ihcah 485 Posted September 30, 2022 Author Share Posted September 30, 2022 (edited) Trained it on some more basic pullups, and only had a few minutes left to waste tonight so I tried only one prompt -- Princess Peach wearing white pullups (with some other tags added). More of this later, but here's a very quick show off of the new model. Enjoy. note: usually I play with things more to fix up the problems, but these are "raw" 1.0 outputs right after I finished training. I'll make better ones later. Edited September 30, 2022 by ihcah (see edit history) DuffMan, TimmyTrihard69 and rachelkirwan 3 Quote Link to comment
ihcah 485 Posted September 30, 2022 Author Share Posted September 30, 2022 (edited) Honestly, using a custom trained model like this, and given just a few hours, I could create more quality Princess Peach diaper art than currently exists on the entire internet : Edited September 30, 2022 by ihcah (see edit history) Quote Link to comment
NouveauWolf 85 Posted September 30, 2022 Share Posted September 30, 2022 (edited) -snip- Edited September 30, 2022 by NouveauWolf Question answered during the approval period (see edit history) Quote Link to comment
LivingInfinite 712 Posted September 30, 2022 Share Posted September 30, 2022 I have experimented with stable diffusion a bit myself and I have to say the results you're getting now are quite impressive. Any chance you can share your model with me? jennymum 1 Quote Link to comment
ihcah 485 Posted September 30, 2022 Author Share Posted September 30, 2022 (edited) Couldn't stay away, tried my hand at creating a diapered Samus. This takes a lot of time to get right. In this case, I first had the model generate many photorealistic adult diaper views (woman in pullup from behind, with modifiers) until I found one I liked. Then, I locked that seed in and enhanced the prompts and parameters for a long time, until it was a great pose. Then, I ran that through image-to-image on the same model with a new prompt about Samus. After a while I found a good set of prompts and parameters, then tweaked it from there. "Adam, can I finally have authorization to use the bathroom after this?" "As soon as you neutralize Ridley, in the chamber ahead." "... n-nevermind" Edited September 30, 2022 by ihcah (see edit history) Optimalhentai, DuffMan, jennymum and 1 other 4 Quote Link to comment
ihcah 485 Posted September 30, 2022 Author Share Posted September 30, 2022 (edited) 16 hours ago, LivingInfinite said: I have experimented with stable diffusion a bit myself and I have to say the results you're getting now are quite impressive. Any chance you can share your model with me? For now I'd rather walk you through exactly how I'm doing it, so you can make your own. I keep changing my model and trying new things anyway, and it's a ton of trial-and-error each time I actually go to make a good image like the Samus above (at least 1.5 hours of tweaking prompts, running over and over, changing parameters, etc... and I still didn't get her hands right but this output was so good overall that I didn't care. Making her soak her diaper is a whole second process after that, using the targeted tool to highlight her rear and then trying a ton of inpainting prompts until it's exactly right. I looked at probably 200 versions of her soggy butt before I honed in on the right combination. But look closely and you'll see that it's not just coloring it damp, it actually made the plastic sag, heavy, soaked... all which were included in the keywords I added to that targeted area). First off, get a cloud GPU unless you have one locally with 32GB. Then, this guide can get you started better than any: https://cyberes.github.io/stable-diffusion-models/ Basically, install the web ui (https://github.com/AUTOMATIC1111/stable-diffusion-webui) on the machine, which gives you much better tools for playing with your model. Then you can start it up using its script (all explained at its link), and you can get a public URL to work with your model. You can actually password protect that URL as I do (in the command line options), because I don't want to be responsible for other people finding it while I'm using it and then being responsible for whatever they try to create on my machine. Then you'll want to go to that main link up above and download one of the models. You need a "full" model with the original weights in order to do custom training. All of my diapered results are from doing custom training so that the model understands adult diapers and how they should look. Custom training is amazingly powerful with just a short set of new images, if you use Dreambooth (a new method of training from Google Research). You can find the links if you search it with stable diffusion. It lets you use a small set (around 5) of clear images of your desired object (so I did Always Discreet incontinence panties once, then I did another kind of white adult pullup after that), and you can train against a background category of images (in this case, "panties" was my background category). It can be done in under an hour, then your model knows the new term you gave it. However, I can't figure out how to do this multiple times without it messing up, so each time I train I get a whole new model with just that added item / term, rather than doing this on the same model over and over. Then it takes hours of playing with the UI tools to start to produce great images. The best threads for understanding this are over at 4chan unfortunately, but easy to find with a single search. There's a huge group of people there trading hints and ideas, and posting their results. I'll think about how to share one of my models, although they're pretty large. I think it's 18GB after I train it. EDIT: I should add that I've worked on AI related projects professionally in the past--not images, but the principles are the same--so perhaps this is much harder if you're starting from ground zero. But I'm definitely keen to share all that I know and guide others along. Edited September 30, 2022 by ihcah (see edit history) imfac and Wee-wee 1 1 Quote Link to comment
