Text-Guided Visual Prompt Tuning for Vision-Language Models

YueWu, Yunhong Wang, Guodong Wang, Jinjin Zhang, Yingjie Gao, Xiuguo Bao, Di Huang

Keywords: Vision Language Model, Prompt Tuning, Zero-shot Learning, Few-shot Learning

Abstract: Prompt tuning has become a crucial technique for adapting pre-trained vision-language models (VLMs) to various downstream tasks. Recent advancements introduce multi-modal learnable prompts to enhance the creation of task-specific classifiers. Despite their utility, these methods commonly encounter challenges in generalizing to unseen classes, as their symmetrically designed visual prompt struggles to capture task-relevant textual knowledge and lacks the flexibility in adjusting to novel test class distributions. To tackle these obstacles, we propose a novel Text-Guided Visual Prompt Tuning (TGVP) method, which uniquely leverages the robust generalizability of textual knowledge to guide the generation of visual prompt. Our method introduces a simple yet effective Text-Knowledge Guidance Module that dynamically incorporates visual prompt with task-relevant textual knowledge through cross-attention mechanism. The generated text-guided visual prompt endows the visual encoder with semantic awareness and thus enhances both generalization and discriminability of VLMs across various scenarios. Comprehensive experiments demonstrate that TGVP significantly outperforms existing methods in base-to-novel generalization, cross-dataset transfer, and domain generalization tasks, offering a substantial improvement in VLM adaptation.

Primary Area: foundation or frontier models, including LLMs

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 3777