KBStyle: Fast Style Transfer Using a 200 KB Network With Symmetric Knowledge Distillation

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• State-of-the-art (SOTA) artistic style transfer models suffer from immense size, leading to high computational costs and memory demands.
• These limitations hinder real-time performance, high-resolution processing on resource-constrained GPUs, and mobile device applications.

Purpose of the Study:

• To propose a novel, lightweight arbitrary artistic style transfer algorithm named KBStyle.
• To significantly reduce model size while maintaining high-quality stylized image generation and efficient inference.

Main Methods:

• Designed a custom style transfer network with optimized encoders and decoders for low computational cost and shape retention.
• Introduced a weighted style loss function to enhance style migration performance.
• Developed a novel Symmetric Knowledge Distillation (SKD) method to compress encoder-decoder style transfer models.

Main Results:

• The proposed KBStyle algorithm achieves a model size of only 200 KB, a 14x compression via SKD.
• KBStyle demonstrates comparable performance to SOTA knowledge distillation methods.
• Achieved high-quality stylized images with fast inference times (20 ms for 2k resolution on Nvidia TITAN RTX).

Conclusions:

• KBStyle offers a significant reduction in model size compared to SOTA methods.
• The algorithm's efficiency and small footprint enable practical artistic style transfer on mobile devices.
• Symmetric Knowledge Distillation is an effective technique for compressing encoder-decoder style transfer models.