Half a century after everybody else, I finally also found a chance to check out spherical harmonics.
As a first step, these are "just" the irradiance SH's, with no precomputed radiance transfer applied to them.
Grace Cathedral light probe on a bridge model courtesy Larian Studios:
A more realistic outdoor setting, Eucalyptus Grove (overcast forest):
At this point I'm mostly interested in this non-PRT version to make sure that the ambient / shadowed areas have some kind of bumpmapping to them. The screenshots above are 100% ambient, no directional light or other shading is applied, so clearly this is working pretty well.
An easy and fast way to make sure your normals do anything useful even in ambient areas, is to use a skylight -- just blend a sky and ground color based on a normal up-component. As expected, this works as long as the surface isn't already facing mostly up (or down); in those cases, bumpmapping adds little vertical variation and so the skycolor dominates. Here's a skylight set for maximum contrast by making skylight==white and ground==black. As you can see, toward the center it's just a flat white plane.
The wash-out can be fixed by using more directions. For instance, Half Life 2 uses three. The neat thing about SH is that it uses the same amount of space, but looks even better and is reportedly faster.
Compare with the same view under Grove setting; more detail is preserved:
Finally, glossy reflections. The only thing that worries me is that these probes don't look much like the published examples :p Although overall color palette and distribution seem correct, so it could be just tone-mapping (all shots in this post are non-HDR, hence the banding).
At this point it's tempting to forge ahead and add PRT.. but then I might as well do diffuse interreflection too.. and away goes the planned time and money budget :) So back to regular engine work tomorrow, saving the juicy bits for "when there's time left" ;)
Hm, I should check in Max tomorrow but I'm guessing that the UV-mapping looks like that :p
I tested the code by for example setting all coefficients except one to zero and seeing if it gave the typical SH patterns on a sphere. If it fuzzies out the normal map it might also be due to too much downsampling and/or DXT5n compression on the map. But yeah, I'm definitely on the lookout for any bugs.
On the side view of the bridge, it looks like the normal map has some stretches (in the lower right side), they also feel a bit soft (might be because the image is scaled), did you test with very simple light probes?