I cut more than 500 images from the pages of Requiem for the training dataset for this LoRA -- this should make it a little more balanced as far as face styles are concerned. Also, I avoided having multiple frames/panels and speech bubbles, so those should almost completely disappear even without negative prompting.