Automatic Coloring of Anime Line Art Based on Improved CycleGAN

Generative Adversarial Networks as a type of deep learning model, can automatically learn the color and content information in images by learning the distribution of real images, thereby generating realistic color images. In animation production, traditional manual coloring is labor-intensive and challenging to ensure consistency. GAN-based automatic coloring technology can significantly improve production efficiency and image quality. This paper proposes an improved CycleGAN for anime line art coloring, constructing a generator structure that includes an encoder, transformer, and decoder, and designing a corresponding discriminator structure. Experimental results show that the improved model significantly outperforms the Pix 2 Pix model in terms of color restoration, detail preservation, and overall visual effect. Both subjective scores and objective evaluation metrics indicate that the improved model demonstrates higher accuracy and consistency in the image coloring task.