I tried it out, it sorta worked but not well for ABDL. It usually does thin depends-like pull-ups instead of diapers (or just adds a baby in). One of my best results is attached. I want to try Textual Inversion to get around that problem but haven't gotten it working yet. Craiyon seems much better at doing thick diapers, see http://4usoivrpy52lmc4mgn2h34cmfiltslesthr56yttv2pxudd3dapqciyd.onion/abdl/res/13468.html

>>16254 I wonder if feeding it the whole abdreams.com catalog, as well as other diaper content, it would make the quality/accuracy get better.

>>16268 This is interesting- Craiyon doesn't know what ABDL means, but Stable Diffusion does. If you prompt with "ABDL" you get like half baby pics but Craiyon just does random stuff. Prompts with "ABDL diaper" instead of "Adult diaper" are much closer though still not perfect. Also because I was an idiot, a better link to the Craiyon thread >>13468

>>16298 That first image is pretty good. Even though it's not photorealist yet. I imagine technology getting better and more accurate in the years to come.

>>16304 I'm specifically going for a painted style in the previous images because IMO it looks better. See this one, where there's lots of little things that just look wrong instead of it being covered by the art style.

>>16311 >instead of it being covered by the art style. Yeah, I think that is the case with these algorithms in general. Like, if you are going for photorealism, any detail out of place will really stand out an you will notice. Now if you are going for a painting style, there is more room for random inaccuracies, and those algorithms make a lot of those, to be masked by the art style and to not be that noticeable.

>>16311 You know, this is still not great but it's worlds beyond where AI wank material was just a year ago. That actually kind of passes the "squint test" which is pretty impressive. I'm willing to bet that a year from now we have something actually decent

>>16332 I started following a bot on twitter a 2 or 3 months ago where people tweet requests at it and it posts images generated from them. It looked worse than these especially human form and faces. Today it looks damn good most of the time, even with crazy prompts. IMO the best ones are submissions in a certain style or crossovers. Whoever is making these diaper ai should consider feeding it tons of anime and draw styles so you can prompt "[insert name] wearing a diaper under her school uniform at school anime" or stuff like that, I think results would be surprisingly good. I'll add a recent example

>>16334 That bot is run by a real person, so I don't think you can use it to generate porn. The submissions are probably viewed before uploaded.

>>16338 not suggesting that, just using it to illustrate a point for whoever is making this diaper one

>>16298 what website did you use to make these?

For anyone wanting to teste some of those new AIs, more precisely those running Stable Diffusion algorithm: it is still not as good as you running it on a dedicated PC with a powerful hardware where you can adjust all the settings and tweak around, but: https://dezgo.com/ https://stability.ai/beta-signup-form It definitely understand the concept of ABDL. I assume DALL-E 2 would also be able to put up good art creation, maybe even better than this one – if their algorithm wasn't so heavily censored and if there was ABDL content on their database (rather than mostly stockphoto imagery, I'm pretty sure they bought all istockphoto.com catalog to use on when making it)

It's really weird the whole thing of prompting, in a way, it's as if you were talking with someone who is heavily autistic, and any slightly difference between the order in which you arrange the words in your sentence... it can often differ between the machine understanding what you said – sort of – or totally missing the point 100%. Oftentimes you have to arrange words in a way that, it might even sound unnatural for humans, but it seems to make sense for the machine. I assume their models still have a way to go when dealing with natural language processing and fine-tuning the process, but this is just a matter of time. This shit is exponencial, in 5 years I can't even imagine how advanced this thing will be.

>>16341 >I assume DALL-E 2 would also be able to put up good art creation, maybe even better than this one – if their algorithm wasn't so heavily censored really grates me that "open ai" licenses all their shit, puts it behind a paywall, and puts huge artificial limits on what it can do like even google has the chromium project

Need to get lucky and have the right prompt but Stable Diffusion can definitely make some good looking artwork. Not professional level, but close. >>16340 Making them on my own machine using the low-VRAM Optimized Stable Diffusion. Takes ~4 minutes to generate a set of 16 images on an RTX 3070.

>>16379 >Takes ~4 minutes to generate a set of 16 images Interesting. Can you fine tune the experience, like, for instance, asking for the program to generate a large sample size although in the smaller resolution and then choosing 2 or 3 that look promising and asking the algorithm to work on those and increase their resolution? It seems that this would be a good way for getting rid of dead ends prompts.

Some of it is uncanny and freaky.

>>16399 Sometimes you can make that work to your advantage, especially if you're in the mood for a spooky ghost waifu.

>>16385 Some of that is possible but I don't have it set up right now. Often a single prompt will produce a mix of good and bad results based on the random noise that it starts with. That's why I do a batch of 16 per prompt.

>>16408 What were the prompts/settings to these two?

>>16449 Both used the same prompt, seed 6 images 1 and 8. ddim 50 steps, 7.5 guidance scale. > ABDL diaper on high fantasy loli teen ABDL elf princess wearing disposable diaper, Trending ArtStation Pixiv high quality digital artwork by artgerm, wlop, sakimichan, Greg Rutkowsk

Tried to make some fake ads, some of them turned out surprisingly well. Here's some of the better ones. Also three more painted/stylized images. I still haven't found any way to reliably get the style of diapers I want, though "bulky disposable diaper" along with "ABDL" seems to be getting closer, only around 10% of images match. Whereas Craiyon is like 95% that type of diaper, with much worse everything else.

>>16630 Oh, max 5 files per post. Here's the other 2 ads and the 3 paint images.

Right. It is important to try to figure out your prompts and how the AI thinks about them. It can find disposable diapers of the pull-up kind, but it will want to try to change them in to pants.

For those interesting, apparently on DreamStudio AI now you can upload a reference image so and tell the script to try follow that art work. If you set a higher value for image strength it algorithm will give more value to the image: https://beta.dreamstudio.ai/dream

If you set a lower value, you will get more variable/diverse range of results, but the algorithm seems to not understand very well the concept of ABDL, so it ends up just replacing the woman being held by an actual child.

For those trying to understand how the whole prompt thing works, I recommend to check on this site, that has a bunch of works created by AI, as well the prompt used on them. https://lexica.art

Here is other examples, base image and the variations:

If you know a little of photoshop you might photoshop the image before uploading it, so that it gets closer to your main idea, and make it easier for the algorithm.

