Well, the meshes I tried were basically variations of a semisphere that pointed toward the camera. This didn't have any issues with poles or cube faces, but it did have issues with changing orientation, because there's no way to lay out the vertices such that the semicircle can snap to any new rotation without there being an abrupt visual state change. The best was 1 axis of rotation, but you need 2 to cover the sphere. If I understand your concept correctly, you're talking about finding the nearest quad on a low detail cubed sphere (though each quad could be subdivided many times), and replacing it with a pre-generated high resolution mesh? If so, this is would indeed work, but it seems like the worst of both worlds. This makes the code more complex, you would have limited detail up close, and near chunk borders you would have to show the neighbours otherwise the transition would look bad. In any case, my CDLOD version works fast enough at over 1000 FPS with 8 million triangles on a 1:1 scale Earth (though there is basically no detail at this range due to each heightmap texel being 600km^2 lol), so I think this solution works fine. The tree structure also plays very well with spawning objects on the surface and colliders.