Enhancing media image style transfer with advanced StyleGAN2 architectures

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Traditional style transfer methods often suffer from mode collapse, loss of image details, and style distortion.
• Existing models like CycleGAN, CUT, and DCLGAN have limitations in generating high-quality, detailed, and stylistically accurate images.

Purpose of the Study:

• To develop an advanced media image style transfer approach using an enhanced StyleGAN2 framework.
• To address and overcome the limitations of conventional style transfer techniques, focusing on image quality and training stability.

Main Methods:

• Integration of a ResNet-based generator with a PatchGAN discriminator.
• Incorporation of the DCL loss function to enhance image generation quality and stabilize training.
• Utilizing enhanced StyleGAN2 for media image style transfer.

Main Results:

• Superior image quality and significant improvements in training stability were achieved.
• The method demonstrated effective content preservation and impressive stylization effects.
• Quantitative metrics (IS: +6%, FID: -13% on Horse2Zebra) confirm advantages over CycleGAN, CUT, and DCLGAN, showing clearer images with richer colors and reduced texture deformation.

Conclusions:

• The enhanced StyleGAN2 approach offers a robust solution for media image style transfer, outperforming existing methods.
• The DCL loss function is crucial for improving edge detail rendering and overall image quality.
• The model exhibits strong generalization capabilities across diverse datasets and applications.