In this paper, we propose a real-time face retouching framework based on a neutral gray layer representation. Instead of directly predicting the retouched image, our method learns a neutral gray layer using image-to-image translation models derived from CycleGAN and pix2pix. During inference, the input image is processed at a reduced resolution for efficiency, and the predicted gray layer is upsampled and blended with the original image via a soft-light composition strategy. This design effectively decouples appearance enhancement from texture modeling, enabling natural-looking retouching while preserving fine skin details. The proposed method provides an efficient solution for enhancing portrait realism in fashion imaging and virtual try-on systems, where skin texture and garment details must coexist harmoniously. We train a face retouching model on the FFHQR dataset, a large-scale professionally retouched face dataset derived from FFHQ, and further demonstrate the flexibility of the proposed framework by training additional models for wrinkle reduction and skin smoothing on separate datasets. Quantitative evaluations using SSIM and PSNR demonstrate that the proposed method achieves favorable reconstruction fidelity compared to baseline approaches. In addition, GPUbased inference benchmarks show that our framework supports real-time performance, making it suitable for practical applications such as live video beautification, interactive fashion displays, and digital garment exhibitions.

face retouching, generative adversarial networks, neutral gray layer, real-time Image processing, virtual fashion display