Well, its similar to a sobel filter, except I threw out all the cumbersome matrix math. It's actually the same process i was using before except I removed the part that filters based on vector angle. The earlier image was only returning diagonal edges that ran LL to UR. This one actually just returns all edges.
Basically Im taking the pixel gradient and applying a sigmoid clamping function sqrt(sin(atan())), so that hard edges stay hard, but soft edges, like gradual shade or lighting changes get softened even further.