Include "Ahegao tongue drool" in your prompts. I find that adding "extra tongue" as a negative prompt also helps a bit on reducing the occurrence of monstrosities with double mouths and tongues.

Just started fiddling with hypernetworks after many Dreambooth models. I think it's quite difficult and more time consuming than DB training (for now) as there's a lack of coherence in the tutorials I've read/watched.

I tried using hypernetwork on faces and it sucked, so I'll keep using Dreambooth on those. But I was curious if hypernetworks can be used on facial expressions instead of subjects. Turns out it can, though v1 still feels half-baked. I still need to experiment on different layer structures and learning rates.