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

Much better dataset, generated from the initial image. Rank 128 is pretty overkill but the results are pretty great too.

Files with SHA256: 605fe392e5c88e5bf2d7a3d95482ed9964fd6095fe9d1cca73eb1f4d817e9322