|
LoRA is basically "Hey. I want to make a new token the Ai recognizes called "Sam", so instead of having to prompt "Sam" with 10 paragraphs of text, and still get things that are not-Sam, I took about 40 images of "Sam" that were successful, and created a plugin for the Ai so the Ai knows what I mean by "Sam."
Once you have a LoRA, you can create the same character consistently. The power here is your training data can come from a Ai base, with post-generation hand edits. Like special markings, unnatural colors, whatever. If you say a "Snow Leopard" for example, the AI has a hard time with that. It treats "Snow" and "Leopard" as both a "Snow leopard" and also tries to dope the scene with snow. You also might get non-snow leopard markings and colors mixed in because it'll also see "Leopard" as it's own thing.
Likewise with "Fox", good luck getting a consistent "Purple fox", because the Ai knows foxes are supposed to be red or orange.
LoRA lets you break that by decoupling your unique character from the characteristics that make it up.
|