Image Colorization via Textual Embeddings with Grayscale Hints Removal and Large Language Model Rephrasing

The authors introduce a new method that leverages text prompts to control the colorization process, allowing for a more nuanced and accurate outcome. Unlike modern images where color can be easily removed, historical images often lack true color references, making it challenging to colorize them accurately. The proposed method addresses this issue by using text prompts as a control knob that regulates the colors of the result, ensuring a controllable and vivid colorization process.
The authors evaluate their approach through experiments on two benchmark datasets, demonstrating its effectiveness in producing high-quality colorization results. They also compare their method to other state-of-the-art approaches, showcasing its superiority in terms of both visual appeal and semantic coherence.
In summary, the proposed method provides a valuable tool for those working with historical images by enabling controllable and vivid colorization results without relying on extensive ground-truth data. By leveraging text prompts as a control knob, the method overcomes the limitation of data scarcity, producing high-quality colorization results even for out-of-distribution photographs.