Given the ever-increasing computational costs of modern machine learning\nmodels, we need to find new ways to reuse such expert models and thus tap into\nthe resources that have been invested in their creation. Recent work suggests\nthat the power of these massive models is captured by the representations they\nlearn. Therefore, we seek a model that can relate between different existing\nrepresentations and propose to solve this task with a conditionally invertible\nnetwork. This network demonstrates its capability by (i) providing generic\ntransfer between diverse domains, (ii) enabling controlled content synthesis by\nallowing modification in other domains, and (iii) facilitating diagnosis of\nexisting representations by translating them into interpretable domains such as\nimages. Our domain transfer network can translate between fixed representations\nwithout having to learn or finetune them. This allows users to utilize various\nexisting domain-specific expert models from the literature that had been\ntrained with extensive computational resources. Experiments on diverse\nconditional image synthesis tasks, competitive image modification results and\nexperiments on image-to-image and text-to-image generation demonstrate the\ngeneric applicability of our approach. For example, we translate between BERT\nand BigGAN, state-of-the-art text and image models to provide text-to-image\ngeneration, which neither of both experts can perform on their own.\n