Originally Posted by TheANIMAL
This isn't designed for characters, or moving animatable objects really, at least not yet, i'm not sure why they chose to use a character to demonstrate it.
Don't forget this is still cutting edge technology even if this demo was last year, in all likelyhood improvements have been made sinse then anyway.
What this technology shows is a perfect system for detail to increase for objects in the foreground and for detail to decrease for objects as they move away. It. Eliminates. LOD.
This basicly makes John Carmack a plenepotentiary for the future of real time graphics.
Be honest, did you use plenepotentiary so that we would have to google it and bow to your wisdom?
Well, i'm not getting on my knees for you buster.
Anyways, you have it wrong. From everything I've read sparse voxel octree is a highly granular LOD system. Basically the way it works is that it links many many LOD's together in a hierarchy that is view-distance dependent. My guess is that it is automatically generated using perspective calculations. This is a massive amount of data, but the system loads it dynamically as needed, so performance remains steady, and very good.
I think that theoretically you should be loading a very similar amount of information at each LOD. So the hit for having a very expensive model in the world is pretty much just memory based. Since you're dealing with a fixed amount of pixels on screen, there shouldn't really be any fluctuation in computing requirements (besides bandwidth and memory) over having a relatively simple model (though obviously the lowest bound is where each pixel in the sillhoute area represents data describing the model). Voxels are obviously exponentially more expensive than pixels, but I'm not sure how expensive they are relative to polygons. I do know that polygons become very expensive at sub-pixel divisions.
The reason that SVO's are bad for animation is that the octree has to be generated for each frame of the animation, which becomes prohibitively expensive on the memory side (not sure about the calculations involved in generation...some latency incured certainly). Anyways, for character models I don't see much of a point in using voxels. You quickly reach diminishing returns (for instance, a 20-30k model with normals looks very good; 200k cars are increadibly smooth), and I doubt you would notice or care about the difference between a 7M+ triangle voxel model or a 200k polygon model with proper use of normal maps. Animation tools for polygonal models are already in place and refined (come a long way since Studio Max 2, which is 1st proggy I used). Voxels are apparently more efficient with radiosity and raytracing though. But lighting for individual characters is probably relatively cheap compared to lighting the game world. Some clever tricks could integrate them wel with a raytraced SVO world.