>>41624 These are great, what did you use for them?

Question regarding training LoRas specially with flux. How do you prepare the images? Are they just 512x512 close up of the diapers in different angles? and how do you caption them? Trying to dive in to training since there's no guide for diapers lol, any help would be appreciated

>>41904 closeups will end up confusing the AI and where it should put the diaper on the person. Best bet is to just pour as many relatively similar images into it as possible.

>>41904 For Flux, close ups with Joycaptions will work as long as there's enough image to tell it's on a person. The white diaper flux lora did that and the results are fine. Even diaper.1 dev did that, but I know better now. The problem you run into with the only close up approach is that you are training 1024x or some higher resolution of the diaper itself. When flux is actually generating, the area the diaper takes up is really small. More like 200x or 400x in size. You'll find when you want the diaper attached to a person at a reasonable distance, the lora hasn't really seen that even though it knows exactly what the diaper looks like. You can prompt just a diaper and it's a perfect image. You don't need a lot of full body images, I've used a few as 3 (~25 total images). Make sure there's no repeating faces or backgrounds and you won't have a problem. That's usually enough to connect the dots between what it learned from the close ups and how that translates to normal distances. You can get away with training at 512x512. It's a ton faster so you can fail fast and learn what doesn't work. Higher resolution is better only because eventually the lora will learn how to make low resolution looking 1024x images. If you can get a good lora at 512, you don't need to retrain it at higher resolutions. Mainly because increasing the resolution needs adjustments to the learning rates and it's not linear.

>>41919 thanks a lot for the input. I was able to do it earlier using 24 images and i did not crop in. although i'm having issues about the lora not knowing what the back of the diaper looks like (like when prompting facing away from the viewer, it just puts the diaper backwards). is it because i did not prompt the back side of the diaper as back side or something? I did have images showing it. Also, do you think it's possible to get it more detailed? More and better images? more training steps?

>>41921 Calling out the front and back helps. If it's joycaptions, I'll add in something like "The front of her diaper is visible." Make sure to clear out any descriptions of the diapers themselves. Any descriptions tend to push the lora into thinking those aspects are optional. For close ups, I don't even bother with captioning beyond "diaper, front" or "diaper, side, back" Haven't noticed any issues with at strategy. Quality over quantity every time. As far as undertrained or overtrained, it's my biggest gripe with Flux, it seems like it can be both at the same time which makes troubleshooting really difficult. Prompt for just a diaper, either on the floor or a close up shot. That'll show you how well the model knows the target. It's a lot of trial and error. There's no magic formula.

>>41921 I find that removing the background and then inpainting a solid colored wall behind them tends to not get it to train random things that I don't want it to train. Foocus lets you use masks to do this rather easily. Just make sure you only use 1 or 2 pics from the same sets and specify "From the back" or some such in your captions.

I'm new to this, I've followed what I can. I'm using comfyUI with Flux and the White Diaper for Flux lora. Annoyingly I keep getting outputs of like cartoon/anime and the pictures are awful close to loli, how the fuck do I stop it from doing that?

>>42056 Got an example prompt? "photo of a woman" should be more than enough to push it away from anime.

>>42061 Yeah, I'll grab one just now. It seems like if I get hyper-specific it will stay away from it like you said. I tried this before about a year ago, and the examples I copied at the time had a load of negative prompts setup so that it would stay output consistent stuff, but the one I copied recently didn't have any negative prompts/inputs. Anyway, I was doing this before "Emo girl with purple hair sitting in her bedroom, wearing a diaper and knee-high socks" Mainly to test it with something simple, but it kept generating loli and that made me uncomfortable as fuck lmao.

>>42066 Ah. It's Flux vs SDXL. Flux has a T5 text encoder and really understands the prompt while also expecting a pretty long prompt. Smaller ones work, but if you say girl, you will get a girl. Not like the 1girl booru tags of SDXL and co. Flux also doesn't do negative prompts, so that's another issue. SD3.5 should be more familiar to what you did a year ago, but no diaper lora yet.

>>42067 Awesome, that's great info. It's cool tech I like toying with it. The good thing is if I'm pretty specific, I can get it to do the right thing now, just takes fine tuning it. Thanks for the help dude!

So, you guys are the experts in the community... or I'll at least come to you as such lol. What are your experiences with AI Voice generation. Curious how good it can be and what pitfalls there are. I suspect AI's not good at intensity (screaming, angry, strong emotional tones) but perhaps somewhat passable for everyday conversation and dry stuff. What say you, oh interwebz?

>>42082 Beautiful, impressive, inspiring

>>42082 Thanks for the inspiration

>>42082 >>42118 >>42119 We need more AI Videos like these ones. Please do more videos like these. With the girls in school, in office, in their nurseries and in swiming areas.

>>42084 xtts2-ui5 works quite well. It is flat, but if you need a relatively normal sounding speak pattern to read erotica to you, it works.

these little mini videos are fucking awesome. these could actually be really really hot if they didn't involve chick fil a lmao

>>42138 Lol sorry, I was just testing out old images I had generated from that guy who made the chick-fil-a requests >>42132 It's very easy to create these, you just need a starting image and your imagination

>>42139 Trust me, i am REALLY sucky when it comes to make the most basic AI stuff.

I recently cooked a couple of LoRA - one of them is ABU BunnyHopps. For now the version on illustrious-xl, but soon it will also appear for ponydiffusion https://civitai.com/models/921894/illustrious-abuniverse-bunnyhopps-diaper-abdl take a look at my other LoRAs https://civitai.com/user/MikuMasterRace/models

>>40016 I've been thinking about this quite a bit and I dont know why AI prompts dont alloe you to zoom in on oirtions of an image to do correctuons and details. So in Photoshop or similar you would zoom in on the hand holding the keys and say compare to all images of a hand holding keys and correct for anatomically corect matching style of rest of picture. Something like this.

Could somebody generate a photo of a gloating girl in a maga hat and a fresh diaper laughing at a crying girl with a Kamala hat and shirt wearing a wet or messy diaper?

>>42250 Should add a spanking to that mix.

>>42250 Great suggestion lol. I already generated these the other day in anticipation for today's result. Video generations distort the face. Its fixable, I was just too lazy to do it.

>>42259 Kek Considering the tantrum that is coming I predict a lot more shit filled huggies from the Kamala crowd in the next few days.

>>42259 More of these AI Diapered AOC MP4s pleease!

>>42259 Holy shit this is good. I always find the AI videos to be a little unsettling but this is such a hot concept. First election is to get Trump back into office. Next election we install diaper-fascism and get AOC back into Huggies!