I still can't reproduce it
.
Two things to try on your side : add a greyscale() in order the chroma not to trouble what you're seeing ( since with your script, chroma is completely randomized ), and try to add "mmx=false, isse=false, d3now=false" in mt_expand(), to see if results are still the same. If not, screenshots would be welcomed ( with them, with some luck, i might be able to find the possible bug ).