ClsGAN: Selective Attribute Editing Model Based On Classification Adversarial Network

A comprehensive survey on semantic facial attribute editing using generative adversarial networks

“…It is notable that experiments have shown that FID values are generally consistent with human judgment [59]. Many studies have utilized FID score to evaluate the quality of the manipulated face images, such as [3,27,28,37,51,81,91,148,161,198,203,216]. If the feature vectors of the images generated by the GAN model have multinomial Gaussian distribution N (m, C) and the features of the real world images have N (m w , C w ), FID score is defined as:…”

“…This metric is used in many papers such as [19,50,91,100,119,120,144,152,174,198,206]. There are also some other versions of SSIM index such as MSSIM (mean SSIM) and MS-SSIM (multiscale SSIM).…”

“…The editing accuracy of the model is then estimated by comparing the attribute values of the resultant face image with the desired values. Many works such as [19,27,58,84,100,102,119,120,140,174,198,199,203] use this approach as a quantitative metric to assess the performance of their attribute editing models. In the face aging models like [56,141,142,191], an age predictor model (usually the publicly available Face++ API) is used to estimate the age of the edited image, and the aging task is assumed to have been successfully done if the estimated age of the edited face image is the same as the desired age group.…”