The paper introduces Visual Prompt Tuning (VPT) as an efficient and effective alternative to full fine-tuning for large-scale Transformer models in vision tasks. VPT introduces a small number of task-specific learnable parameters in the input space while keeping the model backbone frozen. Extensive experiments on various downstream recognition tasks show that VPT achieves significant performance gains compared to other parameter-efficient tuning protocols, often outperforming full fine-tuning in many cases. VPT is particularly effective in low-data regimes and maintains its advantage across different data scales and model capacities. The code for VPT is available at [github.com/kmpv/vpt](https://github.com/kmpv/vpt).The paper introduces Visual Prompt Tuning (VPT) as an efficient and effective alternative to full fine-tuning for large-scale Transformer models in vision tasks. VPT introduces a small number of task-specific learnable parameters in the input space while keeping the model backbone frozen. Extensive experiments on various downstream recognition tasks show that VPT achieves significant performance gains compared to other parameter-efficient tuning protocols, often outperforming full fine-tuning in many cases. VPT is particularly effective in low-data regimes and maintains its advantage across different data scales and model capacities. The code for VPT is available at [github.com/kmpv/vpt](https://github.com/kmpv/vpt).