How can I generate training data on the fly in TensorFlow?

Question

I want to create the Moving MNIST training set in TensorFlow. I've been thinking about creating a subclass of tf.ReaderBase and emitting tf.Example training sequences in tf.ReaderBase.read or using tf.QueueBase, but I am not sure whether this is the right way to approach it and whether it is a good idea in the first place (regarding CPU load for creating the sequences during training).