Pony Diffusion

The easiest way to get started generating locally is a project called Fooocus. It hides a lot of features behind screens so that it's quick to get it downloaded and running good images quickly. You can explore the advanced features as you go. You need an nvidia graphics card for the smoothest introduction. I started with a GTX 1070 and it was fine. Fooocus does well with lower vram cards. It'll download models from Huggingface which has a notoriously bad connection chance. If you get errors, it's probably just a bad connection, so keep trying or download models directly from huggingface or civitai. Checkpoints are the base models. Loras are additional finetuning that can be used with a checkpoint. Most checkpoints don't generate diapers well without a Lora. Put any downloaded models in the models/checkpoint folder or the models/lora folder. Fooocus by default augments the prompts to get better images. Under advanced options you can turn this off or leverage it for specific styles. You'll also find where you can switch the model and the loras used under the advanced menu. SDXL and SD1.5 are the 2 things to pay attention to when downloading new models or loras. SDXL is the newer option with 1024x1024 resolution and SD1.5 is the older one with 512x512 resolution. You can't usually mix and match, so for a clean Fooocus install, make sure to get SDXL models. You can use the filters on civitai to narrow down the options to just SDXL checkpoints and loras. If you see select a word in a prompt and press ctrl+up it will add weight to that word or words. There's a negative prompt which you can use to put things you do not want to appear, like bad lighting or low quality.

>>35730 Go back to your containment thread please.

>>35730 Nice

I posted these on the last thread but since it is dead. I been working on ABU's King diaper. I got it to be way more versatile from the start, but I can't get fine details reliably. Also I'm still figuring out Prodigy optimiser. Another thing (not visible on the example pictures I put here) is that it struggles with scale. Since the dataset is adult (mainly men), the diaper doesn't always scale proportionally to different body types. I test it from like a child to adult and elderly, fat and thin body types, realistic and anime. And it is always the "standard adult" body which has it the best... Makes sense since the training dataset is adult men and the ABDL community is very much into the "supersized diaper" it is hard to find good pictures. I need the poses and angles to be clean and simple or you just get monstrosities.

The new pony checkpoints are crazy, some of the photorealistic ones are amazing; they are as flexible as pony prompting but can look quite good. Here are some gens i did of a specific scenario with a lot of help from a lora on civitai, Although its really easy to make almost whatever you want. I like chubby girls but made some where they are skinnier as i thought this board might prefer that. Many of these have obvious imperfections but this is without any inpainting or anything, they just came out like this. I encourage you to try this yourself if you have some money to pay for a service and post more here, there is no longer need to train loras i think. Some checkpoints like this one seem to work great even with anime loras if you turn them down a lot. I used realpony cute_jp, "Diaper Change | BDSM | ABDL | SDXL PONY" lora on civitai (this thing should really be archived somewhere other than civitai before they do a purge of all their unpalatable content, its crazy good), pdxl embeddings, and an extension called sd-forge-couple (only on forge). You basically need sd-forge-couple or regional prompter to do any scenario with more than one person, and sd-forge-couple is better. Here was my prompt. prompt: score_9, score_8_up, score_7_up, score_6_up, score_5_up, zPDXL, zPDXLxxx, (diaper change:1.3), (2girls, couple, duo,:1.1) <lora:diaperchange_SDXLPONY_r60_768_dim32-a1-adamW:0.4> (looking at each other:1.2) romantically, beautiful face, perfect eyes, 1girl, cute beautiful sexy (submissive:1.2) (japanese:1.1) girl, laying back on a (thick diaper:1.4), (wet diaper:1.3), (pee puddle,:1.2) (urination:1.2), (huge pee stream,:1.1) (pee stream coming from pussy:1.3), pubic hair, excessive pubic hair, blushing, embarrassed, ashamed, thumbsucker, thumbsucking, hand on mouth, pink maid outfit, pink thighhighs, skindentation, huge breasts, (blonde:1.1), (buttplug, anal toy,:1.1) (dominant:1.3) milf motherly japanese woman with brown apron and jeans, (cruel smile:1.2), evil smirk, clothed, fingering her partner's pussy, explicit, dildo insertion, sex toy insertion, Negative prompt: zPDXL-neg, zPDXLxxx-neg, skinny, ribcage, source_furry, source_pony, source_cartoon, 3d, (looking at viewer,:1.3) underage, child, (plain background,:1.2) makeup, mascara, worst quality, low quality, normal quality, messy drawing, amateur drawing, lowres, bad anatomy, bad hands, comic, source filmmaker, 3d, Steps: 27, Sampler: DPM++ 2M Karras, CFG scale: 8, Seed: 3160048575, Size: 1280x1024, Model: realPony_cuteJpFixedNo03, VAE: xlVAEC_c91.safetensors, Clip skip: 2, forge_couple: True, forge_couple_separator: "\n", forge_couple_mode: Basic, forge_couple_direction: Horizontal, forge_couple_background: First Line, forge_couple_background_weight: 1,

