in this and tried it. Sadly I think a diapered anime model doesn't have a bright future, while I looked for references I found that: - Most of the diaper content is either medium-low quality or very very specific art style. - It's very difficult to find very similar diapers in the okey quality pics. - There is very few pics of solo subjects, not sure if multiple characters in the same pic would fuck up the training. I couldn't find a lot of pics so my results are really hit or miss, below are the bests of the bests, after training 13pics with dreambooth on anything3. I'm now hesitating to train sd1.5 with real pics and then merging it (like, 90%anything3 10%sd1.5), but I'm sure it will have a lot of loss. If anyone could tell me if they tried it already, that would be nice. Spoiler I wonder if the poor quality, especially the last one, is due to the vae or the training pictures. (of course the prompt was pink haired and pose, it's not a bias). Not sure if I was supposed to train on pictures of a whole character with diapers, or zoomed in on diapers. (plus sadly I can't make the dreambooth colabs work today ;_;) Edited December 13, 2022 by Azazbel so I don't double post (see edit history) ihcah 1 Quote Link to comment
ihcah 485 Posted December 19, 2022 Author Popular Post Share Posted December 19, 2022 @Azazbel Good questions and findings (and your results are very nice)—I’m happy to have another person trying this and compare findings. As you probably saw in one of my previous posts, I used real photos with Dreambooth to nail the exact look and everything, then I use anime models in later stages. I did try merging before and that works okay but usually I just use the merged model in between as I try to transition the style without losing the diaper. After reading your post I want to try new things... not sure when I’ll find a chance over the holidays but I’ll post when I do. bonus: I made these recently from my model trained on photos of real women’s pull-ups, don’t think I posted them here yet: dmote232, oienacs, Incontsecret and 5 others 4 1 3 Quote Link to comment
rachelkirwan 13,626 Posted December 19, 2022 🌟 OmoOrg VIP Share Posted December 19, 2022 On 11/28/2022 at 8:31 PM, ihcah said: I worked for a while on a new one today when I got inspired thinking of an old favorite image.I'll show my process a little on this one, for fun. Here's where I started - it's a great shot from Animal House, where they go on a panty raid and he watches this beauty in the window: I started by doing some conversions on that photo using my model (the one mostly based on Always Discreet panties) and just had it paint over her panties to replace them, so it output things like this: Then I asked that to attempt some different illustrations, which it did: And ran those through an anime-trained model to get further in the stylistic direction: Then I moved those back through the higher quality model and upscaled as well, to get things like this: Then after many many iterations I combined my favorite versions together into a photo tool and used the layer blending to highlight a few different features, and finally end up with my final dry(ish): And final wet: And that's it, I have something I love now, that fulfills exactly the fantasy with which I started: "what if that famous panty-raid scene had the main college girl wearing a... different sort of panty?" WOW This is an amazing process, thanks for walking through it! Quote Link to comment
ihcah 485 Posted January 15, 2023 Author Share Posted January 15, 2023 Combined a lot of iterations to get this Daisy. Sort of a plastic pants look in the end but I really liked the overall style. Used about 4 different models. DuffMan 1 Quote Link to comment
ihcah 485 Posted January 15, 2023 Author Share Posted January 15, 2023 And a second version... I usually make like 50 and then pick only my favorite rendering / combination but sometimes I look back the next day and change my mind. DuffMan 1 Quote Link to comment
couchesoleil 18 Posted January 15, 2023 Share Posted January 15, 2023 This is really cool stuff! Thanks for posting! Quote Link to comment
Brittanybunny 1,494 Posted January 16, 2023 Share Posted January 16, 2023 (edited) Please do not use stable diffusion or any AI program to do images…. These AI companies are known for stealing art from well known artists without permission to use for their programs to generate art. Artists are now losing commissions because of this and the AI also butches their signature to try to cover the fact that its stolen. This is seriously hurting many artists around the world who are fighting this. 😢 i know it seems fun and cute but its doing a lot of damage to those who are being effected. Those programs are also known to lighten skin on those with darker skin, make more thicker people look skinnier, and used to sexualize images of women. https://www.smithsonianmag.com/smart-news/is-popular-photo-app-lensas-ai-stealing-from-artists-180981281/ Edited January 16, 2023 by Brittanybunny (see edit history) LoadedMink and Mydnyght68 2 Quote Link to comment