Look, can we all just agree that as amazing as this technology is, AI generated art is not yet producing anything worth fapping to?

>>16865 Or is AI showing us that to it, we all look like deformed mentally ill amputees?

>>16865 It's getting there. Surely someone has low enough standards for the top 1% of generated images. I've found some success with doing prompt blending. Take the learned conditioning of a prompt involving diapers, multiply by 1.2, subtract the conditioning for "naked baby" times 0.2. Stuff like that. Textual Inversion should be better than this, but I still haven't gotten it working (I need to either get better hardware or rent some online)

>>16891 Indeed. There are a lot of people with a very low bar for jerkability, I have no doubt that at this point there was a lot of jerk offing going on already But, I agree that this technology has to be refined on many levels: >better ways for choosing art style/better user interface Like, for instance, it would be nice if there was some sorta menu showing a template image on different art styles, and you would choose a given one, so that the machine would follow it when creating yours. >a way to fix mistakes the machine does Sometimes the drawing is like 80% right, but there a third arm somewhere else, it would be nice (sense the algorithm can't do everything yet) that they offered a way for the user to fix this easily. Maybe you would select the wrong part of the image and tell "remove" or something. It's fascinating because this is only the first generation and it is already good, this will only get better and better, because whenever a technology comes up – even if it is not as good yet – it creates a feedback loop: people start using it, there is demand for it get better, people put more money to it, they start to pushing the limits of that given technology, it gets even better.

It's interesting seeing the variations the algorithm comes up:

Fun fact: I was able to find the database they used to trained Stable Diffusion, it is a public database called "LAION-5B", there is like 1 billion images or so, anyway, there is a lot of images. And I was wondering if there were ABDL images in it, and if so, what images they were, and I was able ti find it: https://haveibeentrained.com/?search_text=ABDL There are very few images, most of them being ABDL adult diapers ads rather than the most artistic-ish ABDL. Which might be one of the reasons why the algorithm struggles with creating ABDL content, there a very few references. Here are some images they used.

>>16993 The problem is not that they aren't in there; they are. Search just for any diaper ABDL or medical, or even kids'; you'll find it. The problem is that what terms you need to use in the prompt to find these. I have tried my absolutely best and I can not stop it from conjuring up cloth nappies. However I can get it to sort of call up some diapers with a landing zone like ABUs. HOWEVER! It turns these in to some sort of a belt every time. Issue is that it is really difficult to conjure these things up and not have it mangle them. It took me 2 days to manage to get these, and it was mainly just adding more and more things in to the negative prompts and forcing prompt edit mid steps. >>16993

This might be interesting: I found this algorithm based on stable diffusion, but it does the reverse, it describes the image into text. Maybe doing that can help us to understand better the best tags to use when you want the algorithm to create a image similar to that given one. https://replicate.com/methexis-inc/img2prompt

>>17078 CLIP interrogators don't really help either. Been there done that. Why? Because none of the interrogators understand what "diaper" is unless it is kid's diaper as a product photo or stock photo of a child. Otherwise it is "panties" "shorts" "underwear". Although with testing of SD I can tell you that if you negative prompt "shorts" or "jeans" your chances of getting anything remotely close to a diaper go to almost none.

Another solution: Multi-prompt conditioning. Take multiple prompts with different guidance scales and sum their results every step. Most of these were done with a "Naked Baby" negative prompt. Normally two prompts are used (the prompt you input and the unconditioned empty string prompt). Adding more slows it down, but I've been able to get some good results (and a lot of bad ones)

AI porn will never be good

anyone train any textual inversion binaries yet willing to share? ie you can train the model a better concept of a diaper. If or if someone could just point me to a set of ADBL diaper transparent templates i probably could do the rest? the end goal would be for the model to output better detailed diapers. its a pain in the ass to wait two and half hours just to find out you fucked it up.

So about one afternoon and evening of going about with img2img, masking things and fiddling around with prompts and settings I have managed to produce this. Basically everything but the tape zone and tapes was run through the AI. The lad was generated with txt2img; although changed drastically during the workflow. Now... This seriously took many attempts, a lot trips to Photoshop and back to just correct minute issues from the output, adjusting the mask. I have to say... It was a lot of fun. More than painting by hand, since now I can work with more realistic subjects. I think it'll be a long time before we can straight up generate stuff outright. Also Stable Diffusion seems to respond best if you combine "Diaper" with "Panties" in the prompt. Figured this out by interrogating loads of pictures along with the ones made during the road that lead to this.

Interesting development on this technology: Apparently Google is testing a new algorithm called "DreamBooth" that allows the computer to repeat a given scene/characters in the exact manner among different scenes. This would be pretty useful, since nowadays these algorithm never generate "exactly" the same art style on sequence, which makes pretty hard for a user to create, for instance, a story or HQ, where it would be necessary to have continuity among the different scenes. https://youtube.com/watch?v=NnoTWZ9qgYg

So i tried out dream booth, making a cpkt, trained on a set of a variety of the front of a variety of abdl diapers not bad, probablly the next step would be to train on a set with a specific diaper.

>>17478 This is a very important development, you could also train such AI for Japanese omutsu and plastic cloth nappies.

>>17478 Please, please, please (!!!) share your cpkt!

sure here, 2 version - one is diffusers version that can only be run on a google collab "diaperdream" and model.cpkt which is my first attempt at making a .cpkt, it came out not as i'd liked, it's not the same as the diffuser's verison. prepare to be dissappointed. I'm also includeing my samples and training parameters for anyone to attempt to make their own .cpkt version of the diffusers version. anyways i dumped it here for now, probably should have went with a mega. will get to that later. tldr: im going to remake it later. https://anonfiles.com/aax3SeA3y6/aifag_better_diaper_SD_zip -- AiFag

>>17478 these images came from the diffuses version just to make things clear.

