A Hybrid Deep Learning Method for Skin Retouching

Skin retouching aims to enhance portrait images while preserving natural textures, yet existing methods often struggle to simultaneously eliminate imperfections and maintain fine details, especially under diverse skin tones and lighting conditions. To address this, we propose a hybrid deep learning framework that synergizes mask-guided local adaptation with global texture consistency. Our method comprises two main components: skin smoothing and defect removal. The skin smoothing component employs a blend layer prediction network which selectively edits target regions via learned blending weights. Meanwhile, the defect removal component uses an encoder-decoder architecture to reconstruct flawed areas while preserving facial contours. Experiments on FFHQ and real-world datasets demonstrate state-of-the-art performance: 0.020 LPIPS, 43.752 PSNR, and 87.71% user preference. Ablation studies validate the necessity of each component, particularly the Mask Prediction Branch's role role in avoiding over-smoothing. The framework's modular design enables potential integration into mobile photo-editing pipelines.