Dual contouring
An isosurface-extraction algorithm that, unlike marching cubes, places vertices inside the cells using surface normals and so can reproduce sharp edges and corners.
Published by Ju, Losasso, Schaefer and Warren in 2002. Instead of putting vertices on the grid edges, dual contouring places one vertex per cell — on the "dual" grid — at the point that minimizes a quadratic error function (QEF) over the intersection points and their normals. Those normals, so-called Hermite data, are the key: they say which way the surface leans, so a crisp edge can be reproduced exactly rather than rounded off.
Marching cubes rounds those very edges, because it has no directional information. Dual contouring is therefore used for hard-edged voxel terrain and for CSG operations that cut and weld solid bodies. The cost: it can produce non-manifold geometry and it requires normal data to be available. Its two-dimensional counterpart solves the same problem in the plane.