[mega] https://mega.nz/file/oSNgwbzT#cjqnBcg2v56_P5lZf73b4F4kLtCmYWpnKmoXQvUjUeA

UPDATE: good news, the converter came out for the google collab version so here's the model for these images as a .cpkt, also sorry for spamming up this thread: i would appreciate if someone could merge this in with some of the other models out there and report back the results. [mega] https://mega.nz/folder/pacmSRzK#GycG1GY4KhO589tuiJwDKA

NovelAI's new anime image generator is good but trying to get it to generate ABDL is like shoving my hand through a blender. Here's a stab at AR that finally turned out somewhat decent.

>>17645 You made a babby. 😀

>>17645 That one actually looks decent.

>>17478 It looks good, especially the second photo. You can clearly see the diaper, the only issue – if I may – is that, it still looks like she is wearing panties under the diaper, especially on number 2, like you can see the blue spots of her panties. Number 3 at least it remained sorta consistent, the same color, but you can see the moment where the diaper sorta morphs into a regular par of panties more tight to her body. Again, thank you so much for your work.

>>17645 Think I'm finally starting to get the hang of it. Turns out the curly braces their model uses to strengthen the input help out a lot. Getting some much nicer looking results now.

>>17645 fuck, that is amazing is that open source or you need to pay for it?

>>17804 Unfortunately, it's paid (I was already subbed to them for their AI text gen when they rolled it out). Also struggles with more complicated scenes and I haven't been able to get it to handle wetting or messing, but at the end of the day, if all you're after is some anime girls in diapers, it's pretty damn solid.

>>17810 Apparently their NOVEL AI models leaked online a couple days ago. magnet:?xt=urn:btih:5bde442da86265b670a3e5ea3163afad2c6f8ecc&dn=novelaileak Don't trust me and install by your own risk, but people were talking about it and seems legit.

>>17825 impressive

>>17825 Based. I'll probably continue subbing anyway since I do like the devs (and my hardware is shit) but here's hoping some braver anons than I get some good stuff out of it.

Does anyone know if someone uploaded the leaked NovelAI stuff on a sharehoster or somewhere without torrent?

>>17825 There's a guide: https://rentry.org/sdg_FAQ https://rentry.org/voldy

I just started to play with stable diffusion and i like it for art and photo stuff but its severely lacking in diapers. I see the weebs solved there problems with training a model on anime with waifu diffusion, there's even a model specific to furries. like WTF we can't be beaten by furries. Ive only got a RTX3090 and i have no clue how to train a model but im willing to learn and encourage the rest of the community to also start getting images data together for it. ABDiffusionL can be a reality if we work together. similarly i see people trying to use AI dungeon and similar for diaper stories when logically it would be better to train a model for local usage with KoboldAI. anyway if you can point me in the right direction i think the diffusion model is a good starting point.

>>17917 >Ive only got a RTX3090 Dude that's not an excuse. I have a RTX3060 and got dreambooth working with the help of a youtube tutorial. just google dreambooth and train your own diaper model!

>>17926 https://bytexd.com/how-to-use-dreambooth-to-fine-tune-stable-diffusion-colab/

I'm working on a dreambooth model based on favorite pixiv artists. Hope it'll go well

Trying out v1.5 and the improved vae, using automatic111's webui. Best step/sampler set I've found is 40 steps with Euler a.

https://www.deviantart.com/thetripledl/gallery how? just how does this dude do these?

>>19266 He does give little tech updates with new posts, which is nice. e.g., for pic 1... "Further work on making a custom model. Same work flow as always: generate base image on text2img; adjust on photoshop, iterate on img2img and photoshop. Two issues to that I keep encountering is that adjusting the output to generate and keep a shirt on seems to be either "Either it has it, or it doesn't". Still struggling to reliably to get faces that are not "Square jaw male underwear model" or "Adult toddler" Like in photobash #1. "Custom trained SD1.5 model on 40 hand cropped, upscaled and adjusted images found from search engines. Automatic1111 repo. All relevant refrences and citations to the technology can be found from the Github pages of the relevant projects. I highly recommend everyone to go read up on the original publications and cite them in further development"

Source: https://www.deviantart.com/strykerabdl/gallery/all

>>19313 Damn, those are good. It is amazing how fast this technology is evolving, thanks for the link.

>>19436 Now if only it could do messy butts

>>19720 >>19721 >AI still can't do diaper's back lmao

>>17917 I had absolutely no idea what I was doing or getting into like 4 days ago. I started with this: https://www.youtube.com/watch?v=1mEggRgRgfg And now I've trained a hypernetwork for novelAI that can do some okay diapers (if you follow the video, what you're doing is training a hypernetwork. I learned that literally as I did it). Despite having no idea what I was doing, it still took way longer to prepare all the images and tag them than everything else, and it took me like a dozen or two hours, not hundreds. And btw, I have a 3070. So a 3090 is no excuse at all. In all, I trained it on about 218 images, and it still struggles with a lot of things. Often the diapers look more like panties, and only rarely are they truly tapey-diapers, most of the time they're pull-ups even when that's not specified, and even though I had a tag to specify that in the training set. I may go back and re-train without pull-up images having the "diaper" tag...

>>19862 same person again - I was gonna ask for advice on samplers and stuff for these anime-type images. I noticed that high CFGs tend to make better images but also make them look a lot more burned/overexposed and oversaturated. I also noticed that when using the DPM Fast sampler, if I set it to 150 steps then interrupt it at 84 steps, I get very nice and clean images. Whereas if I just set it to 84 - or anything, really, they look a bit more burned out and less smooth and clean and anime-like. What's up with that? I like things looking less jpeg-y and cleaner and smoother. I haven't turned on "enabler quantization on k samplers" yet, there's still an old seed I want to poke around with a bit more, but once I do I look forward to seeing if that'll sharpen things. Also, some guide(s) said to set "stop at last layers of CLIP model" to 1 when training. Why is that? Also, I trained on DDIM and default to it in general. Still trying to decide what the best sampler is. Any tips there, for anime-style images? Also, have bucket-loli. She's easy to carry around. Treat her well.

