Colorization Quality Assessment with CLIP

Automatic colorization of grayscale images defies simple evaluation by quantitative metrics because there can be multiple equally good but different “truths,” making consistent evaluation challenging. To address this problem, we propose an evaluation model for image colorization that utilizes prompt learning, with novel components Color-Query Cross Attention Block and addition of Positive and Negative input images, based on the prior knowledge CLIP possesses. This complements conventional metrics by evaluating naturalness of the colorization based on the scene content.