Some input images contain negative values in oversaturated areas (e.g. central pixels of bright point sources), instead of NAN. These get clipped to zero by the algorithm and are the cause of ugly "black holes" in the middle of bright stars in the final image.
Maybe some kind of clustering and interpolation could be used optionally to prevent this...