We tackle the challenge of open-vocabulary segmentation, where we need to\nidentify objects from a wide range of categories in different environments,\nusing text prompts as our input. To overcome this challenge, existing methods\noften use multi-modal models like CLIP, which combine image and text features\nin a shared embedding space to bridge the gap between limited and extensive\nvocabulary recognition, resulting in a two-stage approach: In the first stage,\na mask generator takes an input image to generate mask proposals, and the in\nthe second stage the target mask is picked based on the query. However, the\nexpected target mask may not exist in the generated mask proposals, which leads\nto an unexpected output mask. In our work, we propose a novel approach named\nPrompt-guided Mask Proposal (PMP) where the mask generator takes the input text\nprompts and generates masks guided by these prompts. Compared with mask\nproposals generated without input prompts, masks generated by PMP are better\naligned with the input prompts. To realize PMP, we designed a cross-attention\nmechanism between text tokens and query tokens which is capable of generating\nprompt-guided mask proposals after each decoding. We combined our PMP with\nseveral existing works employing a query-based segmentation backbone and the\nexperiments on five benchmark datasets demonstrate the effectiveness of this\napproach, showcasing significant improvements over the current two-stage models\n(1% ~ 3% absolute performance gain in terms of mIOU). The steady improvement in\nperformance across these benchmarks indicates the effective generalization of\nour proposed lightweight prompt-aware method.\n