Smaller simpler images now complete in a few seconds or less, depending on image size and feature size. Unconstrained 'outside' is still challenging, but is now handled by treating it as spheres seeded from the image borders with a radius equal to the distance to the nearest feature in the image.

Small structures are also problematic as ever. But what's the thickness of a 1-pixel-wide line? The answer is much more meaningful for features represented by at least several pixels' thickness (what is several you may ask, well, try it and see...).

The CPU implementation has a problem of thickness spheres overhanging the input mask by a pixel, introducing some inaccuracy. This GPU implementation has a more accurate Euclidean distance transform giving the sphere-fitting higher fidelity to the input geometry and less orthogonal grid bias.