Deep learning-based color optimization and style transfer model for artistic portrait enhancement with facial identity retention
Chuanming Ma, Xiangjing Tian · Journal of the Chinese Institute of Engineers · 2026
The proposed framework leverages deep learning (DL) and style transfer techniques to enhance the attractiveness of digital art while keeping the subject’s identity intact; it works on color optimization for artistic portraits. This method applies color optimization techniques, such as Adaptive Instance Normalization and style transfer using CycleGAN, to modify artistic characteristics while maintaining facial features. In this way, input portraits are preprocessed for color normalization using face detection Multi-task Cascaded Convolutional Networks (MTCNN) and face detection. Essential facial traits and textures are extracted using a ResNet encoder and then transferred to the intended style using color mapping. Modified from CycleGAN generators, retouched with a process known as total variation minimization, the end regenerative output is a stylized, color-optimized portrait that still contains facial characteristics. This DL-based system provides a flexible tool for enhancing portraiture across different art and creative disciplines. This powerful tool by itself will introduce more in-depth post-production for digital painters and photographers seeking to enhance and modify their subjects while retaining fidelity to actual appearance. The methodology is thus reliable, customizable to different types of artistic interpretation, and can work with potentially mismatched data, making it an essentially powerful solution.