After the skimage 0.5 release, it's time to start planning the Next Big Thing (TM). I hoped there would be a better solution by now, but it seems that one of our main weak points is still that we cannot do fast looping over images (e.g., filtering).
We have multiple computing back-end support available as a pull request from Pieter Holtzhausen's GSOC--the code is fairly simple to understand, and doesn't complicate the framework more than necessary. This will allow us to, when doing convolutions e.g., first try doing it via OpenCL, and then to fall back to the CPU. We also have a full testing framework in place, so that the output from different back-ends are compared.
Have a look at the code here:
If you have a better solution or other alternatives in mind, please let me know; I'm keen to explore this space further.