I am experimenting with procedural meshes. Those are generated at runtime in a compute shader. I'm getting to the point where those geometries are big enough to require some sort of culling/LODing techniques. Both requires the bounding box of the object, but this bounding box cannot be computed from the current LOD section, otherwise the information will be incomplete and thus the culling won't be conservative.
For offline mesh generation, the problem doesn't exist, because the bounding box of the fully detailed object is available. This is not true with procedural geometries, as generating the most detailed version of the mesh is too costly, if not impossible for memory reasons.
Computing the bounding box from the current sections works to some degree, but it's very dependent on the geometry, eg with a big enough frequency, it will be incorrect and sections will be culled when they should be visible.
Depending on how the mesh is generated, it's possible to be conservative by computing the real bounding box by replacing variables by the min/max possible values. But it's only possible to be conservative by limiting how the mesh is generated. If the mesh is generated from complex maths functions such as noises, then the min/max are meaningless and this information is either hard or impossible to compute.
I haven't seen much mentions of this problem before, yet it should affect any acceleration structure algorithm.. Maybe by lack of terminology, or maybe because runtime procedural meshes are not common enough. Anyway, I'm looking for directions on how to solve this problem, or articles where people mention this problem.