I recommend doing a low denoising (~50%) img2img pass on images. I'm aware this lora lowers quality quite a bit, but I've been enjoying the results immensely. This model is trained on 66 images of J-Pop idols found on the J-Pop fandom wiki. Tagging was intentionally sparse (such as only tagging things like "white shirt, black shorts" for the clothes) in order to see if some of the fun details could be baked into the activation tag. It is more of an experiment that I wanted to share. Usage <code>jp_outfit</code> is the activation keyword. I prefer to let the model do the heavy lifting just to see what comes out, so I would recommend only giving very basic details of the outfit. Some example prompts: <code>white shorts, white shirt</code> <code>long sleeves, pink skirt</code> <code>red dress, short dress</code>You can see more in the example images. <code>photorealistic</code> should be in the negative prompt since this was trained on real images, not anime images. Known Issues Obviously, quality is the number one concern. In the future, I may find myself training this model again to try and clean it up, but I want to save this one for posterity as my first result that I'm happy with. This model has been in my brain for well over a year.

Files with SHA256: 4026f601c3894a927a087995dafc1ed8d56f21de5743d2225046add194c29411