Someone made a lora on civit that lets you shrink people. It works great with diapers.

>>37797 Really like the redhead one, do more with her please.

>>37802 Sure

>>37819 Sweet, thank you👍

>>37797 >>37819 man, the redhead with the headphones probably listening to some hypno training is top notch stuff

>>37819 These are great, just a shame about the weird drooling

>hating a cutesy drooly hypnotized/regressed chick bro, what is wrong with you?

>>37849 >>37895 Legitimately I don't understand how you can't appreciate a cute drooling diapergirl.

Pony Realism and Pony in general is top tier.

I'm glad to see that you guys liked my previous batch, so I made a slightly modified version leaning a little bit more into the "hypnosis" factor. Btw if you have any request (within reason) I would like to try it out!

A "first day of school" version

>>38176 My request is prompt info. Is it a custom lora?

>>38183 Nope, I just experiment with resources on Civitai

>>37927 drooling is a response to her brain turning to babymush due to the hypnosis tracks she's listening too, its as if you are actively witnessing someone being brainwashed/regressed down to absolute babyhood in front of you, and they are half acknowledging it and half too lost in diaperwetting pleasure to actually notice

>>38184 What model are you using for that?

>>38190 the main model is called "luminaverse", based on Pony xl. Then I use varios additional resources, mainly from the "Vixon's Pony styles" series

Made on pixai.art

I wanted to see how well outpainting handles massive expanding outpainting. I started at 1024x1024 and just kept going until I got this.

Why do sissies have to be so prevalent and ruin so much stuff in this fetish? I found a DeviantArt with pretty good diaper-girl/bimbo AI art but every single thing they post is covered with sissy and feminization captions.Also, some of the images have male faces on bimbo bodies. Why can't people just make hot women in diapers in these scenarios?

>>38271 "Why must other people possess inner lives"

enough with the AI find a real woman you faggots

>>38274 Where do you think you are right now?

>>38214 shit the image on the wall is insanely good, I could picture a couple of scenarios of a surprised girl being forced into ABDL by a babysitter with some sort of magical power diamonds

>>38281 This is autism

>>38311 Where do you think you are right now?

Anyone else use inpainting for diaper edits? Any pointers or settings for better results?

Threw a bunch of upscaled diapergals images into kohya and I think it did a good job of replicating that late 90s early 00s diapergal charm.

>>38345 What are you using to inpaint? I recently switched to foocus because it seems to handle SDXL inpainting and I could never get that working in A1111. Besides that just make sure whatever checkpoint or lora you use can generate diapers well and you should be good.

>>38347 That era was so focused on photosets I feel like there should be plenty of trainign data so this is a great concept.

had a bit of fun with the civitai generators - still a novice but made some smut that seemed alright

>>38345 I have never messed around with AI images but these are incredible and exactly what I need for my own "purpose". Please could you give me an ELI5 on how to do this?

>>38345 These are amazing. Love seeing AOC diapered. I wish I knew how to do this type of stuff.

Added another 100 pics to my diapergals lora. I might go through and separate the pics into diaper types so the diapers aren't such a jumble.

