We develop a variational approach to solve image inverse problems like super resolution, inpainting and deblurring for flow-based generative models.