Could it be because you are calling withImage in IO whereas generateImage coes through ST? A lot of the nice performance numbers of JuicyPixels come from its carefully tailored ST usage, which in turn comes from theefficiency of unboxed mutable vectors (as in the "vector" package).

So could you post the benchmark result for a version where you runST on the result of withImage? That should be a fairer comparison. Also, writing a criterion benchmark would help and make sure the functions are run properly without any of the two taking advantage of computations previsouly performed by the other.