>>38376 >>38381 I will make a quick guide on the way I do it later this week / weekend. It works on just about any existing image of a person. (with their waist / lower body in the photo of course) >>38348 I am using A1111, seems to work pretty well.

>>38423 Thank you, please do!!

here is your wake-up call to use NovelAI for 2D as it is by far the best, but sadly people continue to use sloppy loras

>>38423 Aw hell yeah, thanks for posting more AOC. Looking forward to the tutorial.

>>38430 Is that the leaked NovelAI source or the site?

Have been playing with krea ai AI upscalling to improve some old adultbabysource.com images (boy, that site had some great photosets...) It's a generative upscalle, so the model takes a lot creative liberty, but all in all I liked the result (i fixed her pacifer on photoshop though, cause the model simply coulndn't get that thing right)

>>38498 Ok, but it also seems to have removed the diaper which is kind of a crucial part of diaper fetish imagery

>>38501 Sadly that's the case.

>>38498 >>38505 Both of these seem like downgrades....

can I get some recommendation on tools to start trying generate some AB scenarios involving a Mommy and a baby girl? preferably free tools...

How to use Inpaint to make a diaper edit of (almost) any image. There are probably many ways to do it, but I am going to be using Stable Diffusion with A1111 along with a few extensions. The exact steps I used to get the AOC edits plus others in this post. I am posting this on one condition, that you share your results! Also, once you have a successful output, you may become cursed. Don't say I didn't warn you. Setup = You will need to: - Have a decent enough graphics card (4GB Ram at least) - Install Stable Diffusion & A1111 (https://www.youtube.com/watch?v=kqXpAKVQDNU) - Control Net Extension - (https://www.youtube.com/watch?v=gkkfdWukb5I) - Be sure to download the "diffusers_xl_depth_full.safetensors" ControlNet model (linked in video description) - Download and install a decent base model from CivitAI, make sure its for SDXL or Pony based (https://civitai.com/models?tag=base+model) - I recommend (https://civitai.com/models/140737/albedobase-xl) or (https://civitai.com/models/372465/pony-realism) - Download and install a diaper model from CivitAI. It must also be SDXL or Pony based. I recommend (https://civitai.com/models/352425/diaper-xl-lora) If you don't have a good enough computer / GPU, there is a cloud-based version you can use: https://www.thinkdiffusion.com/#your-way-automatic1111 Finding a Photo = Just find a picture of your subject standing, sitting or lying down and make sure their lower body is exposed (waist, legs, etc). Standing works the best. Also make sure they are not in some weird or odd pose. What they are wearing also matters. Any clothing that keeps the shape of the body is going to work best (panties, form fitting shorts or jeans, yoga pants, etc). Flowing or loose clothing such as dresses, wide skirts, baggy clothes aren't impossible, but just don't work as well. The higher resolution the photo the better, but you can still get something good from low-res. The Process = Open up a1111 and go to the "img2img" tab. From here, you'll see a "Generation" tab already selected, underneath this, go to "Inpaint" tab. This is where you will drop your image. Just drag and drop it here. Next, you'll now have the ability to paint / brush over the image. In the top right corner, you can click the icons to control the size of your paint brush and the color of your brush for better control over what you're doing. Paint over the portion of the image you want to change. If they are naked, wearing panties or shorts, just cover their waist. If they are wearing pants, cover their entire lower body. Try to stay as close to their body as you can while giving a little room for the "poofiness" of the diaper. See first two images I uploaded. After covering the area, scroll down the the settings. I attached a screenshot of what I use. Sampling method can stay default, but I normally switch to Euler a. Sampling steps should be 30. Press the measurement icon (triangle) to get the resolution of your uploaded image automatically (1440x1800 in this case). Increase batch count to get two outputs and choose the best one. Denoising strength should be 0.9 or above. The larger your output image, the more ram it'll use. If your graphics card is not up to par, I would recommend cutting your image output resolution in half. If I were to do this in my settings screenshot, I'd change the 1440 to 720, and 1800 would become 900 (720x900). Lastly, scroll down to Controlnet, refer to my settings screenshot. Click "Enable" checkbox, "Low Vram" checkbox (again if your graphics card isn't as strong, no noticeable affect on quality). Select "Depth", them make sure "diffusers_xl_depth_full" is selected. Lastly, we need our prompt. I normally use something like this: "female, standing, wearing a white diaper, lower body, exposed thighs, <lora:D077b:1>" That last part triggers the diaper model to activate. Change this prompt up based on your needs (sitting, lying down, etc) but don't over-do it.

