SPAST: Arbitrary style transfer with style priors via pre-trained large-scale model

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Arbitrary style transfer aims to combine content structure from one image with the style of another.
• Existing methods using small models produce low-quality results, while large models are slow and may alter content structure.

Purpose of the Study:

• To develop a novel framework, SPAST, for high-quality arbitrary style transfer with reduced inference time.
• To address limitations of existing methods in terms of image quality, content preservation, and computational efficiency.

Main Methods:

• Proposed a novel Local-global Window Size Stylization Module (LGWSSM) for effective fusion of style and content features.
• Introduced a style prior loss to leverage knowledge from large pre-trained models, enhancing SPAST's performance.
• Developed a new framework named SPAST for efficient arbitrary style transfer.

Main Results:

• SPAST generates high-quality stylized images that effectively preserve content structure.
• The proposed method significantly reduces inference time compared to state-of-the-art techniques.
• Experimental results validate the effectiveness of SPAST in arbitrary style transfer.

Conclusions:

• SPAST offers a superior approach to arbitrary style transfer, balancing image quality, content fidelity, and speed.
• The novel LGWSSM and style prior loss are key components enabling efficient and high-quality stylization.
• This work advances the field by providing a practical and effective solution for arbitrary style transfer.