What they suggest in this paper is not alpha-to-coverage:
Alpha-to-coverage converts the alpha value output by the pixel shader into a coverage
mask. This coverage mask is combined with the standard multisample coverage mask to
determine which samples should be updated.
Alpha-to-coverage, when combined with alpha testing, gives softer edges without
sacrificing the ability to use the z-buffer with unsorted primitives. Although this is an
improvement on simple alpha testing, the resulting alpha gradients can be of poor
quality compared to those obtained in alpha blending. This is particularly true when
using a low number of samples or on hardware that doesn't support flexible coverage
masks.
The cases where alpha-to-coverage delivers suitable results are rather rare, since it works best for translucent surfaces which cover a preferably moderately large and homogeneous colored/translucent screen space area.
The technique described in the paper should be worth a try, though.
