The style shift happens because the character you're prompting (like Rouge the Bat) is typically associated with a different style in the base model's training data, like 3D or semi-realistic. This isn't an issue with the LoRA; it's just how the base model handles certain characters.

To fix it, try:

Adding style keywords (e.g., "anime," "illustration").

Using negative prompts to block styles like "3D render."

Emphasizing the style you want over the character in your prompt.

It's a base model behavior, not a problem with the LoRA itself.

@wingusblingushahalol No worries, I get what you meant. It’s a common issue with AI models in general. Just focus on tweaking prompts like adding style keywords or using negatives for unwanted styles. Glad I could help. I know it can be annoying at times, the only way this could be fixed is to go through all the checkpoint images and for every image make 3 different versions 3d, cartoon, anime so it wouldn't generalize the characters, but I don't have those kind of resources to make it happen. I hope you get around the issue.