CoMoGAN is a continuous GAN relying on the unsupervised reorganization of the\ntarget data on a functional manifold. To that matter, we introduce a new\nFunctional Instance Normalization layer and residual mechanism, which together\ndisentangle image content from position on target manifold. We rely on naive\nphysics-inspired models to guide the training while allowing private\nmodel/translations features. CoMoGAN can be used with any GAN backbone and\nallows new types of image translation, such as cyclic image translation like\ntimelapse generation, or detached linear translation. On all datasets, it\noutperforms the literature. Our code is available at\nhttp://github.com/cv-rits/CoMoGAN .\n