Unsupervised Text Style Transfer (UTST) has emerged as a critical task within the domain of Natural Language Processing (NLP), aiming to transfer one stylistic aspect of a sentence to another without changing its semantics, content, or other attributes. This task is especially challenging given the intrinsic lack of parallel text pairings. Among existing methods for UTST, the mask-filling approach and Large Language Models (LLMs) are considered two pioneering paradigms. However, the former often generates unsmooth sentences while the latter tends to alter the original content. In this paper, we investigate if we can effectively combine these two methods. We propose four interaction methods: a pipeline framework with tuned orders, knowledge distillation from LLMs to a mask-filling model, and in-context learning with constructed parallel examples. Empirically, we demonstrate that these multi-way interactions enhance baseline performance in terms of style strength, content preservation, and text fluency. Our results show that a simple method of prompting followed by mask-filling based revision consistently outperforms other systems, including those for supervised text style transfer. Specifically, these methods achieved new state-of-the-art results on the Yelp-clean and Amazon-clean datasets. The code, and trained models are publicly available at https://github.com/Holmes-lei/UTST .