Trained on RealisticVision5 but should work on other realistic models,

Check image for prompts,

Use face inpaint to further fix faces. I have not used it to keep prompt data.