“Controllable Feature-Preserving Style Transfer”, 2023-11-02 ():
This paper proposes a new style transfer quality assessment approach introducing quantifiable metrics to optimize. First, we use a pre-trained DualStyleGAN model to generate multiple stylized portraits in the style vector space.
Then, we design a custom scoring mechanism that uses the newly proposed CSCI and CCVI metrics to evaluate the results’ structural similarity, color consistency, and edge retention. We select and optimize the top outputs using human esthetic standards to obtain the most natural, beautiful, and artistic results.
Experimental results show that our proposed evaluation pipeline can effectively improve the quality of style transfer.
…4.1 Datasets: In our experiments, we use 3 datasets to evaluate the performance of our method in cartoon stylization. For the Caricature dataset, we collected 199 images from WebCaricature, curated explicitly for studying face caricature synthesis. Additionally, we obtained an Anime dataset from Danbooru Portraits, consisting of 140 pairs of style-corresponding portraits. Furthermore, our cartoon stylization experiments involved a cartoon dataset comprising 317 cartoon face images sourced from Toonify.14 These diverse datasets enable us to evaluate the effectiveness and versatility of our method across different stylization tasks, including sketch stylization, caricature synthesis, and cartoon stylization