Simple LORA trained on art of an original character of artist Kotoba Noriaki. The model is supposed to generate (with varying success) Karahara Shima, a gloomy girl with burns across the left half of her face and torso.This is my first attempt at training a LORA so I will definitely try and improve on this once I get a better grip on training.<h2>v1 Release</h2>This model works reasonably well (I think). Only requires the listed activation tag and works best at a weight between 0.8-1.0 (try using 1.0 in most cases, but consider going down to 0.8 if there is some bad interaction with another LORA, which should hopefully be rare).The burn scars (at least for a weight of 1.0) should work reasonably well now, but if they don't just add another "burn scars" into the prompt (and play around with the tag weight until it looks nice).

Karahara Shima - 唐原志麻

Preliminary version, may be somewhat wonky. General features seem to be replicated relatively well but may need some nudging in the right direction. Adding an additional "burn scars" seems to help bring out the burns a bit better.Works reasonably well on all of the models I've tested and should work decent-ish with other LORAs (although I've mostly only played around with styles so far).Trained on a set of 29 images that were extracted from various sources (Gelbooru, Twitter, pixiv) and lightly edited to get rid of easy to remove artifacts such as signatures, text and speech bubbles. Most of the images I used are in a standing pose, either from a cowboy or upper body shot. This seems to be reflected in the default pose of the generated images, but other poses should still be doable.Training was done for 10 epochs and 10 repeats each with a batch size of 2, an input resolution of 512x512 and a network dimension of 16 (and consequently an alpha of 8).There are some other images that I might still absorb into the training set that may require some additional clean up to improve results. I've trained a prior version with the same training data but a higher network dimension, which turned out somewhat worse (presumably due to overfitting on the small sample size).Other than increasing the sample count I'd also like to try training at a higher resolution next.

Files with SHA256: 68329353042699dfb0d2666670a1532dd41f8f741d54cc78369004d223dba574