I find the best chance of having a recognisable face is to have the subject facing the viewer. With subject looking away or at a weird angle it has a hard time for sure. Have part of your prompt reference eyes or face "gold eyes" or "portrait" may help.

Turning on face restore helps.

Embeddings and Loras also help to some extent.

embedding: ng_deepnegative_v1_75t / EasyNegative / badhandv4,

lora : epiCRealismHelper

Models: I am finding the "realcartoon" set to be pretty reliable.

You can also try adetailer, which after the base image is generated will re-generate in more detail either the face or body (your choice). Note: I am not using this at the moment as it slows down generation considerably.

What model(s) are you using? Does the same happen if you copy this prompt for example? Happy to try one of your prompts out to see what the result is.

Edit: final thought, some models struggle with resolutions they weren't trained in, so stick to 512x512 unless you're sure then upscale your choice.