Honestly, this refinement seems like it’s gaming the metric a bit. I have no doubt there’s innovation here, don’t get me wrong, but the first two examples are not just retaining the character, they seem like overfitting.

The model in many of the images reproduces the exact same armor angle, shield angle, head/helmet angle, etc, without any modification - i.e., it isn’t learning how to reproduce a persistent character in multiple angles, scenes, etc, just the exact illustration fed into it, while changing the setting and positioning.