This work introduces a new method to learn image restoration methods from only corrupted data sets. It is an exciting idea that could potentially open up new applications for deep learning methods in settings where it is not possible to obtain ground truth data. Three reviewers initially assessed the work as 5/9/6. Based on a strong author rebuttal all reviewers took part in a discussion and two reviewers revised their score upwards, for a final assessment of 6/9/7. Overall this paper contains an exciting idea and is likely to stimulate the NeurIPS community to further consider the setting of learning only from corrupted data.