[Expand Post]

Once you're done, just press the big orange "Generate" button in the top right corner. Each time you generate, you will get different results in batches of 2, so keep going until you get an output you're satisfied with. Additional Notes = You will likely want to experiment with: Switching the "Inpaint Area" and trying "Only Masked" versus "Whole Area" Experiment with resolution when using "Only Masked" Turn off ControlNet and lower your denoising strength to see what outputs you get. If you are handling multiple subjects in a single photo, I've only tried inpainting their diapers one at a time. Switch your base model to see the different in styles So far, only Pony models handle wetting and pee stains There's a lot to this and you can find guides online about changing clothes with Stable Diffusion, this is just how I've done it. Good luck.

If you want slightly easier inpainting and setup with the downside of not being able to install add-ons checkout foocus. https://github.com/lllyasviel/Fooocus With foocus there are only 3 options for inpainting. Inpainting or outpainting, Improve Detail (face, hand, eyes, etc.), or Modify Content (add objects, change background, etc.). No need to mess with sampling, or denoising, or any of that BS. It just does it all for you. I'm sure I could probably change that but I haven't needed to. All 4 of these images were created with inpainting/outpainting and a custom lora I baked overnight on top of the juggernautxl checkpoint.

>>38517 >>38518 wow, just WOW thanks anon :)

First try with LivePortrait.

Anybody managed to get ai to do big diapers bulging under clothes. Big fan of the "girl tring to hide the beachball under her skirt" look.

>>38581 Sort of. My biggest issue was not having enough good quality images to train from so I can get diaper bulges under clothes but not to a comically huge level. If anyone has a handful of photos of super big diaper bulges I'd give training another go though.

>>38575 fuck, this shit is good, I cant wait to see what happens once we can all generate our own porn and we dont depend on studios generating content

>>38517 Incredible thank you so much. I don't have access to a good GPU on my laptop and my PC with the GPU is shared at the moment so it may be a while until I can do this but I will share my results when I can.

>>38595 Wow these are incredible, especially picture 3. What's your process?

>>38619 It's just regular stable diffusion with a custom lora I created. I think I might have used 30 pics trained with Kohya_ss.

adding a mask to just the diaper in one-trainer seems to really improve the quality of the Lora, on did 24 epochs with 19 images for this fresh one.

>>36201 >thanks for the recommendations been playing around with these

>>38661 and some slighly more hypnokinky

Someone for the love of everything please train a messy diaper LoRA for SDXL ;___;

>>38607 Thanks, had to do the wetting animation myself, but it was easy enough with some transparent PNGs and fading them in at the right time. Having the right idea is still crucial, and it's easy to spend a lot of time generating lots of mediocre pictures for what you are trying to do. Anime faces don't seem to work too well with LivePortrait, and most porn videos look stupid because the mouth turns into an O shape for 80% of the animation. I hope using your own webcam to create these animations becomes more streamlined soon.

>>38662 fuck, I was trying to alter the prompts for these images to generate different scenarios and I completely fucked up and got my account restricted, I get it that some prompts will raise alerts but darn it... creating abdl stuff can surely be extra tough

>>38682 This is easier said than done. Do you happen to have 30-40 unique hi-res messy diaper pic images? I have tried to create one but it just came out meh because I couldn't get enough good training images.

Tried another Cushies lora and it looks pretty good although it likes to sometimes mix up colors of the tape zone.

Has anyone been able to try the new Flux AI model from Fal? https://fal.ai/models/fal-ai/flux/dev The NSFW filter seems to be pretty strict on the site, but if you have a proper GPU, 3090/4090 it seems like you can run it without any restrictions. It does really good text for AI and sticks to prompt adherence better than anything I have tried yet. Hopefully someone can figure out how to get this working on lesser machines.

