Diffusion-based generative models have gained prominence thanks to their impressive ability in video style transfer and shape-aware editing. These successor models adopt manual masks to focus on specific regions during video synthesis. However, these methods depend on user-supplied masks and require considerable training time to identify the desired region. In this paper, we introduce RAViS, the first framework for prompt-based localized video style transfer in diffusion models. Our method automatically produces accurate editing masks to regulate the stylization process by leveraging features from pre-trained diffusion models and utilizing a straightforward clustering technique. Specifically, our framework accumulates attention maps during the inversion steps of the diffusion model, while self-attention maps are utilized to access semantic maps, and cross-attention maps refine these segments to identify the localized region. Finally, we synthesize the stylized video based on the edit mask using a blending algorithm. To comprehensively evaluate our framework, in addition to using CLIPScore and conducting user studies on the same video dataset as previous video editing works, we also apply ImageReward, which is based on human preference feedback, to assess the quality of frames. Extensive experiments demonstrate that RAViS, without re-training, fine-tuning, or additional user input, significantly outperforms existing methods in terms of both accuracy in region identification and stylization quality.