Making the ultimate Diaper AI (Patreon)
Content
Hi all.
Earlier in the month I spent a few weeks struggling with various AI tools to create ABDL-themed images. A dozen images just didn't work out.
Eventually I decided to go all in and build an entirely new ABDL-focused AI to rival all the big image generation AIs.
It involved collecting a few thousand decent reference images, and then giving them lengthy descriptions using consistent terms for outfits, view angles, poses, etc. This turned out to be a way bigger task than it sounded, and I ended up making new software just for quickly adding a bunch of descriptive tags to bunches of images at once. e.g. Selecting a few hundred images and giving them all the 'disposable diaper' tag at once. Once I made that tool things sped up considerably. I also rated the images 1 to 5 stars. On a side note, I now have an amazing gallery of thousands of ABDL images which can be easily searched with 4977 different description tags, and it's the best.
Then there was a lot of experimentation with how to actually train a model on this data. Each experiment took 10-30 hours without being able to use my computer in the meantime. I was definitely getting better results each time, though simply ran out of time in the month.
Right now it's close to being usable, being able to create images in different styles and contexts. I've decided not to use other artist's styles directly, since it feels like their identity, though may sometimes mix elements in to try to get the perfect image. It generally does better on cartoony art, because the details are larger and more obvious. One issue is that during training the images are downscaled to a very low resolution, which means faces, and things like pacifiers, get very messed up, and then get trained as messed up versions. I've been thinking about ways to solve this and think I know how from some earlier tests with Ahsoka, but need to do another few days of data collection and training to see if it works. An alternative option is to drop a few thousand dollars on a high end graphics card which has enough memory to train on high resolution images, and do it faster, which is very tempting right now...
Right now things are close, but not quite there. There's always something just a little bit wrong with each image, when you look closely, or sometimes not so closely.
It struggles with more unusual scenes like those involving giants, due to a lack of good training data. A lot of them are turning out blurry due to using low quality pasties as training data, but I've realized that it is probably better to just train on a few high quality pictures multiple times each, and will change that in the next training run. I can also potentially start using the best outputs for the next training loop as well, building up a library of better giant images if even only a few of them work well, and eventually having a big library of great reference images.
All in all it's a process, but hopefully the quality of the sample images is enough to show it's worth it, when a month ago it was hard to even generate a person crawling. It might be very possible to soon write out some scene descriptions for a diaper dimension story line, and immediately get photo-realistic images. With VR tech, voice synthesis, etc, you might even soon be able to walk around in one of these stories.