>>39079 Can you please post more Flux samples? Sadly I don't have a powerful GPU

>>39079 It's not really unrestricted locally because it's not trained on any pornography. As people report it refuses to show genetalia, so it's the same as any other first-version model out there, it's a better base-model but needs community training to become useable. Also it does not know what disposable diapers looks like, no matter how much I try to nudge it, it seems to insist on the cheap adult pullups types. Trying to dick around with different samplers and schedulers, but 4+ seconds per iteration means it's very exhausting to check out and test.

>>39123 I'm afraid that story with sdxl will repeat here - it's good for "generic" images, but not too good for nsfw, because censorship level after sd 1.5 became higher, and there is much less data from community (as there are no big events like NAI leak). Also the creators themselves say Flux would be impossible to finetune.

>>39124 What are you talking about? SDXL blows SD1.5 out of the water, the Pony finetune especially caused it to be a nobrainer better choice, on top of a ton of good non-pony models. But for our fetish Pony-based is the way to go. SD3 is more comparable because of it's higher requirement and unclear licensing terms. But yes they say so that it can't be trained. But we'll see, that might just be something they say and others may or may not try to prove that wrong.

>>39124 Combined with good sdxl loras it isn't half bad. It does hands and feet so much better than any models 1.5 or sdxl models out now.

Pretty nice but it's always the same diaper model sadly

I love seeing my chubby diaper boy!

>>39280 >>39282 cute babs at the adult daycare's sports day

>>39280 >>39282 Wasn't there an article about female athletes especially gymnasts having poor bladder control?

>>39079 What if we use Flux results for later inpaiting to fit nappies?

While flux is promising im still hooked on pony and making comics / hentia >>39302 I just trained a lora for Flux.dev with Aitoolkit, was easier then setting up onetrainer or Kohya - https://github.com/ostris/ai-toolkit it's meh my potato pc can barely handle running it. >>38773 That looks like a good lora, are you willing to upload it somewhere on hugging face? >>39202 Would you kindly resist using up all our thread space, thanks ...

>>39296 Yes, there was: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9021096/ https://www.telegraph.co.uk/womens-sport/2022/02/23/special-report-urine-flying-air-incontinence-crisis-blighting/ This gymnast didn't listen to the studies and had an accident that left her deeply humiliated and embarrassed in front of the world. The next day, she showed up with a new found confidence in her proper attire to earn the gold medal.

>>39124 >>39127 Just like I said, not only have people already trained LORAs on Flux, but they've figured out how to finetune it. (requires 24GB VRAM) https://www.reddit.com/r/StableDiffusion/comments/1erj8a1/flux_full_fine_tuning_achieved_with_24gb_gpu/

>>39127 >SDXL blows SD1.5 out of the water Try to use SDXL without checkpoints nor loras then.

>>39402 Without checkpoints? That's the model. What are you talking about?

>>39403 Clearly you have not been following how this technology has advanced.

>>39404 Uh... I don't think you know what the fuck a checkpoint is

>>39401 >but they've figured out how to finetune it. Ok, now Im impressed. >>39405 A blob of different models (CLIP, unet, VAE) that are trained to a specific state - to a checkpoint.

>>39406 Correct SD1.5 is a checkpoint. SDXL 1.0 is a checkpoint. Pony V6 is a checkpoint Betterdays is a checkpoint NovelAI is a checkpoint So how am I supposed to use SDXL without using SDXL?!

>>39407 Ok, you got me theret. But still, censorship in SDXL was more than SD 1.5 had a the beggining.

>>39408 Yes it was, then people finetuned it, and thanks to a better CLIP architecture and support for larger amounts of tokens it has vastly improved quality over the SD1.X architecture Likewise SD3 is better than SDXL but few people want to commit to training it because of licensing. And Flux is as good but probably better than SD3 (it has a shitton of UNETs) and I dunno how the licensing situation is but I think people are going to jump on board that one thanks to the already proven impressive output just on the base model

>>39409 They also got it running on lower vram too. I am able to run it on my 8gb 3070 and it only takes about 90 seconds to generate an image. Hopefully someone puts out a good training guide for it.

