firstly, I recommend this guide for the training.

https://civitai.com/models/22530/guide-make-your-own-loras-easy-and-free

secondly, I don't really do anything special,

To start most of my image is trained on several images ranged from 22-42 Images and in some special case (mostly anime), on 100 images.

When I'm creating a Lora, a search for any image from the show (like png or artwork)?

And I add some fanart (just a few).

Most of the fanart have a style that it's close to the show style.

Sometimes it's very hard to do when the show is not widely known, I have very hard time finding good images. In those case I had to go to some unholy website to find some of the art, but the least of those in your dataset, the better.

notes:

 1-More images don’t really mean a better Lora, it really depends on what is the subject.

 2-No-human (furry, animals) subjects are more complicated to make, some time you will need to restart your training multiple times.

 3-When you are training your subject, put in the dataset multiple “full body” images of your subject.

 4-If your subject is a character that rarely change outfit, you should make sure that the dataset of said character doesn’t have too much different outfit.

5-upscale your training images if their quality is low.

All images were generated using abyssorangemixs-2-Sfw