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:
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.
Description
V1 Log- 2/23/23:
This covers the two most known poses, those being upper body booba style and pov ass focus style, each respectively triggered by "boobapad" and "asspad". Testing concluded that it's ultimately better at getting the latter right, while booba mousepads are a bit trickier. It was trained to provide white or simple backgrounds and to include the whole pad, however, possibly due to differing sizes in the training data, it usually crops the image just slightly, even if neg prompted. You can get an uncropped image if you're persistent enough, though.
Oftentimes, I got the best results usually by using a semirealistic model, and my resolution size was consistently kept at 512x600.
This is only V1 and for now, it does a good enough job to be released. The overall dataset was actually remarkably small despite scouring the internet for as many images as I could. If I remake this, it'll probably with added generation images to help the dataset size.
Trained on 74 images, tested across multiple models. Tokens are recommended to keep the AI focused on what you want.
LORA compatibility is entirely random. Some work and some don't, and in the latter case it could be either LORA's/embedding’s fault.
THIS DOES NOT WORK WITH LOLI UNLESS IT'S ASSPAD