Query-Selected Global Attention for Text guided Image Style Transfer using Diffusion Model

Diffusion models have gained tremendous interest in image generation. Additionally, guided text methods for manipulating source images have shown successful progress. However, research on style transfer using diffusion models is still ongoing to address the trade-off between style transfer and content preservation. One representative solution to the issue is contrastive learning in a self-supervised manner, which is useful for extracting specific features from the same location on source and generated images for every pixel. However, there are instances where it is necessary to preserve certain areas, which contain more information from the source image compared to other areas in the image. Therefore, we propose anchoring the areas for preservation and intentionally selecting features at the anchor points through a query-selected global attention method. This enables our method to generate an image that preserves the content of the source while transferring the style without the need for additional fine-tuning or auxiliary network. Our diffusion model follows a simple architecture to enhance image quality and speed up inference time, in comparison to other diffusion methods. Our experimental results also demonstrate superior performance.