UPDATE: V2 IS OUT AND AVAILABLE FOR DOWNLOAD

V2 Log:

Now you can truly and freely create a mousepad of any waifu you want! Newest version is capable of breast, ass, and even crotch mousepad styles!

For V2, I resorted to the laboriousness of cropping, framing, and even making new images from scratch. V1's cropping issue mostly stemmed from all the images being different sizes, backgrounds included. This time, all images, collected or generated, were placed on a blank 1000x1000 canvas so that all outputs during generation would remain consistent in size. Cropping will now only happen if you change the image resolution too much (the further from 512x512 you go, the more it will begin to deviate and distort).

V1 was also incapable of another pose, the legs-up pose, for which on the internet there were not many good training images. The majority of that data for V2 was personally generated by me, and then framed using mousepad templates. Sometimes, when prompting for it, it might put out a pair of mangled-looking legs, but most times it should give you what you want.

Sample prompts were used in the training for this version:

asspad, white background, from behind, ass focus, 1girl, looking back, ass

asspadup, white background, from above, legs up, leg hold, on back, 1girl

boobapad, white background, upper body, portrait, 1girl, breasts

oppai mousepad, white background, upper body, portrait, 1girl, breasts

On very rare occasions, it will completely ignore you and give you a pose you didn't ask for. This seems to be mostly tied to the mousepad token.

THIS WAS TRAINED AS MY FIRST LOCON MODEL, IT WILL NOT WORK WITHOUT IT

Luckily, all you need do is copy-paste the URL to the extensions tab, install and restart, and then load like a normal LORA: https://github.com/KohakuBlueleaf/a1111-sd-webui-locon

Dataset, cuz there's actually some good shit in there if you wanna use it:

asspad

boobapad

crotchpad

Trained on 140 images, tested across multiple models. Tokens are recommended to keep the AI focused on what you want.

Acceptable compatibility with other LORAs and embeddings.