This LoRA, as the name suggests, is focused on learning style through light and shadow. But as a bit of an experiment, it also includes poses and compositions that are hard to produce by default—though this varies by version. Version 1 was trained mainly on distant perspectives, which sometimes causes the image size to be forcibly altered—like adding a frame around the picture. Version 2 fixed that issue and was trained on images featuring swords in unrelated scabbards, and it seems to work reasonably well.