Princess Celestia from Friendship is Magic
-Model is trained only on her in episode appearances
Description
Comments (7)
Don't get me wrong, i'm just curious.🤔 What is the point to make a lora with MLP characters today? Fandoms like sonic universe, MLP, pokemon, or others that have a huge fanbase, they have a lot of data that already included in illustrious dataset. And we can easily generate the most of their characters well, without a lora. 🧐
A LoRA can help issues with character traits bleeding over from other characters in a fandom, or style differences from fanart in datasets.
@Alavatar Can only speak for myself but 1) I find these pretty relaxing to do and I do characters I like, 2) I like having a personal library of models that I have control over to fix, remix or otherwise mess with, and 3) I personally only use source material so I try to keep them mostly on model with the series they are from. I started doing these in general to build an offline library just in case and I didnt like the idea of saving other peoples model work to do local generation in the future
@Coranendon thanks for the answer. Good luck on future trainings! 👌
@Alavatar Np, and Thanks!
Loras, especially with unique trigger words, will push only the character/concept/etc with curated training data. While you could certainly use a checkpoint's base knowledge, a lora helps focus on whatever it is trained on. Additionally, in a case like this you can still opt to use the base model's knowledge by using its tags. This, along with lora weight, and emphasis (can be high or even lower emphasis), can give you much more control on how generations can come out.
@llxkatz Great explanation