There are two obstacles you need to deal with in order to achieve similar quality like the one in the image, the first is artistic and the second is technical (memory, processing). First I assume that you already solved your artistic problem, you can make the models, the art and the shaders etc. (Partially because I can't answer art problems)
The major technical problem is that when voxelizing those models you created in a high level of detail you will end up with huge number of tiny voxels. Processing and rendering huge number of voxels is not trivial. The answer to this is called Sparse Voxel Octrees
SVO will give you the ability to render highly detailed models without processing all the voxels, but only the visible ones. Note that in order to render SVO you will need to use ray-casting rather than the usual rasterization technique.
More details on SVO in this paper
Here is an open-source with BSD License implementation of SVO https://code.google.com/p/efficient-sparse-voxel-octrees/
Here is a video that explains polygon vs SVO based rendering (pros and cons).