Immersing in A Film: A Colorization Method Based on Chinese Film Aesthetics

Colorization is the process of transforming grayscale images into vibrant and realistic ones by adding color. To enhance the aesthetics of colorization methods, this paper proposes an improved image colorization model based on the GAN generative prior, which considers the colorization task from the perspective of Chinese film style for the first time. The model follows an "encoder-generator-discriminator" architecture and incorporates a style injection block, which conditions the generation process on film styles from different classes. By using a pretrained generation model, our method can manipulate latent code for both prior generation and colorization purposes. Moreover, we have introduced a weighted layered feature loss, which is particularly effective in coloring portraits. We have collected and constructed a high-resolution image dataset called " ChineseFilm " , specifically designed to capture the unique characters of Chinese films. This dataset serves for training and testing. Experiments demonstrate the superiority of our method, especially in conveying the distinctive style of Chinese films to faded images. We believe that our work contributes to the promotion and preservation of Chinese culture.