Made another test flux Lora with a pretty bad imageset, result are interesting, could have sworn in flux dev they had better pacifiers, but maybe it's suffering from the full body view / close up issue.

>>39424 These are great. When you have a LoRa ready for release, can you also release it somewhere else that's not CivitAI? They block tor now.

>>38773 is this lora downloadable somewhere? if not, could u upload it somewhere maybe? looks great, thanks

>>39424 Can you post your process for training? I'd love to throw the training images I have for my sdxl loras at flux.

>>39079 Last photo is pretty cute. Would love to see a a whole photoset of it.

>>39424 These are INSANELY good. Please good sir, make that lora downloadable.

Anyone got more Emma Watson?

A few weeks/ couple of months back there was a really good ai pic of a cop girl in a diaper on rule 34, but it sadly got deleted, would anyone have happened to be able to save it or happen to know the source?

>>39505 Her correct title is "Enema" Watson.

>>39468 alright dumping my v2,v3, and v5 flux lora's here, trigger is "ddiaper" v2 - random set including anime drawings from various artist [it's ok] from above v3 - Vroid Posing set with a custom diaper model i customized from gumroad, lower less 3D. v5 - ABDL Kings Diaper Set [i think this one is the best] https://mega.nz/folder/lb03FTTZ#rGry9y4vLDZfinZUIzwH6Q example prompt : Macro photo HBO Cinematic of a adultfemale wearing massive diaper and tanktop, detail lighting, riding carousel horse , bubbles, rainbows, sundogs, fair skin, straight blonde hair, hairclip,print DDiaper with stars, full body view, petite pastel 25 yearold, looking at the view, pink princess castle, solo,

>>39546 Thank u!! How many pics did u use in your set?

>>39546 I dig it

>>38517 Can you do Lauren Boebert?

>>39552 v2 ~ 20 v3 ~ 72 v5 ~ 42

Someone uploaded a half decent flux white diaper lora to civit. digging how great they look.

>>39631 Can confirm, the white diaper lora is excellent. I've even had success with different printed patterns/logos with the right prompts.

the flux white diaper lora from civit ai is very solid, best ive found for sure

the civit trainer is dead simple to use as well. Just throw some images at it, add some captions, and let it cook. $5 will get you about 2 loras plus change left over. These are from a onesie lora and A diapers under clothes lora I trained in about 4-5 hours on civit.

>>39678 do you mind sharing your settings? i couldn't get the lora to work. my result always looks like shit

"Baby made a boom boom!"

>>39631 I was working on updating my SDXL one for flux when I saw it and it's as good as anything I was planning to make. Bittersweet to be beaten to the punch, but saves me a hell of a lot of time fiddling around with kohya.

>>39709 Yo that was me! It's nice to see something I made being enjoyed :) I reckon your XL lora was fantastic, would you want to merge training data? I can take the best of both of ours and improve the flux one, idm doing the training

>>39728 Any chance you could spill any details regarding how you've prepared and captioned the images? I'm experimenting with a training diaper LoRA myself and haven't nailed down a good preparation/captioning pipeline yet, but I saw how well your LoRA performed and wondered what the dataset roughly looked like.

>>39729 training a diaper LoRA*

>>39729 Simple really, presize.io to crop everything, zoomed in on the diaper as much as possible without going over 100% zoom and while also leaving the body position identifiable Then i used joycaption to caption them all. Its fairly good at identifying diapers but I changed any complex descriptions of the diapers to just “wearing a white diaper”. It will probably think the diaper is underwear or shorts a lot, you’ll need to individually edit each. Then civitai trainer 1000 steps, 1024x1024 size

I made a quick and dirty barebum lora on civit and I think it came out decent. If you play with prompts you can get pretty creative with them. One thing I found out is you really do need to keep the images at 1024x1024 as civit likes to crop them however they feel.

