We do all our image processing using floats (as memory and disk storage - not as computation accumulators obviously). It helps a lot if you keep a dozen seventy-megapixel images in memory, because it means you can run twice as many jobs on your cluster.