MAGAN-RT: A Lightweight Adversarial Style Transfer Network for Real-Time Cartoonization on Low-Power Edge Devices

Recent advances in neural style transfer (NST) and generative adversarial networks (GANs) have enabled photorealistic and artistic image stylization. However, deploying such models on resource-constrained edge devices remains challenging due to their high computational and memory demands. In this paper, we propose MAGAN-RT, a lightweight adversarial style transfer framework optimized for real-time cartoon-style transformation on low-power mobile and embedded platforms. MAGAN-RT integrates depthwise separable convolutions, inverted bottleneck residual blocks, and a multi-scale perceptual distillation strategy with auxiliary RGB supervision to enable efficient and expressive stylization. Furthermore, a real-image-based adversarial loss is employed to enhance realism while avoiding the artifacts commonly inherited from teacher models. Experimental results demonstrate that MAGAN-RT outperforms existing lightweight and mobile-compatible style transfer networks in both visual quality and runtime efficiency. It achieves state-of-the-art LPIPS, FID, and SSIM scores, while maintaining sub-10 ms inference latency on commercial smartphones, making it suitable for real-time applications such as mobile AR and video filters.