I'm in the process of making a point based GI/lighting technique.

It's pretty similar to this:
http://graphics.pixar.com/library/PointBasedGlobalIlluminationForMovieProduction/Slides.pdf

Which I guess is pretty similar to the NVidia doc you linked.

Mine works instead with a constant buffer filled with the most relevant vertices to whatever is being rendered.

Instead of turning models into disks, I just work with vertices.

Mine is all a vertex shader, so it's vertex lighting only. As a plus though, I can free up the pixel shader for other effects.

You could make some fairly dense vertex models to get better lighting.

I'll probably share my shader when I'm done with it, I think it still needs some polish.

Something I just thought about. My vertex shader is primarily for doing GI/Shadows. Direct diffuse light could be done in the pixel shader mixed with the vertex shading. So normal maps could still be used.

Ruled out light propagation volumes? That's shipped in games and so long as you have sufficient voxel resolution you can get pretty nice results.

imperfect shadow map reference:

Notice the severe Peter Panning effect at 3 minutes. The result is that it looks as if the ring is flying above the table, never touching it.

Edit: The paper actually states that "While ISMs may contain incorrect depth values, the resulting errors in the indirect illumination are
small but the computational gains...". Inaccurate depth values leads to the requirements of high depth test delta, which leads to Peter Panning. But I suppose the purpose is to use this for indirect illumination, which is indeed what is stated, in which case it may be ok? There are some cases that have to be avoided, like thin geometry.

Whoops, forgot to get back to you! LPVs aren't *quite* the same thing as point-cloud GI. You're basically chopping up the world into a voxel grid and figuring out how much light is flowing through each voxel; you inject some source light via reflective shadow maps or just plain point emitters (as I understand it, this was experimented with in Crysis, though I guess it was dropped. Apparently Crytek don't know how do do specular with it or something?) and then diffuse everything out across the volume.

In practice, you can get pretty fantastic results, though there are still a few areas that can stand to be improved over what shipped in Crysis/possibly the video. The big one, like you point out, is the lack of specular highlights-- this isn't so much a shortcoming of the technique inasmuch as Crytek never seem to have put the effort into doing it. While the idea had been around since around the turn of the millennium, you can project BRDFs into the spherical harmonic basis and use that to compute reflectance by convolving it with the sampled LPV result. Halo 3/Reach actually do something fairly similar, though they actually sample the lightmap at a character's feet-- things are also a little smoothed over since you're using the same lighting data for the whole model.

The other issue is that light can bleed around corners somewhat, since you can't actually linearly blend between SH samples! The Dangerous Curves breakdown by the ATI demo team goes into detail on this, but you actually need to use irradiance gradients for proper blending.