Mink is the main character from Dragon Half.The main trigger for V3.0 is <code>mink</code>, with either <code>dragon_half</code> or <code>dragon horns</code> to specify a variant.Default anime appearance: <code>mink, dragon_half, bikini armor, detached sleeves, vambraces, dragon wings, dragon tail, knee boots, dragon charm</code>. Weight 0.9 - 1.0 is recommended.<h3 id="activation-tags-rxblxj1yj">Activation tags</h3>1) <code>mink</code> is the main trigger. This won't add any horns, just Mink with her spiky hair and red eyes. 2) <code>dragon_half</code> or <code>dragon horns</code>. This might be a bit confusing, but to keep some compatibility with the previous versions <code>dragon_half</code> adds a standard pair of horns, like in the anime. While <code>dragon horns</code> is used to add 3 pairs of horns from her post-transformation manga appearance. 3) <code>bikini armor, detached sleeves, vambraces, knee boots, dragon charm</code> creates the outfit Mink is wearing in the anime. I spent a lot of time training each keyword separately, so you should be able to mix and match them, but some other concepts might bleed into the image from time to time. 4) <code>dragon wings</code> and <code>dragon tail</code> are self explanatory.Bonus tags: a) <code>anime artstyle</code> forces the anime aesthetics. b) <code>official art</code> can be used to shift the art style toward the official illustrations.<h3 id="misc.-2ferk0ukk">Misc.</h3>Versions 1.0 and 2.0 are obsolete and shouldn't be used.Version 3.0 took a lot of time to make. I went over every image and manually edited them to unify the appearance, and create multiple variants for higher flexibility.Trained on 215 images (x5 repeats), with person used as a class. 13 epochs, 32 DIM / 16 Alpha, Prodigy, 0.25/0.01/0.01 dropouts. Base model used for training was animefull_final, but the example images were created on AbyssOrangeMix.Feedback is welcome.

Mink (Dragon Half)

This version has significantly better consistency. I've been preparing 2.0 for over a month now, crafting a perfect dataset and settings resulting in what I think is an ideal LoRA. Additionally v2.0 no longer uses <code>dragon necklace</code> and <code>bare shoulders</code> activation tags.

Files with SHA256: 64b63f772b201507db9bcbcf3d3d0d2046a0710d989ca1424d476db628c71f63