Recent achievements of the team - Geodesic feature augmentation for zero-shot text-guided diffusion style transfer
Our team has published the paper “Geodesic feature augmentation for zero-shot text-guided diffusion style transfer” in the international journal VISUAL COMPUTER (Impact Factor: 2.9, Q3 in CAS). The School of Computer Engineering and Science at Shanghai University is listed as the first affiliation.
Image style transfer aims to combine the content of one image with another artistic style, yet has long faced numerous challenges. Traditional approaches predominantly rely on specific “style reference images”, which severely restricts user creativity and necessitates time-consuming searches for suitable references. While the recently emerging text-guided methods offer greater flexibility, they universally confront a core dilemma: how to inject new styles while preventing the core content and structure of the original image from being destroyed or distorted—the problem of “balancing content and style”.
To address this challenge, our team proposes FAGStyle, a text-guided diffusion-based image style transfer method requiring no sample images. This approach innovatively applies shape space theory to state-of-the-art diffusion generative models. Our method enhances information exchange between image patches by integrating sliding window cropping techniques and feature enhancement on geodesic surfaces into the style control loss function. Furthermore, we integrate a pre-shaping autocorrelation consistency loss function to ensure content coherence.
Extensive qualitative and quantitative experimental results demonstrate that, when confronted with diverse abstract composite styles, the FAGStyle approach overcomes the reliance on reference images and content preservation shortcomings inherent in traditional style transfer methods. It outperforms current mainstream methods in stylisation intensity, fidelity, and content fidelity.
Eassy：Geodesic feature augmentation for zero-shot text-guided diffusion style transfer
Code：https://github.com/han-yuexing/FAGStyle
