Generative modeling for continuous non-linearly embedded visual inference