Then, we repeat the branching and bounding steps on subproblems that still have a negative lower bound. We terminate when all unstable neurons are split or all subproblems are verified.

In this supplement, we provide additional empirical visualizations of our approach in Section A.1. Next, we provide details on experimental setup in Section A.2. Mean and standard deviation (s.d.) metric scores across 3 random seeds In COMET we utilize a residual network to parameterize an underlying energy function. We illustrate the underlying architecture of the energy function in Figure 2. The energy function takes as input an image at We remove normalization layers from our residual network. To infer global factors from an input image, we utilize a convolutional encoder in Figure 3. We illustrate the overall architecture in Figure 4. Training Details.

Tables 4 and 5 provide descriptions of the network architectures we use in each of the experiments. In Tab. 4 both networks consist of We provide additional visualization results for DSD as in Sec. Figure 11 shows an additional visualization of the effect of the DSD loss. Figure 13 shows a visual comparison between different losses that utilize feature distributions. Figure 16 shows additional motion-debluring comparisons. We do this for two pairs of scales (between full res.