ihcah 485 Posted September 30, 2022 Author Share Posted September 30, 2022 (edited) These tools can kill your free time... I'll post more later but need to stop using this thing. Edited September 30, 2022 by ihcah (see edit history) Quote Link to comment
Azazbel 14 Posted October 1, 2022 Share Posted October 1, 2022 Thanks for sharing all these infos, I'm impressed by your dedication, I thought it was a much simpler process. Right now the tutorials for AI (being texts or images), for example how to train a model, are still aimed toward professionals, not average people, which is sad. Continue the good work for all of us too dumb to do it ourself! Downjacket 1 Quote Link to comment
ihcah 485 Posted October 2, 2022 Author Share Posted October 2, 2022 Finished up a Princess Peach... while watching football in my home office, so it took roughly the whole 3 hours to get all the details (used lots of techniques here... not perfect still but I like it). Teaxeru, John_John and DuffMan 2 1 Quote Link to comment
Tailsuser 120 Posted October 2, 2022 Share Posted October 2, 2022 Damn, that's really cool stuff! I have been watching the AI field for a few years casually and have been impressed, even more lately as I just found an AI image upscaling software and now also what you did here. Needless to say, I put your images through the upscaler I found. And the results look great! I really wished I could do those kinds of things locally on my own GPU. Quote Link to comment
ihcah 485 Posted October 4, 2022 Author Share Posted October 4, 2022 Belle almost made it This one took me an eternity. I wanted a very specific thing, and had to use all the steps (outpaint, inpaint, resample over and over, etc) to get there. Optimalhentai, DuffMan, Azazbel and 2 others 3 1 1 Quote Link to comment
NouveauWolf 85 Posted October 6, 2022 Share Posted October 6, 2022 Out of curiosity, have you tried training textual inversion on diapers? I understand it's mostly used for styles as opposed to subjects, but some people seem to have had luck. Asking in case I'd be definitely barking up the wrong tree to try it. I did try dreambooth training my own diaper model, but ultimately it was without much success. I was training it on drawn images of worn diapers, the dataset didn't seem to work out very well. Going to try again at some point, but I'm only running a 3080ti locally, and dreambooth is a hassle to set up remote. When you trained yours, did was your dataset pictures of the subject diapers being worn, or on their own? Better still would probably be to run an actual labeled finetune on a padded dataset but so far I've found exactly zero information on how that could be done. Quote Link to comment
ihcah 485 Posted October 6, 2022 Author Share Posted October 6, 2022 (edited) 1 hour ago, NouveauWolf said: Out of curiosity, have you tried training textual inversion on diapers? I understand it's mostly used for styles as opposed to subjects, but some people seem to have had luck. Asking in case I'd be definitely barking up the wrong tree to try it. I did try dreambooth training my own diaper model, but ultimately it was without much success. I was training it on drawn images of worn diapers, the dataset didn't seem to work out very well. Going to try again at some point, but I'm only running a 3080ti locally, and dreambooth is a hassle to set up remote. When you trained yours, did was your dataset pictures of the subject diapers being worn, or on their own? Better still would probably be to run an actual labeled finetune on a padded dataset but so far I've found exactly zero information on how that could be done. Good question — I used only high quality professional photos. I’ve found that dreambooth does better with training on real things, then inferring backwards to artistic styles on its own. for instance, here is one of my training sources, the professional shots of women in pull-ups from this shop (cropping just the close-up of the diaper area): https://www.etsy.com/listing/526586072/pink-safety-mittens-soft-padded-mittens My “background training” photos for category regularization were largely generated by the model itself. I basically generated “panties” shots with many added variables for angles etc (“rear view of..”) and collected them in a dataset. I trained the pull-ups against that regularization category, and also made sure I selected a batch of pull-up photos with a few different angles represented. Edited October 6, 2022 by ihcah (see edit history) Quote Link to comment
ThatYellowSuit 103 Posted October 7, 2022 Share Posted October 7, 2022 Do you think this would work with regular omo? Like, women embarrassed peeing panties public? Or is the prompt of liquid to complicated for the ai to handle. Love your work, seems like it takes a long time to get exactly right. Quote Link to comment
ihcah 485 Posted October 7, 2022 Author Share Posted October 7, 2022 10 minutes ago, ThatYellowSuit said: Do you think this would work with regular omo? Like, women embarrassed peeing panties public? Or is the prompt of liquid to complicated for the ai to handle. Love your work, seems like it takes a long time to get exactly right. It’s definitely on my list of things to do next. I think I can figure out how to train it but it will require just the right photos for the input. Probably I’ll want something like HD screenshots from high quality, professionally produced porn with panty wetting, where the thighs and wet area are clearly visible. Also the base models out there are getting better every day, so I want to redo my training against a better hentai/waifu variation model. But my normal career and real wife is getting the way this week, and after finishing Belle I haven’t had time to dive in again. Quote Link to comment
