Image-to-image translation has recently achieved remarkable results. But\ndespite current success, it suffers from inferior performance when translations\nbetween classes require large shape changes. We attribute this to the\nhigh-resolution bottlenecks which are used by current state-of-the-art\nimage-to-image methods. Therefore, in this work, we propose a novel deep\nhierarchical Image-to-Image Translation method, called DeepI2I. We learn a\nmodel by leveraging hierarchical features: (a) structural information contained\nin the shallow layers and (b) semantic information extracted from the deep\nlayers. To enable the training of deep I2I models on small datasets, we propose\na novel transfer learning method, that transfers knowledge from pre-trained\nGANs. Specifically, we leverage the discriminator of a pre-trained GANs (i.e.\nBigGAN or StyleGAN) to initialize both the encoder and the discriminator and\nthe pre-trained generator to initialize the generator of our model. Applying\nknowledge transfer leads to an alignment problem between the encoder and\ngenerator. We introduce an adaptor network to address this. On many-class\nimage-to-image translation on three datasets (Animal faces, Birds, and Foods)\nwe decrease mFID by at least 35% when compared to the state-of-the-art.\nFurthermore, we qualitatively and quantitatively demonstrate that transfer\nlearning significantly improves the performance of I2I systems, especially for\nsmall datasets. Finally, we are the first to perform I2I translations for\ndomains with over 100 classes.\n