The model was trained on lyriel. Works well with this model.
Recommendations:
Lora scales 0.77-0.88
28 steps, Euler a/ 2M Karras