Here's an image from the CNC-milling simulator I am working on. It stores a signed distance field in an adaptively subdivided octree. To render a surface of the stock we then run an isosurface extraction algorithm. The most well-known of these is the Marching Cubes (MC) algorithm, used here also. One problem with MC is that it doesn't preserve fine detail (like sharp edges) very well, and many times there are also gaps or cracks in the surface. Both of these problems are visible below.
From about 2000 onward there are a number of papers that try to improve on MC. The most promising seems to be Dual Contouring, by Tao Ju. There is both c++ and Java LGPL code for DC on sourceforge, and I made some slight changes to make it compile with cmake on Ubuntu: https://github.com/aewallin/dualcontouring
So far so good. Now the task is to either rewrite my octree + boolean CSG operations code to use the data structure used in the DC code, or vice versa, change the DC code to use my datastructure.