Azazbel 14 Posted January 21, 2023 Share Posted January 21, 2023 (edited) On 1/16/2023 at 3:43 PM, Brittanybunny said: Please do not use stable diffusion or any AI program to do images…. These AI companies are known for stealing art from well known artists without permission to use for their programs to generate art. Artists are now losing commissions because of this and the AI also butches their signature to try to cover the fact that its stolen. This is seriously hurting many artists around the world who are fighting this. 😢 i know it seems fun and cute but its doing a lot of damage to those who are being effected. Those programs are also known to lighten skin on those with darker skin, make more thicker people look skinnier, and used to sexualize images of women. https://www.smithsonianmag.com/smart-news/is-popular-photo-app-lensas-ai-stealing-from-artists-180981281/ I hope your post was sarcarstic or a joke, because if it isn't then I'm saddened to tell you: everything you said is false. The only ambiguous thing is the "stealing art" part, but the most widespread model (stable diffusion) is, supposedly, trained on public domain images (such as stock photos). On most platforms (Facebook, deviantart,...) you agreed that your pictures were to be scrapped so bots and algorithms could use them ( before StableDiffusion, now you can opt-out). At best the AI generate pictures in the same style as some artists if they were trained on it, but at most it's "inspiration", you wouldn't find a copy of an existing picture (except public domain pictures, again, for example well known paintings). Your other points are the most bullshit I ever read, and I even been harassed by "artists" on deviantart for posting AI generated images (labeled as such). Your arguments are even worse than what they send me. We are just using StableDiffusion to generate content that please us and others, why would it be bad? Do I harass content creators (games/images/stories/videos/pictures) because they want us to pay for our very niche fetichism? No. So, why would you come annoy us because we want to generate images we can't draw, and post those to please other people? On another note, some kind anon posted a model on 8moe/abdl, sadly they didn't give the keywords, so I assume the only one is "diaper", it works, not bad not stellar, it's a .ckpt so check for pickle if you try it, I only tried it on a colab notebook. And, of course it's a Anything-dreambooth). As I said in my album's comment I lost my hdd and my trained model, which is very sad because it was really good. I'm training another one but I'm not really satisfied with results so far (compared to the previous one). The best results I have are: Spoiler But most of the results I get are a mix between a diaper and a pantie. For those interested (and for myself, kind of memo) using thelastben's dreambooth colab: I used 30 instances images for clean diaper and 30 instances for wet diaper, with 140 concepts images. I trained it 3 times for a total of 1300 UNet steps (1e-5), about 2500 Text concept, and 700 Text encoder steps (1e-6). I'm surprised because I used very puffy diapers as the base images (concept and instance), so I guess the training steps are too low? Even tho it should be overtrained by the standards they wrote. Edited January 21, 2023 by Azazbel (see edit history) ihcah, Garrus and HentaiKami 3 Quote Link to comment
Azazbel 14 Posted January 22, 2023 Share Posted January 22, 2023 (edited) 2 hours ago, John said: I'm new to this so forgive me if I mess up terminology. I've just installed stable diffusion and I'm using openai anime as a model for it. It's as you said, it doesn't know about pee at all. But I really like making anime images of desperate women. So often I'm making prompts to show embarrassed or desperate faces, squeezed or crossed legs together and small puddles underneath them. It understands sitting on toilet too lol so I've been doing that. Anyone have any tips or good prompts to use? I'll give prompts too soon, I'm still learning this. I can't make the puddles yellow and I can never get any peeing or pee streams from their body. Lol this is too addicting. I have extra computers with good gpus I didn't get a chance to sell after mining ended so I let them run ai diffusion for days and some of the results are pretty nice. Here's a fun one I made. Is there a site where it's acceptable to upload ai art? Hey! I hope you will have fun with that and don't give up, it's fun to use but a bit of learning curve if you want