I've ended up doing quite a bit of image processing in Mathematica for which one of my first steps is often applying the BilateralFilter function to de-noise the images while preserving edge detail.

Unfortunately, this rapidly gets very slow with both size and number of images.

Is there any way to speed this up using CUDALink? I've been looking through the documentation, and haven't found a dedicated implementation of BilateralFilter, but maybe it's possible to compose an equivalent function using CUDAImageConvolve and the other built-in functions?

If so, how, and is is there any way to guesstimate how much of a speed-up might be possible? I don't have an NVIDIA graphics card right now, but would likely buy dedicated hardware for this if I could get a serious speed-up.