>>19863 Nice anon! Despite the quirks, definitely looks like an enormous improvement over what vanilla NovelAI produces. My (unfortunately very limited) understanding is that there's a definite sweetspot for CFG. A low value means the image tends to follow the prompt very little and have more deviation (not always in a positive way) whereas a high value tends to follow pieces of a prompt a bit too closely. Specifying a purple shirt, for instance, might cause purple to bleed into other parts of the image as well. Likewise, results tend to be a lot more similar. As for samplers, those get into pretty gnarly territory since my understanding is that they're basically methods of seeding the noise you use for generation, since stable diffusion tries to start from noise, then work backward to an actual image. I'd expect that for some samplers, the number of steps may factor into this - if you specify a high number of steps, you begin from a much more "noisy" seed since you have more steps to go from noise to image and so you often end up with smoother results. For a smaller number of steps, this initial seed may be a little less noisy initially, which converges faster but may result in more blurry looking images. The comments in this twitter post are a good read: https://twitter.com/iScienceLuvr/status/1564847717066559488?ref_src=twsrc%5Etfw%7Ctwcamp%5Etweetembed%7Ctwterm%5E1564847717066559488%7Ctwgr%5E984eb22e760b6672815e026a4f84e75655268642%7Ctwcon%5Es1_&ref_url=https%3A%2F%2Fwandb.ai%2Fagatamlyn%2Fbasic-intro%2Freports%2FStable-Diffusion-and-the-Samplers-Mystery--VmlldzoyNTc4MDky I believe NovelAI used the so-called "Clip Skip" when they trained their model whereas normal SD didn't, which itself was actually inspired by another paper. They discuss it in more detail here: https://blog.novelai.net/novelai-improvements-on-stable-diffusion-e10d38db82ac Gist of it seems to be that getting the premature outputs tends to give better results than the final outputs themselves, since the final layers usually shove image information into a much smaller set of outputs which can lead to loss of information (think image compression). By skipping that preparation step, the model just has more to work with. As for samplers, k_lms, Euler and Euler-Ancestral seem to be the ones I see recommended the most often on the NAI discord, so any of those would likely be a safe bet.

>>19868 Ah, I've already seen that thread on samplers, actually. But I haven't seen that thing about "Clip Skip" before. That makes it sound like I actually want the "stop at last layers of CLIP model" to 2 or even 3 when training in order to get the hypernetwork to better understand the difference between, say, "pull-diaper" and "diaper", since those are separate tags. I definitely got that there's a sweet spot for CFG. But the two things I'd like to know more about are 1. why high CFGs tend to over-expose/"burn out" the image, and if there's anything I can do about that other than turning the CFG down. There's a lot of seeds - most, in fact, seem to do best with a CFG above 14, but that results in them being kinda burned out. See attached images as an example. You can see it resolves fantastically better and is way more coherent with less random broken bits at the higher, CFG, but the lower cfg one has way better style and colors, and the higher cfg looks burned out. 2. how to consistently recreate that smoother look that I get when interrupting DPM fast at 84 steps with a target of 120. I imagine if I try other numbers I might find other combinations that work, but that looked particularly nice. I mean, I guess I can do it consistently, but it's annoying to have to time something every time. It's a lot more work to generate images when I have to babysit them like that. But surely somewhere there's some more consistent way to get the images to have less intense colors? Something in the prompt?

>>19876 It's not too hard to have a custom script to save the image at a consistent point. As for the high-cfg oversaturation problem, there isn't a good solution to that right now that works with stable diffusion reliably. Adding a normalization/dynamic thresholding step between each sampling seems like it helps, see https://www.reddit.com/r/StableDiffusion/comments/xalo78/fixing_excessive_contrastsaturation_resulting/

>>19879 >What I'm currently using for "normalization" at high CFG scales is this code, inspired by the "dynamic thresholding" method but without any clamping. That's a nice bit of code I'd like to try that he links there, I just have no idea where to put it. >>19942 I'd really like to know how they got that diaper model. That does the diapers incredibly well. Fwiw, my model actually stabilized and I wasn't able to continue training anymore at around step 32,500. Used 218 images to train on. Tagged them all by hand with my own system for diapers. Having "front diaper", "back diaper", "side diaper," "top diaper" and "Bottom diaper" all seemed to help a bit. Still need to get around to trying training a new one where instead of having both "diaper" and "pull-up diaper", images with pull-ups only have "pull-up diaper". Maybe that'll help the diapers to not be pull-ups as much. But the real issue is how panty-like they often end up being. Putting "panties" in the negative prompt helped a little bit but also made the images often a bit less coherent. I wonder if all the various diaper-related tags aren't confusing it, though. "diaper", "pull-up diaper", ["front diaper", "side diaper", "back diaper", "bottom diaper", "top diaper"], "clean diaper", "damp diaper", "wet diaper", "leakguard stripes", "wetness indicator", "pampers", "colored wings", "big diaper". Next time around I think I'll do without "clean diaper", and just be sure to use "damp diaper" or "wet diaper" when they apply. Not sure how I feel about the viewpoint tags. They seem to have helped quite a bit, though. Elsewhere I saw people complaining that AIs can't do diapers from behind, but mine... kinda can. Also I probably need to get the cropped images out of the training pool. Maybe I'll go through the loli thread for the uncropped versions of stuff by 初夢 and 午後 from Pixiv. I have no idea what those names are in not-kanji, lol. Why do you think this other model does diapers so much better, though? Is the artist using inpaint and doing another pass to make the diapers even more diaper-y? Or maybe I need to be able to train my model more without it stabilizing?

>>19953 They did mention that people would need a manual to use it, whatever that means. Could be that it's more of a pipeline in that there's inpainting and some additional edits involved. But I'll stop putting words in their mouth. As for the tags and potential improvements, if in doubt, simpler is better. The less tags you have to work with, the more likely the model will be able to recognize what individual tags mean. Nice perk of training hypernets over a full model tune is that the original model should be mostly in-tact. It might be worth experimenting and seeing if any combination of the above tags and vanilla tags make others redundant. Another question is how well-represented is each tag in your training set? If there are any tags that are barely present, it's unlikely the model will perform well with those tags. If the software or scripts you're using for training have a way to assign loss weights to samples during training, you can always mess around with those. There's also data augmentation if you're not doing that already. That is, given your 218 images, do something like flipping all your images horizontally, and suddenly you go from 218 to 436 images to work with. It's not quite the same as hand curating 218 new images, obviously, but it can still add enough variety to be useful, and in general more data to work with is always better. I believe in you anon!

