[UPDATE]

Hah. I guess that whole thing with closed-source Illustrious was kinda wrong :/
OnomaAI is still doing kinda shady moves for no reason imo.

Nevertheless, I will train v1.0 for NAI eps soon.

FINALLY. This LoRA gives normal output 90% of the time.

The "side view" tag gives the best results, and the "from behind" tag makes the LoRA misbehave. There's still room for A LOT of improvement, but at least it's stable now.