Sampling and correspondence are combined without normalization, so the balance depends on the data rather than on relative_weighting

[akenmorris]

The optimizer combines the two terms as sampling + relative_weighting * correspondence without normalizing either, so what relative_weighting = 1 means changes completely from one set of shapes to another. For sampling, we have the sampling_scale auto factor that multiplies the sampling term by surface area over a hard-coded reference