>>19978 The effort to create a functional abdl model is greatly appreciated, I just have one question, could this model allow me to merge the merunya and pieceofsoap drawing style?

>>19978 Thank you anon. I know you don't want to divulge too much, but your advice is really appreciated. To add to that, my previous post here >>19953 wasn't entirely accurate. I stated that hypernets leave the original model in-tact, but it would have been more accurate to say that they don't train on the original model. Rather, hypernets serve to modify the weights of the original model at runtime, so you can still get behavior the original model didn't exemplify, it's just significantly cheaper. Even then though, hypernets will be secondary to a full model tune, and if in doubt and you have the necessary compute, it absolutely doesn't hurt to experiment with that either. This thread: https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/2284 has some good advice for hypernet tuning and good hyperparameters to try, though I'm less familiar with recommended hyperparameters for a full NAI model tune. Also, if you'd like to further experiment with parameters, I'd fully recommend Weights and Biases: https://wandb.ai/site . It has a really nice suite for hyperparameter optimization and is great for result visualization. You do have to be a bit comfortable with Python to get started, but if you are, it's an amazing tool that can save a significant amount on time and organization. If folks decide to crowdsource images and tags for an ABDL model, I'd be 100% behind that and down to help.

>>19988 Whoops - meant to reference >>19967 .

>>19992 the merunyaa one is incredible. I think nozumi would be a great style to train on https://www.pixiv.net/en/users/32169882/illustrations

>>19967 >inpainting and edits One thing I've noticed is the AI does way better when given a simple sketch, even, to build off of. I'd noticed that when doing some vanilla fantasy stuff earlier when I first got SD, haven't really thought of it again since then. But it might be interesting to draw something with my crappy skills then use img2img, or even just look up really lackluster diaper pics and see how the AI handles them in img2img. Most of the tags are pretty common. I made sure every image with a diaper had 1-3 tags on the diaper's angle, "front diaper", "side diaper", "bottom diaper", etc, and either "clean diaper", "damp diaper" or "wet diaper". Yeah, "auto-flip images" was an option I used. There were 872 files in the reference folder. Half of those were text files for the tags on each image, and half of the remaining 436 images were just flipped versions of the original 218 (and of course the flipped images had the exact same tags as the non-flipped versions). >I believe in you anon! Thanks! Truth be told I think I'm beginning to burn out a little on this, though. Or at least, I'm recognizing that staying up all night tagging images just isn't tenable long-term, lol. But when you don't stay up all night doing something it's a lot harder to find time for it. Fortunately I only got 67 images I'm gonna add and tag to the batch. Gonna need to go through those first 218 and remove the "diaper" tag from all the pull-up images, though. But all-in-all it shouldn't take like the 6+ hours it took to tag the first 218. It took a long time because I was very careful with the tags, and had to do a lot of manual cropping. One issue I've noticed a lot is one of the main ways, if not the main way that image generations end up being failures is actually it generating weird stuff at their stomach/chest. I wonder if "flat chest" will help with this issue? Or if there's anything I can do in training to help it. (first three attached images are examples of this) >>19978 Many thanks for all the advice and such! >that I spent the last 29 days on 10hrs/d >background knowledge on AI stuff That... explains a lot. Your model is fantastic and shows that a lot more is possible than I would otherwise think is possible with current tech. >I'd be very surprised It's a bit rare, but be very surprised because this does actually produce decent-ish results sometimes. See my last two attached images. Not perfect by any means, but if I spent more time in inpaint I could probably fix them up nicely. But, your advice does tell me a lot. Maybe I'll try some other keyword for diapers in training. This might be why my diapers look so panty-like, if the base model's getting confused with the original tag that was applied to many images where the diapers aren't really visible or drawn as we think of them... The other thing it tells me is I have a lot to learn about AI, lol. That's part of why I'm doing this, actually. I'll be studying physics again soon and in general I'd like to learn more about AI and I learn best by doing/practicing. Also, don't know if this will horrify you or what, but this hypernetwork actually stabilized in training - it hit step 32,500 and said training could not continue. >I hope that helped. It does, a lot, thanks! >>19988 I might give hypernetworks another try or two. I'm encouraged by how close a lot of these results are getting, and it's something I'm a bit more familiar with. But after that I'll look into model fine-tuning. Not familiar with python, except for a single lab class where I used it at uni. But I'm very familiar with C# in a Unity environment and c++. So knowing some languages should make it a lot easier.

>>19988 Have you considered ripping tags from a site using Hyrdus? That would probably really speed the process up. I haven't used the more advanced options before, but I'm sure you could figure out a way to even share tags/galleries here for a decentralized method of tagging a bunch of images if you asked around the hydrus general over at >>>/t/9763

I used the AI of diferent dimenssion me and edited the result to fit the nappy.

>>20749 Thanks for sharing this AI art. It's fascinating how far it's come so far.

>>20749 Nice, it looks much better then my original

https://www.deviantart.com/thetripledl/ can someone figure out how they do this stuff? all they say is "custom db model" but if I train something on DB I get shit that doesn't work? they apparently even use the new model now which is supposed to be bad for this stuff because the developers censored all lewd stuff and styles.

Slightly related chat-gpt has been kind of fun Here is a sample daily guide for a week-long ABDL role play scenario based on different stages of potty training: Day 1: As a baby, you would not have any potty training and would wear diapers at all times. You may have multiple accidents throughout the day, as is normal for a baby. Day 2: On this day, you would simulate the start of potty training. You might wear training pants or pull-ups during the day, but switch to diapers at night. You may still have some accidents during the day, but you would also be encouraged to try using the toilet or potty chair whenever you feel the urge to go. Day 3: On this day, you would be close to the end of potty training and would wear training underwear during the day. You may still need to wear pull-ups or training pants at night to help prevent accidents. You should aim to have fewer accidents during the day as you continue to practice using the toilet or potty chair. Day 4: On this day, you would be considered fully potty trained and would wear regular underwear during the day. You may still need to wear pull-ups or training pants at night for a while, until you are able to stay dry through the night. Day 5-7: You would continue to wear regular underwear during the day and may or may not need to wear pull-ups or training pants at night, depending on your level of continence. You should aim to have fewer and fewer accidents as you continue to practice good bathroom habits. It's important to note that this is just a sample guide and that everyone's potty training experience is different. The key is to find a schedule that works for you and to make sure that you are comfortable and having fun during the role play.

