Supervised methods for vehicle re-identification (Re-ID) often encounter performance decline when applied to new datasets, emphasizing the importance of unsupervised domain adaptive (UDA) methods. The goal of these methods is to facilitate the transfer of knowledge from a source domain to an unlabeled target domain, but are challenged by substantial differences in images across datasets due to varying backgrounds, illuminations, camera views, and resolutions. Recent advancements in pre-trained vision-language models like CLIP have excelled in diverse applications. However, applying these models directly to UDA Re-ID tasks remains complex because the data annotation is typically a serial number, lacking meaningful semantics. These challenges limit the effective deployment of vision-language models in this domain. To address this issue, we introduce an iterative two-stage strategy that fully utilizes CLIP’s capabilities in visual-text representation. Initially, in the Constrained Prompt Learning (CPL) stage, we optimize the text prompts while the image and text encoders remain fixed. The prompt learning can enhance the image encoder’s ability to extract transferable representations. Subsequently, in the Joint-Domain Image Encoder Optimization (JIEO) stage, we merge source and target domains, maintaining static text tokens and text encoder to provide constraints to fine-tuning the image encoder. Through iterative optimization, our approach effectively represents data as feature embeddings, achieving competitive performance across multiple datasets in vehicle Re-ID under UDA scenarios. For example, on VehicleX-to-VeRi, we achieve 43.1% in terms of mAP, and a 2.5% improvement over state-of-the-art approaches.