Conventional image segmentation methodologies based solely on visual input often exhibit limited efficacy in complex scenes, particularly when objects possess similar visual attributes or are partially occluded. To overcome these limitations, this study introduces a novel multimodal segmentation framework that systematically integrates heterogeneous information sources, including both visual and textual modalities. The proposed pipeline enables comprehensive and accurate scene segmentation through the use of unimodal prompts as well as their combinations. Furthermore, an architectural optimization of the VRP-SAM model is presented, significantly improving inference efficiency. Experimental evaluations demonstrate that the proposed approach yields substantial performance gains over standard baseline methods, without the need for ensemble techniques or additional model fine-tuning.