ihcah 485 Posted October 11, 2022 Author Share Posted October 11, 2022 Spent quite a while again... finally achieved the balance between character and realism I wanted. This is how Anna keeps warm on a frozen night. Wee-wee, PattyWets, Teaxeru and 1 other 3 1 Quote Link to comment
jennymum 161 Posted October 19, 2022 Share Posted October 19, 2022 I've been trying out Waifu Diffusion 1.2 and got some "ok results", but it's really hard to convince the ai to use diapers and it's even harder to get it to actually make a form that doesn't look grotesque. I'm stupid and don't know what I'm doing so I just copied someone else's code and changed the prompt/some settings, I have no idea how to train it on images or anything. Heres a few decent examples though; I just threw in as many abdl prompts as I could along with samus aran & zelda Quote Link to comment
ihcah 485 Posted October 20, 2022 Author Share Posted October 20, 2022 (edited) New variations.... Edited October 20, 2022 by ihcah (see edit history) yellowsailgg and jennymum 2 Quote Link to comment
Azazbel 14 Posted October 23, 2022 Share Posted October 23, 2022 On 10/19/2022 at 12:20 PM, jennymum said: I've been trying out Waifu Diffusion 1.2 and got some "ok results", but it's really hard to convince the ai to use diapers and it's even harder to get it to actually make a form that doesn't look grotesque. I'm stupid and don't know what I'm doing so I just copied someone else's code and changed the prompt/some settings, I have no idea how to train it on images or anything. Heres a few decent examples though; I just threw in as many abdl prompts as I could along with samus aran & zelda I tried with https://github.com/ShivamShrirao/diffusers/tree/main/examples/dreambooth because my computer is weak and you can try that in GoogleColab. It was somewhat successful with few pics of diapers, but it was very limited and when I tried complex things, it completly ignored the diapers. Maybe someone will have better success with that. jennymum and Wee-wee 1 1 Quote Link to comment
ihcah 485 Posted November 1, 2022 Author Share Posted November 1, 2022 Worked on this for hours off and on. Tried to get them to look like Aerith & Tifa but it’s not a perfect resemblance. rachelkirwan and PattyWets 2 Quote Link to comment
ihcah 485 Posted November 2, 2022 Author Share Posted November 2, 2022 That last image by itself probably cost me $10 in cloud compute... I ran it through so many times, probably thousands of accumulated sampling steps in the final output. I should move hosting providers or build a better setup. Quote Link to comment
Showerguy 2 Posted November 4, 2022 Share Posted November 4, 2022 As a software engineer I consider it my duty to be up to date with things like this. I have a powerful enough PC to not only run these things, but to also host a server for other people to use. I've got stable diffusion and novelAI installed. NovelAI is significantly better at generating anime art, and I've attached a few images I've produced in the past. I've not refined any of these, they were straight out of the machine. I won't post a link here because A: It could look suspicious and B: I'm not going to have it running all the time because it's intensive. However, if you're interested in using it just DM me and I'll drop you a temporary link when I'm around. Diapheure and ihcah 1 1 Quote Link to comment
ihcah 485 Posted November 6, 2022 Author Share Posted November 6, 2022 Using a lot more editing now, so these are based on about 5 different final outputs that I merged together in photo editing, taking what I liked most of each and blending or masking to get the perfect output. Then of course I added the fun logos : ) One thing I tend to ignore though: the way her diaper and skirt interact is illogical here if you look closely, particularly on the far right of the image. Doesn’t really bother me at first but it’s one of those AI things you start to see after a minute. jennymum, Optimalhentai and Diapheure 3 Quote Link to comment
Optimalhentai 39 Posted November 7, 2022 Share Posted November 7, 2022 Wow, that's really good. I've been messing around with stable diffusion's hypernetwork training trying to generate decent outputs. So far some progress, not going to be able to work on it for a bit but I'll drop these okay ones here. Using Waifu diffusion 1.3 and training on 140 images ihcah and Diapheure 2 Quote Link to comment
ihcah 485 Posted November 7, 2022 Author Share Posted November 7, 2022 (edited) @Optimalhentai I still haven’t tried using hypernetwork training, thanks for the reminder. Those results are promising... what kind of inputs did you use, Abdl art? So far I use Dreambooth mainly because it works so well with a small set of images (5 or 6 high quality photos from different angles is perfect), but it takes a lot of GPU and storage space compared to hypernetwork so I should try that too. But I use photos only, not art in my training set, because I found that training on “reality” and then converting to art is better understood by the model. Currently I use a combination of just about everything to get my final output: - often start with a generic (not Waifu) model and just use prompts to get the overall pose I want, making it photorealistic at this point and not necessarily using diaper weights yet; - then I put that through img2img on one of my diaper models, or often I’ll put it through several different models in sequence as it shapes up; - then I make it conform more to the character I want in the different Waifu based models; - along the way I’ll use masked in painting and sometimes outpainting to edit parts (making diaper wet) or to expand the scene - then at the end I may do upscaling to slightly higher res with some style tags - finally I take a set of best outputs and put them into a photo editor as layers, then blend as needed, always keeping one dry and one with the wet layer showing Edited November 7, 2022 by ihcah (see edit history) 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.