>>39739 Ah okay, so pretty much what I was already doing, lol. Main problem I had was JoyCaption kept giving overly verbose descriptions and hallucinating a ton, so I tried switching to other VLMs but those didn't fare too well either (CogVLM/CogAgent got close, but wasn't as good as JoyCaption). I've modified my JoyCaption script to use the LLaMA 3.1 instruct prompt and that seems to have helped quite a bit, as the descriptions are a lot less verbose and seem to be a little more accurate now.

>>39739 Wow, that's so much different than SDXL training. No wonder I was having so much trouble. SDXL needs a careful mix of close and far with different poses thrown in. I'll see if I can make something different enough to warrant publishing.

>>39690 i use comfy UI but the settings should be transferrable

>>39741 Damn I love the diaper packaging one in this set.

>>39741 If anyone makes a Pampers lora I would love u long time. Also please upload somewhere except civitai. It requires registration to download nsfw stuff.

>>39781 I'll happily make any lora if you can provide me with high quality photos lol

>>39784 How many photos and how high quality are we talking? I could probably find a few in the baby diapers thread or the pull ups thread for a pull ups lora.

>>39785 I use between 20-30 in my Loras, although I am starting to think with flux that you can do with less as some of my first epochs look decent. Here is a goodnites lora that I made on civit over night. Amazed with the quality.

>>39728 >>39797 Could you please share instructions on how you're doing these? They're fucking sick, and I would love to use the image editor to pamper some girls I know IRL

>>39812 Your best bet is to lurk on the stablediffusion subreddit to learn, but the quick guide is to download whatever can run the model you want, I use webui forge for the flux model and foocus for inpainting. Then hit up civitai.com for a lora to be able to generate specific diaper stuff. Keep in mind you will need a GPU with at least 8gb of ram to start.

>>39816 Sorry, my specific question is: what checkpoint are you using for the white diaper LORA? Can you share a sample prompt?

>>39797 These are great! Any chance you could upload somewhere that isn't civit?

>>39818 The loras that civit spits out for flux are literally 18mb. They can be uploaded here assuming it doesn't block safetensor files.

>>39836 How'd you get yours to work so well? I just started today and i got shitty results from in painting

>>39851 Not really sure how i got here, but i found the best setting is to use the diaper lora (SDXL) found on civitAi, with RealVisXL inpainting version. Use Euler A (uniform) as the sampling method with 35 steps, and set the inpainting mask mode according to your needs. For the prompt, just describe the original photo and add the diaper lora.

>>39883 Holy hell. Who is this and where can I find more of her?

>>39882 I used realvis yesterday and it was a night and day difference for me. It's just that the diaper is a mess most of the time (missing or no tapes, really disfigured). Would you mind sharing a sample prompt for >>39836

>>39650 >>39836 >>39883 Always great to see fellow inpaint enjoyers in this thread. Looking forward to more!

I keep thinking, as a big fan of Sebtomato... I wondwr how many more years we need to be able to feed a age regression story to AI fir it to actually create a movie or video about it properly depicting the whole "becoming a baby " process

When training Flux loras, captions are far less critical. Used the same 5 images to train, the top row had written out captions and the bottom row just had the word diaper. The captions do have an impact though. If you use Joycaption and leave in the details about the diaper, you'd need to prompt those details or it might leave them out like the tapes in the second picture. It'll also say stuff about about the body size of the person. I found that if it calls a normal person slim, the resulting lora makes chubby people by default. Less is more it seems.

Flux is cool.

>>39897 Holy shit these are masterful.

>>39893 try this prompt. Positive: 1girl, portrait, southeast asian girl, petite, slim, skinny, white skin, small tight tits, light skin, messy black hair, age 24, wearing bra, bedroom background, <lora:D077b:1>, diaper, white diaper, wearing a diaper, perfect tapes, <lora:polyhedron_all_sdxl:0.35> detailed skin, blemishes, perfect eyes, Negative: saggy boobs, blonde hair, monochrome, body hair, cleft chin, bad anatomy, bad proportions, cropped, deformed, dehydrated, low quality, ugly, unrealistic anatomy, shiny skin, anorexic, dirty body, wet body, pubic hair, short hair, shoes Use "RealVisXL" with "Diaper XL Lora" and "Polyhedron ALL SDXL" Set the VAE to automatic. Do let me know of the results.

