Hello! It has been some time. I train very infrequently, please excuse it.(Trained at 768px with 77 images for ~6000 steps using Prodigy optimizer on terminal SNR FluffyRock e103; May be overbaked, but testing seems to work okay. Lower weight to 0.8 or less if results are undesirable. PDXL version trained on same dataset for ~2000 steps.)Something about this character was not cooperating very well with my usual training methods, but I really wanted to create this LoRA, as I enjoy their design. Material was limited, so some issues may arise. (Behind views are particularly bad, there is next to no data available.)Activation keyword is feirunekumodesu; As mentioned, while the LoRA works at weight 1, it may be desirable to lower it to 0.8 or more.Recommended positive keywords: black body, tan countershading, purple sclera, black eyes, long ears, nose horn, tail markingsOptional keywords: gloves (marking), eye markings (If undesired, keywords should be added to negative prompt, as they may bleed through while exempt from positive prompt.)Recommended negative keywords: lagomorph, slim, tan tail, digitigrade (Rare situations may occur with facial expressions on SD1.5 version, which may require angry, sad, annoyed, or similar keywords added to negative prompt as well. Not as problematic for PDXL.)

Feirune [Earth Wyrm Form] (Kumo Desu ga, Nani Ka?) LoRA

Retrained on same dataset with different settings and rank. Should be more flexible than previous version, less stubborn with poses and body. May yield less on-model results on average, but still controllable.

Files with SHA256: b2a8130315b65b03b9726a230e5db40c576c84ff5c939c3d8aa55e39658e92ab