Some stuff I've generated with Hentai Diffusion. It's not perfect for diapers but generally good for other sorts of hentai

>>20848 Don't deserve the AI to understand the design of the hands, besides I hadn't noticed that detail to edit it.

>>16220 Have you tried stable difussion 2.1? maybe with that version you will understand much better how to draw a nappy once it is trained?

>>17254 That's incredible. Do you have any more pictures with Pampers?

>>20931 That last photo is just adorable. Thank for posting. It's like an alternate reality where women aren't potty trained until they were 20 years old or so.

>>19978 Pretty knowledgeable post.

Is anyone here experiencing with generating photorealistic AI content? Like, I don't know, training an algorithm on all the thousands of abdreams.com photos catalog or so. There has been some cool models I saw popping out online, like Analog Diffusion model. It would be cool to have something like this, but with girls wearing diapers.

>>19978 NAI does diapers just fine, but (un)fortunately only with the furry model. Which is great for furs like me. I have a feeling you can still invoke some diapered humans with the furry model though, I'll give it a shot in a bit.

>>21050 Just did a quick test, 8 out of 10 results were humans(albeit some had animal-ish noses)

>>20844 I'm impressed with how far this tech has gone, especially comparing with the first few posts on this thread. A shame I couldn't get the diaper to look more poofy, even with using the tag ((puffy diaper))

has anyone tried with this AI is somewhat rudimentary modifies the original image to something you want to add I do not know if they can be of help https://colab.research.google.com/drive/15MQTRMqc7yn626PiC3KPb-DvVx_FrBAr

>>21067 Yeah, you're not going to get shit with Looking Glass. It generates low quality variations of images and that's about it.

>>21051 Is this with the "animefinal-full-pruned" model from the NAI leak and the "furry" module loaded as a hypernetwork (There's also furry_2, furry_3 and some others)? Or is this a newer version of NAI? What's the negative prompt? CFG? CLIPS? Sampler? Number of steps? Default eta noise seed delta of 31337? I've been trying to recreate this for hours and haven't even gotten close.

>>21233 This is actual NAI(not the leak), Furry Beta 1.3. Had my usual negatives for when I'm doing furry porn (extra hands, ugly, watermarks, etc). 50 steps, 11 scale, k_euler. You won't get anywhere close with the leak unfortunately, they've made leaps and bounds since their last update.

>>21238 I thought their current stuff was like, "pay for access to our server where you can send in requests to our machines" so basically your prompts and outputs would be monitored so you couldn't do stuff like this. Does it actually allow you to get a model on your machine?

Textual inversion is powerful. I've uploaded two versions of an ABDL textual inversion that I trained to: https://mega.nz/file/AIoBCDra#N2GRsWu3WBYBi0nCbVPhO8zp4HjRHw8rb41xbf6NpoE Both were trained on waifu diffusion 1.3 but I found the embeddings work surprisingly well with the novelai leak or the redshift style model. I highly recommend putting something like the bad-artist and bad-prompt (v2) embeddings in the negative prompt space (possibly with 0.6 multiplier) to compensate for the abdl embeddings causing bad artifacts. abdl-girl-wd16 is the best, often you'll want to put it at both the start and end of your prompt. -wd16b was trained with much higher learning rate and causes much worse artifacts because of it, but putting it at the end of a long description seems to help make sure a diaper is on if -wd16 isn't enough. The attached images used these embeddings, though they were the best of the ones I generated. Next up is to try using IRL ABDL images to train an embedding, my first test showed that I didn't have enough images to train it effectively.

>>21257 NTA, but NovelAI doesn't store anything image-related on their side and they delete any images you generate once you logout. They're fine with NSFW or NSFW-adjacent stuff for either image gen or text gen, even have a separate Discord server for it IIRC. They don't let you download their models though, and unless you're a furry I'm not sure the subscription is worth for image gen alone, especially post-leak where the furry model seems like the only real improvement. Might change on that stance if they improve the anime one though.

>>21064 A few of my better results. It seems you need to combine multiple tags relating to diapers + including "abdl", but they still come out looking like frilly briefs sometimes

>>21275 Oh damn, human and anime tags doing way better than I expected using their furry model. Some curation is still definitely necessary, but in a lot of cases I'd even say it's better than the anime model itself.

>>21341 Damn this is pretty impressive. Gotta say, using mittens and booties to hide the horrific hands and feet the AI produces is pretty smart.

>>21389 AI-chan can also do hands! (If you give her cookies and ask pretty please...) You can see some examples on Twitter: https://twitter.com/AzusaChansAIArt

>>21394 In your experience, how much diaper closeup photos should I include in my training dataset versus full-body shots and such? With dreambooth the model overfits and creates weird results really fast if I only train on closeups.

>>21562 I have pretty much no close-ups, I just feed whole illustrations. If you're giving it a mix of close-ups and other stuff you'll probably get a mixture of a full body and a screen diaper (which I'd call a blurry mess). Just give it the types of images you want it to generate and it'll do just that... if your parameters are in line. ~Azusa's wisdom

>>21625 surprisingly nice diaper and uniform

>>21643 >>21337 >>20844 Some of these are quite impressive, exciting to see how fast AI image generation has advanced over just the last year or so, hopefully it continues apace

It's not perfect but its pretty good and definitely better than nothing. Not sure how to fix the faces but you can tag the ones you want. https://anonfiles.com/P1OajbR8y9/diaper3_diaper3_5900_0.85-SD15NewVAEpruned_0.15-Weighted_sum-merged_ckpt Won't be doing a tutorial on how to train but no one else was sharing a model so I decided to make one and share it. It's trained off of stuff I found on kemono.

can we get some AI generated diaper hyper messing content? pretty please?

>>21681 Took a stab at it.

what a performance

>>21935 You might want to crop the toes

Embeddings available here: https://mega.nz/folder/gUpGzABa#tWTCHka0Wh5-wkMkmjBaDA/folder/FZIAHRzB abdl-irl12 is the better of the two. It's overfit on zoomed-in pictures of the butt area and of being photographs, but repetition of other prompt words and/or negative prompting fixes that.

