This time, I'm taking a different approach by focusing on close-ups of faces. I think the results are better than the previous version, but still not quite excellent. The dataset I gathered is not of very good quality either.

Also trained it on RealisticVision_v51