Lora is trained in the works of a Golden Age artist. It's realism, of course. Anime hadn't been invented yet.
It will try to make your image look realistic. Use the scientific method and adjust the weight of LoRa for a particular character. Different checkpoints will have different results, different datasets will have different results.
It looks like the hatching is perceived by SD as part of the frame. Try non-standard resolutions, for example 1920x1080 with upscaling x2 as in one of the previews, you may get very interesting results.
Description
This is the first, early version, trained for testing on a small volume of images. It requires a precise selection of weights and a balance between your composition and the weights of other lora. It reduces the image to realism by default - and that's fine.