Jack Hoxley recently started a thread on parallax occlusion mapping, asking if the algorithm is fundamentally flawed or not. I have to say it is not, and posted several additions to the algorithm for ensuring that the image does not get clipped. Here is what needs to be done:

1) Clip any pixels that have their texture coordinates extend beyond the heightmap boundary (texture coordinates cannot go outside [0,1], or whatever other range you want). This can also be specified per-vertex for added adjustability.

2) Render a box instead of just a plane. This will allow the front edge of the height map to be rendered as well.

Here are the results of the first part:


I think it turned out pretty good...