First of all, I don't use Civitai to start with -- I use Wombo -- but even the free version should be enough for this.
The reason I use Wombo is that it seems to work really well with negative prompts of famous people -- something I haven't seen work very well here, but I've mostly used Pony, maybe SD is different.
So, on Wombo, I enter a prompt like "portrait, young woman, auburn hair, $african, asian,Sartre, Audrey Hepburn$", where the part between dollar signs ($) is the negative prompt. By using two or three different famous people, one can get a fairly consistent face -- though one may need to add prompts for (or against) certain ethnicities or hair colors/styles.
One problem with this system is that it tends to produce overly beautiful characters -- especially if one uses very masculine and/or ugly faces to produce a female character, for example. Though it is useful to use these.
Once I have a "beautiful" character, then I try to think of the famous person they look most like. Maybe the character looks a bit similar to Megan Fox -- then I will use her as the last negative prompt to get a character that is not bland, but unique, someone who doesn't look too perfect.
You can even ask for different ages and get a reasonably consistent character from a child to an elderly person. (Though I haven't tried training a LoRA with such images ... and maybe wouldn't suggest it here.)
Now I produce several portrait images using the more realistic styles on Wombo -- being a bit selective. I very much discourage using the "Cartoonist" styles. I thought it might be good to have some exaggeration in the facial characteristics, but the style tends to conflict with the styles here producing a kind of bad Norman Rockwell. Use the more realistic styles -- along with the semi-realistic anime style.
In addition to the portraits, which are generally either frontal or 3/4, I also ask for a few profile portraits (from the side). I generally get around 20 to 30 images of the character.
When training the character LoRA on Civitai with these images, 10 generations is plenty. I generally use flip augmentation. Make sure to tag each of your pictures with the name of your character -- and try to choose a relatively unique name. Also, when you set the prompt for you showcase images, make them simple, but use the character name, for example "1girl, joanne, portrait, by the sea, highest quality, illustration".
Best of luck - joh