something not included in any model. There is multiple things you are not really specific about so it's complicated to give you tips, but by "installing stable diffusion" you mean AUTOMATIC1111 or InvokeAI? Or maybe another user interface? Openai anime model isn't really a thing either so I guess it's nAI or AnythingV3? For the tips eh.... depend on what you already know? CFG between 3-7-11 is good, steps above 30-40 is useless, some sampler give very good results at 20-30. For good prompts, I very often use "cinematic, cinematic light, ultradetailed, by manga artist" because it really enhance my results, some other things are really hit or miss (eg: "intricate", "insane detail", "detailed hands"). Advice: join a discord about image generation and keep an eye in channels where people post their images, sometime they also specify the prompts, you can learn for that. Similar, don't underestimate the power of negative prompts, my goto is " disfigured, kitsch, ugly, oversaturated, grain, low-res, Deformed, blurry, bad anatomy, poorly drawn face, mutation, mutated, extra limb, poorly drawn hands, missing limb, blurry, floating limbs, disconnected limbs, malformed hands, blur, out of focus, long neck, long body, disgusting, poorly drawn, childish, mutilated, mangled, text, watermark, caption". Don't hesitate to put things in negative prompt (eg clothing, place) when you don't want them to appear. Most of the time I found that a lenghty negative prompt is better than a long positive prompt. Last thing: the position of your keywords. Put first in your prompt want you want to see the most. It really help (most of the time). Look, it's my trained model (on diaper), but it's AnythingV3 model: "1girl, diaper, inside, bedroom, ultradetailed, by manga artist" WITHOUT and WITH negative prompt: Spoiler Both images are meh, but you see how we change from a picture a bit saturated with a psychopath and weird eyes/hands, to something a bit better. Now, negative prompt and better positive prompt to get a better result BUT trying to get the same general appearance: "1girl, white shirt with ribbon, legs spread, (diaper), lying on back, looking at viewer, inside, bedroom, ultradetailed, by manga artist, detailed hands, black hairs, pink eyes, blue skirt, white bed" Spoiler When you try to get a very specific result, remember that colours bleed. The order of the keywords in my prompt was annoying to find because the colour of the hairs and eyes was bleeding onto the skirt, the skirt on the bed, etc... It's complicated to see from one picture, but don't forget to use VAE, it really help with colours ( it look like you don't have it on but it's difficult to say from one picture ). WITHOUT and WITH VAE: Spoiler Try to test the samplers, Euler_a and DPM++ 2 a work wonder for me, but it changes from interface from interface. If you really want to generate pee images you should check few vids about embedding and dreambooth (https://www.youtube.com/@Aitrepreneur make good videos with step-to-step). Dunno where you can post you pictures but most sites accept it if you tag it as AI work. I did a bit on deviantART, but you should expect to have ""artists"" commenting on your pics about how "ai generation is bad, it's taking their jobs, it will lead to the end of society" (damn, that's the same argument people use against immigration) . I guess it's the same with any image sharing site. Better place would be fetish sites, people are more welcoming there (8moe/4chan/reddit/here). Edited January 22, 2023 by Azazbel (see edit history) ihcah 1 Quote Link to comment
Azazbel 14 Posted January 23, 2023 Share Posted January 23, 2023 (edited) 11 hours ago, John said: Thanks for your response. It's not Automatic1111 or InvokeAI, I have a folder called stable diffusion webui. For the models I'm trying both of nAI and Anything V3. Thanks for all of your tips, I've read through them. I have to fix the order of my prompts. I'm going to check out the youtube channel you linked soon, thanks for the resources! Ok yeah to my knowledge only AUTOMATIC1111 create a directory named "stable-diffusion-webui", so it should be that one. All user interfaces use Stable Diffusion as a base, but Stable Diffusion itself hasn't a user interface, only cmds to run. If it's that one I would advise to not update it a lot (if you don't use the extensions), pull-requests are pushed very often and it tend to break things a lot (lately Dreambooth extension, LoRA extension, and samplers were broken because of bad pushs). Edited January 23, 2023 by Azazbel (see edit history) Quote Link to comment
