v1 of this LoRA was trained on a single image dataset of this image: <a target="_blank" rel="ugc" href="https://civarchive.com/posts/3746856">https://civarchive.com/posts/3746856</a> v2 was trained on images generated by v1, see the included training data for details.It will produce this specific blue dragon woman character if you use the prompt `dragon woman`. It also responds to the prompts `face focus, dragon ears, horns, nose scales, smile, blue skin, black bodysuit, dark blue hair, thick eyebrows, red-orange eyes, looking at viewer, blue forehead crystal, parted lips, fangs, sharp teeth, from front, simple background`I picked her for her distinctive look, it makes it easy to tell when the LoRA is well trained to have distinctive elements such as her ears/horns, her nose scales, and her forehead crystal.It is regularized on the <a target="_blank" rel="ugc" href="https://civarchive.com/models/524110?modelVersionId=582312">Oatd Style</a> dataset (and the original image was generated by that LoRA). This makes it far more flexible than a pure single image LoRA would be. However, it also makes the two largely incompatible, especially at higher strengths.Works at a wide variety of strengths up to 1.0. Works best at 1024x1024, 1024x1536, and 1536x1536 resolutions.For v1, as a single image LoRA, it is super biased towards producing the original image or a close variation of it, though you can look at the gallery for examples of breaking out of that tendency. In particular I recommend using `face focus, from front, simple background` as a negative prompt (unless you want those things very strongly) and including prompts such as a body position, body parts, clothing, or background/setting elements. v2 should have no such issues.

Blue Dragon Lady (OC) LoRA

Same as v2.0 but resized with sv_fro 0.9 so the file is much smaller. Compare the two showcases to see the minor differences for vastly less disk space.

Files with SHA256: 74468213795309aa2dbf12bbcb0672be4821fccde52f6cf1c77b26041f64066b