You can see the code for both the distance field generation and the real time rendering in the downloads and the demo I have posted on this site. Neither is a big problem, and both are useful. I encourage you to read the code. It is the best source of detailed information, and I published it so that people should be able to read it and make use of it, instead of having to re-implement everything in their own code.
Green used pixel images and a distance transform, just like I do. Neither Green's work nor my work is limited to any particular class of images, except that we both assume that the image is a strictly binary contour (i.e. only two colors, foreground and background), and that the contour can be represented reasonably well by a bitmap of limited resolution.
The difference is that I take better care to do the distance transform with good accuracy using an anti-aliased input image. That is my scientific contribution. Green used very high resolution binary images as input, which is a problem in production. He also used a very crude (but good enough) algorithm for the distance transform. My method is a lot faster and can be performed on the fly if needed. It can even be implemented on the GPU. I have a GPU version of the distance transform without the antialiasing done, and I started doing the anti-aliased version but never got around to finishing it. The demo without AA is here:
Jump flooding algorithm in GLSL
I can't offer any support for that code, but the method is explained in the article by Danielsson (a PDF is on the link above), and the GLSL details are presented in a fairly recent PhD dissertation by Guodong Rong. Google for "Jump Flooding Algorithm" and you will find it. As I said, I did a preliminary GPU implementation of my anti-aliased version. I did not finish it well enough to release it, but I can find the code and put it online if you are interested. It's just the simple version of the AA EDT, the one with the linear approximation to the distance function d_f, but at least it's more accurate than using a binary image as input.
The real time rendering is basically the same in Green's work and mine, but I use a better (and more standard) method for finding the gradient vectors in the shader. An alternative is also to store the gradient explicitly with the distance image. Green used an explicit finite difference instead of the dFdx() and dFdy() functions in GLSL. His method is slower, and has accuracy problems on ATI hardware where texture interpolation is performed in a fixed point texture coordinate system.
I hope that answered your questions. If not, ask again!
