Creative Style Transfer

Jun Li, Li Shi Min, Ze-dong Zhang, Gan Sun, Hongfu Liu, Li-wei H. Lehman, Jian Yang

Keywords: style transfer; creative learning;

TL;DR: We propose a creative style transfer method to produce new and meaningful artistic styles by creatively transferring the style of one image to another.

Abstract: Generating novel artistic styles from a single style is a formidable challenge for for traditional style transfer techniques, which typically focus on emulating the provided style without introducing fresh and meaningful elements. In this paper, we propose a creativity process for producing new and meaningful artistic styles, called creative style transfer (CSFer). We first introduce a neural permutation network (PerNet) to rearrange the feature maps of a single style image, thereby adapting them to the feature maps of a content image, resulting in the desired stylization content. Essentially, this permutation process enriches the bases of the single style within a high-dimensional feature space, departing from the conventional linear combination of multiple styles. To gauge the quality of our stylized content, we leverage metrics encompassing content structure, style perception, and artistic aesthetics. These metrics enable us to assess our stylized content in comparison to the output produced by traditional style transfer methods. In the training phase, PerNet learns to generate high-quality stylized content by randomly sampling permutation matrices that yield high-quality stylization outcomes. Experimental results demonstrate that our CSFer can create novel and original stylization outcomes. Furthermore, CSFer exhibits robust generalization capabilities by simply inserting the PerNet into the style transfer methods.

Primary Area: transfer learning, meta learning, and lifelong learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 4898