We distill commonsense spatial priors from large language models into a flow-based generative model to help agents imagine unseen parts of indoor scenes for Object Goal Navigation task.