A Nice Waifu from Shijou Saikyou no Deshi Kennichi. This is the Locon version based on my Shigure KohsakaV2 LORA.Works fine with: &lt;lyco:KousakaShigureV3:0.7&gt;, kousaka_shigure, I was only able to isolate one outfit:<ul><li>ShigurePinkOutfit (It her typical pink kimono with an under mesh top)</li></ul>As most of my LOCON this is a product of my laziness as instead of cleaning a new dataset I just did a retrain of one I have already collected, cleaned and tagged. It is nice though. I find them marginally better than normal LORA, but just a bit.I have also added the dataset and the tags summary file so people can experiment.

Shigure Kosaka (LOCON version) - HSDK

data10b.zip

KousakaShigureV3.safetensors

v1.0

A Nice Waifu from Shijou Saikyou no Deshi Kennichi.Works fine with: &lt;lora:KousakaShigureV2:0.7&gt;, kousaka_shigure, ponytail,I was only able to isolate one outfit:<ul><li>ShigurePinkOutfit (It her typical pink kimono with an under mesh top)</li></ul>This was supposed to be an easy retrain but instead i had to clean up the dataset. My oldest loras are really showing their age, I am surprised how nice they came out in retrospect.I have also added the dataset and the tags summary file so people can experiment.Note stroked though text are for V1<s>Might be a bit of a failure. I was only able to get her zigzagged eyebrows at high weight .65+and high steps 30+</s><s>She looks perfectly fine at .4 weight but she will be missing her characteristic eyebrows.</s><s>Got more accurate results with: &lt;lora:Kousaka_Shigure:.65&gt;, Kousaka_Shigure, big_breasts, eyebrows_visible_through_hair</s><s>Got better and consistent results but normal eyebrows result with: &lt;lora:Kousaka_Shigure:.45&gt;, Kousaka_Shigure, big_breasts,</s><s>Note: eyebrows_visible_through_hair is not a trigger but without it images are less likely to show the zigzag pattern in her eyebrows.</s>

Shigure Kosaka - HSDK

Was supposed to do a small retraining with updated parameters. Turns out my first loras are rougher than i remember beneath the curtain(I am surprised how good they turned out, LORAs can be really forgiving). Ended up deleting mirrored images because it was something i did back then(and then turn on mirror augment). Clean up and upscale some images. And basically completely redo the tagging as the original was made in deepdanbooru instead of WD.

Files with SHA256: 898b829bcb000aa1cfb72706d0fe6468d2481744a13d0c6dd4fde2ae329ff40d