The First thing I have been looking at lately is voxels. Harmless has done some work with them in ways that most of you are unfamiliar with. What I am talking about is in fact voxels to represent a model in full rotational 3 space using a few co-planar layers, with holes to represent transperancy.
First things to note: This can be hardware accelerated, the Permedia 3 supports 3d volumetric rendering using OpenGL Texture 3d. So, it is possible. Hardware rendering can also be approximated by making use of lod and triangles.
Okay, the way of doing things is sorta like this for a layer is like this. You have your heightfield, and you get exact height field position points, by finding lines of skew through the camera, and then linearly interpolating your points along them. From each point you extend another point, in the direction of the normal. You can then treat these extended points as a triangle mesh, or use voxels or what ever. Pretty simple, but it gets great lighting, because you can use a phong map and the typical environment bumpmapping to do your lighting very quick. I would recommend voxels personally. But, the point is not that, you could use lod triangles anyway in software. The point is in up and coming hardware, this would be very cool.
This is just another way to look at things I have been thinking of. The first one is polarised voxels. Every point has a distance and angles it is centre. What if you used multiple sphere layers and just saved distances into the sphere map? This would cain for rotating large models. Also I can thinking of lighting speed ups and a style of rendering that would make the object inherintly curved. The important thing to realise is I am of course not talking of USUAL polar coordinates. Just those that have a unit vector and an offset (not a longitude and latitude and a distance). Of course, saving a sphere map, and using a few cheats, we never have to save the unit vectors, as they are mapped at runtime. This could be really efficient for hi-end ray-tracers and hardware alike. I have to work on the algos, and I might see about trying the OCTREE, as it is efficient for these things. This is probably not new, but anyway, for lack of a better name, I'll call it Spherical Volumetric mapping, which is just a description really.
I have to think about this one, but I will keep you guys updated.
Also, I welcome Jaap and Jacco (Aghh, the idea has caught on... Read these guys files, I talk to them often, they are both very smart people, and I use them as target practice for my ideas.)