Type | |
Stats | 5,289 |
Reviews | (317) |
Published | Mar 21, 2023 |
Base Model | |
Training | Steps: 16,812 Epochs: 100 |
Trigger Words | diaper pullups goodnites back tena pattern [17] abena pattern [13] abu preschool pattern [5] pampers pattern [5] bambino classico pattern [4] lfb pattern [4] lfb vintagebaby pattern [4] +34 more |
Hash | AutoV2 D8DFFC133E |
The stylized model are based on the 8chan and 4chan models made by two different anons
4chan model https://8chan.moe/abdl/res/16220.html#17641
8chan model https://8chan.moe/abdl/res/16220.html#21661
This is my attempt at making a model that can generate women wearing diapers for any diaper fetishist out there.
Since Stable Diffusion works a lot with broad associations and "diaper" being associated with a lot of toddler and baby apparel images, it is advised in most cases to put emphasis on the age of the character you want, this is less of a problem with the newer models but is still adviseable.
The models have not been trained on any sexual activity nor exposed genitals, so images of a sexual nature combined with a diaper CAN work, but also has a high probability of giving distorted genitals and warped results.
--- Update Log ---
24/3 - Updated with V2, trained with a curated and hand captioned set of 168 images, it can sort of... do pullups, it has been tagged with a lot of patterns but it doesn't really get them. The dreamboothed model out of the gate wasn't very good on anatomy and such on it's own, and it still faces some issues with crushed faces and whatnot, but I did some merges with other abdl models of mine and others and at least you can get a lot of doable generations with it while most of the uniqueness of the model is retained.
23/2 - Gotten back into training again now that most of the issues with D8's dreambooth seems to be gone and I can find the energy to do so. Did a test run with auto-generated prompts for the training images and got pretty good results with that, so did a 200 epoch run on 50 images I got done manually prompting for training today and merged that in with the Semi model and an overfit model of mine with a Reverse Smoothstep 4 blockmerge, keeping the alpha/clip of the original models, then merged those two models together 50/50 with the alpha/clip of the Semi model. Since this is just the spawn of a test run I call it 1.5 since it's intended to be much more versatile rather than just a side-grade/minor upgrade.
12/2 - Updated with two different levels of cartoon/anime styled models in a long attempted and trialed merge with the anime novelai 4chan model and the cartoony 8chan model. Both who are overfit in their own very different and respective ways that made it hard to get a good result I could be satisfied with.
30/1 - updated with broken clip values fixed, updated with new images to be representative of the new prompting understanding of the model, easier to work with now. Wish I knew about this issue before permanently deleting a bunch of my overfit training attempts.