Shinomiya Natsuha

Shinomiya Natsuha [pronounced 'she-know-me yah' 'gnats-ooh-ha']

Saw a bounty for this LoRA, so I made this on a whim, since oddly enough I've seen this character before while looking for dataset images--never used them, of course, but they definitely left a lasting impression on me. Other than that, I have absolutely no idea who she is.

The images I found of Natsuha were... intense to put it lightly, both in content, and in format (low quality GIFs with plenty of visual artifacts that were also mislabeled). Because Monet likes vanilla above all, I tried to make the dataset as vanilla as possible. Since the images in the dataset were mostly low-quality GIFs, I had to improvise, and refine each image both manually and via inpainting. That being said, the dataset only has 20 images total; 14 of which are actual images of Natsuha. If you know me, you know I can't just have one thing, so I included 6 different images of similar-but-different characters doing various... acts. For the images of Natsuha herself, I manually edited each image so that only Natsuha is in the image--even going as far as completely painting over other people in her various images. Since there wasn't a lot of images for me to find, I didn't have much of a luxury to include raw images. Arguably, I could have tried to find a video online, but I like the challenge of turning GIFs into high quality PNGs. lol

I also experimented with the tagging by completely removing the "1girl" tag that almost always populates my other datasets. However, I kept the "solo" tag in.

The trigger word for Natsuha is her full name: Shinomiya Natsuha

For the other trigger, it is simply "ekiben". This is an experiment to see if it can't override "suspended congress". Just like how I'd personally say パイズリ "paizuri" over its English counterpart, I also prefer saying 対面座位 "taimenzai", 正常位 "seijoui", and 駅弁 "ekiben" over their English counterparts, among other positions XD. Speaking of which, there's exactly 1 image of taimenzai in the dataset, but since it's coupled with a hug tag, I'm not sure how effective it will be.

If it turns out this first iteration isn't good, I'll use this model to refine the dataset and add more to build a second version. As of writing this, the first sample image has completed (training it locally), and besides the fact the hairstyle is wrong, it otherwise looks like Natsuha--at least from what the dataset has shown her to be.

Since I trained this LoRA locally, I opted to train it on the Prefect Pony checkpoint... though in hindsight, I probably could have trained it on LimestoneYen and it would have saved me a lot of time (5 minutes @ 8-steps for LimestoneYen, over 15-40 minutes @ 30 steps for Prefect; LimestoneYen is what I like to call a Hyper P model). Regardless, I chose Prefect because I like the style, though I have no clue what effect this would have on the LoRA itself.