Hi! I've played around with mesh LOD (edge-collapse), and it works out quite fine. but the problem is that all the models it should work with are bumpmapped. Of course those normalmaps are invalidated as soon as i change the mesh LOD level. How do other games manage this? Do they just store a number of precalculated LOD levels per mesh and their according normalmaps, or do they automatically recalculate the normalmap from a high-res mesh? or do they just leave it like it is? i haven't implemented a normalmap generation algo on my own, but from what i've heard it takes some time to compute (and i can't think of a game that takes 5 minutes or so for every change of graphics options). Any suggestions?

If you are talking about discrete LOD steps then I'd say if your model is far enough to use LOD you might as well not render with normalmaps from that point on. On dynamic LOD you would have to generate normal maps on the fly which isn't feasable in realtime I think. (and if it would be the cost of generating would cancle out the gained speed of the LOD).

it wouldn't be that expensive to have a copy of the normal map for each LOD, just make each sucessive normal map half the size. Sortof like mipmapping but with LODs. at most it would be twice the texture space... Thats not that much, especially if you use DXTC or the like on the LOD versions.