>>21935 Fuck yes, I need more of this celeb stuff

I mustache you a question, anyone got a working pacifier embedding? i have trained the concept into some of my models but it's messy has hell sometimes, Thanks for the fish btw

that's good and all but i'm looking for something like this

>>16220 Inspired by the other thread, https://8chan.moe/abdl/res/317.html have you tried making packs of baby nappies but for adult size?

>>22087 I've tried, but results are usually mediocre. Stable Diffusion doesn't do text.

Whoever uploaded their ckpt is awesome!

>>22166 Just too bad it is struggling with the eyes big time and wants everyone to have the blue PieceofSoap eyes, and the model is very clearly trained mainly on PieceofSoap and God Hand Mar and seems a bit overfit as evident both by their style being applied to everyting and the noisy backgrounds and the eyes getting all fucked up. I think if it was retrained with "blue eyes" included in the training caption and doing a few less epochs of training it would be quite golden, because it is actually quite good as it is in many ways. Just a bit overfit and unresponsive in certain ways. At least we also have 4chan anon https://mega.nz/folder/pacmSRzK#GycG1GY4KhO589tuiJwDKA and https://civitai.com/models/4714/diaperai And mergin them together in different variations can yield great results as you combate some of the overfitting and hangups of the respective models.

>>22186 >Jack Black My pamps are in orbit! Great stuff, anon.

>>22186 Outside of Jack Black and Taylor Swift, who are the others?

>>22200 i recognize anne hattaway and scarlet johnason (no idea about the spelling for both those names an too lazy to google) absolutly no idea who the one next to black is

>>22200 >>22204 There is a filename

>>16220 Have you considered training a model, but one that does AB/DL calart style drawings?

>>22205 Wow on these just wow great artwork

>>22205 f8f83d2f23cbaeb6220ed09a622127487b3d8744c7e8ec56b50fe9d3f1822d2b is my favorite actress!!

Is it possible to train the AI to put Ocasio-Cortez in diaps?

>>22223 Fucking newfag

>>22225 Hahahahaha whoops

The prompt that you use mage a huge difference on weather these look like generic diaper art or something more interesting. That diaperAI model works pretty great so far. Can't wait to see it get better.

I started my own model using about 60 pics of various women in diapers in all types of positions and these were the results after about 2500-3000 UNet training steps. You can do this for free via google colab if you have a potato PC as well. Probably going to keep training it until it starts giving me garbage.

>>16220 ¿Has anyone on this thread tried training and testing with photoshopped images like diaperedxtreme and kevin abdl?

>>21661 thanks, works well

>>22351 You typically crop the images before you train them and also have a descriptor explaining what the image is so there is no need to get subpar photoshops of celebrities to scan. You are better off to just have good quality photos of people in diapers in a ton of positions.

>>22336 olso started my own model, but with very few pics so far

>>22336 >>22362 >>22364 I'm a bit new to this. How are you training a model using Google Colab? Dreambooth? I found this tutorial but the script isn't succeeding: https://www.youtube.com/watch?v=HVOvL2CyBT0 If one or both of you could provide a rundown that would be great.

>>22366 I use dreambooth and colab. Specifically this one. https://colab.research.google.com/github/TheLastBen/fast-stable-diffusion/blob/main/fast-DreamBooth.ipynb

Started trying to get a new AI model to recognize diapers underneath clothing. It works surprisingly well with pajamas. pretty much all models I have tried before this fail at that.

>>16220 Have you tried training Protogen to generate abdl photos?

some pictures I managed to snap at this year's protest against toilets, unbelievable turn up! Model: diaperAI

>>22364 Based, thank you! >>22416 It is a porn revolution.

Has anyone come across, or have any, hyper realistic AI generated or art that has the girl(a) wearing baby diapers (Pampers/Luvs)? I’ve seen some come somewhat close but not entirely to that description.

>>22453 AI still isn't there yet. I have tried to train a model just on ABDL diapers and it still can't quite generate the pattern perfectly. This is my results for a cushie diaper model.

>>22623 I actually kinda like the designs on these, and how they're not quite what you'd expect. The AI uncanny valley style works from the pov of a baby

oy vey!

Played around with outpainting today and turned an image of just a naked torso wearing a diaper into a full scene.

Spent the night finishing up merging the DiaperAI model with basis in the two above posted checkpoints from 8chan anon and 4chan anon (been working on it since forever but never satisfied) So new cartoon-based versions of diaperAI is up and running https://civitai.com/models/4714/diaperai

>>22453 Depends on how picky you are

>>22671 Damn some of those are great (2nd and last especially.) Maybe there's hope for this whole AI thing after all.

>>22677 Why were you ever in doubt? Have you seen the progress made by the community in just a few months? Can you imagine this shit in 5 years when it's all gotten streamlined, is really user friendly, fast and much easier to use and sort? Right now the only two things holding us back is that the easy-to-use dreambooth extension by D8 keeps having problems so training is a bit of a pain to get into and that some of the people that have got it to work seized the opportunity to keep it for themselves so they can try to profit it off of it. Just silly temporary problems. It was obviously going to work from the moment it was released to the public in October, the image quality was nothing like today, yet it was already extremely impressive on it's own. Right now we're in the boom of creating a billion small additions through custom models, LORAs and textual inversions which all have their own respective files you have to download, sort, and keep track of their respective keywords and shit. Exhausting. At some point that will all be automated and done by the program itself.

>>22224 ....yes Or just slap a diaper onto a pic of her and then start working with the AI to deepfake it. really not hard ...but yes, if you trained an AOC dreambooth and then gave it some merges with diaper checkpoints you'd have a checkpoint that could shit out AOC in diapers. for the curious all I did for these was go for the anime filter thing snapchat uses but of course not janky as fuck all of em came from babyfae, dunno why it made her into a tanned girl, prolly shite light

>>22224 Forgot about this one until above post. She's already recognized in Stable Diffusion by default, just use the DiaperAI model my dude.

>>22696 I have a doubt so that you could have the original image with the anime filter, did you extract the frame directly from the girl who uploaded the video or you uploaded some video outside your Snapchat account and you managed to place a filter and then extract the specific frames, I don't really know how snapchat works.

