Text-driven image synthesis has emerged as a popular research area. Existing approaches to invert image to text face challenges like requiring multiple images, slow convergence, or overfitting thus suffering from editing capability. In this paper, we propose a novel initialization method for inverting text using off-the-shelf classification or captioning models. This approach enables multi-token embedding learning from a single input image while eliminating the need for fine-tuning and ensuring faster convergence. We demonstrate a significant improvement in convergence speed compared to vanilla TI.