GAN based Hairstyle Generation Framework for Standardization of Lightweight-model

Recently GAN-based image generation models are not properly generating local image accurately, especially for a hairstyle generation task. For the reason, we propose a new hairstyle generation framework based on GAN method. The framework consists of the Face segmentation module (F) with the Hairstyle Generator (HG) module, and network interface to control internal modules. The F module detects hairstyle and facial features, and then extracts global and facial feature masks. The HG module using a GAN method based on transformer to generate hairstyle image, while fixing the detailed fusion in the hairstyle generation process. We used FFHQ and CelebA-HQ datasets for training and evaluation of our framework. The evaluation process used FID, PSNR, SSIM methods and we compared our model with other hairstyle generation models. We get higher test score and robust generated image than other model, especially in lightness of model size.