I can't use anything undersampled more than /2 for this case. Given how fine details are often produced by ship occlusion (which are using sharp occlusion mask) it simply start losing too many details. Thus only thing I can realistically do when talking quality and optimization is use nonuniform sampling and then drag it through bilateral blur.
Also not sure how well would control over density and weight work in multiple passes, maybe I can try to compare results of both and get back with results once I have some time. My HDR bloom is ofc using multiple passes but I am considering getting rid of the lowest undersample lately.