This is quite similar to what VCT can learn on the synthesized dataset Objects-Room. As the real-world dataset is more diverse, we observe several failure cases shown in Figure 8. We suppose those failure cases are due to VCT, trained withreconstruction loss,isnotgoodatsynthesizing counterfactual samples which arefarfromthe data distribution.

To handle this setting, we use the encoderΘofthe autoencoder trained onAnimalsalso forΦ,butalways apply a spatial deformation to its inputs.

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.