Lens flare is a common visual artifact in photographic imaging, particularly prominent when capturing scenes containing intense light sources. The formation mechanisms of lens flare involve complex optical interactions, which are difficult to prevent during practical photography and make the flare diverse in form. Acquiring large-scale paired datasets of flare-corrupted and flare-free images for deep learning applications through real-world photography proves challenging, the training datasets for lens flare removal are usually obtained through synthetic methods currently. We explain the limitations of the existing datasets in fitting the real-world situation from two aspects: one is the brightness distribution of background images is concentrated; the other is the type of flare images is single. To solve these problems, this paper studies and synthesizes the datasets in the flare removal task. For background images, we introduce the Bag of Curves (BoC) method, which adjusts image brightness in accordance with physical principles to better reflect the diverse luminance conditions encountered in real-world scenes. For flare images, in order to obtain a richer type of flare images, we use a stable diffusion model to design and generate flare images corresponding to the specified light source images. Experimental results demonstrate that compared with existing datasets, our dataset offers greater diversity, enabling models trained with the same methods to remove flare artifacts more effectively and naturally.