This paper proposes a framework for building trainable variants of image priors and recipes to facilitate their training. The reviewers appreciate the flexibility of the framework, the viewpoint of noncooperative game to formulate the problem, and promising empirical results. Overall, reviewers are positive about this work, and based on their given score, I recommend accept. However, please note that all reviewers believed the presentation can improve. Please apply reviewers' feedback to the final draft. In particular, R1 has concrete suggestions that are easy to incorporate.