High-quality fundus images are crucial in the diagnosis of ophthalmic diseases. However, these images in real-world settings often suffer from degradation due to motion blur, illumination irregularities, and artifacts. Existing enhancement methods that rely on paired datasets or simplified degradation models have difficulty addressing the complex degradations commonly observed in clinical realities. We state that pre-trained large-scale text-to-image models contain rich image priors to enhance fundus images to high-quality ones, and we can take low-quality images directly as input with the skip-connection maintaining structure consistency, reducing the ambiguity brought from random noise sampling while simultaneously eliminating the need for additional controlling modules. We then fine-tune the pre-trained network in a single step with a small fraction of trainable parameters to adapt it to the fundus image enhancement task, and show the superiority of BEAM through extensive experiments over other state-of-the-art approaches. Moreover, we expanded our framework to an unpaired scheme and showcased its capacity to generate a realistic paired simulation dataset. The source code and dataset are available at https://github.com/wangzh1/BEAM .