Efficient generative model for motion deblurring