Usually works better with correct prompts and an image as reference + 0.5 Denoise.

It's tricky and needs further experimentation.