Optimizing GAN Architecture for Enhanced Image Style Transfer Performance
Year: 2025
Type: article
Abstract: Despite recent advances in GAN-based image stylization technologies, critical limitations persist—namely, instability during training, mode collapse, and challenges in balancing style and content features. These issues often result in degraded visual quality, lack of stylistic fidelity, and limited applicability in real-world creative contexts. To address these challenges, this study systematically proposes an optimization strategy aimed at diagnosing and overcoming such technical constraints. An enhanced GAN model architecture was introduced, incorporating a U-Net-based generator and a PatchGAN-based discriminator. Additionally, advanced training techniques—such as learning rate adjustment, gradient clipping, and optimizer selection—were integrated to effectively reduce style degradation and content loss. The results of both quantitative and qualitative evaluations demonstrated the reproducibility and practical applicability of the proposed approach. Ultimately, the improvement in stylization quality is expected to contribute to the broader adoption of this technology in digital content production and artistic creation.
Authors: Rongbin Chen, Sangguk Lim
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7128239696