Multi-Scale Attentional Networks with Adaptive Normalization for Real-to-Cartoon Facial Image Translation

Facial cartoonization, crucial for applications in virtual reality, multimedia entertainment, and online communication, faces challenges in maintaining semantic coherence and high-fidelity outputs under hardware constraints. Traditional Cycle-GAN-based methods often suffer from mode collapse and discriminators’ overemphasis on local features, leading to artifacts and semantic fragmentation. This paper proposes an enhanced framework, U-GAT-IT (Unsupervised Generative Attentional Networks with Adaptive Layer-Instance Normalization), integrating multi-scale attention mechanisms and adaptive normalization (AdaLIN) to address these limitations. The attention module dynamically prioritizes critical facial features across hierarchical levels, balancing local details and global structures. AdaLIN adaptively fuses Instance and Layer Normalization, enabling content-aware stylization intensity adjustment. Experiments on a dataset of 10,000 Asian faces and 10,000 anime images demonstrate U-GAT-IT’s superiority: it achieves a 25% improvement in eye transformation accuracy (SSIM) and 30% faster convergence compared to Cycle-GAN. Results show realistic cartoon avatars with preserved geometric consistency, complete background removal, and reduced noise artifacts. Additionally, U-GAT-IT exhibits robustness in cross-domain tasks, such as medical imaging. Future directions include 3D expression synchronization, lightweight deployment, and multi-task collaboration with speech-driven modules. This work establishes a novel paradigm for GANs in complex cross-domain image translation, offering theoretical and practical advancements.