Diverse Image Colorization Based on Diffusion Model

Abstract

Image colorization refers to the process of coloring a grayscale image. In the field of computer vision, image colorization is a very crucial task, because it can transform grayscale images into real and natural color images, which can provide better visualization effects and more visual information. This paper proposes a method based on the pre-trained diffusion model and control network, so that the model can obtain coloring results with bright colors, natural textures, and diverse results, and can handle a variety of control input conditions. Furthermore, we innovatively proposes a spatial attention module to process reference map control inputs and improve coloring results. Through the test results of a large number of various types of data, such as validation datasets with GT images and old photos without GT images, we show that the model in this paper has excellent performance in grayscale image colorization and strong versatility, which can adapt to different scenarios and different types of test datasets, demonstrating that our algorithm has practical application value.