SupplementaryMaterial

Due to the page limits, the main manuscript has not presented the sufficient visual results of the11 reported tasksoverthereported baselines. In our work, the default initialization strategy is Kaiming initialization.17 To ensure a fair comparison, the other factor keeps the same. In detail, we change the default three-layer Central Difference Convolution and Invertible Neural41 Networkbysevenlayers.42 Modelnumbers.