Yeah. And having no need to repeat stuff on purpose, level editing work could be much easier, and procedural generation too.
But at this point i don't know how to solve the storage space problem. Current compression methods are not nearly good enough.
Recently i have implemented Voxel DAG compression, which is said to do 'wonders'. But it does wonders only at really high resolutions where any geometry becomes pretty flat at finest voxel levels. At lower resolution it is not much better than regular octree.
Still, the idea is interesting: It works by searching branches of the tree that are equal. So if you have two equal houses bath aligned to the voxel grid, you only need the space of one. You get automatic 'instancing' at all levels of the tree, from a single 4^3 cube of voxels up to a full house.
Limitation here is: Grid alignment must match, and rotation is not supported at all. (Thinking of Bruce Dell again
)
But could this idea be used for lossy texture compression? Searching for similar 8*8 quads of texture, make a dictionary of quads, and compose a huge virtual texture from a small dictionary? Maybe even hierarchical, so the result is a sum of multiple mip levels, each level compressed on its own this way?
I think this could work. There is the problem of discontinuity between quads, but this may be solvable or acceptable.
However, if you want rotation too, then the compression algo would likely take years to complete.
No idea. That's why i think storage space is the main future limitation for games.