How much do I need to worry about style bleeding into a lora if I train on mostly images with the same distinct style?

I want to train a model of Amane Kanata's outfit from the Oracle music video, but all of the images have a certain style and color grading to them that I don't want to capture, I've tried to even out the white balancing, e.g.
https://files.catbox.moe/hxi9r2.png
but I don't want to capture the simple, vector art style, though I don't have many training images that are not in that style
https://files.catbox.moe/s4iv2k.png
I feel like I should probably exclude the 4th and 7th images for having too radically different colors. Tagging "blue theme" might make the lora avoid picking up that styling, but idk