> I would have expected something more algorithmic out of experimentation

Faces are a great test case. They represent a large fraction of the images we need to compress, and humans are very good at picking out issues with faces.