It couldn’t at all just a little while ago, and even now that it can it’s rather questionable how well it works since the whole nanite system is predicated on modifying geometry…
Shifting vertex that arent there is kinda impossible…
Realistically manual tessellation has always been a better choiche because of 2 reasons.
- you actually control the vertex positions.
- it has the same load in the end result.
The one point you have right is that the mesh is slightly more expensive to store.
But really, by how much?
It’s less than peanuts in most if not all cases.
Next, tessellation was/is a multiplier, and alwyas produced questionable slices by design.
The tris count would be multipled, but to add vertex to a square mesh like you posit, you’ have to crank that well past the maxmum that was allowed.
For instance the landscape with a 1m square needs around an 8 multiplier to even start to do anything visual via shader.
You can arguably achieve a better result with less tris manually by paying attention to how you split up the square for best possible vertex distribution.
On the rest, only you can possibly know what end result you want.
I just seriously doubt run time tessellation makes any difference to the end result - in all cases.
This is particulalry true for stuff that needs to LOD down.
Lod’s have to be particulalry crafted for things like water and terrain(snow) that needs to displace without creaing gaps - and, procedural tessellation never did that well…