Feature Augmentation via Dirichlet Mixup for Text-guided Diffusion Image Style Transfer

Image style transfer aims to generate an image that reflects a desired style while preserving the original content. Traditional image-guided methods rely on specific style reference images, limiting their versatility and sometimes compromising the quality of the results. In contrast, text-guided methods offer greater flexibility by allowing users to describe the target style through text prompts. However, these methods often face challenges in maintaining style consistency, accurately reflecting the described style, and preserving the content of the target image. To overcome these limitations, we introduce the Feature Augmentation via Dirichlet Mixup module into the style transfer process. Our method outperforms existing approaches, consistently achieving high-quality stylization while preserving the semantic content of the source image. Experimental results demonstrate the effectiveness of our method across a wide variety of source images and style prompts.