So there are two main issues with this, and I suspect it's because I know what your dataset looks like. First, you baked in the fact that she has a tanto on her back; that means that it's attached to one of the other tags, and generates if you generate the outfit. Second, I'm guessing you only used front shots, because it really struggles to do rear shots while clothed, and also loves to copy the sash pattern. So my guess is that the dataset isn't diverse enough, not that Anima is that good at learning things, as I tried myself to get it to learn Samui and not amount of messing with datasets or settings got it to figure out how a sash works properly.