ihcah 485 Posted February 3, 2023 Author Share Posted February 3, 2023 (edited) A little generation I did today... if you know the classic scene of Deanna Troi and Doctor Crusher working out, you know. Otherwise this particular fantasy may be confusing. I’ve been mostly playing with different styles lately, like the simpler cartoon look here, just for fun. Edited February 4, 2023 by ihcah (see edit history) Diapheure 1 Quote Link to comment
ihcah 485 Posted February 24, 2023 Author Share Posted February 24, 2023 So, the next great advancement in Stable Diffusion is ControlNet. I strongly recommend looking into it if you're doing AI fetish art. It lets you transfer pose, position, or any other level of information from one image to a new prompt, waaaay better than the former img-2-img methods in its comprehension. Here's a 5-minute test I did today. I took this lingerie photo: And used a LORA diaper network (this time, on publicly available... search civitai.com) along with ControlNet to easily generate: I didn't have time today to fix the little details, but that result was so quick and easy while retaining the exact pose. It's fantastic, I look forward to testing more. nicdiaper 1 Quote Link to comment
nicdiaper 8 Posted February 25, 2023 Share Posted February 25, 2023 Some on deviant art have figured out a lot. I have been playing around with it and just found a helpful guide that I thought I'd share. https://www.deviantart.com/pinkthedinosaur/art/A-basic-how-to-guide-for-diaper-AI-art-937958320 Seems like this artist (and others) has suggested Novel AI as a jumping off point. Happy prompting. Quote Link to comment
mr.pee 114 Posted February 26, 2023 Share Posted February 26, 2023 nappypants 1 Quote Link to comment
ihcah 485 Posted February 27, 2023 Author Share Posted February 27, 2023 Kind of lazy (did this one on my phone after booting a WebUI session remotely) but this image shows how well ControlNet handles composition of multiple people. Quote Link to comment
ihcah 485 Posted March 1, 2023 Author Share Posted March 1, 2023 haven't found a lot of time, but this is based on me playing with making LORA networks finally, which is more flexible, and then using ControlNet in the final work. I could have made this one much better with time to spare, but it was a quick 10 minute test of my LORA model once it finished a first couple of epochs of training. Tailsuser 1 Quote Link to comment
ihcah 485 Posted March 1, 2023 Author Share Posted March 1, 2023 (edited) One more … targeting a Depend-Silhouette look ControlNet source image attached below. Using ControlNet (instead of needing img2img) works so well… her pose and proportions were kept perfectly even though i was able to go a very different direction with the image overall. Edited March 2, 2023 by ihcah (see edit history) nappypants 1 Quote Link to comment
ihcah 485 Posted March 2, 2023 Author Share Posted March 2, 2023 (edited) Edited March 3, 2023 by ihcah (see edit history) Quote Link to comment
323014025 28 Posted March 3, 2023 Share Posted March 3, 2023 You guys have managed in to a proper art form with what seems near the same hours and effort. Proper respect to you lot. ihcah 1 Quote Link to comment
ihcah 485 Posted March 4, 2023 Author Share Posted March 4, 2023 On 2/24/2023 at 8:13 PM, nicdiaper said: Some on deviant art have figured out a lot. I have been playing around with it and just found a helpful guide that I thought I'd share. https://www.deviantart.com/pinkthedinosaur/art/A-basic-how-to-guide-for-diaper-AI-art-937958320 Seems like this artist (and others) has suggested Novel AI as a jumping off point. Happy prompting. Thanks for sharing. I’ve had NovelAI subscription since even before they started doing images, because it’s a pretty great (and totally uncensored) erotic story generator too. I play with their image tool occasionally and it’s pretty neat; in my case I won’t be as interested in their diaper tags because it’s not the style I like (it’s very much all taken from furry abdl content, and I’m only into women wearing realistic adult incontinence products). But it’s another option out there and a fun tool to try if you want to play with stable diffusion on an uncensored site without having to set up your own workspace. nicdiaper 1 Quote Link to comment
ihcah 485 Posted March 6, 2023 Author Share Posted March 6, 2023 (edited) I tried hard to get a soft painterly style, like a portrait of domestic life on canvas. Not perfect but it fills the intended fantasy well Here’s the wet edit: one of the amazing things about AI is how it unlocks the ability to dream up a specific fantasy and then bring it to life — for anyone, even if you lack the artistic skills. That’s a good thing, in general. Edited March 6, 2023 by ihcah (see edit history) Xeno_ 1 Quote Link to comment
Recommended Posts
Join the conversation
You can post now and register later. If you have an account, sign in now to post with your account.
Note: Your post will require moderator approval before it will be visible.