I don't have a good idea of how to combine to focused refinement maps so that the overlap region is of best quality.  But here are some ideas about what is causing the degradation in quality when you take the maximum at each grid point.   If one map has low resolution or has high enough intensity noise where the other map has sharp well-defined density you would hope that the second map has higher intensity at each grid point so that taking a maximum value would ignore the poor quality map.  But the low resolution noisy map has broader blobs than the sharp map or blobs of density that don't even overlap the sharp map structure so when you take the maximum all that noise gets added outside the contours of the sharp map density and degrades the quality.  I think the only way to remedy that is you can't combine noisy poor quality density when it extends beyond the sharp density of the other map.  But the low quality density is probably always going to extend beyond the sharp density.  So my conclusion is you to get rid of that poor quality density before you combine the two with maximum, or you combine them using some other function for instance that only puts high density where both maps have high density.  The most obvious way to achieve this is to carefully by hand erase the low quality density in both maps near the overlap reagions for example with the ChimeraX Map Eraser tool.  But you said you want a more automated way since you have many maps to combine.  Possibly if you have local resolution estimates for both maps you could use that to combine them more sensibily than using maximum, but that would require some custom code and I think it is harder than it looks.  Another vague idea that would need special purpose code is to define a plane at the hinge and only blend the two maps within a small region (10A either side) near that plane.

  Combining focused refinement maps has been a common problem for several years now, so perhaps there is literature that describes better techniques, or exisiting software which does it in better ways than per-grid-point maximum.

Tom