It varies on the character and art style I guess, If its simply generating images, then it will depend on the checkpoint partially, rest comes down to the negatives and prompts, distance of the character helps too (close up usually can be better than full body). When it comes to training, as long as there isn't too many oddball eyes in some images that don't look similar to other images, then it should train pretty good results, example if the eyes have same shape/pupil design/color/shading, then it should turn out nice, depends too on if the majority of training images if they are the same, maybe 2-10 images out of 100 can be a bit different, so on my headshot of yunyun, you see 3 white spots on her eye in the dark and red areas, I wanna say most of the images I used only had 2 white spots, a few had a third, so they kinda blended in, a few others were kinda blank or not as detailed. If you can sharpen the training images somehow, that helps or redraw the eyes and make them more detailed, coloring would be a little bit tough, as long as the color shapes and shades match.

I hope that answers your question

@xSatouMatsuzakax Characters like Hibana from Fire force tend to not work all that well, no matter what I tried there, I cant get the pupils correct since that character has that flower pattern in her eyes.