And before you ask what the other problems are, --that I believe I could fix easier than I can explain--, here's a couple:
1) It's okay to use floating point arithmetic in the sampling loop for the filters; but the sample accumulators need to be double precision. I believe from looking at the results, that you guys are using floats for the accumulators. Floats are terrible when used as accumulators; you get aliasing precision losses (the first sample accumulated gets full precision; the last gets almost NO precision).
2) Lots of choices for filters... Cone, disc, gaussian, cosine... But none of them is the "right" filter for shininess modulation. The right filter would be a "Phong" filter:
Each mip level corresponds to a blur angle, which in turn corresponds, mathematically, to a shininess exponent. Call it S.
Then, the right filter would be,
sample_coefficient = pow( cos(angle), S );
Other features we'd like to add are:
1) Being able to save settings to a file
2) Being able to save side order and axis flips, when working with large numbers of existing cubemaps made up of individual textures.
3) Maybe... Automatic detection of face order and rotations.
4) Optional Auto adjustment of gains and gammas, for when a large set of existing cubemaps need to be "normalized" in brightness and contrast.
5) Optional color range stretching, to maximize precision, with correction factors used by CMG saved to a text file with the same name as the cubemap, so that correction data can be sent to the shader by code.
6) Batch operations.