>>22696 My other question is, can the same technique be applied to tiktok anime filters?

>>22703 wat i was styling after it, not using it. i just used some of the abyss checkpoints and for the diaper all i did was clip it with a mask and then run a diaper checkpoint on it so it looked a bit more anime. 2-step process from original image of the girl to what you see- never had to leave stable diffusion

Luz in diapers

>>22700 Is it easy to do stuff like this if you have no prior experience with AI stuff?

>>22761 Yes. If using the CMD window sounds to scary use the automatic installer for Automatic111. https://github.com/EmpireMediaScience/A1111-Web-UI-Installer Download the model(s) you want from https://civitai.com and place them in the ...\stable-diffusion-webui\models\Stable-diffusion folder Also download (optional) https://civitai.com/models/7808/easynegative https://civitai.com/models/5224/bad-artist-negative-embedding And put those in the ...\stable-diffusion-webui\embeddings folder. These are easy textual inversions that improve image quality when prompted in the "negative" prompt. Run that sucker up, in the top left choose the downloaded model you want to generate an image with and start making. Start your prompt with "high quality" (an allrounder good quality modifier) and in the negative prompt have at least "easynegative, by_bad_artist, (worst quality, low quality, normal quality:1.2)" (the two first there are the name of the files downloaded, if you rename the files, also rename the name in the prompt, and don't include if you didn't download them of course) Then just input whatever you're trying to create in the positive prompt section. Then after some attempts you can play around with all the buttons and sliders and figure out what they do or read up on it here and there. The Unstable Diffusion discord is a resource rich place. https://top.gg/servers/1010980909568245801

It's easy enough if you can find the right settings and models

Elsa sure is looking good this winter

So now that I have an infinite diapered waifu generation machine, has anyone figured out a reliable method of getting messy images, or does a LoRA need to be cooked up for that?

Here's some movie posters I cooked up. Why does stable diffusion just shit the bed with words?

>>22918 The ultimate hope in the end is that I manage to train DiaperAI to do all without confusing them, at least mostly not confusing them since they share a lot of the same space both in tokens and visual placement for the model. For now it's just blind random chance when you pile on words like scat, dirty and pooped. Though you'll generally find everything else getting dirtier and broken way before the diaper.

Started playing with ControlNet and OpenPose Editor inside of stable diffusion tonight. Works good if you want your checkpoint to pose the subject in a particular way.

Trying out SD 2.1 and the WD 1.5 beta. The former definitely wasn't trained on many ABDL images, TI on it gave worse results than with SD 1. WD is definitely more powerful than 1.3 if you can prompt it right, it even can do IRL pics surprisingly well. Hands are still fucked though. Using the v2 embeddings in https://mega.nz/folder/oERgRK4a#TjpfF4To9CcN-vXswmXBGA

>>22999 WOW that thing is amazing, i was playing around with it and i made the discover that if you use higher quality, not only the pixels but the quality of the composition improves... i know im an idiot. This is are images i made with 512x512 and then sent to img2img in 1024x1024

Finetuned DiaperAI (by lifania) on tanukeart artwork. I think it is not bad. Mostly babyfurs, because i think 8chan does not like shota.

>>23023 These are super cute, good work!

>>23007 A little bit more

>>16220 Anyone here could generate AB/DL images of Equestria Girls.

>>23027 Here's a few I threw together quick. Couldn't get the others to look reasonable within a couple tries so I included 2 other images too.

>>23027 yes no not really

>>23038 thank you very much

>>16220 I'm not being asked, nor am I anyone to suggest or recommend, but I'll still mention a few suggestions for future models to train that could be added with AB/DL themed AI and thus have some extra variety 1. equestria girls and monster high 2. panty stocking 3. monster musune 4. mecha musune (Google search result and danbooru) 5. studio ghibli designs 6. 32-bit pixel art 7. liminal spaces and AB/DL rooms 8. Animatronics and fnaf robots 9. the x men of the 90s 10. some manga panels where there are nappy scenes or show nappy scenes.

>>23023 This is very cute! How did you merge the one lifania with tanukeart or furry? :3 Can you provide model link pls? <3

Wanted to see if I could edit this photo to remove watermarks and the old dudes in the pool. Too much jpg to really change anything else though.

>>23085 You would have to remove the noise from the image, I would first opt for Tencent Arc of all the online pages to remove objects and watermarks for free, this is the one that has worked best for me, as it lets you zoom in to a maximum of 500% so you can select the watermark you want to remove with greater precision https://www.vidmore.com/es/watermark-remover/#

>>23093 >>23085 The best I could get out of gigapixel. Not great

>>23085 I let controlnet look at the image to see if it could recreate it and it did a half decent job. I'm sure with some inpainting I could get the diaper to look better.

huggies

controlnet is awsome,

minus the body deformities and wearing the diapers backward this will look good once im done processing it.

>>16220 I think I found a Stable Diffusion model called ControlNet that allows you to generate a "copy" (so to speak) that allows you to generate variables from the same image, this could help AB/DL models to have less trouble generating more consistent nappies and hands here the GitHub of the https://github.com/lllyasviel/ControlNet model.

>>23133 NTA, but my understanding is it can work from preprocessed features from a particular image (i.e., canny edges) but it seems that in some cases it can also take some manual features (scribbles, etc.) as input. A more recent update also added something called "guidance start", which allows for correction of particular parts of the image (like SD's notorious abomination hands). Diapers are probably worth a shot too.

>>23152 Maybe the piecesofsoap and red moda drawings can help?

>>23163 Also the layout of merunya ?

>>23069 I promised Tanukeart that I would not distribute this model, unfortunately I cannot give you the link

Waifu Diffusion updated again, so I did a bit of fine tuning of my TIs. The aesthetic model is less general but has a pretty good fancy art style baked in.

Dumped a couple of my models on the diaper2000 mega https://mega.nz/folder/pacmSRzK#GycG1GY4KhO589tuiJwDKA/folder/9O0gyDyJ Have fun, they're not the greatest, but one of them is useful for inpainting pacifiers at least. I'm not uploading them to civai

Examples ----- ^

>>23259 Yeah the diaper models are just meh, but the paci model is half decent if you combine it with a better diaper model.