I've learned that tag order has a lot of influence over this one, try to put the things you're looking for specifically closer to the start.

This version seems to be on par with V1, but with the extra control and compatibility I was aiming for with V2. It's still annoyingly inconsistent with textures and transparency, though... I'd suggest using this if you want finer control over positions, more variety from "girl sitting on slime blob", or furry stuff, but otherwise I think V1 still has the best overall quality.

(This was trained on Colab using only ~730 steps, so a retrain (when I get enough buzz), along with a few tweaks based on things learned from testing, should hopefully make it even better.)