Recent large-scale supervised low-light image enhancers based on Transformer and Mamba architectures have achieved leading performance on specific datasets. However, adapting these well-trained supervised enhancers to unlabeled data from novel scenes remains a critical challenge. In this paper, we propose a source-free fast scene adaptation framework to revitalize supervised enhancers. Specifically, we propose an Efficient Adaptation Module (EAM) to rapidly adapt a pre-trained supervised model to novel scene data. Since the EAM struggles to fully recover details, we design a frequency-domain prior-guided refinement module to refine the EAM’s output by adjusting the input’s amplitude in Fourier domain to brighten it and then combining its detail component with the approximate component of the EAM’s output in wavelet domain. Our framework is plug-and-play, requiring only dozens of unlabeled target-domain data and no source-domain data. Extensive experiments and ablation studies demonstrate the superiority of our framework compared to SOTA unsupervised and zero-shot methods.