Visual Place Recognition (VPR) is crucial in computer vision, aiming to retrieve database images similar to a query image from an extensive collection of known images. However, like many vision-related tasks, learning-based VPR often experiences a decline in performance during nighttime due to the scarcity of nighttime images. Specifically, VPR needs to address the cross-domain problem of night-to-day rather than just the issue of a single nighttime domain. In response to these issues, we present NocPlace, which leverages a generated large-scale, multi-view, nighttime VPR dataset to embed resilience against dazzling lights and extreme darkness in the learned global descriptor. Firstly, we establish a day-night urban scene dataset called NightCities, capturing diverse nighttime scenarios and lighting variations across 60 cities globally. Following this, an unpaired image-to-image translation network is trained on this dataset. Using this trained translation network, we process an existing VPR dataset, thereby obtaining its nighttime version. The NocPlace is then fine-tuned using night-style images, the original labels, and descriptors inherited from the Daytime VPR model. Comprehensive experiments on various nighttime VPR test sets reveal that NocPlace considerably surpasses previous state-of-the-art methods.