Thanks for this perspective. I've never tried this kind of automated visual regression testing except in very specialized circumstances, but I've sometimes thought that I should.

In analogy with scientific modeling, there's a risk of over-fitting your tests against the exact circumstances of your product as it exists today. You want tests to capture the important aspects of your product, while having a little bit of flexibility for the "noise" related to differences between platforms, and changes to mostly but not entirely unrelated parts of the product.