You can apply the exact same algorithm to any starting pattern (it does not have to be random dots). When a pixel is unconstrained by the algorithm, just pick a pixel from an image (rather than a random value). Of course, to work well the image needs a lot of small details. I actually tried that here. But the result "leaks information" into the 2D image (as you move around, the deformations of the 2D image reflects the 3D geometry). It seemed less interesting.
Does anyone know alternative non-random dot approaches to autostereogram generation?
Seeing the random dot approach in video highlights the lack of temporal coherence across frames.