Image-to-image (i2i) networks struggle to capture local changes because they\ndo not affect the global scene structure. For example, translating from highway\nscenes to offroad, i2i networks easily focus on global color features but\nignore obvious traits for humans like the absence of lane markings. In this\npaper, we leverage human knowledge about spatial domain characteristics which\nwe refer to as 'local domains' and demonstrate its benefit for image-to-image\ntranslation. Relying on a simple geometrical guidance, we train a patch-based\nGAN on few source data and hallucinate a new unseen domain which subsequently\neases transfer learning to target. We experiment on three tasks ranging from\nunstructured environments to adverse weather. Our comprehensive evaluation\nsetting shows we are able to generate realistic translations, with minimal\npriors, and training only on a few images. Furthermore, when trained on our\ntranslations images we show that all tested proxy tasks are significantly\nimproved, without ever seeing target domain at training.\n