LORA trained on 253 sprites from initial Higurashi no Naku Koro ni release, drawn by Ryukishi07. Works best for portraits as it was trained on upperbody images only. Use 1.0 weight.

No trigger keyword.

Training and dataset parameters:

• Images are all game sprites, transparency changed to black background, tagged by WD tagger;

• Trained using kohya-ss LORA training GUI, rank 32, resolution 512x512, no bucketing, other settings were default;

• Trained on anything-v4.5-pruned-fp16 model;

• 16 epochs, 1 step per image.

Boccher on preview image was created using this LORA https://civarchive.com/models/7624/hitori-gotoh-bocchi-the-rock