The light emitted by ubiquitous strong light sources in nighttime scenes will inevitably produce flare artifacts in the captured images after entering the camera lens. Flare artifacts of various shapes, positions, and colors not only severely interfere with downstream computer vision applications but also greatly increase the difficulty and challenge of restoring the visual quality of the images. Existing methods rely on predefined fixed prior information, including intensity, position, and shape, to synthesize and remove flare artifacts. They may not be able to deal with flare situations where multiple types of light sources coexist, resulting in limited generalization performance of deep models. In this paper, we propose a flare removal method that adopts a distribution restoration strategy. Through the light source preservation (LSP) module for recycling light sources and the flare prior rendering (FPR) module based on the point spread function (PSF), we extract and learn the flare distribution in the input images. Then, by utilizing the latent relationships between the flare/flare-free regions mapped by the cycle consistency constraint, our method can manipulate the removal and generation of multiple types of flare. In addition, by comparing the actual size of the light source with the flare range obtained through visual segmentation, we contribute an objective and comprehensive unpaired dataset for flare removal. Different from the paired flare that is difficult to collect and the synthetic flare formed by simple addition, numerous experiments have demonstrated the effectiveness of our framework trained on unpaired data and its superiority compared with baseline methods.