Supplementary Materials-Unsupervised Domain-Specific Deblurring via Disentangled Representations

1. Paramter selection for λp As we mentioned in the main submission, the weight for perceptual loss λp needs to be tuned so that the deblurred image neither stays too close to the original blurred image, nor contains many artifacts. The quantitative performance and qualitative visualizations are shown in Table 1 and Fig. 1 respectively. If setting the λp too high (λp = 1), the deblurred images become very blurred (Fig. 4(b)), and both the quantitative performance and visualization results are poor. In contrast, if λp is set too low (λp = 0.01), the deblurred images contain many artifacts (Fig. 1(d)).