I'm not sure about the 'backend' solution but it seems that Forest Pack uses bitmap as the sampling source rather than a mask to hide already sampled instances. This makes for a much more predictable sample pattern in relation to the bitmap, and also seems to by a lot more efficient.
Currently Density Mapping is difficult to control with exponent parameter, it behaves very abruptly, and becomes slow when many instances are hidden.
Ideally I would like to treat the bitmap as a base distribution source, where it is projected onto domain geometry, and then for every pixel with value above treshold( no filtering or interpolation involved) an instance is created.
Optionally density parameter hides random instances.