Change Browsing Level to "R" if you want to see the full set of example images.

A very simple character LoRA for the indie VTuber Mipyun. Trained on a dataset of only 16 images that largely consists of screenshots and reference sheets so it may not be accurate and have inconsistent results especially in regards to the outfit and especially the legwear. This is largely due to the dataset containing reference sheets of three different outfits without the express goal of generating any of them accurately. I'm certain that it possible to engineer prompts for each of the outfits but that's not my area of interest nor is it the intention of this LoRA. Play around with it to see what works for you.

Probably won't be updated.

All example images were generated without a refiner step. You can probably make your generations look much better with improved workflows.

V1.5 Changes:

Updated dataset. It's still small at only 19 images, however, I made all images within the dataset to be of her default outfit. This leads to zero outfit mixing and more consistent legwear generation through higher CLIP strength and proper weighting. Low CLIP strength still results in inconsistent legwear. It's not perfect but it's an improvement nonetheless and, again, 100% accurate outfit replication isn't the main purpose of this LoRA. Other than that, it's pretty much the same. No real improvement to overall generation quality, thus the V1.5 designation as opposed to V2.

Triggers:

Mipyun, snail antennae, orange hair, short hair, hair buns, blue eyes, ,

white shirt, short sleeves, dark blue skirt, frilled skirt hem, pink skirt hem, white hair ribbons, dark blue sailor collar, midriff, pink bow, mismatched socks, stripped thighhigh, striped sock, pink and white stripes, single thighhigh, single yellow mid-calf sock, single sock,