Generative Flows as a General Purpose Solution for Inverse Problems

Abstract

Due to the success of generative flows to model data distributions, they have been explored in inverse problems. Given a pre-trained generative flow, previous work proposed to minimize the 2-norm of the latent variables as a regularization term in the main objective. The intuition behind it was to ensure high likelihood latent variables, however this does not ensure the generation of realistic samples as we show in our experiments. We therefore propose a regularization term to directly produce high likelihood reconstructions. Our hypothesis is that our method could make generative flows a general-purpose solver for inverse problems. We evaluate our method in image denoising, image deblurring, image inpainting, and image colorization. We observe a compelling improvement of our method over prior works in the PSNR and SSIM metrics.