In computer vision, fine-tuning is the de-facto  approach to leverage pre-trained vision models to perform  downstream tasks. However, deploying it in practice is quite  challenging, due to adopting parameter inefficient global update  and heavily relying on high-quality downstream data. Recently,  prompt-based learning, which adds the task-relevant prompt to  adapt the pre-trained models to downstream tasks, has drastically  boosted the performance of many natural language downstream  tasks. In this work, we extend this notable transfer ability  benefited from prompt into vision models as an alternative to  fine-tuning. To this end, we propose parameter-efficient Prompt  tuning (Pro-tuning) to adapt diverse frozen pre-trained models to  a wide variety of downstream vision tasks. The key to Pro-tuning  is prompt-based tuning, i.e., learning task-specific vision prompts  for downstream input images with the pre-trained model frozen. By only training a small number of additional parameters, Protuning  can generate compact and robust downstream models both  for CNN-based and transformer-based network architectures. Comprehensive experiments evidence that the proposed Protuning  outperforms fine-tuning on a broad range of vision  tasks and scenarios, including image classification (under generic  objects, class imbalance, image corruption, adversarial robustness,  and out-of-distribution generalization), and dense prediction  tasks such as object detection and semantic segmentation.