Novel Instance Segmentation is a specialized area that focuses on accurately distinguishing and understanding objects in complex images or videos. It is a fundamental prerequisite for 6D pose estimation and robotic manipulation. In this paper, we present a unified, training-free instance segmentation framework that can be instantaneously applied to any novel object. Specifically, we combine open-set object detection with a designed prompt filtering mechanism to integrate a large vision-language model (LVLM) with the Segment Anything Model (SAM). The proposed pipeline initiates with scene understanding through a LVLM, followed by our novel prompt filter that distills semantic cues to guide Grounding DINO for bounding box generation. These refined detection results subsequently inform SAM for precise mask extraction. We further introduce a matching strategy that unifies semantic, appearance, and geometric cues into a single score for robust instance association and confidence estimation. Experimental results demonstrate that our proposed pipeline leverages the potential of large vision-language models (LVLMs) and vision foundation models, achieving a significant performance improvement. Specifically, it elevates the average precision (AP) to over 50% on the Model-based 2D segmentation of unseen objects task of the BOP challenge, achieving state-of-the-art performance.