If you try to apply the suggested recipe as it is,you will get a lot
of uninteresting images like the ones which have the same color
everywhere.
I have used some cuts to eliminate these images. These cuts are
also necessary to speed up the process of discovery of new
formulas, since it is applied after the random algorithm is
generated but before it is used to compute the complete image.
I keep only those formulas where a large percentage of points
are not equal and have a value bigger than a certain threshold.
Another possible cut is used to have more symmetrical
images. For this I tried the following test:
I accept only algorithms that in the single iteration will
give the same results for (x,y) and (-x,-y) for a few random
set of of values.
Of course these cuts are all a matter of personal taste and
the best way is to play with them.