>>39904 Is this with image gen? I was trying my luck on inpainting as i have lots of photos i would like to be diapered lmao. Tho i've tried your prompt on image gen and they work freaking great.

>>39919 Yeah, you could use the same prompt for inpainting. your image looks like it's lacking sampling steps or on of your lora strength is set too high. can you share your setting so that i could take a look? a screenshot of the webui will do.

>>39904 this is what i got after playing around a couple of times. Kinda got better and kinda not. I want it to look natural :/ any help i can get is a huge thanks!

>>39922 this is good, your prompt is also good, better than mine by the looks of it. mine had such few details. I noticed that your polyhedron lora strength is quite high. i suggest lower the strength to 0.35 and try again with these settings. Also, you could try to simplify your prompt. not sure it that will make any difference.

A woman is standing outside on a sunny day. She is wearing a white diaper. The left side of the diaper is plain blue and the right side of the diaper is red with white polka dots. There are four diaper tapes. She is wearing a green hat on her head and a yellow crop top with the words "fuckmuffin" in glittery blue letters. She has short blonde hair.

>>39919 For inpainting in my experience, you don't need a very complicated prompt. Something like: "female, standing, waist, wearing a white diaper, perfect tapes, <lora:D077b:1>" Is perfectly fine. I got the 2nd image above from: - Use an SDXL model (I use Zavy Chroma, but most work great) - img2img > Inpaint - Upload Image - Paint over area where diaper will be (refer to 1st image) - Sampling Steps to 30 - Set width and height to the width and height of the image (if your GPU can't handle it, size it down but keep the aspect ratio. For example, 1440x1800 can size down to 1024x1280) - Denoising strength to 0.8 - 1 (I always do 0.95) - Control Net enabled, set to "Depth" using a depth model from the dropdown, ideally "diffusers_xl_depth_full" but it can be any - Do a batch of 2 - 4 so you can keep the best result and improve it further - Generate I got the 3rd image above from: - Take the best diapered image result from your first inpainting - Upload it to Inpaint replacing the non-diapered image - Re-inpaint over the diaper but cover more area, maybe "wider" or a "bulkier" crotch area - Turn off control net, lower denoising strength to 0.55 - 0.7 (I used 0.65) - Switch to a realistic Pony model (I used Zonkey Realism) - Experiment with "Only Masked" and "Whole picture" Inpaint areas - Generate until you are happy with a result, but this isn't the final output - Take the image you are satisfied with and then upload it (it should keep your same inpainted area mask) - Turn control net back on, set denoising back to 0.95 - Generate (again experiment with Only Masked and Whole Picture) I don't really know what I'm doing from a tech perspective, but this is my workflow. Though sometimes the first result from control net is perfectly fine without re-inpainting with Pony. Your preference. Would love to see more inpainting results in this thread.

>>39928 Holy shit are you diapering Lorena Abreu's Dora cosplay? Please do more. Probably my top internet crush I want to see diapered.

One thing I have noticed when training is that you can train/prompt for wet diapers much much easier with flux. I captioned my images blue line/orange line and prompting works with it almost every time. This was from about 15 vintage attends pics.

>>39949 Doesn't really look wet though. You should get images where the wetness indicator is partially dissolved/changing colors. Bonus points for visibility wet without an indicator. (Image not AI generated :P)

progress is being made on a lora that understands different parts of the diaper.

aifog, shame on you. For what it's worth, you made an image using the dude who just makes random concept loras for points and gave nothing to the guy/gal that made a great white diaper lora.

>>39994 You know you're right,let me fix that by posting this, i did mess up there, it really was an oversight on my part. I gave buzz to the guy/gal that made the white diaper lora, but contributing an image does go much longer way. I'm very sorry for that.

>>40012 Wholesome af